build/android-arm64v8a/src/core/CL/cl_kernels/gemm.clembed - platform/external/ComputeLibrary - Git at Google

 R"(

 /*
  * Copyright (c) 2017-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 /*
  * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 /*
  * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */

 /*
  * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #ifndef ARM_COMPUTE_HELPER_H
 #define ARM_COMPUTE_HELPER_H

 /*
  * Copyright (c) 2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */

 /** Store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_n
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE(N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_n

 /** Convert and store the 0th to (n-1)th rows of the given variables
  * @name CONVERT_STORE_ROW_n
  *
  * @param[in] N0        The size of the vectors
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                     \
     (CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));

 /** @} */ // end of groupd CONVERT_STORE_ROW_n

 /** Store a block of the given size M0xN0
  * @name STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group STORE_BLOCK

 /** Convert and store a block of the given size M0xN0
  * @name CONVERT_STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group CONVERT_STORE_BLOCK

 /** Partially store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_PARTIAL_n
  * Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
  *
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] STORE_N0  The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_PARTIAL_n

 /** Partially store a block of the given size STORE_M0xSTORE_N0
  * @name STORE_BLOCK_PARTIAL
  *
  * @note The vector width @p N0 is also required for correct partial storing behaviour.
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] STORE_M0  The number of rows to store. Supported: 1-16
  * @param[in] STORE_N0  The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
  * @param[in] N0        The size of each vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** Store a block that can be partial in both x and y dimensions
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                            \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                           \
     }                                                                                                                                                     \
     else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else                                                                                                                                                  \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                               \
     }
 /** Store a block that can only be partial in x but not y.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** Store a block that can only be partial in y but not x.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  */
 #define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
     if(!(PARTIAL_COND_Y))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** @} */ // end of group STORE_BLOCK_PARTIAL

 #if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)

 /** Boundary-aware GEMM block store
  * @name STORE_BLOCK_BOUNDARY_AWARE
  * This macro assumes the following schemes to achieve boundary-awareness:
  *  - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
  *  - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
  *  - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
  * The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
  *
  * In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
  * blocks **at the end**.
  * Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
  * "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
  *
  *  *--x-->                         x == 0                        x == 1
  *  |                  |<------------------------------N-------------------------->|
  *  y                  |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
  *  |     -------------#############################################################
  *  *     |          | |...............................|...........................|
  * y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
  *        |          | |...............................|...........................|
  *        M          --#############################################################
  *        |          | |                               |...........................|
  * y == 1 |         M0 |      Non-boundary block       |....Boundary block in x....|
  *        |          | |                               |...........................|
  *        |------------#############################################################
  *
  * Then @p PARTIAL_STORE_M0 = M % M0      and @p PARTIAL_STORE_N0 = N % N0
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
  * and select corresponding store methods such that the boundary detection logic is only added when needed.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  * @{
  */
 #if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case1: No partial blocks in either x or y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

 #elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
 // Case2: Partial blocks in y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)

 #elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
 // Case3: Partial blocks in x
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X)

 #else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case4: Partial blocks in both x and y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X)

 #endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0

 #endif    // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
 /** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE

 #if defined(PARTIAL_STORE_M0)
 /** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
  * @name COMPUTE_M0_START_ROW
  * If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
  * This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
  * blocks in the y dimension to avoid any padding.
  * EG: M0=4, PARTIAL_STORE_M0=1:
  *                  | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
  * block 0 (partial)| start row = 0   | start row = 0
  * block 1 (full)   | start row = 4   | start row = 1
  * block 2 (full)   | start row = 8   | start row = 5
  *
  * @param[in] y                Global id of current block in y.
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @{
  */
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
 #else // defined(PARTIAL_STORE_M0)
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(y * M0))
 #endif    // defined(PARTIAL_STORE_M0)
 /** @} */ // end of group COMPUTE_M0_START_ROW

 /** Store a vector that can only be partial in x.
  *
  * @note in case @p vec_size or @p leftover != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to end in a 0.
  * E.g., for basename=c, the expected name is c0.
  *
  * @param[in] basename  The name of the variable without trailing 0
  * @param[in] data_type The data type of the vector
  * @param[in] ptr       The base pointer
  * @param[in] vec_size  The vector size if cond = false. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] leftover  The vector size if cond = true. Supported range: [1, @p vec_size0)
  * @param[in] cond      Condition to select either vec_size0 or vec_size1
  * @{
  */
 #define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
     STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond)
 /** @} */ // end of group STORE_VECTOR_SELECT

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)

 #if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
 #pragma OPENCL EXTENSION cl_arm_printf : enable
 #endif // defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)

 #define GPU_ARCH_MIDGARD 0x100
 #define GPU_ARCH_BIFROST 0x200

 /** Concatenate two inputs.
  *
  * @param[in] a The first input to be concatenated
  * @param[in] b The second input to be concatenated
  *
  * @return The concatenated output
  */
 #define CONCAT(a, b) a##b

 /** Expand the given vector
  *
  * @param[in] x The vector to be expanded
  *
  * @return The expanded output
  */
 #define EXPAND(x) x

 /** Clamp the given value between an upper and lower bound.
  *
  * @param[in] x       The value to be clamped
  * @param[in] min_val The lower bound
  * @param[in] max_val The upper bound
  *
  * @return The clamped value.
  */
 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)

 /** REVn reverses the given vector whose size is n.
  * @name REVn
  *
  * @param[in] x The vector to be reversed
  *
  * @return The reversed vector
  * @{
  */
 #define REV1(x) ((x))
 #define REV2(x) ((x).s10)
 #define REV3(x) ((x).s210)
 #define REV4(x) ((x).s3210)
 #define REV8(x) ((x).s76543210)
 #define REV16(x) ((x).sFEDCBA9876543210)
 /** @} */ // end of group REVn

 /** Reverse the given vector.
  * @name REVERSE
  *
  * @param[in] x The vector to be reversed
  * @param[in] s The size of the vector
  *
  * @return The reversed vector
  * @{
  */
 #define REVERSE_STR(x, s) REV##s((x))
 #define REVERSE(x, s) REVERSE_STR(x, s)
 /** @} */ // end of group REVERSE

 /** Circular-right-shift (rotate-right) the vector of size s by the amount of n.
  * @name ROTs_n
  *
  * @param[in] x The vector to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROT1_0(x) ((x))

 #define ROT2_0(x) ((x))
 #define ROT2_1(x) ((x).s10)

 #define ROT3_0(x) ((x))
 #define ROT3_1(x) ((x).s201)
 #define ROT3_2(x) ((x).s120)

 #define ROT4_0(x) ((x))
 #define ROT4_1(x) ((x).s3012)
 #define ROT4_2(x) ((x).s2301)
 #define ROT4_3(x) ((x).s1230)

 #define ROT8_0(x) ((x))
 #define ROT8_1(x) ((x).s70123456)
 #define ROT8_2(x) ((x).s67012345)
 #define ROT8_3(x) ((x).s56701234)
 #define ROT8_4(x) ((x).s45670123)
 #define ROT8_5(x) ((x).s34567012)
 #define ROT8_6(x) ((x).s23456701)
 #define ROT8_7(x) ((x).s12345670)

 #define ROT16_0(x) ((x))
 #define ROT16_1(x) ((x).sF0123456789ABCDE)
 #define ROT16_2(x) ((x).sEF0123456789ABCD)
 #define ROT16_3(x) ((x).sDEF0123456789ABC)
 #define ROT16_4(x) ((x).sCDEF0123456789AB)
 #define ROT16_5(x) ((x).sBCDEF0123456789A)
 #define ROT16_6(x) ((x).sABCDEF0123456789)
 #define ROT16_7(x) ((x).s9ABCDEF012345678)
 #define ROT16_8(x) ((x).s89ABCDEF01234567)
 #define ROT16_9(x) ((x).s789ABCDEF0123456)
 #define ROT16_10(x) ((x).s6789ABCDEF012345)
 #define ROT16_11(x) ((x).s56789ABCDEF01234)
 #define ROT16_12(x) ((x).s456789ABCDEF0123)
 #define ROT16_13(x) ((x).s3456789ABCDEF012)
 #define ROT16_14(x) ((x).s23456789ABCDEF01)
 #define ROT16_15(x) ((x).s123456789ABCDEF0)
 /** @} */ // end of group ROTs_n

 /** Circular-right-shift (rotate-right) the given vector by the given amount.
  * @name ROTATE
  *
  * @param[in] x The vector to be shifted
  * @param[in] s The size of the vector
  * @param[in] n The amount to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROTATE_STR(x, s, n) ROT##s##_##n(x)
 #define ROTATE(x, s, n) ROTATE_STR(x, s, n)
 /** @} */ // end of group ROTATE

 /** Creates a vector of size n filled with offset values corresponding to the location of each element.
  * @name V_OFFSn
  *
  * @param[in] dt The data type of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define V_OFFS1(dt) (dt##1)(0)
 #define V_OFFS2(dt) (dt##2)(0, 1)
 #define V_OFFS3(dt) (dt##3)(0, 1, 2)
 #define V_OFFS4(dt) (dt##4)(0, 1, 2, 3)
 #define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7)
 #define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
 /** @} */ // end of group V_OFFSn

 /** Create a vector filled with offset values corresponding to the location of each element.
  * @name VEC_OFFS
  *
  * @param[in] dt The data type of the output vector
  * @param[in] s  The size of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define VEC_OFFS_STR(dt, s) V_OFFS##s(dt)
 #define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s)
 /** @} */ // end of group VEC_OFFS

 #define VLOAD_STR(size) vload##size
 #define VLOAD(size) VLOAD_STR(size)

 #define PIXEL_UNIT4 1
 #define PIXEL_UNIT8 2
 #define PIXEL_UNIT16 4

 /** Utility macro to convert a vector size in pixel unit.
  *
  * @name CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT
  *
  * @param[in] vec_size Vector size. Only 4,8 and 16 is supported
  *
  * @return The pixel unit (number of pixels)
  * @{
  */
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size)
 /** @} */ // end of group CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT

 #define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord)));
 #define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord)));

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord)));
 #define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord)));
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 /** Utility macro to read a 2D OpenCL image object.
  *
  * @note Coordinates are not normalized
  *
  * @param[in] data_type Data type
  * @param[in] n0        Number of pixel to read. Only 1,2 and 4 is supported
  * @param[in] img       OpenCL image object
  * @param[in] x_coord   The x coordinate for the top-left pixel
  * @param[in] y_coord   The y coordinate for the top-left pixel
  *
  * @return Pixels from the 2D OpenCL image object
  * @{
  */
 #define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord)
 #define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord)

 #define VSTORE_STR(size) vstore##size
 #define VSTORE(size) VSTORE_STR(size)

 #define float1 float
 #define half1 half
 #define char1 char
 #define uchar1 uchar
 #define short1 short
 #define ushort1 ushort
 #define int1 int
 #define uint1 uint
 #define long1 long
 #define ulong1 ulong
 #define double1 double

 #define vload1(OFFSET, PTR) *(OFFSET + PTR)
 #define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA

 /** Extended partial vstore that correctly handles scalar values as well.
  * Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name VSTORE_PARTIAL
  *
  * @note With this macro, the passed data can be both a vector and a scalar
  * @note @p store_size needs to be <= @p size
  * eg 1: Valid
  * VSTORE_PARTIAL(16, 15) ...;
  * eg 2: Invalid
  * VSTORE_PARTIAL(4, 7) ...;
  *
  * @param[in] size       The width of @p DATA. Supported values: 1(scalar), 2, 3, 4, 8, 16
  * @param[in] store_size The number of lower elements to store. Supported values: 1-16, but has to be <= @p size
  * @{
  */
 #define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
 #define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)

 #define NO_STORE(data, offs, ptr) \
     {                             \
     }

 // Size == 1 (scalar)
 #define vstore_partial_1_0 NO_STORE
 #define vstore_partial_1_1 vstore1
 #define vstore_partial_1_2 NO_STORE
 #define vstore_partial_1_3 NO_STORE
 #define vstore_partial_1_4 NO_STORE
 #define vstore_partial_1_5 NO_STORE
 #define vstore_partial_1_6 NO_STORE
 #define vstore_partial_1_7 NO_STORE
 #define vstore_partial_1_8 NO_STORE
 #define vstore_partial_1_9 NO_STORE
 #define vstore_partial_1_10 NO_STORE
 #define vstore_partial_1_11 NO_STORE
 #define vstore_partial_1_12 NO_STORE
 #define vstore_partial_1_13 NO_STORE
 #define vstore_partial_1_14 NO_STORE
 #define vstore_partial_1_15 NO_STORE
 #define vstore_partial_1_16 NO_STORE
 // Size == 2
 #define vstore_partial_2_0 NO_STORE
 #define vstore_partial_2_1 vstore_partial_1
 #define vstore_partial_2_2 vstore_partial_2
 #define vstore_partial_2_3 NO_STORE
 #define vstore_partial_2_4 NO_STORE
 #define vstore_partial_2_5 NO_STORE
 #define vstore_partial_2_6 NO_STORE
 #define vstore_partial_2_7 NO_STORE
 #define vstore_partial_2_8 NO_STORE
 #define vstore_partial_2_9 NO_STORE
 #define vstore_partial_2_10 NO_STORE
 #define vstore_partial_2_11 NO_STORE
 #define vstore_partial_2_12 NO_STORE
 #define vstore_partial_2_13 NO_STORE
 #define vstore_partial_2_14 NO_STORE
 #define vstore_partial_2_15 NO_STORE
 #define vstore_partial_2_16 NO_STORE
 // Size == 3
 #define vstore_partial_3_0 NO_STORE
 #define vstore_partial_3_1 vstore_partial_1
 #define vstore_partial_3_2 vstore_partial_2
 #define vstore_partial_3_3 vstore_partial_3
 #define vstore_partial_3_4 NO_STORE
 #define vstore_partial_3_5 NO_STORE
 #define vstore_partial_3_6 NO_STORE
 #define vstore_partial_3_7 NO_STORE
 #define vstore_partial_3_8 NO_STORE
 #define vstore_partial_3_9 NO_STORE
 #define vstore_partial_3_10 NO_STORE
 #define vstore_partial_3_11 NO_STORE
 #define vstore_partial_3_12 NO_STORE
 #define vstore_partial_3_13 NO_STORE
 #define vstore_partial_3_14 NO_STORE
 #define vstore_partial_3_15 NO_STORE
 #define vstore_partial_3_16 NO_STORE
 // Size == 4
 #define vstore_partial_4_0 NO_STORE
 #define vstore_partial_4_1 vstore_partial_1
 #define vstore_partial_4_2 vstore_partial_2
 #define vstore_partial_4_3 vstore_partial_3
 #define vstore_partial_4_4 vstore_partial_4
 #define vstore_partial_4_5 NO_STORE
 #define vstore_partial_4_6 NO_STORE
 #define vstore_partial_4_7 NO_STORE
 #define vstore_partial_4_8 NO_STORE
 #define vstore_partial_4_9 NO_STORE
 #define vstore_partial_4_10 NO_STORE
 #define vstore_partial_4_11 NO_STORE
 #define vstore_partial_4_12 NO_STORE
 #define vstore_partial_4_13 NO_STORE
 #define vstore_partial_4_14 NO_STORE
 #define vstore_partial_4_15 NO_STORE
 #define vstore_partial_4_16 NO_STORE
 // Size == 8
 #define vstore_partial_8_0 NO_STORE
 #define vstore_partial_8_1 vstore_partial_1
 #define vstore_partial_8_2 vstore_partial_2
 #define vstore_partial_8_3 vstore_partial_3
 #define vstore_partial_8_4 vstore_partial_4
 #define vstore_partial_8_5 vstore_partial_5
 #define vstore_partial_8_6 vstore_partial_6
 #define vstore_partial_8_7 vstore_partial_7
 #define vstore_partial_8_8 vstore_partial_8
 #define vstore_partial_8_9 NO_STORE
 #define vstore_partial_8_10 NO_STORE
 #define vstore_partial_8_11 NO_STORE
 #define vstore_partial_8_12 NO_STORE
 #define vstore_partial_8_13 NO_STORE
 #define vstore_partial_8_14 NO_STORE
 #define vstore_partial_8_15 NO_STORE
 #define vstore_partial_8_16 NO_STORE
 // Size == 16
 #define vstore_partial_16_0 NO_STORE
 #define vstore_partial_16_1 vstore_partial_1
 #define vstore_partial_16_2 vstore_partial_2
 #define vstore_partial_16_3 vstore_partial_3
 #define vstore_partial_16_4 vstore_partial_4
 #define vstore_partial_16_5 vstore_partial_5
 #define vstore_partial_16_6 vstore_partial_6
 #define vstore_partial_16_7 vstore_partial_7
 #define vstore_partial_16_8 vstore_partial_8
 #define vstore_partial_16_9 vstore_partial_9
 #define vstore_partial_16_10 vstore_partial_10
 #define vstore_partial_16_11 vstore_partial_11
 #define vstore_partial_16_12 vstore_partial_12
 #define vstore_partial_16_13 vstore_partial_13
 #define vstore_partial_16_14 vstore_partial_14
 #define vstore_partial_16_15 vstore_partial_15
 #define vstore_partial_16_16 vstore_partial_16

 /** Partial vstore. Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name vstore_partial_n
  *
  * @note @p DATA needs to be a vector not a scalar
  * @note n needs to be <= the vector width of the input variable @p DATA
  * eg 1: Valid
  * vstore_partial_15(var:float16, 0, 0xabcd);
  * eg 2: Invalid
  * vstore_partial_7(var:float4, 0, 0xabcd);
  *
  * @note in cases n == 1, 2, 3, 4, 8, 16, no extra vstore is invoked, thus there's no performance penalty.
  *
  * @param[in] DATA   The name of the variable
  * @param[in] OFFSET Offset in n
  * @param[in] PTR    The base pointer
  * @{
  */
 #define vstore_partial_1(DATA, OFFSET, PTR) \
     vstore1(DATA.s0, OFFSET, PTR);

 #define vstore_partial_2(DATA, OFFSET, PTR) \
     vstore2(DATA.s01, OFFSET, PTR);

 #define vstore_partial_3(DATA, OFFSET, PTR) \
     vstore3(DATA.s012, OFFSET, PTR);

 #define vstore_partial_4(DATA, OFFSET, PTR) \
     vstore4(DATA.s0123, OFFSET, PTR);

 #define vstore_partial_5(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore1(DATA.s4, OFFSET, PTR + 4);

 #define vstore_partial_6(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_2(DATA.s45, OFFSET, PTR + 4);

 #define vstore_partial_7(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_3(DATA.s456, OFFSET, PTR + 4);

 #define vstore_partial_8(DATA, OFFSET, PTR) \
     vstore8(DATA.s01234567, OFFSET, PTR);

 #define vstore_partial_9(DATA, OFFSET, PTR)        \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore1(DATA.s8, OFFSET, PTR + 8);

 #define vstore_partial_10(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_2(DATA.s89, OFFSET, PTR + 8);

 #define vstore_partial_11(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_3(DATA.s89a, OFFSET, PTR + 8);

 #define vstore_partial_12(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8);

 #define vstore_partial_13(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_14(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_15(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_16(DATA, OFFSET, PTR) \
     vstore16(DATA, OFFSET, PTR);
 /** @} */ // end of groupd vstore_partial_n
 /** @} */ // end of groupd VSTORE_PARTIAL

 // Convert built-in functions with _sat modifier are not supported in floating point so we create defines
 // without _sat to overcome this issue
 #define convert_float_sat convert_float
 #define convert_float1_sat convert_float
 #define convert_float2_sat convert_float2
 #define convert_float3_sat convert_float3
 #define convert_float4_sat convert_float4
 #define convert_float8_sat convert_float8
 #define convert_float16_sat convert_float16
 #define convert_half_sat convert_float
 #define convert_half1_sat convert_half
 #define convert_half2_sat convert_half2
 #define convert_half3_sat convert_half3
 #define convert_half4_sat convert_half4
 #define convert_half8_sat convert_half8
 #define convert_half16_sat convert_half16

 #define convert_float1 convert_float
 #define convert_half1 convert_half
 #define convert_char1 convert_char
 #define convert_uchar1 convert_uchar
 #define convert_short1 convert_short
 #define convert_ushort1 convert_ushort
 #define convert_int1 convert_int
 #define convert_uint1 convert_uint
 #define convert_long1 convert_long
 #define convert_ulong1 convert_ulong
 #define convert_double1 convert_double

 #define convert_char1_sat convert_char_sat
 #define convert_uchar1_sat convert_uchar_sat
 #define convert_short1_sat convert_short_sat
 #define convert_ushort1_sat convert_ushort_sat
 #define convert_int1_sat convert_int_sat
 #define convert_uint1_sat convert_uint_sat
 #define convert_long1_sat convert_long_sat
 #define convert_ulong1_sat convert_ulong_sat
 #define convert_double1_sat convert_double_sat

 #define VEC_DATA_TYPE_STR(type, size) type##size
 #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)

 #define CONVERT_STR(x, type) (convert_##type((x)))
 #define CONVERT(x, type) CONVERT_STR(x, type)

 #define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x)))
 #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)

 #define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x)))
 #define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round)

 #define select_vec_dt_uchar(size) uchar##size
 #define select_vec_dt_char(size) char##size
 #define select_vec_dt_ushort(size) ushort##size
 #define select_vec_dt_short(size) short##size
 #define select_vec_dt_half(size) short##size
 #define select_vec_dt_uint(size) uint##size
 #define select_vec_dt_int(size) int##size
 #define select_vec_dt_float(size) int##size
 #define select_vec_dt_ulong(size) ulong##size
 #define select_vec_dt_long(size) long##size

 #define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size)
 #define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size)
 #define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1)

 #define sum_reduce_1(x) (x)
 #define sum_reduce_2(x) ((x).s0) + ((x).s1)
 #define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2)
 #define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23)
 #define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567)
 #define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF)

 #define SUM_REDUCE_STR(x, size) sum_reduce_##size(x)
 #define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size)

 #define max_reduce_1(x) (x)
 #define max_reduce_2(x) max(((x).s0), ((x).s1))
 #define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2))
 #define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23))
 #define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567))
 #define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF))

 #define MAX_REDUCE_STR(x, size) max_reduce_##size(x)
 #define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size)

 #define VECTOR_DECLARATION(name)     \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_offset_first_element_in_bytes

 #define IMAGE_DECLARATION(name)      \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR3D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR4D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_stride_w, \
     uint        name##_step_w,   \
     uint        name##_offset_first_element_in_bytes

 #define CONVERT_TO_VECTOR_STRUCT(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)

 #define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0)

 #define CONVERT_TO_IMAGE_STRUCT(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)

 #define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                           \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0)

 #define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size)                                                                                                 \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)

 #define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name)                                                                                       \
     tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                            name##_stride_z, name##_step_z)

 /** Structure to hold Vector information */
 typedef struct Vector
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
 } Vector;

 /** Structure to hold Image information */
 typedef struct Image
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
 } Image;

 /** Structure to hold 3D tensor information */
 typedef struct Tensor3D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
 } Tensor3D;

 /** Structure to hold 4D tensor information */
 typedef struct Tensor4D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
     int             stride_w;                      /**< Stride of the image in W dimension (in bytes) */
 } Tensor4D;

 /** Wrap vector information into an Vector structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
  * @param[in] stride_x                      Stride of the vector in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x)
 {
     Vector vector =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
     };
     vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x;
     return vector;
 }

 /** Wrap image information into an Image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y;
     return img;
 }

 /** Wrap 3D tensor information into an image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return img;
 }

 /** Wrap 3D tensor information into an tensor structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return tensor;
 }

 /** Wrap 3D tensor information into an tensor structure.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     return tensor;
 }

 inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w,
                                              uint step_w,
                                              uint mod_size)
 {
     Tensor4D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z,
         .stride_w                      = stride_w
     };

     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w;
     return tensor;
 }

 /** Get the pointer position of a Vector
  *
  * @param[in] vec Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  */
 inline __global const uchar *vector_offset(const Vector *vec, int x)
 {
     return vec->ptr + x * vec->stride_x;
 }

 /** Get the pointer position of a Image
  *
  * @param[in] img Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  * @param[in] y   Relative Y position
  */
 inline __global uchar *offset(const Image *img, int x, int y)
 {
     return img->ptr + x * img->stride_x + y * img->stride_y;
 }

 /** Get the pointer position of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  */
 inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z;
 }

 /** Get the pointer position of a Tensor4D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  * @param[in] w      Relative W position
  */
 inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w;
 }

 /** Get the offset for a given linear index of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] width  Width of the input tensor
  * @param[in] height Height of the input tensor
  * @param[in] depth  Depth of the input tensor
  * @param[in] index  Linear index
  */
 inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index)
 {
     uint num_elements = width * height;

     const uint z = index / num_elements;

     index %= num_elements;

     const uint y = index / width;

     index %= width;

     const uint x = index;

     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes;
 }

 #endif // _HELPER_H

 #if GPU_ARCH == GPU_ARCH_BIFROST
 #define MLA(a, b, c) (fma(c, b, a))
 #else // GPU_ARCH == GPU_ARCH_BIFROST
 #define MLA(a, b, c) ((b) * (c) + (a))
 #endif // GPU_ARCH == GPU_ARCH_BIFROST

 // Hard-Swish
 #define hard_swish_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (x * ((min(max((x + (DATA_TYPE)3.0), (DATA_TYPE)0.0), (DATA_TYPE)6.0)) * (DATA_TYPE)0.166666667))

 // Logistic Activation
 #define logistic_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ((DATA_TYPE)1.0 / ((DATA_TYPE)1.0 + exp(-x)))

 // Hyperbolic Tangent Activation
 #define tanh_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ((DATA_TYPE)A_VAL * tanh((DATA_TYPE)B_VAL * x))

 // RELU Tangent Activation
 #define relu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (max((DATA_TYPE)0.0, x))

 // Bounded RELU Activation
 #define brelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)0.0, x)))

 // Lower Upper Bounded RELU Activation
 #define lu_brelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))

 // Leaky RELU Activation
 #define lrelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ((min(x, (DATA_TYPE)0.0) * (DATA_TYPE)A_VAL) + max(x, (DATA_TYPE)0.0))

 // Soft RELU Activation
 #define srelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (log((DATA_TYPE)1.0 + exp(x)))

 // ELU Activation
 #define elu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (select(((DATA_TYPE)A_VAL * (exp(x) - (DATA_TYPE)1.0)), x, (SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))isgreaterequal(x, (DATA_TYPE)0.0)))

 // Absolute Activation
 #define abs_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (fabs(x))

 // Square Activation
 #define square_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (x * x)

 // Square-root Activation
 #define sqrt_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (sqrt(x))

 // Linear Activation
 #define linear_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (MLA((DATA_TYPE)B_VAL, (DATA_TYPE)A_VAL, x))

 // Identity Activation
 #define identity_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (x)

 #define ACT_OP(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) op##_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL)

 #define ACTIVATION(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ACT_OP(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL)
 /*
  * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #ifndef ARM_COMPUTE_HELPER_H
 #define ARM_COMPUTE_HELPER_H

 /*
  * Copyright (c) 2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */

 /** Store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_n
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE(N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_n

 /** Convert and store the 0th to (n-1)th rows of the given variables
  * @name CONVERT_STORE_ROW_n
  *
  * @param[in] N0        The size of the vectors
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                     \
     (CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));

 /** @} */ // end of groupd CONVERT_STORE_ROW_n

 /** Store a block of the given size M0xN0
  * @name STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group STORE_BLOCK

 /** Convert and store a block of the given size M0xN0
  * @name CONVERT_STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group CONVERT_STORE_BLOCK

 /** Partially store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_PARTIAL_n
  * Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
  *
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] STORE_N0  The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_PARTIAL_n

 /** Partially store a block of the given size STORE_M0xSTORE_N0
  * @name STORE_BLOCK_PARTIAL
  *
  * @note The vector width @p N0 is also required for correct partial storing behaviour.
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] STORE_M0  The number of rows to store. Supported: 1-16
  * @param[in] STORE_N0  The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
  * @param[in] N0        The size of each vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** Store a block that can be partial in both x and y dimensions
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                            \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                           \
     }                                                                                                                                                     \
     else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else                                                                                                                                                  \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                               \
     }
 /** Store a block that can only be partial in x but not y.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** Store a block that can only be partial in y but not x.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  */
 #define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
     if(!(PARTIAL_COND_Y))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** @} */ // end of group STORE_BLOCK_PARTIAL

 #if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)

 /** Boundary-aware GEMM block store
  * @name STORE_BLOCK_BOUNDARY_AWARE
  * This macro assumes the following schemes to achieve boundary-awareness:
  *  - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
  *  - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
  *  - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
  * The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
  *
  * In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
  * blocks **at the end**.
  * Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
  * "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
  *
  *  *--x-->                         x == 0                        x == 1
  *  |                  |<------------------------------N-------------------------->|
  *  y                  |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
  *  |     -------------#############################################################
  *  *     |          | |...............................|...........................|
  * y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
  *        |          | |...............................|...........................|
  *        M          --#############################################################
  *        |          | |                               |...........................|
  * y == 1 |         M0 |      Non-boundary block       |....Boundary block in x....|
  *        |          | |                               |...........................|
  *        |------------#############################################################
  *
  * Then @p PARTIAL_STORE_M0 = M % M0      and @p PARTIAL_STORE_N0 = N % N0
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
  * and select corresponding store methods such that the boundary detection logic is only added when needed.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  * @{
  */
 #if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case1: No partial blocks in either x or y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

 #elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
 // Case2: Partial blocks in y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)

 #elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
 // Case3: Partial blocks in x
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X)

 #else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case4: Partial blocks in both x and y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X)

 #endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0

 #endif    // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
 /** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE

 #if defined(PARTIAL_STORE_M0)
 /** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
  * @name COMPUTE_M0_START_ROW
  * If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
  * This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
  * blocks in the y dimension to avoid any padding.
  * EG: M0=4, PARTIAL_STORE_M0=1:
  *                  | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
  * block 0 (partial)| start row = 0   | start row = 0
  * block 1 (full)   | start row = 4   | start row = 1
  * block 2 (full)   | start row = 8   | start row = 5
  *
  * @param[in] y                Global id of current block in y.
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @{
  */
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
 #else // defined(PARTIAL_STORE_M0)
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(y * M0))
 #endif    // defined(PARTIAL_STORE_M0)
 /** @} */ // end of group COMPUTE_M0_START_ROW

 /** Store a vector that can only be partial in x.
  *
  * @note in case @p vec_size or @p leftover != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to end in a 0.
  * E.g., for basename=c, the expected name is c0.
  *
  * @param[in] basename  The name of the variable without trailing 0
  * @param[in] data_type The data type of the vector
  * @param[in] ptr       The base pointer
  * @param[in] vec_size  The vector size if cond = false. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] leftover  The vector size if cond = true. Supported range: [1, @p vec_size0)
  * @param[in] cond      Condition to select either vec_size0 or vec_size1
  * @{
  */
 #define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
     STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond)
 /** @} */ // end of group STORE_VECTOR_SELECT

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)

 #if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
 #pragma OPENCL EXTENSION cl_arm_printf : enable
 #endif // defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)

 #define GPU_ARCH_MIDGARD 0x100
 #define GPU_ARCH_BIFROST 0x200

 /** Concatenate two inputs.
  *
  * @param[in] a The first input to be concatenated
  * @param[in] b The second input to be concatenated
  *
  * @return The concatenated output
  */
 #define CONCAT(a, b) a##b

 /** Expand the given vector
  *
  * @param[in] x The vector to be expanded
  *
  * @return The expanded output
  */
 #define EXPAND(x) x

 /** Clamp the given value between an upper and lower bound.
  *
  * @param[in] x       The value to be clamped
  * @param[in] min_val The lower bound
  * @param[in] max_val The upper bound
  *
  * @return The clamped value.
  */
 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)

 /** REVn reverses the given vector whose size is n.
  * @name REVn
  *
  * @param[in] x The vector to be reversed
  *
  * @return The reversed vector
  * @{
  */
 #define REV1(x) ((x))
 #define REV2(x) ((x).s10)
 #define REV3(x) ((x).s210)
 #define REV4(x) ((x).s3210)
 #define REV8(x) ((x).s76543210)
 #define REV16(x) ((x).sFEDCBA9876543210)
 /** @} */ // end of group REVn

 /** Reverse the given vector.
  * @name REVERSE
  *
  * @param[in] x The vector to be reversed
  * @param[in] s The size of the vector
  *
  * @return The reversed vector
  * @{
  */
 #define REVERSE_STR(x, s) REV##s((x))
 #define REVERSE(x, s) REVERSE_STR(x, s)
 /** @} */ // end of group REVERSE

 /** Circular-right-shift (rotate-right) the vector of size s by the amount of n.
  * @name ROTs_n
  *
  * @param[in] x The vector to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROT1_0(x) ((x))

 #define ROT2_0(x) ((x))
 #define ROT2_1(x) ((x).s10)

 #define ROT3_0(x) ((x))
 #define ROT3_1(x) ((x).s201)
 #define ROT3_2(x) ((x).s120)

 #define ROT4_0(x) ((x))
 #define ROT4_1(x) ((x).s3012)
 #define ROT4_2(x) ((x).s2301)
 #define ROT4_3(x) ((x).s1230)

 #define ROT8_0(x) ((x))
 #define ROT8_1(x) ((x).s70123456)
 #define ROT8_2(x) ((x).s67012345)
 #define ROT8_3(x) ((x).s56701234)
 #define ROT8_4(x) ((x).s45670123)
 #define ROT8_5(x) ((x).s34567012)
 #define ROT8_6(x) ((x).s23456701)
 #define ROT8_7(x) ((x).s12345670)

 #define ROT16_0(x) ((x))
 #define ROT16_1(x) ((x).sF0123456789ABCDE)
 #define ROT16_2(x) ((x).sEF0123456789ABCD)
 #define ROT16_3(x) ((x).sDEF0123456789ABC)
 #define ROT16_4(x) ((x).sCDEF0123456789AB)
 #define ROT16_5(x) ((x).sBCDEF0123456789A)
 #define ROT16_6(x) ((x).sABCDEF0123456789)
 #define ROT16_7(x) ((x).s9ABCDEF012345678)
 #define ROT16_8(x) ((x).s89ABCDEF01234567)
 #define ROT16_9(x) ((x).s789ABCDEF0123456)
 #define ROT16_10(x) ((x).s6789ABCDEF012345)
 #define ROT16_11(x) ((x).s56789ABCDEF01234)
 #define ROT16_12(x) ((x).s456789ABCDEF0123)
 #define ROT16_13(x) ((x).s3456789ABCDEF012)
 #define ROT16_14(x) ((x).s23456789ABCDEF01)
 #define ROT16_15(x) ((x).s123456789ABCDEF0)
 /** @} */ // end of group ROTs_n

 /** Circular-right-shift (rotate-right) the given vector by the given amount.
  * @name ROTATE
  *
  * @param[in] x The vector to be shifted
  * @param[in] s The size of the vector
  * @param[in] n The amount to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROTATE_STR(x, s, n) ROT##s##_##n(x)
 #define ROTATE(x, s, n) ROTATE_STR(x, s, n)
 /** @} */ // end of group ROTATE

 /** Creates a vector of size n filled with offset values corresponding to the location of each element.
  * @name V_OFFSn
  *
  * @param[in] dt The data type of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define V_OFFS1(dt) (dt##1)(0)
 #define V_OFFS2(dt) (dt##2)(0, 1)
 #define V_OFFS3(dt) (dt##3)(0, 1, 2)
 #define V_OFFS4(dt) (dt##4)(0, 1, 2, 3)
 #define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7)
 #define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
 /** @} */ // end of group V_OFFSn

 /** Create a vector filled with offset values corresponding to the location of each element.
  * @name VEC_OFFS
  *
  * @param[in] dt The data type of the output vector
  * @param[in] s  The size of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define VEC_OFFS_STR(dt, s) V_OFFS##s(dt)
 #define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s)
 /** @} */ // end of group VEC_OFFS

 #define VLOAD_STR(size) vload##size
 #define VLOAD(size) VLOAD_STR(size)

 #define PIXEL_UNIT4 1
 #define PIXEL_UNIT8 2
 #define PIXEL_UNIT16 4

 /** Utility macro to convert a vector size in pixel unit.
  *
  * @name CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT
  *
  * @param[in] vec_size Vector size. Only 4,8 and 16 is supported
  *
  * @return The pixel unit (number of pixels)
  * @{
  */
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size)
 /** @} */ // end of group CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT

 #define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord)));
 #define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord)));

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord)));
 #define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord)));
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 /** Utility macro to read a 2D OpenCL image object.
  *
  * @note Coordinates are not normalized
  *
  * @param[in] data_type Data type
  * @param[in] n0        Number of pixel to read. Only 1,2 and 4 is supported
  * @param[in] img       OpenCL image object
  * @param[in] x_coord   The x coordinate for the top-left pixel
  * @param[in] y_coord   The y coordinate for the top-left pixel
  *
  * @return Pixels from the 2D OpenCL image object
  * @{
  */
 #define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord)
 #define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord)

 #define VSTORE_STR(size) vstore##size
 #define VSTORE(size) VSTORE_STR(size)

 #define float1 float
 #define half1 half
 #define char1 char
 #define uchar1 uchar
 #define short1 short
 #define ushort1 ushort
 #define int1 int
 #define uint1 uint
 #define long1 long
 #define ulong1 ulong
 #define double1 double

 #define vload1(OFFSET, PTR) *(OFFSET + PTR)
 #define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA

 /** Extended partial vstore that correctly handles scalar values as well.
  * Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name VSTORE_PARTIAL
  *
  * @note With this macro, the passed data can be both a vector and a scalar
  * @note @p store_size needs to be <= @p size
  * eg 1: Valid
  * VSTORE_PARTIAL(16, 15) ...;
  * eg 2: Invalid
  * VSTORE_PARTIAL(4, 7) ...;
  *
  * @param[in] size       The width of @p DATA. Supported values: 1(scalar), 2, 3, 4, 8, 16
  * @param[in] store_size The number of lower elements to store. Supported values: 1-16, but has to be <= @p size
  * @{
  */
 #define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
 #define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)

 #define NO_STORE(data, offs, ptr) \
     {                             \
     }

 // Size == 1 (scalar)
 #define vstore_partial_1_0 NO_STORE
 #define vstore_partial_1_1 vstore1
 #define vstore_partial_1_2 NO_STORE
 #define vstore_partial_1_3 NO_STORE
 #define vstore_partial_1_4 NO_STORE
 #define vstore_partial_1_5 NO_STORE
 #define vstore_partial_1_6 NO_STORE
 #define vstore_partial_1_7 NO_STORE
 #define vstore_partial_1_8 NO_STORE
 #define vstore_partial_1_9 NO_STORE
 #define vstore_partial_1_10 NO_STORE
 #define vstore_partial_1_11 NO_STORE
 #define vstore_partial_1_12 NO_STORE
 #define vstore_partial_1_13 NO_STORE
 #define vstore_partial_1_14 NO_STORE
 #define vstore_partial_1_15 NO_STORE
 #define vstore_partial_1_16 NO_STORE
 // Size == 2
 #define vstore_partial_2_0 NO_STORE
 #define vstore_partial_2_1 vstore_partial_1
 #define vstore_partial_2_2 vstore_partial_2
 #define vstore_partial_2_3 NO_STORE
 #define vstore_partial_2_4 NO_STORE
 #define vstore_partial_2_5 NO_STORE
 #define vstore_partial_2_6 NO_STORE
 #define vstore_partial_2_7 NO_STORE
 #define vstore_partial_2_8 NO_STORE
 #define vstore_partial_2_9 NO_STORE
 #define vstore_partial_2_10 NO_STORE
 #define vstore_partial_2_11 NO_STORE
 #define vstore_partial_2_12 NO_STORE
 #define vstore_partial_2_13 NO_STORE
 #define vstore_partial_2_14 NO_STORE
 #define vstore_partial_2_15 NO_STORE
 #define vstore_partial_2_16 NO_STORE
 // Size == 3
 #define vstore_partial_3_0 NO_STORE
 #define vstore_partial_3_1 vstore_partial_1
 #define vstore_partial_3_2 vstore_partial_2
 #define vstore_partial_3_3 vstore_partial_3
 #define vstore_partial_3_4 NO_STORE
 #define vstore_partial_3_5 NO_STORE
 #define vstore_partial_3_6 NO_STORE
 #define vstore_partial_3_7 NO_STORE
 #define vstore_partial_3_8 NO_STORE
 #define vstore_partial_3_9 NO_STORE
 #define vstore_partial_3_10 NO_STORE
 #define vstore_partial_3_11 NO_STORE
 #define vstore_partial_3_12 NO_STORE
 #define vstore_partial_3_13 NO_STORE
 #define vstore_partial_3_14 NO_STORE
 #define vstore_partial_3_15 NO_STORE
 #define vstore_partial_3_16 NO_STORE
 // Size == 4
 #define vstore_partial_4_0 NO_STORE
 #define vstore_partial_4_1 vstore_partial_1
 #define vstore_partial_4_2 vstore_partial_2
 #define vstore_partial_4_3 vstore_partial_3
 #define vstore_partial_4_4 vstore_partial_4
 #define vstore_partial_4_5 NO_STORE
 #define vstore_partial_4_6 NO_STORE
 #define vstore_partial_4_7 NO_STORE
 #define vstore_partial_4_8 NO_STORE
 #define vstore_partial_4_9 NO_STORE
 #define vstore_partial_4_10 NO_STORE
 #define vstore_partial_4_11 NO_STORE
 #define vstore_partial_4_12 NO_STORE
 #define vstore_partial_4_13 NO_STORE
 #define vstore_partial_4_14 NO_STORE
 #define vstore_partial_4_15 NO_STORE
 #define vstore_partial_4_16 NO_STORE
 // Size == 8
 #define vstore_partial_8_0 NO_STORE
 #define vstore_partial_8_1 vstore_partial_1
 #define vstore_partial_8_2 vstore_partial_2
 #define vstore_partial_8_3 vstore_partial_3
 #define vstore_partial_8_4 vstore_partial_4
 #define vstore_partial_8_5 vstore_partial_5
 #define vstore_partial_8_6 vstore_partial_6
 #define vstore_partial_8_7 vstore_partial_7
 #define vstore_partial_8_8 vstore_partial_8
 #define vstore_partial_8_9 NO_STORE
 #define vstore_partial_8_10 NO_STORE
 #define vstore_partial_8_11 NO_STORE
 #define vstore_partial_8_12 NO_STORE
 #define vstore_partial_8_13 NO_STORE
 #define vstore_partial_8_14 NO_STORE
 #define vstore_partial_8_15 NO_STORE
 #define vstore_partial_8_16 NO_STORE
 // Size == 16
 #define vstore_partial_16_0 NO_STORE
 #define vstore_partial_16_1 vstore_partial_1
 #define vstore_partial_16_2 vstore_partial_2
 #define vstore_partial_16_3 vstore_partial_3
 #define vstore_partial_16_4 vstore_partial_4
 #define vstore_partial_16_5 vstore_partial_5
 #define vstore_partial_16_6 vstore_partial_6
 #define vstore_partial_16_7 vstore_partial_7
 #define vstore_partial_16_8 vstore_partial_8
 #define vstore_partial_16_9 vstore_partial_9
 #define vstore_partial_16_10 vstore_partial_10
 #define vstore_partial_16_11 vstore_partial_11
 #define vstore_partial_16_12 vstore_partial_12
 #define vstore_partial_16_13 vstore_partial_13
 #define vstore_partial_16_14 vstore_partial_14
 #define vstore_partial_16_15 vstore_partial_15
 #define vstore_partial_16_16 vstore_partial_16

 /** Partial vstore. Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name vstore_partial_n
  *
  * @note @p DATA needs to be a vector not a scalar
  * @note n needs to be <= the vector width of the input variable @p DATA
  * eg 1: Valid
  * vstore_partial_15(var:float16, 0, 0xabcd);
  * eg 2: Invalid
  * vstore_partial_7(var:float4, 0, 0xabcd);
  *
  * @note in cases n == 1, 2, 3, 4, 8, 16, no extra vstore is invoked, thus there's no performance penalty.
  *
  * @param[in] DATA   The name of the variable
  * @param[in] OFFSET Offset in n
  * @param[in] PTR    The base pointer
  * @{
  */
 #define vstore_partial_1(DATA, OFFSET, PTR) \
     vstore1(DATA.s0, OFFSET, PTR);

 #define vstore_partial_2(DATA, OFFSET, PTR) \
     vstore2(DATA.s01, OFFSET, PTR);

 #define vstore_partial_3(DATA, OFFSET, PTR) \
     vstore3(DATA.s012, OFFSET, PTR);

 #define vstore_partial_4(DATA, OFFSET, PTR) \
     vstore4(DATA.s0123, OFFSET, PTR);

 #define vstore_partial_5(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore1(DATA.s4, OFFSET, PTR + 4);

 #define vstore_partial_6(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_2(DATA.s45, OFFSET, PTR + 4);

 #define vstore_partial_7(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_3(DATA.s456, OFFSET, PTR + 4);

 #define vstore_partial_8(DATA, OFFSET, PTR) \
     vstore8(DATA.s01234567, OFFSET, PTR);

 #define vstore_partial_9(DATA, OFFSET, PTR)        \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore1(DATA.s8, OFFSET, PTR + 8);

 #define vstore_partial_10(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_2(DATA.s89, OFFSET, PTR + 8);

 #define vstore_partial_11(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_3(DATA.s89a, OFFSET, PTR + 8);

 #define vstore_partial_12(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8);

 #define vstore_partial_13(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_14(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_15(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_16(DATA, OFFSET, PTR) \
     vstore16(DATA, OFFSET, PTR);
 /** @} */ // end of groupd vstore_partial_n
 /** @} */ // end of groupd VSTORE_PARTIAL

 // Convert built-in functions with _sat modifier are not supported in floating point so we create defines
 // without _sat to overcome this issue
 #define convert_float_sat convert_float
 #define convert_float1_sat convert_float
 #define convert_float2_sat convert_float2
 #define convert_float3_sat convert_float3
 #define convert_float4_sat convert_float4
 #define convert_float8_sat convert_float8
 #define convert_float16_sat convert_float16
 #define convert_half_sat convert_float
 #define convert_half1_sat convert_half
 #define convert_half2_sat convert_half2
 #define convert_half3_sat convert_half3
 #define convert_half4_sat convert_half4
 #define convert_half8_sat convert_half8
 #define convert_half16_sat convert_half16

 #define convert_float1 convert_float
 #define convert_half1 convert_half
 #define convert_char1 convert_char
 #define convert_uchar1 convert_uchar
 #define convert_short1 convert_short
 #define convert_ushort1 convert_ushort
 #define convert_int1 convert_int
 #define convert_uint1 convert_uint
 #define convert_long1 convert_long
 #define convert_ulong1 convert_ulong
 #define convert_double1 convert_double

 #define convert_char1_sat convert_char_sat
 #define convert_uchar1_sat convert_uchar_sat
 #define convert_short1_sat convert_short_sat
 #define convert_ushort1_sat convert_ushort_sat
 #define convert_int1_sat convert_int_sat
 #define convert_uint1_sat convert_uint_sat
 #define convert_long1_sat convert_long_sat
 #define convert_ulong1_sat convert_ulong_sat
 #define convert_double1_sat convert_double_sat

 #define VEC_DATA_TYPE_STR(type, size) type##size
 #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)

 #define CONVERT_STR(x, type) (convert_##type((x)))
 #define CONVERT(x, type) CONVERT_STR(x, type)

 #define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x)))
 #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)

 #define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x)))
 #define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round)

 #define select_vec_dt_uchar(size) uchar##size
 #define select_vec_dt_char(size) char##size
 #define select_vec_dt_ushort(size) ushort##size
 #define select_vec_dt_short(size) short##size
 #define select_vec_dt_half(size) short##size
 #define select_vec_dt_uint(size) uint##size
 #define select_vec_dt_int(size) int##size
 #define select_vec_dt_float(size) int##size
 #define select_vec_dt_ulong(size) ulong##size
 #define select_vec_dt_long(size) long##size

 #define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size)
 #define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size)
 #define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1)

 #define sum_reduce_1(x) (x)
 #define sum_reduce_2(x) ((x).s0) + ((x).s1)
 #define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2)
 #define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23)
 #define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567)
 #define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF)

 #define SUM_REDUCE_STR(x, size) sum_reduce_##size(x)
 #define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size)

 #define max_reduce_1(x) (x)
 #define max_reduce_2(x) max(((x).s0), ((x).s1))
 #define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2))
 #define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23))
 #define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567))
 #define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF))

 #define MAX_REDUCE_STR(x, size) max_reduce_##size(x)
 #define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size)

 #define VECTOR_DECLARATION(name)     \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_offset_first_element_in_bytes

 #define IMAGE_DECLARATION(name)      \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR3D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR4D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_stride_w, \
     uint        name##_step_w,   \
     uint        name##_offset_first_element_in_bytes

 #define CONVERT_TO_VECTOR_STRUCT(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)

 #define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0)

 #define CONVERT_TO_IMAGE_STRUCT(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)

 #define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                           \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0)

 #define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size)                                                                                                 \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)

 #define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name)                                                                                       \
     tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                            name##_stride_z, name##_step_z)

 /** Structure to hold Vector information */
 typedef struct Vector
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
 } Vector;

 /** Structure to hold Image information */
 typedef struct Image
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
 } Image;

 /** Structure to hold 3D tensor information */
 typedef struct Tensor3D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
 } Tensor3D;

 /** Structure to hold 4D tensor information */
 typedef struct Tensor4D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
     int             stride_w;                      /**< Stride of the image in W dimension (in bytes) */
 } Tensor4D;

 /** Wrap vector information into an Vector structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
  * @param[in] stride_x                      Stride of the vector in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x)
 {
     Vector vector =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
     };
     vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x;
     return vector;
 }

 /** Wrap image information into an Image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y;
     return img;
 }

 /** Wrap 3D tensor information into an image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return img;
 }

 /** Wrap 3D tensor information into an tensor structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return tensor;
 }

 /** Wrap 3D tensor information into an tensor structure.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     return tensor;
 }

 inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w,
                                              uint step_w,
                                              uint mod_size)
 {
     Tensor4D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z,
         .stride_w                      = stride_w
     };

     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w;
     return tensor;
 }

 /** Get the pointer position of a Vector
  *
  * @param[in] vec Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  */
 inline __global const uchar *vector_offset(const Vector *vec, int x)
 {
     return vec->ptr + x * vec->stride_x;
 }

 /** Get the pointer position of a Image
  *
  * @param[in] img Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  * @param[in] y   Relative Y position
  */
 inline __global uchar *offset(const Image *img, int x, int y)
 {
     return img->ptr + x * img->stride_x + y * img->stride_y;
 }

 /** Get the pointer position of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  */
 inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z;
 }

 /** Get the pointer position of a Tensor4D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  * @param[in] w      Relative W position
  */
 inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w;
 }

 /** Get the offset for a given linear index of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] width  Width of the input tensor
  * @param[in] height Height of the input tensor
  * @param[in] depth  Depth of the input tensor
  * @param[in] index  Linear index
  */
 inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index)
 {
     uint num_elements = width * height;

     const uint z = index / num_elements;

     index %= num_elements;

     const uint y = index / width;

     index %= width;

     const uint x = index;

     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes;
 }

 #endif // _HELPER_H

 /** Utility macro to access a vector with the scalar positions
  *
  * Supported cases are: Offset can only be of the same size of the OpenCL vector (2,3,4,8,16)
  *
  * @param[in] offset The offset within the vector. Offset can only be of the same size of the OpenCL vector (2,3,4,8,16)
  * @param[in] n0     The number of consecutive columns to access. n0 + offset must be <= 16
  * @param[in] x      Vector to access
  * @{
  */
 #define SCALAR_ACCESS_STR(offset, n0, x) scalar_access_##offset##_##n0(x)
 #define SCALAR_ACCESS(offset, n0, x) SCALAR_ACCESS_STR(offset, n0, x)

 // offset == 0
 #define scalar_access_0_1(x) ((x).s0)
 #define scalar_access_0_2(x) ((x).s01)
 #define scalar_access_0_3(x) ((x).s012)
 #define scalar_access_0_4(x) ((x).s0123)
 #define scalar_access_0_8(x) ((x).s01234567)
 #define scalar_access_0_16(x) ((x).s0123456789ABCDEF)

 // offset == 1
 #define scalar_access_1_1(x) ((x).s1)
 #define scalar_access_1_2(x) ((x).s12)
 #define scalar_access_1_3(x) ((x).s123)
 #define scalar_access_1_4(x) ((x).s1234)
 #define scalar_access_1_8(x) ((x).s12345678)

 // offset == 2
 #define scalar_access_2_1(x) ((x).s2)
 #define scalar_access_2_2(x) ((x).s23)
 #define scalar_access_2_3(x) ((x).s234)
 #define scalar_access_2_4(x) ((x).s2345)
 #define scalar_access_2_8(x) ((x).s23456789)

 // offset == 3
 #define scalar_access_3_1(x) ((x).s3)
 #define scalar_access_3_2(x) ((x).s34)
 #define scalar_access_3_3(x) ((x).s345)
 #define scalar_access_3_4(x) ((x).s3456)
 #define scalar_access_3_8(x) ((x).s3456789A)

 // offset == 4
 #define scalar_access_4_1(x) ((x).s4)
 #define scalar_access_4_2(x) ((x).s45)
 #define scalar_access_4_3(x) ((x).s456)
 #define scalar_access_4_4(x) ((x).s4567)
 #define scalar_access_4_8(x) ((x).s456789AB)

 // offset == 8
 #define scalar_access_8_1(x) ((x).s8)
 #define scalar_access_8_2(x) ((x).s89)
 #define scalar_access_8_3(x) ((x).s89A)
 #define scalar_access_8_4(x) ((x).s89AB)
 #define scalar_access_8_8(x) ((x).s89ABCDEF)

 // offset == 12
 #define scalar_access_12_1(x) ((x).sC)
 #define scalar_access_12_2(x) ((x).sCD)
 #define scalar_access_12_3(x) ((x).sCDE)
 #define scalar_access_12_4(x) ((x).sCDEF)

 // offset == 16
 #define scalar_access_16_1(x) ((x).sF)

 /** Loads the rows from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1) without allocating variables.
  * @name LOAD_TENSOR_ROW_n
  *
  * @param[in] N0         The number of columns to load
  * @param[in] DATA_TYPE  The data type of variables
  * @param[in] BASENAME   The basename of the destination variables for the loaded rows
  * @param[in] PTR        The base pointer
  * @param[in] COL_OFFSET The column vector offset. COL_OFFSET + N0 must be <= 16
  * @param[in] STRIDE_Y   The stride value in y-axis direction
  * @param[in] Z          The z-axis offset vector
  * @{
  */
 #define LOAD_TENSOR_ROW_0(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     ({})

 #define LOAD_TENSOR_ROW_1(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##0) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define LOAD_TENSOR_ROW_2(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_1(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##1) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define LOAD_TENSOR_ROW_3(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_2(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##2) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define LOAD_TENSOR_ROW_4(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_3(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##3) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define LOAD_TENSOR_ROW_5(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_4(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##4) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define LOAD_TENSOR_ROW_6(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_5(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##5) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define LOAD_TENSOR_ROW_7(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_6(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##6) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define LOAD_TENSOR_ROW_8(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_7(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##7) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define LOAD_TENSOR_ROW_9(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_8(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##8) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define LOAD_TENSOR_ROW_10(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_9(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)      \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##9) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define LOAD_TENSOR_ROW_11(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_10(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##A) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define LOAD_TENSOR_ROW_12(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_11(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##B) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define LOAD_TENSOR_ROW_13(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_12(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##C) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define LOAD_TENSOR_ROW_14(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_13(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##D) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define LOAD_TENSOR_ROW_15(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_14(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##E) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define LOAD_TENSOR_ROW_16(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) \
     LOAD_TENSOR_ROW_15(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)     \
     SCALAR_ACCESS(COL_OFFSET, N0, BASENAME##F) = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @}*/ // end of group LOAD_TENSOR_ROW_n

 /** Load tensor (consecutive rows and columns) with Z offset.
  * @name LOAD_TENSOR
  *
  * Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
  * The data to load is expected to have consecutive names for each row.
  * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3, and Z=zin, the expected Z offsets are zin0, zin1 and zin2.
  *
  * @param[in] M0         The number of consecutive rows
  * @param[in] N0         The number of consecutive columns
  * @param[in] DATA_TYPE  The data type of the target
  * @param[in] BASENAME   The basename of the result variables
  * @param[in] PTR        The base pointer for the data
  * @param[in] COL_OFFSET The column vector offset. COL_OFFSET + N0 must be <= 16
  * @param[in] STRIDE_Y   The stride in y-axis direction
  * @param[in] Z          The z-axis offset vector
  * @{
  */
 #define LOAD_TENSOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) LOAD_TENSOR_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)
 #define LOAD_TENSOR(M0, N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z) LOAD_TENSOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, COL_OFFSET, STRIDE_Y, Z)
 /** @} */ // end of group LOAD_TENSOR

 /** Load 2D tensor (consecutive rows and columns) with Z offset.
  * @name LOAD_TENSOR_M0Xn
  *
  * @param[in] M0        The number of rows to load [0-16]
  * @param[in] N0        The number of columns to load [0-16]
  * @param[in] DATA_TYPE The data type of variables
  * @param[in] BASENAME  The basename of the destination variables for the loaded rows
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The z-axis offset vector
  * @{
  */
 #define LOAD_TENSOR_M0X0(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     ({})

 #define LOAD_TENSOR_M0X1(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);

 #define LOAD_TENSOR_M0X2(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);

 #define LOAD_TENSOR_M0X3(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);

 #define LOAD_TENSOR_M0X4(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);

 #define LOAD_TENSOR_M0X5(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);       \
     LOAD_TENSOR(M0, 1, DATA_TYPE, a, input_ptr + 4 * sizeof(DATA_TYPE), 4, src_stride_y, zin);

 #define LOAD_TENSOR_M0X6(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);       \
     LOAD_TENSOR(M0, 2, DATA_TYPE, a, input_ptr + 4 * sizeof(DATA_TYPE), 4, src_stride_y, zin);

 #define LOAD_TENSOR_M0X7(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);       \
     LOAD_TENSOR(M0, 3, DATA_TYPE, a, input_ptr + 4 * sizeof(DATA_TYPE), 4, src_stride_y, zin);

 #define LOAD_TENSOR_M0X8(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);

 #define LOAD_TENSOR_M0X9(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr 0, src_stride_y, zin);        \
     LOAD_TENSOR(M0, 1, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin);

 #define LOAD_TENSOR_M0X10(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);        \
     LOAD_TENSOR(M0, 2, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin);

 #define LOAD_TENSOR_M0X11(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);        \
     LOAD_TENSOR(M0, 3, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin);

 #define LOAD_TENSOR_M0X12(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);        \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin);

 #define LOAD_TENSOR_M0X13(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin)                  \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);                         \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin); \
     LOAD_TENSOR(M0, 1, DATA_TYPE, a, input_ptr + 12 * sizeof(DATA_TYPE), 12, src_stride_y, zin);

 #define LOAD_TENSOR_M0X14(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin)                  \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr 0, src_stride_y, zin);                          \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin); \
     LOAD_TENSOR(M0, 2, DATA_TYPE, a, input_ptr + 12 * sizeof(DATA_TYPE), 12, src_stride_y, zin);

 #define LOAD_TENSOR_M0X15(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin)                  \
     LOAD_TENSOR(M0, 8, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);                         \
     LOAD_TENSOR(M0, 4, DATA_TYPE, a, input_ptr + 8 * sizeof(DATA_TYPE), 8, src_stride_y, zin); \
     LOAD_TENSOR(M0, 3, DATA_TYPE, a, input_ptr + 12 * sizeof(DATA_TYPE), 12, src_stride_y, zin);

 #define LOAD_TENSOR_M0X16(M0, N0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \
     LOAD_TENSOR(M0, N0, DATA_TYPE, a, input_ptr, 0, src_stride_y, zin);
 /** @}*/ // end of group LOAD_TENSOR_M0Xn

 /** Load 2D tensor (consecutive rows and columns) with Z offset.
  * @name LOAD_TENSOR_M0XN0
  *
  * @param[in] M0        The number of consecutive rows [0-16]
  * @param[in] N0        The number of consecutive columns [0-16]
  * @param[in] DATA_TYPE The data type of the target
  * @param[in] BASENAME  The basename of the result variables
  * @param[in] PTR       The base pointer for the data
  * @param[in] STRIDE_Y  The stride in y-axis direction
  * @param[in] Z         The z-axis offset vector
  * @{
  */
 #define LOAD_TENSOR_M0XN0_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) LOAD_TENSOR_M0X##N0(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define LOAD_TENSOR_M0XN0(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) LOAD_TENSOR_M0XN0_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

 /** Loads the rows from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
  * @name LOAD_ROW_n
  *
  * @param[in] N0        The number of columns to load
  * @param[in] DATA_TYPE The data type of variables
  * @param[in] BASENAME  The basename of the destination variables for the loaded rows
  * @param[in] PTR       The base pointer
  * @param[in] OFFSET    The offset within a row
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The z-axis offset vector
  * @{
  */
 #define LOAD_ROW_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##0 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y + Z##0));

 #define LOAD_ROW_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##1 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y + Z##1));

 #define LOAD_ROW_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##2 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y + Z##2));

 #define LOAD_ROW_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##3 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y + Z##3));

 #define LOAD_ROW_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##4 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y + Z##4));

 #define LOAD_ROW_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##5 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y + Z##5));

 #define LOAD_ROW_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##6 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y + Z##6));

 #define LOAD_ROW_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##7 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y + Z##7));

 #define LOAD_ROW_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                      \
     BASENAME##8 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y + Z##8));

 #define LOAD_ROW_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)      \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##9 = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y + Z##9));

 #define LOAD_ROW_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##A = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y + Z##A));

 #define LOAD_ROW_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##B = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y + Z##B));

 #define LOAD_ROW_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##C = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y + Z##C));

 #define LOAD_ROW_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##D = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y + Z##D));

 #define LOAD_ROW_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##E = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y + Z##E));

 #define LOAD_ROW_16(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \
     LOAD_ROW_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##F = VLOAD(N0)(0, (__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y + Z##F));

 /** @}*/ // end of group LOAD_ROW_n

 /** Load Blocks (consecutive rows and columns) with Z offset.
  * @name LOAD_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
  * The data to load is expected to have consecutive names for each row.
  * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3, and Z=zin, the expected Z offsets are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of consecutive rows
  * @param[in] N0        The number of consecutive columns
  * @param[in] DATA_TYPE The data type of the target
  * @param[in] BASENAME  The basename of the result variables
  * @param[in] PTR       The base pointer for the data
  * @param[in] OFFSET    The offset within a row
  * @param[in] STRIDE_Y  The stride in y-axis direction
  * @param[in] Z         The z-axis offset vector
  * @{
  */
 #define LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
 #define LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
 /** @} */ // end of group LOAD_BLOCK

 /** Loads the rows from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
  * @name LOAD_TEXTURE2D_ROW_n
  *
  * @param[in] N0         The number of pixels to read
  * @param[in] DATA_TYPE  The data type of variables
  * @param[in] BASENAME   The basename of the destination variables for the loaded rows
  * @param[in] IMG        The 2D OpenCL image object
  * @param[in] X_COORD    The x coordinate for the top-left pixel
  * @param[in] Y_COORD    The y coordinate for the top-left pixel
  * @param[in] X_STEP_ROW The incremental step row for the x coordinate (in pixels)
  * @param[in] Y_STEP_ROW The incremental step row for the y coordinate (in pixels)
  * @{
  */
 #define LOAD_TEXTURE2D_ROW_1(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     BASENAME##0 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 0 * X_STEP_ROW), (Y_COORD + 0 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_2(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_1(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##1 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 1 * X_STEP_ROW), (Y_COORD + 1 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_3(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_2(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##2 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 2 * X_STEP_ROW), (Y_COORD + 2 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_4(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_3(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##3 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 3 * X_STEP_ROW), (Y_COORD + 3 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_5(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_4(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##4 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 4 * X_STEP_ROW), (Y_COORD + 4 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_6(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_5(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##5 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 5 * X_STEP_ROW), (Y_COORD + 5 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_7(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_6(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##6 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 6 * X_STEP_ROW), (Y_COORD + 6 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_8(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_7(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##7 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 7 * X_STEP_ROW), (Y_COORD + 7 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_9(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_8(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##8 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 8 * X_STEP_ROW), (Y_COORD + 8 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_10(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_9(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)      \
     BASENAME##9 = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 9 * X_STEP_ROW), (Y_COORD + 9 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_11(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_10(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##A = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 10 * X_STEP_ROW), (Y_COORD + 10 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_12(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_11(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##B = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 11 * X_STEP_ROW), (Y_COORD + 11 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_13(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_12(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##C = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 12 * X_STEP_ROW), (Y_COORD + 12 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_14(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_13(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##D = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 13 * X_STEP_ROW), (Y_COORD + 13 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_15(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_14(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##E = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 14 * X_STEP_ROW), (Y_COORD + 14 * Y_STEP_ROW))

 #define LOAD_TEXTURE2D_ROW_16(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) \
     LOAD_TEXTURE2D_ROW_15(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)     \
     BASENAME##F = READ_IMAGE2D(DATA_TYPE, N0, IMG, (X_COORD + 15 * X_STEP_ROW), (Y_COORD + 15 * Y_STEP_ROW))
 /** @} */ // end of group LOAD_TEXTURE2D_ROW_n

 /** Load a 2D texture in unit of pixel. A pixel is made of 4 floating point values
  * @name LOAD_TEXTURE2D
  *
  * Supported cases are M0=1,2,3,...,16 and N0=1
  * The data to load is expected to have consecutive names for each row.
  * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
  *
  * @param[in] M0         The number of consecutive rows
  * @param[in] N0         The number of consecutive pixels. Only 1, 2 and 4 are supported
  * @param[in] DATA_TYPE  The data type of the target
  * @param[in] BASENAME   The basename of the result variables
  * @param[in] IMG        The 2D OpenCL image object
  * @param[in] X_COORD    The x coordinate for the top-left pixel
  * @param[in] Y_COORD    The y coordinate for the top-left pixel
  * @param[in] X_STEP_ROW The incremental step row for the x coordinate (in pixels)
  * @param[in] Y_STEP_ROW The incremental step row for the y coordinate (in pixels)
  * @{
  */
 #define LOAD_TEXTURE2D_STR(M0, N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) LOAD_TEXTURE2D_ROW_##M0(N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)
 #define LOAD_TEXTURE2D(M0, N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW) LOAD_TEXTURE2D_STR(M0, N0, DATA_TYPE, BASENAME, IMG, X_COORD, Y_COORD, X_STEP_ROW, Y_STEP_ROW)
 /** @} */ // end of group LOAD_TEXTURE2D

 /** Loads the elements from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
  * @name LOAD_ELEMENT_n
  *
  * @param[in] N0        The number of rows to load
  * @param[in] DATA_TYPE The data type of variables
  * @param[in] BASENAME  The basename of the destination variables for the loaded rows
  * @param[in] PTR       The base pointer
  * @param[in] OFFSET    The offset within a row
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @{
  */
 #define LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##0 = *((__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y));

 #define LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##1 = *((__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y));

 #define LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##2 = *((__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y));

 #define LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##3 = *((__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y));

 #define LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##4 = *((__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y));

 #define LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##5 = *((__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y));

 #define LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##6 = *((__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y));

 #define LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##7 = *((__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y));

 #define LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
     BASENAME##8 = *((__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y));

 #define LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)      \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##9 = *((__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y));

 #define LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##A = *((__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y));

 #define LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##B = *((__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y));

 #define LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##C = *((__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y));

 #define LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##D = *((__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y));

 #define LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##E = *((__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y));

 #define LOAD_ELEMENT_16(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
     LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
     VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
     BASENAME##F = *((__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y));

 /** @}*/ // end of group LOAD_ELEMENT_n

 /** Load Scalar as Vector (consecutive elements).
  * @name LOAD_SCALAR_AS_VECTOR
  *
  * Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
  * The data to load is expected to have consecutive names for each row.
  * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
  *
  * @param[in] M0        The number of consecutive rows
  * @param[in] N0        The number of consecutive columns
  * @param[in] DATA_TYPE The data type of the target
  * @param[in] BASENAME  The basename of the result variables
  * @param[in] PTR       The base pointer for the data
  * @param[in] OFFSET    The offset within a row
  * @param[in] STRIDE_Y  The stride in y-axis direction
  * @{
  */
 #define LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_ELEMENT_##M0(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
 #define LOAD_SCALAR_AS_VECTOR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
 /** @} */ // end of group LOAD_SCALAR_AS_VECTOR

 /** Basic macros to calculate Z offset values from Z0 to Zn-1
  * @name CALCULATE_Z_OFFSET_n
  *
  * @param[in] M0              The number of offset values to calculate
  * @param[in] DATA_TYPE       The data type of the results
  * @param[in] Z               The basename of the result variables
  * @param[in] Y               The work-itme ID of y-axis
  * @param[in] HEIGHT_GEMM3D   The height of GEMM3D
  * @param[in] DEPTH_GEMM3D    The depth of GEMM3D
  * @param[in] CROSS_PLANE_PAD The padding required for plane changes accross the z-dimension
  * @param[in] STRIDE_Y        The stride value in y-axis direction
  *
  * @{
  */
 #define CALCULATE_Z_OFFSET_1(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     Z##0 = (0 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##0 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##0);                                                      \
     Z##0 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_2(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_1(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##1 = (1 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##1 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##1);                                                      \
     Z##1 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_3(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_2(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##2 = (2 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##2 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##2);                                                      \
     Z##2 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_4(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_3(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##3 = (3 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##3 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##3);                                                      \
     Z##3 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_5(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_4(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##4 = (4 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##4 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##4);                                                      \
     Z##4 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_6(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_5(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##5 = (5 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##5 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##5);                                                      \
     Z##5 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_7(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_6(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##6 = (6 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##6 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##6);                                                      \
     Z##6 *= (CROSS_PLANE_PAD * STRIDE_Y);

 #define CALCULATE_Z_OFFSET_8(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) \
     CALCULATE_Z_OFFSET_7(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)     \
     Z##7 = (7 + (DATA_TYPE)(Y)) / (DATA_TYPE)HEIGHT_GEMM3D;                               \
     Z##7 = min((DATA_TYPE)(DEPTH_GEMM3D - 1), Z##7);                                                      \
     Z##7 *= (CROSS_PLANE_PAD * STRIDE_Y);

 /** @} */ // end of group CALCULATE_Z_OFFSET_n

 /** Calculate Z offset values from Z0 to Zn-1
  * @name CALCULATE_Z_OFFSET
  *
  * The Z offsets are expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected names of Z offsets are zin1, zin2, zin3.
  * Note that, CROSS_PLANE_PAD (cross plain padding) is required to take into account
  * the possible cross plane paddings in case of the plance changes across the z-dimension.
  *
  * <!--
  * |                  |
  * |      plane0      |
  * |                  |
  * |__________________|
  * |******************|
  * |  cross_plane_pad |
  * |******************|
  * |                  |
  * |      plane1      |
  * |                  |
  * |__________________|
  * -->
  *
  * @param[in] M0              The number of offset values to calculate
  * @param[in] DATA_TYPE       The data type of the results
  * @param[in] Z               The basename of the result variables
  * @param[in] Y               The work-itme ID of y-axis
  * @param[in] HEIGHT_GEMM3D   The height of GEMM3D
  * @param[in] DEPTH_GEMM3D    The depth of GEMM3D
  * @param[in] CROSS_PLANE_PAD The padding required for plane changes accross the z-dimension
  * @param[in] STRIDE_Y        The stride value in y-axis direction
  * @{
  */
 #define CALCULATE_Z_OFFSET_STR(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) CALCULATE_Z_OFFSET_##M0(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)
 #define CALCULATE_Z_OFFSET(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y) CALCULATE_Z_OFFSET_STR(M0, DATA_TYPE, Z, Y, HEIGHT_GEMM3D, DEPTH_GEMM3D, CROSS_PLANE_PAD, STRIDE_Y)
 /** @} */ // end of group CALCULATE_Z_OFFSET

 /** Scale the rows in the given variables (BASENAME0 to BASENAMEn-1)
  * @name SCALE_ROW_n
  *
  * @param[in] DATA_TYPE The data type of the variables
  * @param[in] BASENAME  The basename of the variables
  * @param[in] SCALE     The scale factor
  * @{
  */
 #define SCALE_ROW_1(DATA_TYPE, BASENAME, SCALE) \
     BASENAME##0 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_2(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_1(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##1 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_3(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_2(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##2 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_4(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_3(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##3 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_5(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_4(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##4 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_6(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_5(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##5 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_7(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_6(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##6 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_8(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_7(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##7 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_9(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_8(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##8 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_10(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_9(DATA_TYPE, BASENAME, SCALE)      \
     BASENAME##9 *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_11(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_10(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##A *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_12(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_11(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##B *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_13(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_12(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##C *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_14(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_13(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##D *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_15(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_14(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##E *= (DATA_TYPE)SCALE;

 #define SCALE_ROW_16(DATA_TYPE, BASENAME, SCALE) \
     SCALE_ROW_15(DATA_TYPE, BASENAME, SCALE)     \
     BASENAME##F *= (DATA_TYPE)SCALE;
 /** @} */ // end of group SCALE_ROW_n

 /** Scale elements stored in a block (BASENAME)
  * @name SCALE_BLOCK
  *
  * Supported cases are N=1,2,3,...,16
  *
  * @param[in] N         The number of rows in the block
  * @param[in] DATA_TYPE The data type of the block
  * @param[in] BASENAME  The basename of the block
  * @param[in] SCALE     The scale factor
  * @{
  */
 #define SCALE_BLOCK_STR(N, DATA_TYPE, BASENAME, SCALE) SCALE_ROW_##N(DATA_TYPE, BASENAME, SCALE)
 #define SCALE_BLOCK(N, DATA_TYPE, BASENAME, SCALE) SCALE_BLOCK_STR(N, DATA_TYPE, BASENAME, SCALE)
 /** @} */ // end of group SCALE_BLOCK

 /** Create a new vector containing the values at the given index for a set of given vectors
  * @name COLUMN_VECTORn
  *
  * @param[in] IDX_COL  The index value
  * @param[in] BASENAME The basename of the destination vectors
  * @param[in] X        The basename of the source vectors
  * @param[in] TYPE     The data type of the destination vectors
  * @{
  */
 #define COLUMN_VECTOR1(IDX_COL, BASENAME, X, TYPE) \
     TYPE BASENAME##IDX_COL = (TYPE)((X##0).s##IDX_COL);
 #define COLUMN_VECTOR2(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 2)                         \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 2))((X##0).s##IDX_COL, (X##1).s##IDX_COL);
 #define COLUMN_VECTOR3(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 3)                         \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 3))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL);
 #define COLUMN_VECTOR4(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 4)                         \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 4))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL);
 #define COLUMN_VECTOR8(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 8)                         \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 8))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL, (X##4).s##IDX_COL, (X##5).s##IDX_COL, (X##6).s##IDX_COL, (X##7).s##IDX_COL);
 #define COLUMN_VECTOR16(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 16)                         \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 16))((X##0).s##IDX_COL, (X##1).s##IDX_COL, (X##2).s##IDX_COL, (X##3).s##IDX_COL, (X##4).s##IDX_COL, (X##5).s##IDX_COL, (X##6).s##IDX_COL, (X##7).s##IDX_COL, (X##8).s##IDX_COL, (X##9).s##IDX_COL, (X##A).s##IDX_COL, (X##B).s##IDX_COL, (X##C).s##IDX_COL, (X##D).s##IDX_COL, (X##E).s##IDX_COL, (X##F).s##IDX_COL);
 /** @} */ // end of group COLUMN_VECTORn

 /** Create a new vector containing the values at the given index. Utility macros for transposing a colum-vector
  * @name COLUMN_VECTOR_SCALARn
  *
  * @param[in] IDX_COL  The index value
  * @param[in] BASENAME The basename of the destination vectors
  * @param[in] X        The basename of the source vectors
  * @param[in] TYPE     The data type of the destination vectors
  * @{
  */
 #define COLUMN_VECTOR_SCALAR1(IDX_COL, BASENAME, X, TYPE) \
     TYPE BASENAME##IDX_COL = (TYPE)((X##0));
 #define COLUMN_VECTOR_SCALAR2(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 2)                                \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 2))((X##0), (X##1));
 #define COLUMN_VECTOR_SCALAR3(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 3)                                \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 3))((X##0), (X##1), (X##2));
 #define COLUMN_VECTOR_SCALAR4(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 4)                                \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 4))((X##0), (X##1), (X##2), (X##3));
 #define COLUMN_VECTOR_SCALAR8(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 8)                                \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 8))((X##0), (X##1), (X##2), (X##3), (X##4), (X##5), (X##6), (X##7));
 #define COLUMN_VECTOR_SCALAR16(IDX_COL, BASENAME, X, TYPE) \
     VEC_DATA_TYPE(TYPE, 16)                                \
     BASENAME##IDX_COL = (VEC_DATA_TYPE(TYPE, 16))((X##0), (X##1), (X##2), (X##3), (X##4), (X##5), (X##6), (X##7), (X##8), (X##9), (X##A), (X##B), (X##C), (X##D), (X##E), (X##F));
 /** @} */ // end of group COLUMN_VECTORn

 /** Create transposed vectors of the given vectors
  * @name TRANSPOSE_K0Xn
  *
  * @param[in] K0       The size of the source vectors
  * @param[in] BASENAME The basename of transposed vectors
  * @param[in] B        The basename of source vectors for transposition
  * @param[in] TYPE     The data type of the transposed vectors
  * @{
  */
 #define TRANSPOSE_K0X1(K0, BASENAME, B, TYPE) \
     COLUMN_VECTOR_SCALAR(K0, 0, BASENAME, B, TYPE);
 #define TRANSPOSE_K0X2(K0, BASENAME, B, TYPE) \
     COLUMN_VECTOR(K0, 0, BASENAME, B, TYPE);  \
     COLUMN_VECTOR(K0, 1, BASENAME, B, TYPE);
 #define TRANSPOSE_K0X3(K0, BASENAME, B, TYPE) \
     TRANSPOSE_K0X2(K0, BASENAME, B, TYPE);    \
     COLUMN_VECTOR(K0, 2, BASENAME, B, TYPE);
 #define TRANSPOSE_K0X4(K0, BASENAME, B, TYPE) \
     TRANSPOSE_K0X3(K0, BASENAME, B, TYPE);    \
     COLUMN_VECTOR(K0, 3, BASENAME, B, TYPE);
 #define TRANSPOSE_K0X8(K0, BASENAME, B, TYPE) \
     TRANSPOSE_K0X4(K0, BASENAME, B, TYPE);    \
     COLUMN_VECTOR(K0, 4, BASENAME, B, TYPE);  \
     COLUMN_VECTOR(K0, 5, BASENAME, B, TYPE);  \
     COLUMN_VECTOR(K0, 6, BASENAME, B, TYPE);  \
     COLUMN_VECTOR(K0, 7, BASENAME, B, TYPE);
 #define TRANSPOSE_K0X16(K0, BASENAME, B, TYPE) \
     TRANSPOSE_K0X8(K0, BASENAME, B, TYPE);     \
     COLUMN_VECTOR(K0, 8, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, 9, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, A, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, B, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, C, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, D, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, E, BASENAME, B, TYPE);   \
     COLUMN_VECTOR(K0, F, BASENAME, B, TYPE);

 /** @} */ // end of group TRANSPOSE_K0Xn

 /** Create column vectors to contain the values at the given index for a set of given vectors
  *
  * @param[in] K0       The number of source vectors
  * @param[in] IDX_COL  The index value
  * @param[in] BASENAME The basename of the destination vectors
  * @param[in] B        The basename of the source vectors
  * @param[in] TYPE     The data type of the destination vectors
  */
 #define COLUMN_VECTOR(K0, IDX_COL, BASENAME, B, TYPE) \
     CONCAT(COLUMN_VECTOR, K0)                         \
     (IDX_COL, BASENAME, B, TYPE);

 /** Create column vectors to contain the values at the given index. Utility macro for transposing a column-vector
  *
  * @param[in] K0       The number of source vectors
  * @param[in] IDX_COL  The index value
  * @param[in] BASENAME The basename of the destination vectors
  * @param[in] B        The basename of the source vectors
  * @param[in] TYPE     The data type of the destination vectors
  */
 #define COLUMN_VECTOR_SCALAR(K0, IDX_COL, BASENAME, B, TYPE) \
     CONCAT(COLUMN_VECTOR_SCALAR, K0)                         \
     (IDX_COL, BASENAME, B, TYPE);

 /** Create transposed vectors form the given source vectors
  *
  * @param[in] K0       The size of source vectors
  * @param[in] N0       The number of source vectors
  * @param[in] BASENAME The basename of transposed vectors
  * @param[in] B        The basename of source vectors for transposition
  * @param[in] TYPE     The data type of the transposed vectors
  *
  */
 #define TRANSPOSE_K0XN0(K0, N0, BASENAME, B, TYPE) \
     CONCAT(TRANSPOSE_K0X, N0)                      \
     (K0, BASENAME, B, TYPE);

 /** Add the variables (BIAS0 to BIASn-1) to the others (BASENAME0 to BASENAMEn-1)
  * @name ADD_ROW_n
  *
  * @param[in] BASENAME The basename of the destination variables
  * @param[in] BIAS     The basename of the added variables
  * @{
  */
 #define ADD_ROW_1(BASENAME, BIAS) \
     BASENAME##0 += BIAS##0;

 #define ADD_ROW_2(BASENAME, BIAS) \
     ADD_ROW_1(BASENAME, BIAS)     \
     BASENAME##1 += BIAS##1;

 #define ADD_ROW_3(BASENAME, BIAS) \
     ADD_ROW_2(BASENAME, BIAS)     \
     BASENAME##2 += BIAS##2;

 #define ADD_ROW_4(BASENAME, BIAS) \
     ADD_ROW_3(BASENAME, BIAS)     \
     BASENAME##3 += BIAS##3;

 #define ADD_ROW_5(BASENAME, BIAS) \
     ADD_ROW_4(BASENAME, BIAS)     \
     BASENAME##4 += BIAS##4;

 #define ADD_ROW_6(BASENAME, BIAS) \
     ADD_ROW_5(BASENAME, BIAS)     \
     BASENAME##5 += BIAS##5;

 #define ADD_ROW_7(BASENAME, BIAS) \
     ADD_ROW_6(BASENAME, BIAS)     \
     BASENAME##6 += BIAS##6;

 #define ADD_ROW_8(BASENAME, BIAS) \
     ADD_ROW_7(BASENAME, BIAS)     \
     BASENAME##7 += BIAS##7;

 #define ADD_ROW_9(BASENAME, BIAS) \
     ADD_ROW_8(BASENAME, BIAS)     \
     BASENAME##8 += BIAS##8;

 #define ADD_ROW_10(BASENAME, BIAS) \
     ADD_ROW_9(BASENAME, BIAS)      \
     BASENAME##9 += BIAS##9;

 #define ADD_ROW_11(BASENAME, BIAS) \
     ADD_ROW_10(BASENAME, BIAS)     \
     BASENAME##A += BIAS##A;

 #define ADD_ROW_12(BASENAME, BIAS) \
     ADD_ROW_11(BASENAME, BIAS)     \
     BASENAME##B += BIAS##B;

 #define ADD_ROW_13(BASENAME, BIAS) \
     ADD_ROW_12(BASENAME, BIAS)     \
     BASENAME##C += BIAS##C;

 #define ADD_ROW_14(BASENAME, BIAS) \
     ADD_ROW_13(BASENAME, BIAS)     \
     BASENAME##D += BIAS##D;

 #define ADD_ROW_15(BASENAME, BIAS) \
     ADD_ROW_14(BASENAME, BIAS)     \
     BASENAME##E += BIAS##E;

 #define ADD_ROW_16(BASENAME, BIAS) \
     ADD_ROW_15(BASENAME, BIAS)     \
     BASENAME##F += BIAS##F;

 /** @} */ // end of group ADD_ROW_n

 /** Add the block (BIAS) to another block (BASENAME)
  * @name ADD_BLOCK
  *
  * Supported cases are N=1,2,3,...,16
  *
  * @param[in] N        The number of vectors in the block
  * @param[in] BASENAME The basename of the destination variables
  * @param[in] BIAS     The basename of the added variables
  * @{
  */
 #define ADD_BLOCK_STR(N, BASENAME, BIAS) ADD_ROW_##N(BASENAME, BIAS)
 #define ADD_BLOCK(N, BASENAME, BIAS) ADD_BLOCK_STR(N, BASENAME, BIAS)
 /** @} */ // end of group ADD_BLOCK

 /** Broadcast (add single value) to the each element of the destination variables
  * @name ADD_ROW_BROADCAST_n
  *
  * @param[in] BASENAME The basename of the destination variables
  * @param[in] BIAS     The variable containing the value to add
  * @{
  */
 #define ADD_ROW_BROADCAST_1(BASENAME, BIAS) \
     BASENAME##0 += BIAS;

 #define ADD_ROW_BROADCAST_2(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_1(BASENAME, BIAS)     \
     BASENAME##1 += BIAS;

 #define ADD_ROW_BROADCAST_3(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_2(BASENAME, BIAS)     \
     BASENAME##2 += BIAS;

 #define ADD_ROW_BROADCAST_4(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_3(BASENAME, BIAS)     \
     BASENAME##3 += BIAS;

 #define ADD_ROW_BROADCAST_5(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_4(BASENAME, BIAS)     \
     BASENAME##4 += BIAS;

 #define ADD_ROW_BROADCAST_6(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_5(BASENAME, BIAS)     \
     BASENAME##5 += BIAS;

 #define ADD_ROW_BROADCAST_7(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_6(BASENAME, BIAS)     \
     BASENAME##6 += BIAS;

 #define ADD_ROW_BROADCAST_8(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_7(BASENAME, BIAS)     \
     BASENAME##7 += BIAS;

 #define ADD_ROW_BROADCAST_9(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_8(BASENAME, BIAS)     \
     BASENAME##8 += BIAS;

 #define ADD_ROW_BROADCAST_10(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_9(BASENAME, BIAS)      \
     BASENAME##9 += BIAS;

 #define ADD_ROW_BROADCAST_11(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_10(BASENAME, BIAS)     \
     BASENAME##A += BIAS;

 #define ADD_ROW_BROADCAST_12(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_11(BASENAME, BIAS)     \
     BASENAME##B += BIAS;

 #define ADD_ROW_BROADCAST_13(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_12(BASENAME, BIAS)     \
     BASENAME##C += BIAS;

 #define ADD_ROW_BROADCAST_14(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_13(BASENAME, BIAS)     \
     BASENAME##D += BIAS;

 #define ADD_ROW_BROADCAST_15(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_14(BASENAME, BIAS)     \
     BASENAME##E += BIAS;

 #define ADD_ROW_BROADCAST_16(BASENAME, BIAS) \
     ADD_ROW_BROADCAST_15(BASENAME, BIAS)     \
     BASENAME##F += BIAS;

 /** Broadcast (add a value) to the each element of the destination block (BASENAME)
  * @name ADD_BLOCK_BROADCAST
  *
  * Supported cases are N=1,2,3,...,16.
  *
  * @param[in] N        The number of vectors in the block
  * @param[in] BASENAME The basename of the destination variables
  * @param[in] BIAS     The variable containing the value to add
  * @{
  */
 #define ADD_BLOCK_BROADCAST_STR(N, BASENAME, BIAS) ADD_ROW_BROADCAST_##N(BASENAME, BIAS)
 #define ADD_BLOCK_BROADCAST(N, BASENAME, BIAS) ADD_BLOCK_BROADCAST_STR(N, BASENAME, BIAS)
 /** @} */ // end of group ADD_BLOCK_BROADCAST

 /** Apply activation to the given variables
  * @name ACTIVATION_ROW_n
  *
  * @param[in] ACTIVATION_TYPE The type of the activation
  * @param[in] DATA_TYPE       The data type of the vectors
  * @param[in] BASENAME        The basename of the variables
  * @param[in] A_VAL           Additional value required by the activation
  * @param[in] B_VAL           Additional value required by the activation
  * @{
  */
 #define ACTIVATION_ROW_1(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     BASENAME##0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##0, A_VAL, B_VAL);

 #define ACTIVATION_ROW_2(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_1(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##1 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##1, A_VAL, B_VAL);

 #define ACTIVATION_ROW_3(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_2(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##2 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##2, A_VAL, B_VAL);

 #define ACTIVATION_ROW_4(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_3(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##3 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##3, A_VAL, B_VAL);

 #define ACTIVATION_ROW_5(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_4(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##4 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##4, A_VAL, B_VAL);

 #define ACTIVATION_ROW_6(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_5(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##5 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##5, A_VAL, B_VAL);

 #define ACTIVATION_ROW_7(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_6(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##6 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##6, A_VAL, B_VAL);

 #define ACTIVATION_ROW_8(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_7(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##7 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##7, A_VAL, B_VAL);

 #define ACTIVATION_ROW_9(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_8(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##8 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##8, A_VAL, B_VAL);

 #define ACTIVATION_ROW_10(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_9(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)      \
     BASENAME##9 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##9, A_VAL, B_VAL);

 #define ACTIVATION_ROW_11(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_10(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##A = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##A, A_VAL, B_VAL);

 #define ACTIVATION_ROW_12(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_11(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##B = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##B, A_VAL, B_VAL);

 #define ACTIVATION_ROW_13(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_12(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##C = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##C, A_VAL, B_VAL);

 #define ACTIVATION_ROW_14(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_13(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##D = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##D, A_VAL, B_VAL);

 #define ACTIVATION_ROW_15(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_14(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##E = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##E, A_VAL, B_VAL);

 #define ACTIVATION_ROW_16(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) \
     ACTIVATION_ROW_15(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)     \
     BASENAME##F = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME##F, A_VAL, B_VAL);
 /** @} */ // end of group ACTIVATION_ROW_n

 /** Apply activation to a block (BASENAME)
  * @name ACTIVATION_BLOCK
  *
  * Supported cases are N=1,2,3,...,16.
  *
  * @param[in] N               The number of vectors in the block
  * @param[in] ACTIVATION_TYPE The type of the activation
  * @param[in] DATA_TYPE       The data type of the vectors
  * @param[in] BASENAME        The basename of the variables
  * @param[in] A_VAL           Additional value required by the activation
  * @param[in] B_VAL           Additional value required by the activation
  * @{
  */
 #define ACTIVATION_BLOCK_STR(N, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) ACTIVATION_ROW_##N(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)
 #define ACTIVATION_BLOCK(N, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL) ACTIVATION_BLOCK_STR(N, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, BASENAME, A_VAL, B_VAL)
 /** @} */ // end of group ACTIVATION_BLOCK

 /** Apply convert_<data_type> to the given variables
  * @name CONVERT_ROW_n
  *
  * @param[in] N            The size of the vectors
  * @param[in] DATA_TYPE    The data type of the vectors
  * @param[in] BASENAME_SRC The basename of the source variables
  * @param[in] BASENAME_DST The basename of the destination variables
  */
 #define CONVERT_ROW_1(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##0 = CONVERT(BASENAME_SRC##0, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_2(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_1(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##1 = CONVERT(BASENAME_SRC##1, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_3(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_2(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##2 = CONVERT(BASENAME_SRC##2, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_4(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_3(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##3 = CONVERT(BASENAME_SRC##3, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_5(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_4(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##4 = CONVERT(BASENAME_SRC##4, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_6(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_5(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##5 = CONVERT(BASENAME_SRC##5, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_7(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_6(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##6 = CONVERT(BASENAME_SRC##6, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_8(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_7(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##7 = CONVERT(BASENAME_SRC##7, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_9(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_8(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                 \
     BASENAME_DST##8 = CONVERT(BASENAME_SRC##8, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_10(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_9(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)      \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##9 = CONVERT(BASENAME_SRC##9, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_11(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_10(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##A = CONVERT(BASENAME_SRC##A, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_12(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_11(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##B = CONVERT(BASENAME_SRC##B, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_13(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_12(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##C = CONVERT(BASENAME_SRC##C, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_14(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_13(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##D = CONVERT(BASENAME_SRC##D, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_15(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_14(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##E = CONVERT(BASENAME_SRC##E, VEC_DATA_TYPE(DATA_TYPE, N));

 #define CONVERT_ROW_16(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) \
     CONVERT_ROW_15(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)     \
     VEC_DATA_TYPE(DATA_TYPE, N)                                  \
     BASENAME_DST##F = CONVERT(BASENAME_SRC##F, VEC_DATA_TYPE(DATA_TYPE, N));
 /** @} */ // end of group CONVERT_ROW_n

 /** Apply convert_<data_type> to a block (BASENAME_SRC) and save to another block (BASENAME_DST)
  * @name CONVERT_BLOCK
  *
  * Supported cases N=1,2,3,...,16.
  *
  * @param[in] M            The number of vectors to convert
  * @param[in] N            The size of the vectors
  * @param[in] DATA_TYPE    The data type of the vectors
  * @param[in] BASENAME_SRC The basename of the source variables
  * @param[in] BASENAME_DST The basename of the destination variables
  */
 #define CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_ROW_##M(N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
 #define CONVERT_BLOCK(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST) CONVERT_BLOCK_STR(M, N, DATA_TYPE, BASENAME_SRC, BASENAME_DST)
 /** @} */ // end of group CONVERT_BLOCK
 /*
  * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #ifndef ARM_COMPUTE_REPEAT_H
 #define ARM_COMPUTE_REPEAT_H

 /*
  * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
 #ifndef ARM_COMPUTE_HELPER_H
 #define ARM_COMPUTE_HELPER_H

 /*
  * Copyright (c) 2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
  * deal in the Software without restriction, including without limitation the
  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  * sell copies of the Software, and to permit persons to whom the Software is
  * furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in all
  * copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */

 /** Store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_n
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE(N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_n

 /** Convert and store the 0th to (n-1)th rows of the given variables
  * @name CONVERT_STORE_ROW_n
  *
  * @param[in] N0        The size of the vectors
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                         \
     (CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE(N0)                                                     \
     (CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE(N0)                                                          \
     (CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));

 /** @} */ // end of groupd CONVERT_STORE_ROW_n

 /** Store a block of the given size M0xN0
  * @name STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group STORE_BLOCK

 /** Convert and store a block of the given size M0xN0
  * @name CONVERT_STORE_BLOCK
  *
  * Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0        The number of rows to store
  * @param[in] N0        The size of each vector
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** @} */ // end of group CONVERT_STORE_BLOCK

 /** Partially store the 0 to (n-1)th rows of the given variables
  * @name STORE_ROW_PARTIAL_n
  * Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
  *
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * @param[in] N0        The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] STORE_N0  The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));

 #define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));

 #define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));

 #define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));

 #define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));

 #define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));

 #define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));

 #define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));

 #define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                 \
     (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));

 #define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)      \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));

 #define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));

 #define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));

 #define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));

 #define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));

 #define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));

 #define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
     STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)     \
     VSTORE_PARTIAL(N0, STORE_N0)                                                  \
     (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
 /** @} */ // end of groupd STORE_ROW_PARTIAL_n

 /** Partially store a block of the given size STORE_M0xSTORE_N0
  * @name STORE_BLOCK_PARTIAL
  *
  * @note The vector width @p N0 is also required for correct partial storing behaviour.
  * @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] STORE_M0  The number of rows to store. Supported: 1-16
  * @param[in] STORE_N0  The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
  * @param[in] N0        The size of each vector. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE The data type of the vectors
  * @param[in] BASENAME  The basename of the variables
  * @param[in] PTR       The base pointer
  * @param[in] STRIDE_Y  The stride value in y-axis direction
  * @param[in] Z         The offset in z-axis direction
  * @{
  */
 #define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 #define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
 /** Store a block that can be partial in both x and y dimensions
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y))                                                                                                            \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                                           \
     }                                                                                                                                                     \
     else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X))                                                                                                        \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                                             \
     }                                                                                                                                                     \
     else                                                                                                                                                  \
     {                                                                                                                                                     \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                               \
     }
 /** Store a block that can only be partial in x but not y.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  */
 #define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \
     if(!(PARTIAL_COND_X))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** Store a block that can only be partial in y but not x.
  *
  * @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  */
 #define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
     if(!(PARTIAL_COND_Y))                                                                                         \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                                   \
     }                                                                                                             \
     else                                                                                                          \
     {                                                                                                             \
         STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z);                     \
     }
 /** @} */ // end of group STORE_BLOCK_PARTIAL

 #if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)

 /** Boundary-aware GEMM block store
  * @name STORE_BLOCK_BOUNDARY_AWARE
  * This macro assumes the following schemes to achieve boundary-awareness:
  *  - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
  *  - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
  *  - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
  * The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
  *
  * In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
  * blocks **at the end**.
  * Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
  * "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
  *
  *  *--x-->                         x == 0                        x == 1
  *  |                  |<------------------------------N-------------------------->|
  *  y                  |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
  *  |     -------------#############################################################
  *  *     |          | |...............................|...........................|
  * y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
  *        |          | |...............................|...........................|
  *        M          --#############################################################
  *        |          | |                               |...........................|
  * y == 1 |         M0 |      Non-boundary block       |....Boundary block in x....|
  *        |          | |                               |...........................|
  *        |------------#############################################################
  *
  * Then @p PARTIAL_STORE_M0 = M % M0      and @p PARTIAL_STORE_N0 = N % N0
  *
  * @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
  * and select corresponding store methods such that the boundary detection logic is only added when needed.
  *
  * The data to store is expected to have consecutive names for each row.
  * E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
  * The Z offset is expected to have consecutive names.
  * E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
  *
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] N0               The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] DATA_TYPE        The data type of the vectors
  * @param[in] BASENAME         The basename of the variables
  * @param[in] PTR              The base pointer
  * @param[in] STRIDE_Y         The stride value in y-axis direction
  * @param[in] Z                The offset in z-axis direction
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
  * @param[in] PARTIAL_COND_Y   Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
  * @param[in] PARTIAL_COND_X   Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
  * @{
  */
 #if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case1: No partial blocks in either x or y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)

 #elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
 // Case2: Partial blocks in y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)

 #elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
 // Case3: Partial blocks in x
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X)

 #else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
 // Case4: Partial blocks in both x and y
 #define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
     STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X)

 #endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0

 #endif    // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
 /** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE

 #if defined(PARTIAL_STORE_M0)
 /** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
  * @name COMPUTE_M0_START_ROW
  * If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
  * This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
  * blocks in the y dimension to avoid any padding.
  * EG: M0=4, PARTIAL_STORE_M0=1:
  *                  | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
  * block 0 (partial)| start row = 0   | start row = 0
  * block 1 (full)   | start row = 4   | start row = 1
  * block 2 (full)   | start row = 8   | start row = 5
  *
  * @param[in] y                Global id of current block in y.
  * @param[in] M0               The number of rows to store, for non-partial blocks. Supported: 1-16
  * @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
  * @{
  */
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
 #else // defined(PARTIAL_STORE_M0)
 #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
     ((uint)(y * M0))
 #endif    // defined(PARTIAL_STORE_M0)
 /** @} */ // end of group COMPUTE_M0_START_ROW

 /** Store a vector that can only be partial in x.
  *
  * @note in case @p vec_size or @p leftover != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
  *
  * The data to store is expected to end in a 0.
  * E.g., for basename=c, the expected name is c0.
  *
  * @param[in] basename  The name of the variable without trailing 0
  * @param[in] data_type The data type of the vector
  * @param[in] ptr       The base pointer
  * @param[in] vec_size  The vector size if cond = false. Supported: 1, 2, 3, 4, 8, 16
  * @param[in] leftover  The vector size if cond = true. Supported range: [1, @p vec_size0)
  * @param[in] cond      Condition to select either vec_size0 or vec_size1
  * @{
  */
 #define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
     STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond)
 /** @} */ // end of group STORE_VECTOR_SELECT

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)

 #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
 #pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)

 #if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
 #pragma OPENCL EXTENSION cl_arm_printf : enable
 #endif // defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)

 #define GPU_ARCH_MIDGARD 0x100
 #define GPU_ARCH_BIFROST 0x200

 /** Concatenate two inputs.
  *
  * @param[in] a The first input to be concatenated
  * @param[in] b The second input to be concatenated
  *
  * @return The concatenated output
  */
 #define CONCAT(a, b) a##b

 /** Expand the given vector
  *
  * @param[in] x The vector to be expanded
  *
  * @return The expanded output
  */
 #define EXPAND(x) x

 /** Clamp the given value between an upper and lower bound.
  *
  * @param[in] x       The value to be clamped
  * @param[in] min_val The lower bound
  * @param[in] max_val The upper bound
  *
  * @return The clamped value.
  */
 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)

 /** REVn reverses the given vector whose size is n.
  * @name REVn
  *
  * @param[in] x The vector to be reversed
  *
  * @return The reversed vector
  * @{
  */
 #define REV1(x) ((x))
 #define REV2(x) ((x).s10)
 #define REV3(x) ((x).s210)
 #define REV4(x) ((x).s3210)
 #define REV8(x) ((x).s76543210)
 #define REV16(x) ((x).sFEDCBA9876543210)
 /** @} */ // end of group REVn

 /** Reverse the given vector.
  * @name REVERSE
  *
  * @param[in] x The vector to be reversed
  * @param[in] s The size of the vector
  *
  * @return The reversed vector
  * @{
  */
 #define REVERSE_STR(x, s) REV##s((x))
 #define REVERSE(x, s) REVERSE_STR(x, s)
 /** @} */ // end of group REVERSE

 /** Circular-right-shift (rotate-right) the vector of size s by the amount of n.
  * @name ROTs_n
  *
  * @param[in] x The vector to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROT1_0(x) ((x))

 #define ROT2_0(x) ((x))
 #define ROT2_1(x) ((x).s10)

 #define ROT3_0(x) ((x))
 #define ROT3_1(x) ((x).s201)
 #define ROT3_2(x) ((x).s120)

 #define ROT4_0(x) ((x))
 #define ROT4_1(x) ((x).s3012)
 #define ROT4_2(x) ((x).s2301)
 #define ROT4_3(x) ((x).s1230)

 #define ROT8_0(x) ((x))
 #define ROT8_1(x) ((x).s70123456)
 #define ROT8_2(x) ((x).s67012345)
 #define ROT8_3(x) ((x).s56701234)
 #define ROT8_4(x) ((x).s45670123)
 #define ROT8_5(x) ((x).s34567012)
 #define ROT8_6(x) ((x).s23456701)
 #define ROT8_7(x) ((x).s12345670)

 #define ROT16_0(x) ((x))
 #define ROT16_1(x) ((x).sF0123456789ABCDE)
 #define ROT16_2(x) ((x).sEF0123456789ABCD)
 #define ROT16_3(x) ((x).sDEF0123456789ABC)
 #define ROT16_4(x) ((x).sCDEF0123456789AB)
 #define ROT16_5(x) ((x).sBCDEF0123456789A)
 #define ROT16_6(x) ((x).sABCDEF0123456789)
 #define ROT16_7(x) ((x).s9ABCDEF012345678)
 #define ROT16_8(x) ((x).s89ABCDEF01234567)
 #define ROT16_9(x) ((x).s789ABCDEF0123456)
 #define ROT16_10(x) ((x).s6789ABCDEF012345)
 #define ROT16_11(x) ((x).s56789ABCDEF01234)
 #define ROT16_12(x) ((x).s456789ABCDEF0123)
 #define ROT16_13(x) ((x).s3456789ABCDEF012)
 #define ROT16_14(x) ((x).s23456789ABCDEF01)
 #define ROT16_15(x) ((x).s123456789ABCDEF0)
 /** @} */ // end of group ROTs_n

 /** Circular-right-shift (rotate-right) the given vector by the given amount.
  * @name ROTATE
  *
  * @param[in] x The vector to be shifted
  * @param[in] s The size of the vector
  * @param[in] n The amount to be shifted
  *
  * @return The shifted vector
  * @{
  */
 #define ROTATE_STR(x, s, n) ROT##s##_##n(x)
 #define ROTATE(x, s, n) ROTATE_STR(x, s, n)
 /** @} */ // end of group ROTATE

 /** Creates a vector of size n filled with offset values corresponding to the location of each element.
  * @name V_OFFSn
  *
  * @param[in] dt The data type of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define V_OFFS1(dt) (dt##1)(0)
 #define V_OFFS2(dt) (dt##2)(0, 1)
 #define V_OFFS3(dt) (dt##3)(0, 1, 2)
 #define V_OFFS4(dt) (dt##4)(0, 1, 2, 3)
 #define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7)
 #define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
 /** @} */ // end of group V_OFFSn

 /** Create a vector filled with offset values corresponding to the location of each element.
  * @name VEC_OFFS
  *
  * @param[in] dt The data type of the output vector
  * @param[in] s  The size of the output vector
  *
  * @return The vector filled with offset values
  * @{
  */
 #define VEC_OFFS_STR(dt, s) V_OFFS##s(dt)
 #define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s)
 /** @} */ // end of group VEC_OFFS

 #define VLOAD_STR(size) vload##size
 #define VLOAD(size) VLOAD_STR(size)

 #define PIXEL_UNIT4 1
 #define PIXEL_UNIT8 2
 #define PIXEL_UNIT16 4

 /** Utility macro to convert a vector size in pixel unit.
  *
  * @name CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT
  *
  * @param[in] vec_size Vector size. Only 4,8 and 16 is supported
  *
  * @return The pixel unit (number of pixels)
  * @{
  */
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size
 #define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size)
 /** @} */ // end of group CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT

 #define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord)));
 #define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord)));

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
 #define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord)));
 #define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)));
 #define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord)));
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)

 /** Utility macro to read a 2D OpenCL image object.
  *
  * @note Coordinates are not normalized
  *
  * @param[in] data_type Data type
  * @param[in] n0        Number of pixel to read. Only 1,2 and 4 is supported
  * @param[in] img       OpenCL image object
  * @param[in] x_coord   The x coordinate for the top-left pixel
  * @param[in] y_coord   The y coordinate for the top-left pixel
  *
  * @return Pixels from the 2D OpenCL image object
  * @{
  */
 #define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord)
 #define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord)

 #define VSTORE_STR(size) vstore##size
 #define VSTORE(size) VSTORE_STR(size)

 #define float1 float
 #define half1 half
 #define char1 char
 #define uchar1 uchar
 #define short1 short
 #define ushort1 ushort
 #define int1 int
 #define uint1 uint
 #define long1 long
 #define ulong1 ulong
 #define double1 double

 #define vload1(OFFSET, PTR) *(OFFSET + PTR)
 #define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA

 /** Extended partial vstore that correctly handles scalar values as well.
  * Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name VSTORE_PARTIAL
  *
  * @note With this macro, the passed data can be both a vector and a scalar
  * @note @p store_size needs to be <= @p size
  * eg 1: Valid
  * VSTORE_PARTIAL(16, 15) ...;
  * eg 2: Invalid
  * VSTORE_PARTIAL(4, 7) ...;
  *
  * @param[in] size       The width of @p DATA. Supported values: 1(scalar), 2, 3, 4, 8, 16
  * @param[in] store_size The number of lower elements to store. Supported values: 1-16, but has to be <= @p size
  * @{
  */
 #define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
 #define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)

 #define NO_STORE(data, offs, ptr) \
     {                             \
     }

 // Size == 1 (scalar)
 #define vstore_partial_1_0 NO_STORE
 #define vstore_partial_1_1 vstore1
 #define vstore_partial_1_2 NO_STORE
 #define vstore_partial_1_3 NO_STORE
 #define vstore_partial_1_4 NO_STORE
 #define vstore_partial_1_5 NO_STORE
 #define vstore_partial_1_6 NO_STORE
 #define vstore_partial_1_7 NO_STORE
 #define vstore_partial_1_8 NO_STORE
 #define vstore_partial_1_9 NO_STORE
 #define vstore_partial_1_10 NO_STORE
 #define vstore_partial_1_11 NO_STORE
 #define vstore_partial_1_12 NO_STORE
 #define vstore_partial_1_13 NO_STORE
 #define vstore_partial_1_14 NO_STORE
 #define vstore_partial_1_15 NO_STORE
 #define vstore_partial_1_16 NO_STORE
 // Size == 2
 #define vstore_partial_2_0 NO_STORE
 #define vstore_partial_2_1 vstore_partial_1
 #define vstore_partial_2_2 vstore_partial_2
 #define vstore_partial_2_3 NO_STORE
 #define vstore_partial_2_4 NO_STORE
 #define vstore_partial_2_5 NO_STORE
 #define vstore_partial_2_6 NO_STORE
 #define vstore_partial_2_7 NO_STORE
 #define vstore_partial_2_8 NO_STORE
 #define vstore_partial_2_9 NO_STORE
 #define vstore_partial_2_10 NO_STORE
 #define vstore_partial_2_11 NO_STORE
 #define vstore_partial_2_12 NO_STORE
 #define vstore_partial_2_13 NO_STORE
 #define vstore_partial_2_14 NO_STORE
 #define vstore_partial_2_15 NO_STORE
 #define vstore_partial_2_16 NO_STORE
 // Size == 3
 #define vstore_partial_3_0 NO_STORE
 #define vstore_partial_3_1 vstore_partial_1
 #define vstore_partial_3_2 vstore_partial_2
 #define vstore_partial_3_3 vstore_partial_3
 #define vstore_partial_3_4 NO_STORE
 #define vstore_partial_3_5 NO_STORE
 #define vstore_partial_3_6 NO_STORE
 #define vstore_partial_3_7 NO_STORE
 #define vstore_partial_3_8 NO_STORE
 #define vstore_partial_3_9 NO_STORE
 #define vstore_partial_3_10 NO_STORE
 #define vstore_partial_3_11 NO_STORE
 #define vstore_partial_3_12 NO_STORE
 #define vstore_partial_3_13 NO_STORE
 #define vstore_partial_3_14 NO_STORE
 #define vstore_partial_3_15 NO_STORE
 #define vstore_partial_3_16 NO_STORE
 // Size == 4
 #define vstore_partial_4_0 NO_STORE
 #define vstore_partial_4_1 vstore_partial_1
 #define vstore_partial_4_2 vstore_partial_2
 #define vstore_partial_4_3 vstore_partial_3
 #define vstore_partial_4_4 vstore_partial_4
 #define vstore_partial_4_5 NO_STORE
 #define vstore_partial_4_6 NO_STORE
 #define vstore_partial_4_7 NO_STORE
 #define vstore_partial_4_8 NO_STORE
 #define vstore_partial_4_9 NO_STORE
 #define vstore_partial_4_10 NO_STORE
 #define vstore_partial_4_11 NO_STORE
 #define vstore_partial_4_12 NO_STORE
 #define vstore_partial_4_13 NO_STORE
 #define vstore_partial_4_14 NO_STORE
 #define vstore_partial_4_15 NO_STORE
 #define vstore_partial_4_16 NO_STORE
 // Size == 8
 #define vstore_partial_8_0 NO_STORE
 #define vstore_partial_8_1 vstore_partial_1
 #define vstore_partial_8_2 vstore_partial_2
 #define vstore_partial_8_3 vstore_partial_3
 #define vstore_partial_8_4 vstore_partial_4
 #define vstore_partial_8_5 vstore_partial_5
 #define vstore_partial_8_6 vstore_partial_6
 #define vstore_partial_8_7 vstore_partial_7
 #define vstore_partial_8_8 vstore_partial_8
 #define vstore_partial_8_9 NO_STORE
 #define vstore_partial_8_10 NO_STORE
 #define vstore_partial_8_11 NO_STORE
 #define vstore_partial_8_12 NO_STORE
 #define vstore_partial_8_13 NO_STORE
 #define vstore_partial_8_14 NO_STORE
 #define vstore_partial_8_15 NO_STORE
 #define vstore_partial_8_16 NO_STORE
 // Size == 16
 #define vstore_partial_16_0 NO_STORE
 #define vstore_partial_16_1 vstore_partial_1
 #define vstore_partial_16_2 vstore_partial_2
 #define vstore_partial_16_3 vstore_partial_3
 #define vstore_partial_16_4 vstore_partial_4
 #define vstore_partial_16_5 vstore_partial_5
 #define vstore_partial_16_6 vstore_partial_6
 #define vstore_partial_16_7 vstore_partial_7
 #define vstore_partial_16_8 vstore_partial_8
 #define vstore_partial_16_9 vstore_partial_9
 #define vstore_partial_16_10 vstore_partial_10
 #define vstore_partial_16_11 vstore_partial_11
 #define vstore_partial_16_12 vstore_partial_12
 #define vstore_partial_16_13 vstore_partial_13
 #define vstore_partial_16_14 vstore_partial_14
 #define vstore_partial_16_15 vstore_partial_15
 #define vstore_partial_16_16 vstore_partial_16

 /** Partial vstore. Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
  * @name vstore_partial_n
  *
  * @note @p DATA needs to be a vector not a scalar
  * @note n needs to be <= the vector width of the input variable @p DATA
  * eg 1: Valid
  * vstore_partial_15(var:float16, 0, 0xabcd);
  * eg 2: Invalid
  * vstore_partial_7(var:float4, 0, 0xabcd);
  *
  * @note in cases n == 1, 2, 3, 4, 8, 16, no extra vstore is invoked, thus there's no performance penalty.
  *
  * @param[in] DATA   The name of the variable
  * @param[in] OFFSET Offset in n
  * @param[in] PTR    The base pointer
  * @{
  */
 #define vstore_partial_1(DATA, OFFSET, PTR) \
     vstore1(DATA.s0, OFFSET, PTR);

 #define vstore_partial_2(DATA, OFFSET, PTR) \
     vstore2(DATA.s01, OFFSET, PTR);

 #define vstore_partial_3(DATA, OFFSET, PTR) \
     vstore3(DATA.s012, OFFSET, PTR);

 #define vstore_partial_4(DATA, OFFSET, PTR) \
     vstore4(DATA.s0123, OFFSET, PTR);

 #define vstore_partial_5(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore1(DATA.s4, OFFSET, PTR + 4);

 #define vstore_partial_6(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_2(DATA.s45, OFFSET, PTR + 4);

 #define vstore_partial_7(DATA, OFFSET, PTR)    \
     vstore_partial_4(DATA.s0123, OFFSET, PTR); \
     vstore_partial_3(DATA.s456, OFFSET, PTR + 4);

 #define vstore_partial_8(DATA, OFFSET, PTR) \
     vstore8(DATA.s01234567, OFFSET, PTR);

 #define vstore_partial_9(DATA, OFFSET, PTR)        \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore1(DATA.s8, OFFSET, PTR + 8);

 #define vstore_partial_10(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_2(DATA.s89, OFFSET, PTR + 8);

 #define vstore_partial_11(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_3(DATA.s89a, OFFSET, PTR + 8);

 #define vstore_partial_12(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8);

 #define vstore_partial_13(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_14(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_15(DATA, OFFSET, PTR)       \
     vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
     vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);

 #define vstore_partial_16(DATA, OFFSET, PTR) \
     vstore16(DATA, OFFSET, PTR);
 /** @} */ // end of groupd vstore_partial_n
 /** @} */ // end of groupd VSTORE_PARTIAL

 // Convert built-in functions with _sat modifier are not supported in floating point so we create defines
 // without _sat to overcome this issue
 #define convert_float_sat convert_float
 #define convert_float1_sat convert_float
 #define convert_float2_sat convert_float2
 #define convert_float3_sat convert_float3
 #define convert_float4_sat convert_float4
 #define convert_float8_sat convert_float8
 #define convert_float16_sat convert_float16
 #define convert_half_sat convert_float
 #define convert_half1_sat convert_half
 #define convert_half2_sat convert_half2
 #define convert_half3_sat convert_half3
 #define convert_half4_sat convert_half4
 #define convert_half8_sat convert_half8
 #define convert_half16_sat convert_half16

 #define convert_float1 convert_float
 #define convert_half1 convert_half
 #define convert_char1 convert_char
 #define convert_uchar1 convert_uchar
 #define convert_short1 convert_short
 #define convert_ushort1 convert_ushort
 #define convert_int1 convert_int
 #define convert_uint1 convert_uint
 #define convert_long1 convert_long
 #define convert_ulong1 convert_ulong
 #define convert_double1 convert_double

 #define convert_char1_sat convert_char_sat
 #define convert_uchar1_sat convert_uchar_sat
 #define convert_short1_sat convert_short_sat
 #define convert_ushort1_sat convert_ushort_sat
 #define convert_int1_sat convert_int_sat
 #define convert_uint1_sat convert_uint_sat
 #define convert_long1_sat convert_long_sat
 #define convert_ulong1_sat convert_ulong_sat
 #define convert_double1_sat convert_double_sat

 #define VEC_DATA_TYPE_STR(type, size) type##size
 #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)

 #define CONVERT_STR(x, type) (convert_##type((x)))
 #define CONVERT(x, type) CONVERT_STR(x, type)

 #define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x)))
 #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)

 #define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x)))
 #define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round)

 #define select_vec_dt_uchar(size) uchar##size
 #define select_vec_dt_char(size) char##size
 #define select_vec_dt_ushort(size) ushort##size
 #define select_vec_dt_short(size) short##size
 #define select_vec_dt_half(size) short##size
 #define select_vec_dt_uint(size) uint##size
 #define select_vec_dt_int(size) int##size
 #define select_vec_dt_float(size) int##size
 #define select_vec_dt_ulong(size) ulong##size
 #define select_vec_dt_long(size) long##size

 #define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size)
 #define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size)
 #define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1)

 #define sum_reduce_1(x) (x)
 #define sum_reduce_2(x) ((x).s0) + ((x).s1)
 #define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2)
 #define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23)
 #define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567)
 #define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF)

 #define SUM_REDUCE_STR(x, size) sum_reduce_##size(x)
 #define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size)

 #define max_reduce_1(x) (x)
 #define max_reduce_2(x) max(((x).s0), ((x).s1))
 #define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2))
 #define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23))
 #define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567))
 #define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF))

 #define MAX_REDUCE_STR(x, size) max_reduce_##size(x)
 #define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size)

 #define VECTOR_DECLARATION(name)     \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_offset_first_element_in_bytes

 #define IMAGE_DECLARATION(name)      \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR3D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_offset_first_element_in_bytes

 #define TENSOR4D_DECLARATION(name)   \
     __global uchar *name##_ptr,      \
     uint        name##_stride_x, \
     uint        name##_step_x,   \
     uint        name##_stride_y, \
     uint        name##_step_y,   \
     uint        name##_stride_z, \
     uint        name##_step_z,   \
     uint        name##_stride_w, \
     uint        name##_step_w,   \
     uint        name##_offset_first_element_in_bytes

 #define CONVERT_TO_VECTOR_STRUCT(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)

 #define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
     update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0)

 #define CONVERT_TO_IMAGE_STRUCT(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)

 #define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z)

 #define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
     update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                           \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
     update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0)

 #define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size)                                                                                                 \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                                  name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)

 #define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
     update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)

 #define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name)                                                                                       \
     tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
                            name##_stride_z, name##_step_z)

 /** Structure to hold Vector information */
 typedef struct Vector
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
 } Vector;

 /** Structure to hold Image information */
 typedef struct Image
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
 } Image;

 /** Structure to hold 3D tensor information */
 typedef struct Tensor3D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
 } Tensor3D;

 /** Structure to hold 4D tensor information */
 typedef struct Tensor4D
 {
     __global uchar *ptr;                           /**< Pointer to the starting postion of the buffer */
     int             offset_first_element_in_bytes; /**< The offset of the first element in the source image */
     int             stride_x;                      /**< Stride of the image in X dimension (in bytes) */
     int             stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
     int             stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
     int             stride_w;                      /**< Stride of the image in W dimension (in bytes) */
 } Tensor4D;

 /** Wrap vector information into an Vector structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
  * @param[in] stride_x                      Stride of the vector in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x)
 {
     Vector vector =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
     };
     vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x;
     return vector;
 }

 /** Wrap image information into an Image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  *
  * @return An image object
  */
 inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y;
     return img;
 }

 /** Wrap 3D tensor information into an image structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Image img =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y
     };
     img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return img;
 }

 /** Wrap 3D tensor information into an tensor structure, and make the pointer point at this workitem's data.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
     return tensor;
 }

 /** Wrap 3D tensor information into an tensor structure.
  *
  * @param[in] ptr                           Pointer to the starting postion of the buffer
  * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
  * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
  * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
  * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
  * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
  *
  * @return A 3D tensor object
  */
 inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
 {
     Tensor3D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z
     };
     return tensor;
 }

 inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w,
                                              uint step_w,
                                              uint mod_size)
 {
     Tensor4D tensor =
     {
         .ptr                           = ptr,
         .offset_first_element_in_bytes = offset_first_element_in_bytes,
         .stride_x                      = stride_x,
         .stride_y                      = stride_y,
         .stride_z                      = stride_z,
         .stride_w                      = stride_w
     };

     tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w;
     return tensor;
 }

 /** Get the pointer position of a Vector
  *
  * @param[in] vec Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  */
 inline __global const uchar *vector_offset(const Vector *vec, int x)
 {
     return vec->ptr + x * vec->stride_x;
 }

 /** Get the pointer position of a Image
  *
  * @param[in] img Pointer to the starting position of the buffer
  * @param[in] x   Relative X position
  * @param[in] y   Relative Y position
  */
 inline __global uchar *offset(const Image *img, int x, int y)
 {
     return img->ptr + x * img->stride_x + y * img->stride_y;
 }

 /** Get the pointer position of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  */
 inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z;
 }

 /** Get the pointer position of a Tensor4D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] x      Relative X position
  * @param[in] y      Relative Y position
  * @param[in] z      Relative Z position
  * @param[in] w      Relative W position
  */
 inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w)
 {
     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w;
 }

 /** Get the offset for a given linear index of a Tensor3D
  *
  * @param[in] tensor Pointer to the starting position of the buffer
  * @param[in] width  Width of the input tensor
  * @param[in] height Height of the input tensor
  * @param[in] depth  Depth of the input tensor
  * @param[in] index  Linear index
  */
 inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index)
 {
     uint num_elements = width * height;

     const uint z = index / num_elements;

     index %= num_elements;

     const uint y = index / width;

     index %= width;

     const uint x = index;

     return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes;
 }

 #endif // _HELPER_H

 /** Macros that help in loop unrolling */
 //Repeat macros with 3 param, excluding the implicit ID param
 #define REPEAT_3_1(P_X, P_A, P_B, P_C) P_X##_DEF(0, P_A, P_B, P_C)
 #define REPEAT_3_2(P_X, P_A, P_B, P_C) \
     P_X##_DEF(1, P_A, P_B, P_C);       \
     REPEAT_3_1(P_X, P_A, P_B, P_C)
 #define REPEAT_3_3(P_X, P_A, P_B, P_C) \
     P_X##_DEF(2, P_A, P_B, P_C);       \
     REPEAT_3_2(P_X, P_A, P_B, P_C)
 #define REPEAT_3_4(P_X, P_A, P_B, P_C) \
     P_X##_DEF(3, P_A, P_B, P_C);       \
     REPEAT_3_3(P_X, P_A, P_B, P_C)
 #define REPEAT_3_5(P_X, P_A, P_B, P_C) \
     P_X##_DEF(4, P_A, P_B, P_C);       \
     REPEAT_3_4(P_X, P_A, P_B, P_C)
 #define REPEAT_3_6(P_X, P_A, P_B, P_C) \
     P_X##_DEF(5, P_A, P_B, P_C);       \
     REPEAT_3_5(P_X, P_A, P_B, P_C)
 #define REPEAT_3_7(P_X, P_A, P_B, P_C) \
     P_X##_DEF(6, P_A, P_B, P_C);       \
     REPEAT_3_6(P_X, P_A, P_B, P_C)
 #define REPEAT_3_8(P_X, P_A, P_B, P_C) \
     P_X##_DEF(7, P_A, P_B, P_C);       \
     REPEAT_3_7(P_X, P_A, P_B, P_C)
 #define REPEAT_3_9(P_X, P_A, P_B, P_C) \
     P_X##_DEF(8, P_A, P_B, P_C);       \
     REPEAT_3_8(P_X, P_A, P_B, P_C)
 #define REPEAT_3_10(P_X, P_A, P_B, P_C) \
     P_X##_DEF(9, P_A, P_B, P_C);        \
     REPEAT_3_9(P_X, P_A, P_B, P_C)
 #define REPEAT_3_11(P_X, P_A, P_B, P_C) \
     P_X##_DEF(A, P_A, P_B, P_C);        \
     REPEAT_3_10(P_X, P_A, P_B, P_C)
 #define REPEAT_3_12(P_X, P_A, P_B, P_C) \
     P_X##_DEF(B, P_A, P_B, P_C);        \
     REPEAT_3_11(P_X, P_A, P_B, P_C)
 #define REPEAT_3_13(P_X, P_A, P_B, P_C) \
     P_X##_DEF(C, P_A, P_B, P_C);        \
     REPEAT_3_12(P_X, P_A, P_B, P_C)
 #define REPEAT_3_14(P_X, P_A, P_B, P_C) \
     P_X##_DEF(D, P_A, P_B, P_C);        \
     REPEAT_3_13(P_X, P_A, P_B, P_C)
 #define REPEAT_3_15(P_X, P_A, P_B, P_C) \
     P_X##_DEF(E, P_A, P_B, P_C);        \
     REPEAT_3_14(P_X, P_A, P_B, P_C)
 #define REPEAT_3_16(P_X, P_A, P_B, P_C) \
     P_X##_DEF(F, P_A, P_B, P_C);        \
     REPEAT_3_15(P_X, P_A, P_B, P_C)

 #define REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_3_##P_NUM(P_OP, P_A, P_B, P_C) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
 #define REPEAT_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C)

 // Repeat macros with 4 param, excluding the implicit ID param
 #define REPEAT_4_1(P_X, P_A, P_B, P_C, P_D) P_X##_DEF(0, P_A, P_B, P_C, P_D)
 #define REPEAT_4_2(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(1, P_A, P_B, P_C, P_D);       \
     REPEAT_4_1(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_3(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(2, P_A, P_B, P_C, P_D);       \
     REPEAT_4_2(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_4(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(3, P_A, P_B, P_C, P_D);       \
     REPEAT_4_3(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_5(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(4, P_A, P_B, P_C, P_D);       \
     REPEAT_4_4(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_6(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(5, P_A, P_B, P_C, P_D);       \
     REPEAT_4_5(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_7(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(6, P_A, P_B, P_C, P_D);       \
     REPEAT_4_6(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_8(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(7, P_A, P_B, P_C, P_D);       \
     REPEAT_4_7(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_9(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(8, P_A, P_B, P_C, P_D);       \
     REPEAT_4_8(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_10(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(9, P_A, P_B, P_C, P_D);        \
     REPEAT_4_9(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_11(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(A, P_A, P_B, P_C, P_D);        \
     REPEAT_4_10(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_12(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(B, P_A, P_B, P_C, P_D);        \
     REPEAT_4_11(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_13(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(C, P_A, P_B, P_C, P_D);        \
     REPEAT_4_12(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_14(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(D, P_A, P_B, P_C, P_D);        \
     REPEAT_4_13(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_15(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(E, P_A, P_B, P_C, P_D);        \
     REPEAT_4_14(P_X, P_A, P_B, P_C, P_D)
 #define REPEAT_4_16(P_X, P_A, P_B, P_C, P_D) \
     P_X##_DEF(F, P_A, P_B, P_C, P_D);        \
     REPEAT_4_15(P_X, P_A, P_B, P_C, P_D)

 #define REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_4_##P_NUM(P_OP, P_A, P_B, P_C, P_D) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
 #define REPEAT_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D)

 // Macro for initializing N variables. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
 #define VAR_INIT_TO_CONST_DEF(ID, TYPE, VAR, VAL) TYPE VAR##ID = VAL
 #define REPEAT_VAR_INIT_TO_CONST(N, TYPE, VAR, VAL) REPEAT_3_N(N, VAR_INIT_TO_CONST, TYPE, VAR, VAL)

 // Macro for initializing N variables by converting the data type. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
 #define VAR_INIT_CONVERT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT(VAR_IN##ID, TYPE_OUT)
 #define REPEAT_VAR_INIT_CONVERT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT, TYPE_OUT, VAR_IN, VAR_OUT)

 // Macro for initializing N variables by converting the data type with saturation. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
 #define VAR_INIT_CONVERT_SAT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT_SAT(VAR_IN##ID, TYPE_OUT)
 #define REPEAT_VAR_INIT_CONVERT_SAT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT_SAT, TYPE_OUT, VAR_IN, VAR_OUT)

 // Macro for adding a constant to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define ADD_CONST_TO_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID += (TYPE)VAL
 #define REPEAT_ADD_CONST_TO_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, ADD_CONST_TO_VAR, TYPE, VAR, VAL)

 // Macro for multiplying N variables (VAR_B) by a constant (VAL) and adding to other N variables (VAR_A). Generates N statements that defines VAR_A##N =RHS_ACCESSOR_DEF(...)
 #define MLA_VAR_WITH_CONST_VEC_DEF(ID, VAR_A, VAR_B, VAL) VAR_A##ID += VAR_B##ID * VAL
 #define REPEAT_MLA_VAR_WITH_CONST_VEC(N, VAR_A, VAR_B, VAL) REPEAT_3_N(N, MLA_VAR_WITH_CONST_VEC, VAR_A, VAR_B, VAL)

 // Macro for adding a vector to N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define ADD_VECTOR_TO_VAR_DEF(ID, TYPE, VAR, VEC) VAR##ID += VEC
 #define REPEAT_ADD_VECTOR_TO_VAR(N, VAR, VEC) REPEAT_3_N(N, ADD_VECTOR_TO_VAR, "", VAR, VEC)

 // Macro for adding a two N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define ADD_TWO_VARS_DEF(ID, TYPE, VAR_A, VAR_B) VAR_A##ID += VAR_B##ID
 #define REPEAT_ADD_TWO_VARS(N, VAR_A, VAR_B) REPEAT_3_N(N, ADD_TWO_VARS, "", VAR_A, VAR_B)

 // Macro for performing Max between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define MAX_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = max(VAR##ID, (TYPE)VAL)
 #define REPEAT_MAX_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MAX_CONST_VAR, TYPE, VAR, VAL)

 // Macro for performing Min between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define MIN_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = min(VAR##ID, (TYPE)VAL)
 #define REPEAT_MIN_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MIN_CONST_VAR, TYPE, VAR, VAL)

 // Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
 #define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)

 // Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
 #define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)

 // Macro for performing per-channel ASYMM_MULT_BY_QUANT_MULTIPLIER to N variables.
 #define ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT)                     \
     ({                                                                                                        \
         VEC_DATA_TYPE(int, N0)                                                                                \
         VAR##ID_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0); \
         VEC_DATA_TYPE(int, N0)                                                                                \
         VAR##ID_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0);    \
         VAR##ID           = select(VAR##ID_shift_lt0, VAR##ID_shift_gt0, RES_SHIFT >= 0);                     \
     })
 #define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL, SIZE, VAR, RES_MUL, RES_SHIFT)

 #endif // ARM_COMPUTE_REPEAT_H

 #if defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0)
 #define INC2 (VEC_DATA_TYPE(uint, 2))(0, 1)
 #define INC3 (VEC_DATA_TYPE(uint, 3))(0, 1, 2)
 #define INC4 (VEC_DATA_TYPE(uint, 4))(0, 1, 2, 3)
 #define INC8 (VEC_DATA_TYPE(uint, 8))(0, 1, 2, 3, 4, 5, 6, 7)
 #define INC16 (VEC_DATA_TYPE(uint, 16))(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
 #define CONCAT_INC(K0) INC##K0
 #define INC(K0) CONCAT_INC(K0)

 #if(SRC_WIDTH % K0)
 #define BOUNDARY_CONDITION_X(x, a)                                                                                                                   \
     ({                                                                                                                                               \
         a = select(0, a, CONVERT(((x * (VEC_DATA_TYPE(uint, K0))K0 + INC(K0)) < (VEC_DATA_TYPE(uint, K0))SRC_WIDTH), VEC_DATA_TYPE(DATA_TYPE, K0))); \
     })
 #else // (SRC_WIDTH % K0)
 #define BOUNDARY_CONDITION_X(x, a) \
     ({})
 #endif // (SRC_WIDTH % K0)

 #define LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin)                     \
     ({                                                                                                           \
         if(y * M0 + M0 >= SRC_HEIGHT && PARTIAL_LOAD_M0 != 0)                                                    \
         {                                                                                                        \
             if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0))                                               \
             {                                                                                                    \
                 LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \
             }                                                                                                    \
             else                                                                                                 \
             {                                                                                                    \
                 LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin);              \
             }                                                                                                    \
         }                                                                                                        \
         else                                                                                                     \
         {                                                                                                        \
             if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0))                                               \
             {                                                                                                    \
                 LOAD_TENSOR_M0XN0(M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin);              \
             }                                                                                                    \
             else                                                                                                 \
             {                                                                                                    \
                 LOAD_TENSOR_M0XN0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin);                           \
             }                                                                                                    \
         }                                                                                                        \
     })

 /** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (not transposed) in
  *  the output matrix unrolling the values.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16)
  * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16)
  * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2).
  * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1)
  * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1)
  * @note Only the following values for M0, K0 and V0 are supported:
  *                                      M0: 2,3,4,5,6,7,8
  *                                      K0: 2,3,4,8,16
  *                                      V0: greater than 0
  * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
  * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
  *
  * @param[in]  src_ptr                           Pointer to the source LHS tensor. Supported data types: All
  * @param[in]  src_stride_x                      Stride of the source LHS tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source LHS tensor in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source LHS tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  * @param[in]  cross_plane_pad                   (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  */
 __kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src),
                                          TENSOR3D_DECLARATION(dst)
 #if defined(REINTERPRET_INPUT_AS_3D)
                                          ,
                                          uint cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
                                         )
 {
     // Block size
 #define BLOCK_SIZE ((M0) * (K0))

     // Output offset X
 #if defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (K0)
 #else // defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (BLOCK_SIZE)
 #endif // defined(INTERLEAVE)

     // Output step X
 #if defined(INTERLEAVE)
 #define OUTPUT_STEP_X (K0) * (V0)
 #else // Do not interleave
 #define OUTPUT_STEP_X (K0)
 #endif // defined(INTERLEAVE)

     // Compute source and destination addresses
     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

     // ------------------ Compute input/output addresses ---------------------------

     // Compute the input address
     __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y;

     // Compute the output address
     __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) *
                                  (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE));

     // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0;
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply src_stride_z by DEPTH_GEMM3D

     input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D;

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y);

 #else // defined(REINTERPRET_INPUT_AS_3D)

     input_ptr += z * (uint)src_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     output_ptr += z * (uint)dst_stride_z;

     // ---------------------------Load input values --------------------------------
     // Load values from the LHS matrix
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0);

     LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin);

     // ---------------------------Store output values ------------------------------
     REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0);
     STORE_BLOCK(M0, K0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout);

 #undef BLOCK_SIZE
 #undef OUTPUT_OFFSET_X
 #undef OUTPUT_STEP_X
 }

 #if M0 == 2
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                  \
     ({                                                                                            \
         VEC_DATA_TYPE(DATA_TYPE, M0)                                                              \
         res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i);                                   \
         VSTORE(M0)                                                                                \
         (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \
     })
 #elif M0 == 3 // M0 == 3
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                  \
     ({                                                                                            \
         VEC_DATA_TYPE(DATA_TYPE, M0)                                                              \
         res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i);                          \
         VSTORE(M0)                                                                                \
         (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \
     })
 #elif M0 == 4 // M0 == 4
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                  \
     ({                                                                                            \
         VEC_DATA_TYPE(DATA_TYPE, M0)                                                              \
         res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i);                 \
         VSTORE(M0)                                                                                \
         (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \
     })
 #elif M0 == 5 // M0 == 5
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                      \
     ({                                                                                                \
         VEC_DATA_TYPE(DATA_TYPE, 4)                                                                   \
         res0           = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i);           \
         DATA_TYPE res1 = a4.s##i;                                                                     \
         VSTORE(4)                                                                                     \
         (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)));    \
         *((__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4) = res1; \
     })
 #elif M0 == 6 // M0 == 6
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                       \
     ({                                                                                                 \
         VEC_DATA_TYPE(DATA_TYPE, 4)                                                                    \
         res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i);                      \
         VEC_DATA_TYPE(DATA_TYPE, 2)                                                                    \
         res1 = (VEC_DATA_TYPE(DATA_TYPE, 2))(a4.s##i, a5.s##i);                                        \
         VSTORE(4)                                                                                      \
         (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)));     \
         VSTORE(2)                                                                                      \
         (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \
     })
 #elif M0 == 7 // M0 == 7
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                       \
     ({                                                                                                 \
         VEC_DATA_TYPE(DATA_TYPE, 4)                                                                    \
         res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i);                      \
         VEC_DATA_TYPE(DATA_TYPE, 3)                                                                    \
         res1 = (VEC_DATA_TYPE(DATA_TYPE, 3))(a4.s##i, a5.s##i, a6.s##i);                               \
         VSTORE(4)                                                                                      \
         (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)));     \
         VSTORE(3)                                                                                      \
         (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \
     })
 #elif M0 == 8 // M0 == 8
 #define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i)                                                      \
     ({                                                                                                                \
         VEC_DATA_TYPE(DATA_TYPE, M0)                                                                                  \
         res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i, a4.s##i, a5.s##i, a6.s##i, a7.s##i); \
         VSTORE(M0)                                                                                                    \
         (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)));                     \
     })
 #else // M0 not supported
 #error "M0 value not supported"
 #endif // N0 conditions

 /** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (transposed) in
  *  the output matrix unrolling the values.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16)
  * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16)
  * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2).
  * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1)
  * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1)
  * @note Only the following values for M0, K0 and V0 are supported:
  *                                      M0: 2,3,4,5,6,7,8
  *                                      K0: 2,3,4,8,16
  *                                      V0: greater than 0
  * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped
  * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
  *
  * @param[in]  src_ptr                           Pointer to the source LHS tensor. Supported data types: All
  * @param[in]  src_stride_x                      Stride of the source LHS tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source LHS tensor in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source LHS tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  * @param[in]  cross_plane_pad                   (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  */
 __kernel void gemm_reshape_lhs_matrix_t(TENSOR3D_DECLARATION(src),
                                         TENSOR3D_DECLARATION(dst)
 #if defined(REINTERPRET_INPUT_AS_3D)
                                         ,
                                         uint cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
                                        )
 {
     // Block size
 #define BLOCK_SIZE ((M0) * (K0))

     // Output offset X
 #if defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (M0)
 #else // defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (BLOCK_SIZE)
 #endif // defined(INTERLEAVE)

     // Output step X
 #if defined(INTERLEAVE)
 #define OUTPUT_STEP_X (M0) * (V0)
 #else // Do not interleave
 #define OUTPUT_STEP_X (M0)
 #endif // defined(INTERLEAVE)

     // Compute source and destination addresses
     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

     // ------------------ Compute input/output addresses ---------------------------

     // Compute the input address
     __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y;

     // Compute the output address
     __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) *
                                  (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE));

     // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0;
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply src_stride_z by DEPTH_GEMM3D

     input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D;

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y);

 #else // defined(REINTERPRET_INPUT_AS_3D)

     input_ptr += z * (uint)src_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     output_ptr += z * (uint)dst_stride_z;

     // ---------------------------Load input values --------------------------------
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0);

     LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin);

     // ---------------------------Transpose and store block -----------------------

     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 0);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 1);
 #if K0 > 2
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 2);
 #endif // K0 > 2
 #if K0 > 3
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 3);
 #endif // K0 > 3
 #if K0 > 4
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 4);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 5);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 6);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 7);
 #endif // K0 > 4
 #if K0 > 8
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 8);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 9);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, A);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, B);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, C);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, D);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, E);
     TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, F);
 #endif // K0 > 8

 #undef BLOCK_SIZE
 #undef OUTPUT_OFFSET_X
 #undef OUTPUT_STEP_X
 }
 #endif // defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0)

 #if defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT)
 /** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (not transposed) in
  *  the output matrix unrolling the values.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16)
  * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2).
  * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
  * @note Only the following values for K0, N0 and H0 are supported:
  *                                      N0: 2,3,4,8,16
  *                                      K0: 1,2,3,4,8,16
  *                                      H0: greater than 0
  *
  * @param[in]  src_ptr                           Pointer to the source RHS tensor. Supported data types: All
  * @param[in]  src_stride_x                      Stride of the source RHS tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source RHS tensor in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source RHS tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  */
 __kernel void gemm_reshape_rhs_matrix_nt(TENSOR3D_DECLARATION(src),
                                          TENSOR3D_DECLARATION(dst))
 {
     // Block size
 #define BLOCK_SIZE ((K0) * (N0))

     // Output offset X
 #if defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (N0)
 #else // defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (BLOCK_SIZE)
 #endif // defined(INTERLEAVE)

     // Output step X
 #if defined(INTERLEAVE)
 #define OUTPUT_STEP_X (N0) * (H0)
 #else // Do not interleave
 #define OUTPUT_STEP_X (N0)
 #endif // defined(INTERLEAVE)

     // Compute source and destination addresses
     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

     // ------------------ Compute input/output addresses ---------------------------

     // Compute the input address
     __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z;

     // Compute the output address
     __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % (uint)H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((
                                      x / (uint)H0)
                                  * (uint)dst_stride_y)
                                  + z * (uint)dst_stride_z;

     // ---------------------------Load input values --------------------------------

     REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); ////uint a0=0, a1=0, a2=0...a(M0-1)=0;

     // Load values from the RHS matrix
     a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
 #if K0 > 1
     if(y * (uint)K0 + 1 < SRC_HEIGHT)
     {
         a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
     }
 #endif // K0 > 1
 #if K0 > 2
     if(y * (uint)K0 + 2 < SRC_HEIGHT)
     {
         a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
     }
 #endif // K0 > 2
 #if K0 > 3
     if(y * (uint)K0 + 3 < SRC_HEIGHT)
     {
         a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y));
     }
 #endif // K0 > 3
 #if K0 > 4
     if(y * (uint)K0 + 4 < SRC_HEIGHT)
     {
         a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y));
     }
     if(y * (uint)K0 + 5 < SRC_HEIGHT)
     {
         a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y));
     }
     if(y * (uint)K0 + 6 < SRC_HEIGHT)
     {
         a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y));
     }
     if(y * (uint)K0 + 7 < SRC_HEIGHT)
     {
         a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y));
     }
 #endif // K0 > 4
 #if K0 > 8
     if(y * (uint)K0 + 8 < SRC_HEIGHT)
     {
         a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y));
     }
     if(y * (uint)K0 + 9 < SRC_HEIGHT)
     {
         a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y));
     }
     if(y * (uint)K0 + 10 < SRC_HEIGHT)
     {
         aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y));
     }
     if(y * (uint)K0 + 11 < SRC_HEIGHT)
     {
         aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y));
     }
     if(y * (uint)K0 + 12 < SRC_HEIGHT)
     {
         aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y));
     }
     if(y * (uint)K0 + 13 < SRC_HEIGHT)
     {
         aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y));
     }
     if(y * (uint)K0 + 14 < SRC_HEIGHT)
     {
         aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y));
     }
     if(y * (uint)K0 + 15 < SRC_HEIGHT)
     {
         aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y));
     }
 #endif // K0 > 8

     // ---------------------------Store output values ------------------------------
     REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0);
     STORE_BLOCK(K0, N0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout);

 #undef BLOCK_SIZE
 #undef OUTPUT_OFFSET_X
 #undef OUTPUT_STEP_X
 }

 #if defined(TRANSPOSE)
 /** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (transposed) in
  *  the output matrix unrolling the values.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16)
  * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2).
  * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time.
  * @note The option -DTRANSPOSE must passed at compile time.
  * @note Only the following values for K0, N0 and H0 are supported:
  *                                      N0: 2,3,4,8,16
  *                                      K0: 2,3,4,8,16
  *                                      H0: greater than 0
  *
  * @param[in]  src_ptr                           Pointer to the source RHS tensor. Supported data types: All
  * @param[in]  src_stride_x                      Stride of the source RHS tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source RHS tensor in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source RHS tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  */
 __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src),
                                         TENSOR3D_DECLARATION(dst))
 {
     // Block size
 #define BLOCK_SIZE ((K0) * (N0))

     // Output offset X
 #if defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (K0)
 #else // defined(INTERLEAVE)
 #define OUTPUT_OFFSET_X (BLOCK_SIZE)
 #endif // defined(INTERLEAVE)

     // Output step X
 #if defined(INTERLEAVE)
 #define OUTPUT_STEP_X (K0) * (H0)
 #else // Do not interleave
 #define OUTPUT_STEP_X (K0)
 #endif // defined(INTERLEAVE)

     // Compute source and destination addresses
     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

     // ------------------ Compute input/output addresses ---------------------------

     // Compute the input address
     __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z;

     // Compute the output address
     __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((x /
                                  (uint)H0) * (uint)dst_stride_y) + z * (uint)dst_stride_z;

     // ---------------------------Load input values --------------------------------
     REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    a0=0, a1=0, ... a(K0-1)=0;

     // Load values from the RHS matrix
     a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y));
     if(y * (uint)K0 + 1 < SRC_HEIGHT)
     {
         a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));
     }
 #if K0 > 2
     if(y * (uint)K0 + 2 < SRC_HEIGHT)
     {
         a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));
     }
 #endif // K0 > 2
 #if K0 > 3
     if(y * (uint)K0 + 3 < SRC_HEIGHT)
     {
         a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y));
     }
 #endif // K0 > 3
 #if K0 > 4
     if(y * (uint)K0 + 4 < SRC_HEIGHT)
     {
         a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y));
     }
     if(y * (uint)K0 + 5 < SRC_HEIGHT)
     {
         a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y));
     }
     if(y * (uint)K0 + 6 < SRC_HEIGHT)
     {
         a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y));
     }
     if(y * (uint)K0 + 7 < SRC_HEIGHT)
     {
         a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y));
     }
 #endif // K0 > 4
 #if K0 > 8
     if(y * (uint)K0 + 8 < SRC_HEIGHT)
     {
         a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y));
     }
     if(y * (uint)K0 + 9 < SRC_HEIGHT)
     {
         a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y));
     }
     if(y * (uint)K0 + 10 < SRC_HEIGHT)
     {
         aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y));
     }
     if(y * (uint)K0 + 11 < SRC_HEIGHT)
     {
         aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y));
     }
     if(y * (uint)K0 + 12 < SRC_HEIGHT)
     {
         aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y));
     }
     if(y * (uint)K0 + 13 < SRC_HEIGHT)
     {
         aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y));
     }
     if(y * (uint)K0 + 14 < SRC_HEIGHT)
     {
         aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y));
     }
     if(y * (uint)K0 + 15 < SRC_HEIGHT)
     {
         aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y));
     }
 #endif // K0 > 8

     // ---------------------------Transpose the block ------------------------------
     REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), res, 0); //VEC_DATA_TYPE(DATA_TYPE, K0)    res0=0, res1=0, res2=0,... res(N0-1)=0;

 #if K0 == 2
     // This part computes the following transpositions:
     // 2x2 -> 2x2
     // 2x4 -> 4x2
     // 2x8 -> 8x2
     // 2x16 -> 16x2
     res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0);
     res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1);
 #if N0 > 2
     res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2);
 #endif // N0 > 2
 #if N0 > 3
     res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3);
 #endif // N0 > 3
 #if N0 > 4
     res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4);
     res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5);
     res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6);
     res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7);
 #endif // N0 > 4
 #if N0 > 8
     res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8);
     res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9);
     resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA);
     resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB);
     resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC);
     resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD);
     resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE);
     resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF);
 #endif // N0 > 8

 #elif K0 == 3 // K0 == 2
     // This part computes the following transpositions:
     // 3x2 -> 2x3
     // 3x4 -> 4x3
     // 3x8 -> 8x3
     // 3x16 -> 16x3
     res0                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0);
     res1                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1);
 #if N0 > 2
     res2                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2);
 #endif // N0 > 2
 #if N0 > 3
     res3                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3);
 #endif // N0 > 3
 #if N0 > 4
     res4                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4);
     res5                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5);
     res6                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6);
     res7                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7);
 #endif // N0 > 4
 #if N0 > 8
     res8                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8);
     res9                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9);
     resA                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA);
     resB                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB);
     resC                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC);
     resD                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD);
     resE                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE);
     resF                      = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF);
 #endif // N0 > 8

 #elif K0 == 4 // K0 == 4
     // This part computes the following transpositions:
     // 4x2 -> 2x4
     // 4x4 -> 4x4
     // 4x8 -> 8x4
     // 4x16 -> 16x4
     res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0);
     res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1);
 #if N0 > 2
     res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2);
 #endif // N0 > 2
 #if N0 > 3
     res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3);
 #endif // N0 > 3
 #if N0 > 4
     res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4);
     res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5);
     res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6);
     res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7);
 #endif // N0 > 4
 #if N0 > 8
     res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8);
     res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9);
     resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA);
     resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB);
     resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC);
     resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD);
     resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE);
     resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF);
 #endif // N0 > 8

 #elif K0 == 8 // K0 == 8
     // This part computes the following transpositions:
     // 8x2 -> 2x8
     // 8x4 -> 4x8
     // 8x8 -> 8x8
     // 8x16 -> 16x8
     res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0);
     res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1);
 #if N0 > 2
     res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2);
 #endif // N0 > 2
 #if N0 > 3
     res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3);
 #endif // N0 > 3
 #if N0 > 4
     res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4);
     res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5);
     res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6);
     res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7);
 #endif // N0 > 4
 #if N0 > 8
     res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8);
     res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9);
     resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA);
     resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB);
     resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC);
     resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD);
     resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE);
     resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF);
 #endif // N0 > 8

 #elif K0 == 16 // K0 == 16

     // This part computes the following transpositions:
     // 16x2 -> 2x16
     // 16x4 -> 4x16
     // 16x8 -> 8x16
     // 16x16 -> 16x16
     res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0,
                                           a8.s0, a9.s0, aA.s0, aB.s0, aC.s0, aD.s0, aE.s0, aF.s0);
     res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1,
                                           a8.s1, a9.s1, aA.s1, aB.s1, aC.s1, aD.s1, aE.s1, aF.s1);
 #if N0 > 2
     res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2,
                                           a8.s2, a9.s2, aA.s2, aB.s2, aC.s2, aD.s2, aE.s2, aF.s2);
 #endif // N0 > 2
 #if N0 > 3
     res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3,
                                           a8.s3, a9.s3, aA.s3, aB.s3, aC.s3, aD.s3, aE.s3, aF.s3);
 #endif // N0 > 3
 #if N0 > 4
     res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4,
                                           a8.s4, a9.s4, aA.s4, aB.s4, aC.s4, aD.s4, aE.s4, aF.s4);
     res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5,
                                           a8.s5, a9.s5, aA.s5, aB.s5, aC.s5, aD.s5, aE.s5, aF.s5);
     res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6,
                                           a8.s6, a9.s6, aA.s6, aB.s6, aC.s6, aD.s6, aE.s6, aF.s6);
     res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7,
                                           a8.s7, a9.s7, aA.s7, aB.s7, aC.s7, aD.s7, aE.s7, aF.s7);
 #endif // N0 > 4
 #if N0 > 8
     res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8,
                                           a8.s8, a9.s8, aA.s8, aB.s8, aC.s8, aD.s8, aE.s8, aF.s8);
     res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9,
                                           a8.s9, a9.s9, aA.s9, aB.s9, aC.s9, aD.s9, aE.s9, aF.s9);
     resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA,
                                           a8.sA, a9.sA, aA.sA, aB.sA, aC.sA, aD.sA, aE.sA, aF.sA);
     resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB,
                                           a8.sB, a9.sB, aA.sB, aB.sB, aC.sB, aD.sB, aE.sB, aF.sB);
     resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC,
                                           a8.sC, a9.sC, aA.sC, aB.sC, aC.sC, aD.sC, aE.sC, aF.sC);
     resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD,
                                           a8.sD, a9.sD, aA.sD, aB.sD, aC.sD, aD.sD, aE.sD, aF.sD);
     resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE,
                                           a8.sE, a9.sE, aA.sE, aB.sE, aC.sE, aD.sE, aE.sE, aF.sE);
     resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF,
                                           a8.sF, a9.sF, aA.sF, aB.sF, aC.sF, aD.sF, aE.sF, aF.sF);
 #endif // N0 > 8

 #else // N0 == 16
 #error "Not supported N0 value"
 #endif // N0 > 2

     // ---------------------------Store the output values ------------------------------
     REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0);
     STORE_BLOCK(N0, K0, DATA_TYPE, res, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout);

 #undef BLOCK_SIZE
 #undef OUTPUT_OFFSET_X
 #undef OUTPUT_STEP_X
 }
 #endif // defined(TRANSPOSE)
 #endif // defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT)

 #if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K)

 #define CONCAT(a, b) a##b

 #define ARM_DOT1(a, b, c) \
     ({                    \
         c = fma(a, b, c); \
     })
 #define ARM_DOT2(a, b, c)       \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
     })
 #define ARM_DOT3(a, b, c)           \
     ({                              \
         ARM_DOT2(a, b, c);          \
         c = fma((a.s2), (b.s2), c); \
     })
 #define ARM_DOT4(a, b, c)           \
     ({                              \
         ARM_DOT3(a, b, c);          \
         c = fma((a.s3), (b.s3), c); \
     })
 #define ARM_DOT8(a, b, c)            \
     ({                               \
         ARM_DOT4((a.lo), (b.lo), c); \
         ARM_DOT4((a.hi), (b.hi), c); \
     })
 #define ARM_DOT16(a, b, c)           \
     ({                               \
         ARM_DOT8((a.lo), (b.lo), c); \
         ARM_DOT8((a.hi), (b.hi), c); \
     })

 #if N0 == 2
 #define ARM_DOT_K0XN0(k0, a, b, c) \
     ({                             \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##0), (c.s0));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##1), (c.s1));     \
     })
 #elif N0 == 3 // N0 == 3
 #define ARM_DOT_K0XN0(k0, a, b, c) \
     ({                             \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##0), (c.s0));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##1), (c.s1));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##2), (c.s2));     \
     })
 #elif N0 == 4 // N0 == 4
 #define ARM_DOT_K0XN0(k0, a, b, c) \
     ({                             \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##0), (c.s0));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##1), (c.s1));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##2), (c.s2));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##3), (c.s3));     \
     })
 #elif N0 == 8 // N0 == 8
 #define ARM_DOT_K0XN0(k0, a, b, c) \
     ({                             \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##0), (c.s0));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##1), (c.s1));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##2), (c.s2));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##3), (c.s3));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##4), (c.s4));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##5), (c.s5));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##6), (c.s6));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##7), (c.s7));     \
     })
 #elif N0 == 16 // N0 == 16
 #define ARM_DOT_K0XN0(k0, a, b, c) \
     ({                             \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##0), (c.s0));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##1), (c.s1));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##2), (c.s2));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##3), (c.s3));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##4), (c.s4));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##5), (c.s5));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##6), (c.s6));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##7), (c.s7));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##8), (c.s8));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##9), (c.s9));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##A), (c.sA));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##B), (c.sB));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##C), (c.sC));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##D), (c.sD));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##E), (c.sE));     \
         CONCAT(ARM_DOT, k0)        \
         ((a), (b##F), (c.sF));     \
     })
 #else // N0 not supported
 #error "N0 value not supported"
 #endif // N0 conditions

 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix is NOT reshaped
  *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
  *
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90)
  * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64)
  * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
  * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS matrix. Supported data type: F16/F32
  * @param[in]  lhs_stride_x                       Stride of the LHS matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS matrix
  * @param[in]  rhs_ptr                            Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  rhs_stride_x                       Stride of the RHS reshaped matrix in X dimension (in bytes)
  * @param[in]  rhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  rhs_stride_y                       Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes  The offset of the first element in the RHS reshaped matrix
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  lhs_stride_z                       Stride of the LHS matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  lhs_cross_plane_pad                (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
                                           IMAGE_DECLARATION(rhs),
 #if defined(BETA)
                                           IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                           IMAGE_DECLARATION(dst),
                                           uint lhs_stride_z,
                                           uint rhs_stride_z,
 #if defined(BETA)
                                           uint bias_stride_z,
 #endif //defined(BETA)
                                           uint dst_stride_z
 #if defined(REINTERPRET_INPUT_AS_3D)
                                           ,
                                           uint lhs_cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                           ,
                                           uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                          )
 {
     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (K0)
 #define RHS_STEP_X ((K0) * (H0))
 #define RHS_STEP_LOOP (1)
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (K0)
 #define RHS_STEP_LOOP (H0)
 #endif // defined(RHS_INTERLEAVE)

     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
 #else  // defined(MATRIX_B_DEPTH)
     rhs_offset += z * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)

     REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
     lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     lhs_offset += z * lhs_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(M0-1)=0;

     int i = 0;
     for(; i <= (K - K0); i += K0)
     {
         // Supported cases (M0, K0):
         // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
         // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
         // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
         // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
         // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
         // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
         // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
         // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);

         // Load values from RHS reshaped matrix
         LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);

         // Accumulate
         ARM_DOT_K0XN0(K0, a0, b, c0);
 #if M0 > 1
         ARM_DOT_K0XN0(K0, a1, b, c1);
 #endif // M0 > 1
 #if M0 > 2
         ARM_DOT_K0XN0(K0, a2, b, c2);
 #endif // M0 > 2
 #if M0 > 3
         ARM_DOT_K0XN0(K0, a3, b, c3);
 #endif // M0 > 3
 #if M0 > 4
         ARM_DOT_K0XN0(K0, a4, b, c4);
 #endif // M0 > 4
 #if M0 > 5
         ARM_DOT_K0XN0(K0, a5, b, c5);
 #endif // M0 > 5
 #if M0 > 6
         ARM_DOT_K0XN0(K0, a6, b, c6);
 #endif // M0 > 6
 #if M0 > 7
         ARM_DOT_K0XN0(K0, a7, b, c7);
 #endif // M0 > 7

         lhs_offset += K0 * sizeof(DATA_TYPE);
         rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
     }

     // Left-over accumulations
     for(; i < K; ++i)
     {
         // Load values from LHS matrix
         LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);

         // Load values from RHS reshaped matrix
         LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);

         // Accumulate
         ARM_DOT_K0XN0(1, a0, b, c0);
 #if M0 > 1
         ARM_DOT_K0XN0(1, a1, b, c1);
 #endif // M0 > 1
 #if M0 > 2
         ARM_DOT_K0XN0(1, a2, b, c2);
 #endif // M0 > 2
 #if M0 > 3
         ARM_DOT_K0XN0(1, a3, b, c3);
 #endif // M0 > 3
 #if M0 > 4
         ARM_DOT_K0XN0(1, a4, b, c4);
 #endif // M0 > 4
 #if M0 > 5
         ARM_DOT_K0XN0(1, a5, b, c5);
 #endif // M0 > 5
 #if M0 > 6
         ARM_DOT_K0XN0(1, a6, b, c6);
 #endif // M0 > 6
 #if M0 > 7
         ARM_DOT_K0XN0(1, a7, b, c7);
 #endif // M0 > 7

         lhs_offset += sizeof(DATA_TYPE);
         rhs_offset += sizeof(DATA_TYPE);
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
     ADD_BLOCK_BROADCAST(M0, c, bias0);

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
     ADD_BLOCK(M0, c, bias);

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = y == 0;
     const bool cond_x = ((x + 1) * N0 >= N);

     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);

 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }

 #if defined(OPENCL_IMAGE_SUPPORT)
 /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image
  *  The LHS matrix is NOT reshaped
  *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
  *
  * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90)
  * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
  *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
  *       could be different from the value returned by get_image_height(rhs_img).
  * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
  * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 4, 8, 16
  *  - K0 = 4, 8, 16
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS matrix. Supported data type: F32
  * @param[in]  lhs_stride_x                       Stride of the LHS matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS matrix
  * @param[in]  rhs_img                            The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  lhs_stride_z                       Stride of the LHS matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  lhs_cross_plane_pad                (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
                                                   __read_only image2d_t rhs_img,
 #if defined(BETA)
                                                   IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                                   IMAGE_DECLARATION(dst),
                                                   uint lhs_stride_z,
                                                   uint rhs_stride_z,
 #if defined(BETA)
                                                   uint bias_stride_z,
 #endif //defined(BETA)
                                                   uint dst_stride_z
 #if defined(REINTERPRET_INPUT_AS_3D)
                                                   ,
                                                   uint lhs_cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                                   ,
                                                   uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                                  )
 {
     // Pixel unit
 #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)

 #define LEFTOVER_K (K % K0)

     // Block size
 #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (PIXEL_UNIT)
 #define RHS_STEP_X (PIXEL_UNIT * (H0))
 #define RHS_STEP_LOOP (1)
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X PIXEL_UNIT
 #define RHS_STEP_LOOP (H0)
 #endif // defined(RHS_INTERLEAVE)

     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
 #else  // defined(MATRIX_B_DEPTH)
     const uint z_rhs = get_global_id(2);
 #endif // defined(MATRIX_B_DEPTH)

     // Compute RHS matrix coordinates
     uint       x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
     const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
     lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     lhs_offset += z * lhs_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0);

     int i = 0;
     for(; i <= (K - K0); i += K0)
     {
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);

         // Load values from RHS matrix stored in a cl_image
         REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
         LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);

         // Accumulate
         ARM_DOT_K0XN0(K0, a0, b, c0);
 #if M0 > 1
         ARM_DOT_K0XN0(K0, a1, b, c1);
 #endif // M0 > 1
 #if M0 > 2
         ARM_DOT_K0XN0(K0, a2, b, c2);
 #endif // M0 > 2
 #if M0 > 3
         ARM_DOT_K0XN0(K0, a3, b, c3);
 #endif // M0 > 3
 #if M0 > 4
         ARM_DOT_K0XN0(K0, a4, b, c4);
 #endif // M0 > 4
 #if M0 > 5
         ARM_DOT_K0XN0(K0, a5, b, c5);
 #endif // M0 > 5
 #if M0 > 6
         ARM_DOT_K0XN0(K0, a6, b, c6);
 #endif // M0 > 6
 #if M0 > 7
         ARM_DOT_K0XN0(K0, a7, b, c7);
 #endif // M0 > 7

         lhs_offset += K0 * sizeof(DATA_TYPE);
         x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
     }

 #if LEFTOVER_K != 0
     // Note: We cannot read out-of-bound elements from the RHS matrix because
     // the RHS width is always multiple of K0. This is not be true for the LHS matrix

     union UNION_VEC_TYPE
     {
         DATA_TYPE s[K0];
         VEC_DATA_TYPE(DATA_TYPE, K0)
         v;
     };

     union UNION_VEC_TYPE a0 = {.v = 0 };
 #if M0 > 1
     union UNION_VEC_TYPE a1 = {.v = 0 };
 #endif // M0 > 1
 #if M0 > 2
     union UNION_VEC_TYPE a2 = {.v = 0 };
 #endif // M0 > 2
 #if M0 > 3
     union UNION_VEC_TYPE a3 = {.v = 0 };
 #endif // M0 > 3
 #if M0 > 4
     union UNION_VEC_TYPE a4 = {.v = 0 };
 #endif // M0 > 4
 #if M0 > 5
     union UNION_VEC_TYPE a5 = {.v = 0 };
 #endif // M0 > 5
 #if M0 > 6
     union UNION_VEC_TYPE a6 = {.v = 0 };
 #endif // M0 > 6
 #if M0 > 7
     union UNION_VEC_TYPE a7 = {.v = 0 };
 #endif // M0 > 7

     REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);

     // Load from RHS matrix
     LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);

     // Load from LHS matrix
     for(int k = 0; k < LEFTOVER_K; ++k)
     {
         a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0);
 #if M0 > 1
         a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1);
 #endif // M0 > 1
 #if M0 > 2
         a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2);
 #endif // M0 > 2
 #if M0 > 3
         a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3);
 #endif // M0 > 3
 #if M0 > 4
         a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4);
 #endif // M0 > 4
 #if M0 > 5
         a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5);
 #endif // M0 > 5
 #if M0 > 6
         a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6);
 #endif // M0 > 6
 #if M0 > 7
         a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7);
 #endif // M0 > 7

         lhs_offset += sizeof(DATA_TYPE);
     }

     // Accumulate
     ARM_DOT_K0XN0(K0, a0.v, b, c0);
 #if M0 > 1
     ARM_DOT_K0XN0(K0, a1.v, b, c1);
 #endif // M0 > 1
 #if M0 > 2
     ARM_DOT_K0XN0(K0, a2.v, b, c2);
 #endif // M0 > 2
 #if M0 > 3
     ARM_DOT_K0XN0(K0, a3.v, b, c3);
 #endif // M0 > 3
 #if M0 > 4
     ARM_DOT_K0XN0(K0, a4.v, b, c4);
 #endif // M0 > 4
 #if M0 > 5
     ARM_DOT_K0XN0(K0, a5.v, b, c5);
 #endif // M0 > 5
 #if M0 > 6
     ARM_DOT_K0XN0(K0, a6.v, b, c6);
 #endif // M0 > 6
 #if M0 > 7
     ARM_DOT_K0XN0(K0, a7.v, b, c7);
 #endif // M0 > 7

 #endif // LEFTOVER_K != 0

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
     ADD_BLOCK_BROADCAST(M0, c, bias0);

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
     ADD_BLOCK(M0, c, bias);

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = y == 0;
     const bool cond_x = ((x + 1) * N0 >= N);

     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);

 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 #undef LEFTOVER_K
 #undef PIXEL_UNIT
 }
 #endif // defined(OPENCL_IMAGE_SUPPORT)

 #define VFMA(a, b, c)     \
     ({                    \
         c = fma(a, b, c); \
     })

 #if M0 == 1
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
     })
 #elif M0 == 2 // M0 == 2
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
     })
 #elif M0 == 3 // M0 == 3
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
     })
 #elif M0 == 4 // M0 == 4
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
     })
 #elif M0 == 5 // M0 == 5
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
     })
 #elif M0 == 6 // M0 == 6
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
     })
 #elif M0 == 7 // M0 == 7
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
     })
 #elif M0 == 8 // M0 == 8
 #define VFMA_M0xN0(i, a, b, c)                                        \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \
     })
 #else // M0 not supported
 #error "M0 not supported"
 #endif // M0 not supported

 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix is NOT reshaped
  *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
  *
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90).
  * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
  * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS matrix. Supported data type: F16/F32
  * @param[in]  lhs_stride_x                       Stride of the LHS matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS matrix
  * @param[in]  rhs_ptr                            Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  rhs_stride_x                       Stride of the RHS reshaped matrix in X dimension (in bytes)
  * @param[in]  rhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  rhs_stride_y                       Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes  The offset of the first element in the RHS reshaped matrix
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  lhs_stride_z                       Stride of the LHS matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  lhs_cross_plane_pad                (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
                                            IMAGE_DECLARATION(rhs),
 #if defined(BETA)
                                            IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                            IMAGE_DECLARATION(dst),
                                            uint lhs_stride_z,
                                            uint rhs_stride_z,
 #if defined(BETA)
                                            uint bias_stride_z,
 #endif //defined(BETA)
                                            uint dst_stride_z
 #if defined(REINTERPRET_INPUT_AS_3D)
                                            ,
                                            uint lhs_cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                            ,
                                            uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                           )
 {
     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (N0)
 #define RHS_STEP_X ((N0) * (H0))
 #define RHS_STEP_LOOP (1)
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (N0)
 #define RHS_STEP_LOOP (H0)
 #endif // defined(RHS_INTERLEAVE)

     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

     // Compute RHS reshaped matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
 #else  // defined(MATRIX_B_DEPTH)
     rhs_offset += z * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)

     REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0);   //uint zin0=0,zin1=0,zin2=0,... zin7=0;
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0;

 #if defined(REINTERPRET_INPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
     lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     lhs_offset += z * lhs_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(N0-1)=0;

     int i = 0;
     for(; i <= (K - K0); i += K0)
     {
         // Supported cases (M0, K0):
         // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
         // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
         // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
         // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
         // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
         // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
         // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
         // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);

         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;

         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(0, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(1, a, b0, c);
 #if K0 > 2
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(2, a, b0, c);
 #endif // K0 > 2
 #if K0 > 3
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(3, a, b0, c);
 #endif // K0 > 3
 #if K0 > 4
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(4, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(5, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(6, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(7, a, b0, c);
 #endif // K0 > 4
 #if K0 > 8
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(8, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(9, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(A, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(B, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(C, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(D, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(E, a, b0, c);
         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(F, a, b0, c);
 #endif // K0 > 8

         lhs_offset += K0 * sizeof(DATA_TYPE);
         rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE);
     }

     // Left-over accumulations
     for(; i < K; ++i)
     {
         // Load values from LHS matrix
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
 #if M0 > 1
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
 #endif // M0 > 1
 #if M0 > 2
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
 #endif // M0 > 2
 #if M0 > 3
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
 #endif // M0 > 3
 #if M0 > 4
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
 #endif // M0 > 4
 #if M0 > 5
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
 #endif // M0 > 5
 #if M0 > 6
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
 #endif // M0 > 6
 #if M0 > 7
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
 #endif // M0 > 7

         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;

         b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
         VFMA_M0xN0(0, a, b0, c);

         lhs_offset += sizeof(DATA_TYPE);
         rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;

 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
     ADD_BLOCK_BROADCAST(M0, c, bias0);

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
     ADD_BLOCK(M0, c, bias);

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = y == 0;
     const bool cond_x = ((x + 1) * N0 >= N);

     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);

 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }

 #if defined(OPENCL_IMAGE_SUPPORT)
 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix is NOT reshaped
  *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
  *
  * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90).
  * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
  *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
  *       could be different from the value returned by get_image_height(rhs_img).
  * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
  * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 4, 8, 16
  *  - K0 = 4, 8, 16
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS matrix. Supported data type: F32
  * @param[in]  lhs_stride_x                       Stride of the LHS matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS matrix
  * @param[in]  rhs_img                            The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  lhs_stride_z                       Stride of the LHS matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  lhs_cross_plane_pad                (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
                                                    __read_only image2d_t rhs_img,
 #if defined(BETA)
                                                    IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                                    IMAGE_DECLARATION(dst),
                                                    uint lhs_stride_z,
                                                    uint rhs_stride_z,
 #if defined(BETA)
                                                    uint bias_stride_z,
 #endif //defined(BETA)
                                                    uint dst_stride_z
 #if defined(REINTERPRET_INPUT_AS_3D)
                                                    ,
                                                    uint lhs_cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                                    ,
                                                    uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                                   )
 {
     // Pixel unit
 #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)

     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (PIXEL_UNIT)
 #define RHS_STEP_X ((PIXEL_UNIT) * (H0))
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (PIXEL_UNIT)
 #endif // defined(RHS_INTERLEAVE)

     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     const uint z_rhs = (z % MATRIX_B_DEPTH);
 #else  // defined(MATRIX_B_DEPTH)
     const uint z_rhs = z;
 #endif // defined(MATRIX_B_DEPTH)

     // Compute RHS matrix coordinates
     uint       x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
     const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;

     REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0);
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
     lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     lhs_offset += z * lhs_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0);

     int i = 0;
     for(; i <= (K - K0); i += K0)
     {
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);

         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;

         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(0, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(1, a, b0, c);
 #if K0 > 2
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(2, a, b0, c);
 #endif // K0 > 2
 #if K0 > 3
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(3, a, b0, c);
 #endif // K0 > 3
 #if K0 > 4
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(4, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(5, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(6, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(7, a, b0, c);
 #endif // K0 > 4
 #if K0 > 8
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(8, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(9, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(A, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(B, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(C, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(D, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(E, a, b0, c);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));
         VFMA_M0xN0(F, a, b0, c);
 #endif // K0 > 8

         lhs_offset += K0 * sizeof(DATA_TYPE);
         x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP;
     }

     // Left-over accumulations
     for(; i < K; ++i)
     {
         // Load values from LHS matrix
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
 #if M0 > 1
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
 #endif // M0 > 1
 #if M0 > 2
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
 #endif // M0 > 2
 #if M0 > 3
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
 #endif // M0 > 3
 #if M0 > 4
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
 #endif // M0 > 4
 #if M0 > 5
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
 #endif // M0 > 5
 #if M0 > 6
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
 #endif // M0 > 6
 #if M0 > 7
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
 #endif // M0 > 7

         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));

         VFMA_M0xN0(0, a, b0, c);

         lhs_offset += sizeof(DATA_TYPE);
         x_rhs += RHS_STEP_X;
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;

 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
     ADD_BLOCK_BROADCAST(M0, c, bias0);

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
     ADD_BLOCK(M0, c, bias);

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = y == 0;
     const bool cond_x = ((x + 1) * N0 >= N);

     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);

 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }
 #endif // defined(OPENCL_IMAGE_SUPPORT)
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K)

 #if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N)

 #if defined(MIXED_PRECISION)
 #if K0 == 2
 #define ARM_DOT_K0(a, b, c) \
     ({                      \
         c += a.s0 * b.s0;   \
         c += a.s1 * b.s1;   \
     })
 #elif K0 == 3 // K0 == 3
 #define ARM_DOT_K0(a, b, c) \
     ({                      \
         c += a.s0 * b.s0;   \
         c += a.s1 * b.s1;   \
         c += a.s2 * b.s2;   \
     })
 #elif K0 == 4 // K0 == 4
 #define ARM_DOT_K0(a, b, c) \
     ({                      \
         c += a.s0 * b.s0;   \
         c += a.s1 * b.s1;   \
         c += a.s2 * b.s2;   \
         c += a.s3 * b.s3;   \
     })
 #elif K0 == 8 // K0 == 8
 #define ARM_DOT_K0(a, b, c) \
     ({                      \
         c += a.s0 * b.s0;   \
         c += a.s1 * b.s1;   \
         c += a.s2 * b.s2;   \
         c += a.s3 * b.s3;   \
         c += a.s4 * b.s4;   \
         c += a.s5 * b.s5;   \
         c += a.s6 * b.s6;   \
         c += a.s7 * b.s7;   \
     })
 #elif K0 == 16 // K0 == 16
 #define ARM_DOT_K0(a, b, c) \
     ({                      \
         c += a.s0 * b.s0;   \
         c += a.s1 * b.s1;   \
         c += a.s2 * b.s2;   \
         c += a.s3 * b.s3;   \
         c += a.s4 * b.s4;   \
         c += a.s5 * b.s5;   \
         c += a.s6 * b.s6;   \
         c += a.s7 * b.s7;   \
         c += a.s8 * b.s8;   \
         c += a.s9 * b.s9;   \
         c += a.sA * b.sA;   \
         c += a.sB * b.sB;   \
         c += a.sC * b.sC;   \
         c += a.sD * b.sD;   \
         c += a.sE * b.sE;   \
         c += a.sF * b.sF;   \
     })
 #else // K0 not supported
 #error "K0 value not supported"
 #endif // K0 conditions
 #else  // defined(MIXED_PRECISION)
 #if K0 == 2
 #define ARM_DOT_K0(a, b, c)     \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
     })
 #elif K0 == 3 // K0 == 3
 #define ARM_DOT_K0(a, b, c)     \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
         c = fma(a.s2, b.s2, c); \
     })
 #elif K0 == 4 // K0 == 4
 #define ARM_DOT_K0(a, b, c)     \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
         c = fma(a.s2, b.s2, c); \
         c = fma(a.s3, b.s3, c); \
     })
 #elif K0 == 8 // K0 == 8
 #define ARM_DOT_K0(a, b, c)     \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
         c = fma(a.s2, b.s2, c); \
         c = fma(a.s3, b.s3, c); \
         c = fma(a.s4, b.s4, c); \
         c = fma(a.s5, b.s5, c); \
         c = fma(a.s6, b.s6, c); \
         c = fma(a.s7, b.s7, c); \
     })
 #elif K0 == 16 // K0 == 16
 #define ARM_DOT_K0(a, b, c)     \
     ({                          \
         c = fma(a.s0, b.s0, c); \
         c = fma(a.s1, b.s1, c); \
         c = fma(a.s2, b.s2, c); \
         c = fma(a.s3, b.s3, c); \
         c = fma(a.s4, b.s4, c); \
         c = fma(a.s5, b.s5, c); \
         c = fma(a.s6, b.s6, c); \
         c = fma(a.s7, b.s7, c); \
         c = fma(a.s8, b.s8, c); \
         c = fma(a.s9, b.s9, c); \
         c = fma(a.sA, b.sA, c); \
         c = fma(a.sB, b.sB, c); \
         c = fma(a.sC, b.sC, c); \
         c = fma(a.sD, b.sD, c); \
         c = fma(a.sE, b.sE, c); \
         c = fma(a.sF, b.sF, c); \
     })
 #else // K0 not supported
 #error "K0 value not supported"
 #endif // K0 conditions
 #endif // defined(MIXED_PRECISION)

 #if N0 == 2
 #define ARM_DOT_K0XN0(a, b, c)           \
     ({                                   \
         ARM_DOT_K0((a), (b##0), (c.s0)); \
         ARM_DOT_K0((a), (b##1), (c.s1)); \
     })
 #elif N0 == 3 // N0 == 3
 #define ARM_DOT_K0XN0(a, b, c)           \
     ({                                   \
         ARM_DOT_K0((a), (b##0), (c.s0)); \
         ARM_DOT_K0((a), (b##1), (c.s1)); \
         ARM_DOT_K0((a), (b##2), (c.s2)); \
     })
 #elif N0 == 4 // N0 == 4
 #define ARM_DOT_K0XN0(a, b, c)           \
     ({                                   \
         ARM_DOT_K0((a), (b##0), (c.s0)); \
         ARM_DOT_K0((a), (b##1), (c.s1)); \
         ARM_DOT_K0((a), (b##2), (c.s2)); \
         ARM_DOT_K0((a), (b##3), (c.s3)); \
     })
 #elif N0 == 8 // N0 == 8
 #define ARM_DOT_K0XN0(a, b, c)           \
     ({                                   \
         ARM_DOT_K0((a), (b##0), (c.s0)); \
         ARM_DOT_K0((a), (b##1), (c.s1)); \
         ARM_DOT_K0((a), (b##2), (c.s2)); \
         ARM_DOT_K0((a), (b##3), (c.s3)); \
         ARM_DOT_K0((a), (b##4), (c.s4)); \
         ARM_DOT_K0((a), (b##5), (c.s5)); \
         ARM_DOT_K0((a), (b##6), (c.s6)); \
         ARM_DOT_K0((a), (b##7), (c.s7)); \
     })
 #elif N0 == 16 // N0 == 16
 #define ARM_DOT_K0XN0(a, b, c)           \
     ({                                   \
         ARM_DOT_K0((a), (b##0), (c.s0)); \
         ARM_DOT_K0((a), (b##1), (c.s1)); \
         ARM_DOT_K0((a), (b##2), (c.s2)); \
         ARM_DOT_K0((a), (b##3), (c.s3)); \
         ARM_DOT_K0((a), (b##4), (c.s4)); \
         ARM_DOT_K0((a), (b##5), (c.s5)); \
         ARM_DOT_K0((a), (b##6), (c.s6)); \
         ARM_DOT_K0((a), (b##7), (c.s7)); \
         ARM_DOT_K0((a), (b##8), (c.s8)); \
         ARM_DOT_K0((a), (b##9), (c.s9)); \
         ARM_DOT_K0((a), (b##A), (c.sA)); \
         ARM_DOT_K0((a), (b##B), (c.sB)); \
         ARM_DOT_K0((a), (b##C), (c.sC)); \
         ARM_DOT_K0((a), (b##D), (c.sD)); \
         ARM_DOT_K0((a), (b##E), (c.sE)); \
         ARM_DOT_K0((a), (b##F), (c.sF)); \
     })
 #else // N0 not supported
 #error "N0 value not supported"
 #endif // N0 conditions

 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed
  *  The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float)
  * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
  * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
  * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *  - V0 >= 1
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS reshaped matrix. Supported data type: F16/F32
  * @param[in]  lhs_stride_x                       Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS reshaped matrix
  * @param[in]  rhs_ptr                            Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  rhs_stride_x                       Stride of the RHS reshaped matrix in X dimension (in bytes)
  * @param[in]  rhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  rhs_stride_y                       Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes  The offset of the first element in the RHS reshaped matrix
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  k                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
  * @param[in]  lhs_stride_z                       Stride of the LHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
                                             IMAGE_DECLARATION(rhs),
 #if defined(BETA)
                                             IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                             IMAGE_DECLARATION(dst),
                                             uint k,
                                             uint lhs_stride_z,
                                             uint rhs_stride_z,
 #if defined(BETA)
                                             uint bias_stride_z,
 #endif //defined(BETA)
                                             uint dst_stride_z
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                             ,
                                             uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                            )
 {
     // Block size
 #define LHS_BLOCK_SIZE ((K0) * (M0))

 #if defined(LHS_INTERLEAVE)
 #define LHS_OFFSET_X (K0)
 #define LHS_STEP_X ((K0) * (V0))
 #define LHS_STEP_LOOP (1)
 #else // defined(INTERLEAVE)
 #define LHS_OFFSET_X (LHS_BLOCK_SIZE)
 #define LHS_STEP_X (K0)
 #define LHS_STEP_LOOP (V0)
 #endif // defined(INTERLEAVE)

     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (K0)
 #define RHS_STEP_X ((K0) * (H0))
 #define RHS_STEP_LOOP (1)
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (K0)
 #define RHS_STEP_LOOP (H0)
 #endif // defined(RHS_INTERLEAVE)

 #if defined(DUMMY_WORK_ITEMS)
     if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y +
                                (get_global_id(2) * lhs_stride_z);

     // Compute RHS matrix address
     __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (get_global_id(0) % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(0) / (uint)H0) * rhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     rhs_addr += (get_global_id(2) % MATRIX_B_DEPTH) * rhs_stride_z;
 #else  // defined(MATRIX_B_DEPTH)
     rhs_addr += get_global_id(2) * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

     for(int i = 0; i < k; i += K0)
     {
         // Supported cases (M0, K0):
         // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
         // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
         // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
         // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
         // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
         // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
         // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
         // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs);

         // Load values from RHS matrix
         LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X * sizeof(DATA_TYPE), zero);

         // Accumulate
         ARM_DOT_K0XN0(a0, b, c0);
 #if M0 > 1
         ARM_DOT_K0XN0(a1, b, c1);
 #endif // M0 > 1
 #if M0 > 2
         ARM_DOT_K0XN0(a2, b, c2);
 #endif // M0 > 2
 #if M0 > 3
         ARM_DOT_K0XN0(a3, b, c3);
 #endif // M0 > 3
 #if M0 > 4
         ARM_DOT_K0XN0(a4, b, c4);
 #endif // M0 > 4
 #if M0 > 5
         ARM_DOT_K0XN0(a5, b, c5);
 #endif // M0 > 5
 #if M0 > 6
         ARM_DOT_K0XN0(a6, b, c6);
 #endif // M0 > 6
 #if M0 > 7
         ARM_DOT_K0XN0(a7, b, c7);
 #endif // M0 > 7

         lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE);
         rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1) * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += get_global_id(2) * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 #endif // defined(MIXED_PRECISION)

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
                                     2) * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK(M0, c, bias_hp);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK(M0, c, bias);
 #endif // defined(MIXED_PRECISION)

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
 #if defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, VEC_SIZE, c, A_VAL, B_VAL);
 #else  // defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(MIXED_PRECISION)
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);

     // Store output block
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #else  // defined(MIXED_PRECISION)
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #endif // defined(MIXED_PRECISION)

 #undef LHS_BLOCK_SIZE
 #undef LHS_OFFSET_X
 #undef LHS_STEP_X
 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 #undef LHS_STEP_LOOP
 #undef RHS_STEP_LOOP
 }

 #if defined(OPENCL_IMAGE_SUPPORT)
 /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object.
  *  The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed
  *  The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed
  *
  * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
  * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float)
  * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
  * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
  *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
  *       could be different from the value returned by get_image_height(rhs_img).
  * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
  * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 4, 8, 16
  *  - K0 = 4, 8, 16
  *  - V0 >= 1
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS reshaped matrix. Supported data type: F32
  * @param[in]  lhs_stride_x                       Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS reshaped matrix
  * @param[in]  rhs_img                            The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  k                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
  * @param[in]  lhs_stride_z                       Stride of the LHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs),
                                                     __read_only image2d_t rhs_img,
 #if defined(BETA)
                                                     IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                                     IMAGE_DECLARATION(dst),
                                                     uint k,
                                                     uint lhs_stride_z,
                                                     uint rhs_stride_z,
 #if defined(BETA)
                                                     uint bias_stride_z,
 #endif //defined(BETA)
                                                     uint dst_stride_z
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                                     ,
                                                     uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                                    )
 {
     // Pixel unit
 #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)

     // Block size
 #define LHS_BLOCK_SIZE ((K0) * (M0))

 #if defined(LHS_INTERLEAVE)
 #define LHS_OFFSET_X (K0)
 #define LHS_STEP_X ((K0) * (V0))
 #define LHS_STEP_LOOP (1)
 #else // defined(INTERLEAVE)
 #define LHS_OFFSET_X (LHS_BLOCK_SIZE)
 #define LHS_STEP_X (K0)
 #define LHS_STEP_LOOP (V0)
 #endif // defined(INTERLEAVE)

     // Block size
 #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (PIXEL_UNIT)
 #define RHS_STEP_X (PIXEL_UNIT * (H0))
 #define RHS_STEP_LOOP (1)
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X PIXEL_UNIT
 #define RHS_STEP_LOOP (H0)
 #endif // defined(RHS_INTERLEAVE)

 #if defined(DUMMY_WORK_ITEMS)
     if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y +
                                (get_global_id(2) * lhs_stride_z);

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
 #else  // defined(MATRIX_B_DEPTH)
     const uint z_rhs = get_global_id(2);
 #endif // defined(MATRIX_B_DEPTH)

     // Compute RHS matrix coordinates
     uint       x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
     const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

     for(int i = 0; i < K; i += K0)
     {
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs);

         // Load values from RHS matrix stored in a cl_image
         REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
         LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);

         // Accumulate
         ARM_DOT_K0XN0(a0, b, c0);
 #if M0 > 1
         ARM_DOT_K0XN0(a1, b, c1);
 #endif // M0 > 1
 #if M0 > 2
         ARM_DOT_K0XN0(a2, b, c2);
 #endif // M0 > 2
 #if M0 > 3
         ARM_DOT_K0XN0(a3, b, c3);
 #endif // M0 > 3
 #if M0 > 4
         ARM_DOT_K0XN0(a4, b, c4);
 #endif // M0 > 4
 #if M0 > 5
         ARM_DOT_K0XN0(a5, b, c5);
 #endif // M0 > 5
 #if M0 > 6
         ARM_DOT_K0XN0(a6, b, c6);
 #endif // M0 > 6
 #if M0 > 7
         ARM_DOT_K0XN0(a7, b, c7);
 #endif // M0 > 7

         lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE);

         x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1) * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += get_global_id(2) * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 #endif // defined(MIXED_PRECISION)

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
                                     2) * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK(M0, c, bias_hp);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK(M0, c, bias);
 #endif // defined(MIXED_PRECISION)

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
 #if defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, VEC_SIZE, c, A_VAL, B_VAL);
 #else  // defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(MIXED_PRECISION)
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);

     // Store output block
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #else  // defined(MIXED_PRECISION)
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #endif // defined(MIXED_PRECISION)

 #undef LHS_BLOCK_SIZE
 #undef LHS_OFFSET_X
 #undef LHS_STEP_X
 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 #undef PIXEL_UNIT
 #undef LHS_STEP_LOOP
 #undef RHS_STEP_LOOP
 }
 #endif // defined(OPENCL_IMAGE_SUPPORT)

 #if defined(LHS_TRANSPOSE)

 #define VTYPE(TYPE, SIZE) VEC_DATA_TYPE(TYPE, SIZE)

 #if defined(MIXED_PRECISION)

 #if(GPU_ARCH == GPU_ARCH_MIDGARD)
 #define ARM_VFMA(N0, a, b, c) c += (CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))) * (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0)));
 #else // GPU_ARCH == GPU_ARCH_MIDGARD
 #define ARM_VFMA(N0, a, b, c) c = fma((CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (c));
 #endif // GPU_ARCH == GPU_ARCH_MIDGARD

 #else // defined(MIXED_PRECISION

 #if(GPU_ARCH == GPU_ARCH_MIDGARD)
 #define ARM_VFMA(N0, a, b, c) c += (a) * (b);
 #else // GPU_ARCH == GPU_ARCH_MIDGARD
 #define ARM_VFMA(N0, a, b, c) c = fma((a), (b), (c));
 #endif // GPU_ARCH == GPU_ARCH_MIDGARD

 #endif // defined(MIXED_PRECISION)

 #define ARM_VVM_T_NT_1xN0x1(N0, TYPE, a, b, C)         \
     ({                                                 \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a), b, (C##0)); \
     })
 #define ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C)            \
     ({                                                    \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s0), b, (C##0)); \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s1), b, (C##1)); \
     })
 #define ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C)            \
     ({                                                    \
         ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C);           \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s2), b, (C##2)); \
     })
 #define ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C)            \
     ({                                                    \
         ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C);           \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s3), b, (C##3)); \
     })
 #define ARM_VVM_T_NT_8xN0x1(N0, TYPE, a, b, C)            \
     ({                                                    \
         ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C);           \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s4), b, (C##4)); \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s5), b, (C##5)); \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s6), b, (C##6)); \
         ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s7), b, (C##7)); \
     })

 // Factory macro for the column-vector (transposed) by row-vector (not transposed) multiplication. K0 = 1
 // a is the column-vector (transposed)
 // b is the row-vector (not transposed)
 // C is the output matrix
 // Lower case is a vector (a, b)
 // Upper case is a matrix (C)
 #define ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, a, b, C) ARM_VVM_T_NT_##M0##xN0x1(N0, TYPE, a, b, C)

 #define ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C)             \
     ({                                                         \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##0), (B##0), C); \
     })
 #define ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C)             \
     ({                                                         \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C);            \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##1), (B##1), C); \
     })
 #define ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C)             \
     ({                                                         \
         ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C);            \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##2), (B##2), C); \
     })
 #define ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C)             \
     ({                                                         \
         ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C);            \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##3), (B##3), C); \
     })
 #define ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C)             \
     ({                                                         \
         ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C);            \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##4), (B##4), C); \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##5), (B##5), C); \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##6), (B##6), C); \
         ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##7), (B##7), C); \
     })
 #define ARM_MM_T_NT_M0xN0x16(M0, N0, TYPE, A, B, C)           \
     ({                                                        \
         ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C);           \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##8), (B##8), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##9), (B##9), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##A), (B##A), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##B), (B##B), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##C), (B##C), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##D), (B##D), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##E), (B##E), C); \
         ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##F), (B##F), C); \
     })

 // Factory macro for the matrix (transposed) by matrix (not transposed) multiplication.
 // The dimensions for this matrix multiplications are defined through M0, N0 and K0
 // The dimensions supported are:
 // M0: 1, 2, 3, 4, 8
 // N0: 1, 2, 3, 4, 8, 16
 // K0: 1, 2, 3, 4, 8, 16
 // This macro calls the vector-by-matrix macro K0 times
 // A, B and C are matrices
 #define ARM_MM_T_NT(M0, N0, K0, TYPE, A, B, C) \
     CONCAT(ARM_MM_T_NT_M0xN0x, K0)             \
     (M0, N0, TYPE, A, B, C)

 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed
  *  The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed
  *
  * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE).
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
  * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
  * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 2, 3, 4, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *  - V0 >= 1
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS reshaped matrix. Supported data type: F16/F32
  * @param[in]  lhs_stride_x                       Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS reshaped matrix
  * @param[in]  rhs_ptr                            Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  rhs_stride_x                       Stride of the RHS reshaped matrix in X dimension (in bytes)
  * @param[in]  rhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  rhs_stride_y                       Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes  The offset of the first element in the RHS reshaped matrix
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  k                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
  * @param[in]  lhs_stride_z                       Stride of the LHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs),
                                             IMAGE_DECLARATION(rhs),
 #if defined(BETA)
                                             IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                             IMAGE_DECLARATION(dst),
                                             uint k,
                                             uint lhs_stride_z,
                                             uint rhs_stride_z,
 #if defined(BETA)
                                             uint bias_stride_z,
 #endif //defined(BETA)
                                             uint dst_stride_z
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                             ,
                                             uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                            )
 {
     // Block size
 #define LHS_BLOCK_SIZE ((K0) * (M0))

 #if defined(LHS_INTERLEAVE)
 #define LHS_OFFSET_X (M0)
 #define LHS_STEP_X ((M0) * (V0))
 #define LHS_STEP_LOOP (1)
 #else // defined(INTERLEAVE)
 #define LHS_OFFSET_X (LHS_BLOCK_SIZE)
 #define LHS_STEP_X (M0)
 #define LHS_STEP_LOOP (V0)
 #endif // defined(INTERLEAVE)

     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (N0))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (N0)
 #define RHS_STEP_X ((N0) * (H0))
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (N0)
 #endif // defined(RHS_INTERLEAVE)

     const uint x = get_global_id(0);
     const uint y = get_global_id(1);
     const uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z);

     // Compute RHS matrix address
     __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     rhs_addr += (z % MATRIX_B_DEPTH) * rhs_stride_z;
 #else  // defined(MATRIX_B_DEPTH)
     rhs_addr += z * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);

     __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr);
     __global DATA_TYPE *rhs = (__global DATA_TYPE *)(rhs_addr);

     for(int i = 0; i < k; i += K0)
     {
         VEC_DATA_TYPE(DATA_TYPE, M0)
         a0;
         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

 #if K0 > 1
         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;
 #endif // K0 > 1

 #if K0 > 2
         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;
 #endif // K0 > 2

 #if K0 > 3
         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;
 #endif // K0 > 3

 #if K0 > 4
         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;
 #endif // K0 > 4

 #if K0 > 8
         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = VLOAD(N0)(0, rhs);

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
         rhs += RHS_STEP_X;
 #endif // K0 > 8

 #ifndef LHS_INTERLEAVE
         lhs += (M0 * K0 * (V0 - 1));
 #endif // LHS_INTERLEAVE

 #ifndef RHS_INTERLEAVE
         rhs += (N0 * K0 * (H0 - 1));
 #endif // RHS_INTERLEAVE
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, y * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 #endif // defined(MIXED_PRECISION)

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
                                     2) * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK(M0, c, bias_hp);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK(M0, c, bias);
 #endif // defined(MIXED_PRECISION)

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
 #if defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, VEC_SIZE, c, A_VAL, B_VAL);
 #else  // defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(MIXED_PRECISION)
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);

     // Store output block
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #else  // defined(MIXED_PRECISION)
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #endif // defined(MIXED_PRECISION)

 #undef LHS_BLOCK_SIZE
 #undef LHS_OFFSET_X
 #undef LHS_STEP_X
 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }

 #if defined(OPENCL_IMAGE_SUPPORT)
 /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object.
  *  The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed
  *  The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed
  *
  * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
  * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE).
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
  * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
  *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
  *       could be different from the value returned by get_image_height(rhs_img).
  * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
  * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
  * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
  * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
  * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 2, 3, 4, 8
  *  - N0 = 4, 8, 16
  *  - K0 = 4, 8, 16
  *  - V0 >= 1
  *  - H0 >= 1
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS reshaped matrix. Supported data type: F32
  * @param[in]  lhs_stride_x                       Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS reshaped matrix
  * @param[in]  rhs_img                            The RHS reshaped matrix as cl_image 2d. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  k                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
  * @param[in]  lhs_stride_z                       Stride of the LHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs),
                                                     __read_only image2d_t rhs_img,
 #if defined(BETA)
                                                     IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                                                     IMAGE_DECLARATION(dst),
                                                     uint k,
                                                     uint lhs_stride_z,
                                                     uint rhs_stride_z,
 #if defined(BETA)
                                                     uint bias_stride_z,
 #endif //defined(BETA)
                                                     uint dst_stride_z
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                                                     ,
                                                     uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                                                    )
 {
     // Pixel unit
 #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)

     // Block size
 #define LHS_BLOCK_SIZE ((K0) * (M0))

 #if defined(LHS_INTERLEAVE)
 #define LHS_OFFSET_X (M0)
 #define LHS_STEP_X ((M0) * (V0))
 #define LHS_STEP_LOOP (1)
 #else // defined(INTERLEAVE)
 #define LHS_OFFSET_X (LHS_BLOCK_SIZE)
 #define LHS_STEP_X (M0)
 #define LHS_STEP_LOOP (V0)
 #endif // defined(INTERLEAVE)

     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))

     // RHS offset and step X
 #if defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (PIXEL_UNIT)
 #define RHS_STEP_X ((PIXEL_UNIT) * (H0))
 #else // defined(RHS_INTERLEAVE)
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)
 #define RHS_STEP_X (PIXEL_UNIT)
 #endif // defined(RHS_INTERLEAVE)

     const uint x = get_global_id(0);
     const uint y = get_global_id(1);
     const uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z);

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     const uint z_rhs = (z % MATRIX_B_DEPTH);
 #else  // defined(MATRIX_B_DEPTH)
     const uint z_rhs = z;
 #endif // defined(MATRIX_B_DEPTH)

     // Compute RHS matrix coordinates
     uint       x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
     const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);

     __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr);

     for(int i = 0; i < K; i += K0)
     {
         VEC_DATA_TYPE(DATA_TYPE, M0)
         a0;
         VEC_DATA_TYPE(DATA_TYPE, N0)
         b0;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

 #if K0 > 1
         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
 #endif // K0 > 1

 #if K0 > 2
         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
 #endif // K0 > 2

 #if K0 > 3
         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
 #endif // K0 > 3

 #if K0 > 4
         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
 #endif // K0 > 4

 #if K0 > 8
         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;

         a0 = VLOAD(M0)(0, lhs);
         b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));

         ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);

         lhs += LHS_STEP_X;
 #endif // K0 > 8

 #ifndef LHS_INTERLEAVE
         lhs += (M0 * K0 * (V0 - 1));
 #endif // LHS_INTERLEAVE

         x_rhs += K0 * RHS_STEP_X;
 #ifndef RHS_INTERLEAVE
         x_rhs += (PIXEL_UNIT * K0 * (H0 - 1));
 #endif // RHS_INTERLEAVE
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

 #if defined(REINTERPRET_OUTPUT_AS_3D)

     // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, y * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 #endif // defined(MIXED_PRECISION)

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
     ADD_BLOCK(M0, c, bias_hp);
 #else  // defined(MIXED_PRECISION)
     ADD_BLOCK(M0, c, bias);
 #endif // defined(MIXED_PRECISION)

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
 #if defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, VEC_SIZE, c, A_VAL, B_VAL);
 #else  // defined(MIXED_PRECISION)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(MIXED_PRECISION)
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
     const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);

     // Store output block
 #if defined(MIXED_PRECISION)
     CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #else  // defined(MIXED_PRECISION)
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
 #endif // defined(MIXED_PRECISION)

 #undef LHS_BLOCK_SIZE
 #undef LHS_OFFSET_X
 #undef LHS_STEP_X
 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 #undef PIXEL_UNIT
 #undef LHS_STEP_LOOP
 #undef RHS_STEP_LOOP
 }
 #endif // defined(OPENCL_IMAGE_SUPPORT)

 #endif // defined(LHS_TRANSPOSE)

 #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) && defined(DATA_TYPE)

 #if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE)

 #define VFMA(a, b, c)     \
     ({                    \
         c = fma(a, b, c); \
     })

 #if M0 == 1
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
     })
 #elif M0 == 2 // M0 == 2
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
     })
 #elif M0 == 3 // M0 == 3
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
     })
 #elif M0 == 4 // M0 == 4
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
     })
 #elif M0 == 5 // M0 == 5
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
     })
 #elif M0 == 6 // M0 == 6
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
     })
 #elif M0 == 7 // M0 == 7
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
     })
 #elif M0 == 8 // M0 == 8
 #define RHS_VFMA_M0xN0(i, a, b, c)                                    \
     ({                                                                \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
         VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \
     })
 #else // M0 not supported
 #error "M0 not supported"
 #endif // M0 not supported

 /** This OpenCL kernel computes the matrix multiplication between 2 matrices.
  *  The LHS matrix is NOT reshaped
  *  The RHS matrix is NOT reshaped
  *
  * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
  * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90)
  * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64)
  * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
  * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (e.g., -DK0=2)
  * @note The number of N0 columns to process must be passed at compile time using -DN0 (e.g. -DN0=2)
  * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
  * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
  *  - N0 = 2, 3, 4, 8, 16
  *  - K0 = 2, 3, 4, 8, 16
  *
  * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
  *       The activation function is performed after the bias addition
  * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
  *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
  *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
  *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
  * @param[in]  lhs_ptr                            Pointer to the LHS matrix. Supported data type: F16/F32
  * @param[in]  lhs_stride_x                       Stride of the LHS matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                         lhs_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                       Stride of the LHS matrix in Y dimension (in bytes)
  * @param[in]  lhs_step_y                         lhs_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  lhs_offset_first_element_in_bytes  The offset of the first element in the LHS matrix
  * @param[in]  rhs_ptr                            Pointer to the RHS matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  rhs_stride_x                       Stride of the RHS matrix in X dimension (in bytes)
  * @param[in]  rhs_step_x                         rhs_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  rhs_stride_y                       Stride of the RHS matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                         rhs_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes  The offset of the first element in the RHS matrix
  * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
  * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
  * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
  * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data type: same as @p lhs_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  * @param[in]  lhs_stride_z                       Stride of the LHS matrix in Z dimension (in bytes)
  * @param[in]  rhs_stride_z                       Stride of the RHS matrix in Z dimension (in bytes)
  * @param[in]  bias_stride_z                      (Optional) Stride of the bias matrix in Z dimension (in bytes)
  * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  lhs_cross_plane_pad                (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
  * @param[in]  dst_cross_plane_pad                (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
  */
 __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
                              IMAGE_DECLARATION(rhs),
 #if defined(BETA)
                              IMAGE_DECLARATION(bias),
 #endif // defined(BETA)
                              IMAGE_DECLARATION(dst),
                              uint lhs_stride_z,
                              uint rhs_stride_z,
 #if defined(BETA)
                              uint bias_stride_z,
 #endif //defined(BETA)
                              uint dst_stride_z
 #if defined(REINTERPRET_INPUT_AS_3D)
                              ,
                              uint lhs_cross_plane_pad
 #endif // REINTERPRET_INPUT_AS_3D
 #if defined(REINTERPRET_OUTPUT_AS_3D)
                              ,
                              uint dst_cross_plane_pad
 #endif // REINTERPRET_OUTPUT_AS_3D
                             )
 {
     // Block size
 #define RHS_BLOCK_SIZE ((K0) * (N0))

     // RHS offset and step X
 #define RHS_OFFSET_X (RHS_BLOCK_SIZE)

     uint x = get_global_id(0);
     uint y = get_global_id(1);
     uint z = get_global_id(2);

 #if defined(DUMMY_WORK_ITEMS)
     if((x * N0 >= N) || (y * M0 >= M))
     {
         return;
     }
 #endif // defined(DUMMY_WORK_ITEMS)

     // Compute LHS matrix address
     uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;

     // Compute RHS matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);

 #if defined(MATRIX_B_DEPTH)
     // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
     rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
 #else  // defined(MATRIX_B_DEPTH)
     rhs_offset += z * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
     REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);

 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
     lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_INPUT_AS_3D)

     // Add offset for batched GEMM
     lhs_offset += z * lhs_stride_z;

 #endif // defined(REINTERPRET_INPUT_AS_3D)

     // Initialize the accumulators
     REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(M0-1)=0;

     int i = 0;
     for(; i <= (K - K0); i += K0)
     {
         // Supported cases (M0, K0):
         // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
         // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
         // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
         // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
         // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
         // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
         // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
         // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
         // Load values from LHS matrix
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);

         // Load values from RHS matrix
         LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero);

         RHS_VFMA_M0xN0(0, a, b0, c);
         RHS_VFMA_M0xN0(1, a, b1, c);
 #if K0 > 2
         RHS_VFMA_M0xN0(2, a, b2, c);
 #endif // K0 > 2
 #if K0 > 3
         RHS_VFMA_M0xN0(3, a, b3, c);
 #endif // K0 > 3
 #if K0 > 4
         RHS_VFMA_M0xN0(4, a, b4, c);
         RHS_VFMA_M0xN0(5, a, b5, c);
         RHS_VFMA_M0xN0(6, a, b6, c);
         RHS_VFMA_M0xN0(7, a, b7, c);
 #endif // K0 > 4
 #if K0 > 8
         RHS_VFMA_M0xN0(8, a, b8, c);
         RHS_VFMA_M0xN0(9, a, b9, c);
         RHS_VFMA_M0xN0(A, a, bA, c);
         RHS_VFMA_M0xN0(B, a, bB, c);
         RHS_VFMA_M0xN0(C, a, bC, c);
         RHS_VFMA_M0xN0(D, a, bD, c);
         RHS_VFMA_M0xN0(E, a, bE, c);
         RHS_VFMA_M0xN0(F, a, bF, c);
 #endif // K0 > 8

         lhs_offset += K0 * sizeof(DATA_TYPE);
         rhs_offset += K0 * rhs_stride_y;
     }

     // Left-over accumulations
     for(; i < K; ++i)
     {
         // Load values from LHS matrix
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0));
 #if M0 > 1
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1));
 #endif // M0 > 1
 #if M0 > 2
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2));
 #endif // M0 > 2
 #if M0 > 3
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3));
 #endif // M0 > 3
 #if M0 > 4
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4));
 #endif // M0 > 4
 #if M0 > 5
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5));
 #endif // M0 > 5
 #if M0 > 6
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6));
 #endif // M0 > 6
 #if M0 > 7
         VEC_DATA_TYPE(DATA_TYPE, 2)
         a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7));
 #endif // M0 > 7

         VEC_DATA_TYPE(DATA_TYPE, N0)
         b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y));
         RHS_VFMA_M0xN0(0, a, b, c);

         lhs_offset += sizeof(DATA_TYPE);
         rhs_offset += rhs_stride_y;
     }

     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);

     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);

 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
     CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);

     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
     dst_addr += z * dst_stride_z * DEPTH_GEMM3D;

 #else // defined(REINTERPRET_OUTPUT_AS_3D)

     // Add offset for batched GEMM
     dst_addr += z * dst_stride_z;

 #endif // defined(REINTERPRET_OUTPUT_AS_3D)

     // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)

     // Add beta*bias
 #if defined(BETA)
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));

     LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias[broadcasted]
     ADD_BLOCK_BROADCAST(M0, c, bias0);

 #else // defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;

     LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);

 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
 #endif // UNIT_BIAS

     // c = c + bias
     ADD_BLOCK(M0, c, bias);

 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)

 #if defined(ACTIVATION_TYPE)
     ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
 #endif // defined(ACTIVATION_TYPE)

     const bool cond_y = y == 0;
     const bool cond_x = ((x + 1) * N0 >= N);

     // Store output block
     STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);

 #undef RHS_BLOCK_SIZE
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE)

 #if defined(BETA)
 /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
  *
  * @note The beta's value need to be passed at compile time using -DBETA
  *
  * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
  * @param[in]  src_stride_x                      Stride of the source matrix in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source matrix in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source matrix
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  */
 __kernel void gemm_ma_f32(TENSOR3D_DECLARATION(src),
                           TENSOR3D_DECLARATION(dst))
 {
     // Compute source and destination addresses
     Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
     Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);

     // Load values from A x B
     float4 alpha_ab = vload4(0, (__global float *)dst.ptr);

     // Load values from Matrix C
     float4 c = vload4(0, (__global float *)src.ptr);

     // Computes alpha * axb + beta * c
     float4 out = alpha_ab + (float4)BETA * c;

     // Store final result in axb matrix
     vstore4(out, 0, (__global float *)dst.ptr);
 }

 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
 /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
  *
  * @note The beta's value need to be passed at compile time using -DBETA
  *
  * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F16
  * @param[in]  src_stride_x                      Stride of the source matrix in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source matrix in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source matrix
  * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
  * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
  */
 __kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src),
                           TENSOR3D_DECLARATION(dst))
 {
     // Compute source and destination addresses
     Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
     Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);

     // Load values from A x B
     half8 alpha_ab = vload8(0, (__global half *)dst.ptr);

     // Load values from Matrix C
     half8 c = vload8(0, (__global half *)src.ptr);

     // Computes alpha * axb + beta * c
     half8 out = alpha_ab + (half8)BETA * c;

     // Store final result in axb matrix
     vstore8(out, 0, (__global half *)dst.ptr);
 }
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
 #endif // defined(BETA)

 #if defined(WIDTH_VECTOR_A)
 /** This OpenCL kernel computes the vector by matrix multiplication between each row of A (src0) and matrix B (src1) used for locally connected layer
  *
  * @note The width of A need to be passed at compile time using -DWIDTH_VECTOR_A
  *
  * @note The input A and matrix B must not be reshaped
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
  * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
  * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
  * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
  * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
  * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
  * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
  * @param[in]  src1_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
  * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
  * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
  * @param[in]  dst_step_y                         dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
  */
 __kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0),
                              TENSOR3D_DECLARATION(src1),
                              IMAGE_DECLARATION(dst))
 {
     int idx = get_global_id(0) * 4;
     int idy = get_global_id(1);

     // Compute the address for the vector A and matrix B
     int2 src_addr = ((int2)(src0_offset_first_element_in_bytes + src0_stride_y * idy, src1_offset_first_element_in_bytes + src1_stride_z * idy));
     src_addr.s1 += idx * sizeof(float);

     int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(float));

     float4 acc = 0.0f;

     for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(float)); src_addr += (int2)(2 * sizeof(float), 2 * src1_stride_y))
     {
         float2 a0 = vload2(0, (__global float *)(src0_ptr + src_addr.s0));
         float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
         float4 b1 = vload4(0, (__global float *)(src1_ptr + src_addr.s1 + src1_stride_y));

         acc += b0 * (float4)a0.s0;
         acc += b1 * (float4)a0.s1;
     }

     for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(float), src1_stride_y))
     {
         float  a0 = *((__global float *)(src0_ptr + src_addr.s0));
         float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));

         acc += b0 * (float4)a0;
     }

     // Compute destination address
     Image dst = CONVERT_TO_IMAGE_STRUCT(dst);

     vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0)));
 }
 #endif // defined(WIDTH_VECTOR_A)

 )"