Google Git
Sign in
android / platform / external / ComputeLibrary / 39d8017e5604caf0dff5d3ee802827dd4dd9a7a5 / . / build / android-arm64v8a / src / core / CL / cl_kernels / depthwise_convolution_quantized.clembed
blob: 2ae61369c8ad979eca518c0dc95da43528c55168 [file] [log] [blame]
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) 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.
*/
#ifndef ARM_COMPUTE_HELPERS_ASYMM_H
#define ARM_COMPUTE_HELPERS_ASYMM_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
/** Convert the given vector with round to nearest even rounding mode
*
* @param[in] x The target to be converted
* @param[in] type The target type
*
* @return The converted vector
*/
#define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x)))
#define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type)
/** Quantize a floating-point scalar value to 8-bit asymmetric
*
* @param[in] input Input value to quantize
* @param[in] offset Quantization offset
* @param[in] scale Quantization scale
*
* @return quantized value
*/
inline uchar quantize_qasymm8(float input, float offset, float scale)
{
float out_f32 = input / scale + offset;
uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar);
return res_u8;
}
/** Dequantize a scalar value from 8-bit asymmetric to floating-point
*
* @param[in] input Input value to quantize
* @param[in] offset Quantization offset
* @param[in] scale Quantization scale
*
* @return quantized value
*/
inline float dequantize_qasymm8(uchar input, float offset, float scale)
{
return ((float)input - offset) * scale;
}
/** Dequantize a scalar value from signed 8-bit asymmetric to floating-point
*
* @param[in] input Input value to quantize
* @param[in] offset Quantization offset
* @param[in] scale Quantization scale
*
* @return quantized value
*/
inline float dequantize_qasymm8_signed(char input, float offset, float scale)
{
return ((float)input - offset) * scale;
}
/** Quantize a vector of values from floating-point
*
* @param[in] type Output data type.
* @param[in] size Size of vector.
*
* @return quantized values
*/
#define QUANTIZE_IMPL(type, size) \
inline VEC_DATA_TYPE(type, size) quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \
{ \
VEC_DATA_TYPE(float, size) \
out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \
VEC_DATA_TYPE(type, size) \
res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \
return res; \
}
/** Dequantize a vector of values to floating-point
*
* @param[in] type Input data type.
* @param[in] size Size of vector.
*
* @return dequantized values in floating point
*/
#define DEQUANTIZE_IMPL(type, size) \
inline VEC_DATA_TYPE(float, size) dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \
{ \
return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \
}
/** Correctly-rounded-to-nearest division by a power-of-two.
*
* @param[in] size Size of vector.
*
* @return Correctly-rounded-to-nearest division by a power-of-two.
*/
#define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent) \
{ \
const VEC_DATA_TYPE(int, size) \
zero = (VEC_DATA_TYPE(int, size))0; \
const VEC_DATA_TYPE(int, size) \
one = (VEC_DATA_TYPE(int, size))1; \
VEC_DATA_TYPE(int, size) \
mask = (one << exponent) - one; \
VEC_DATA_TYPE(int, size) \
threshold = (mask >> 1) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))(x < 0)); \
return (x >> exponent) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))((x & mask) > threshold)); \
}
/** Product of two numbers, interpreting them as fixed-point values in the interval [-1, 1),
* rounding to the nearest value, and saturating -1 * -1 to the maximum value.
*
* @param[in] size Size of vector.
*
* @return Product of two fixed-point numbers.
*/
#define ASYMM_MULT_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
{ \
VEC_DATA_TYPE(int, size) \
overflow = a == b && a == INT_MIN; \
VEC_DATA_TYPE(long, size) \
a_64 = convert_long##size(a); \
VEC_DATA_TYPE(long, size) \
b_64 = convert_long##size(b); \
VEC_DATA_TYPE(long, size) \
ab_64 = a_64 * b_64; \
/* Revert COMPMID-907 */ \
VEC_DATA_TYPE(long, size) \
mask1 = 1 << 30; \
VEC_DATA_TYPE(long, size) \
mask2 = 1 - (1 << 30); \
VEC_DATA_TYPE(long, size) \
is_positive_or_zero = ab_64 >= 0; \
VEC_DATA_TYPE(long, size) \
nudge = select(mask2, mask1, (SELECT_VEC_DATA_TYPE(long, size))(is_positive_or_zero)); \
VEC_DATA_TYPE(long, size) \
mask = 1ll << 31; \
VEC_DATA_TYPE(int, size) \
ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask); \
return select(ab_x2_high32, INT_MAX, (SELECT_VEC_DATA_TYPE(int, size))(overflow)); \
}
/** Calculates \f$ exp(x) \f$ for x in [-1/4, 0).
*
* @param[in] size Size of vector.
*
* @return Result in fixed-point format Q0.
*/
#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \
{ \
const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \
const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \
const int k_fractional_bits = 31; \
VEC_DATA_TYPE(int, size) \
x = a + (1 << (k_fractional_bits - 3)); \
VEC_DATA_TYPE(int, size) \
x2 = ASYMM_MULT(x, x, size); \
VEC_DATA_TYPE(int, size) \
x3 = ASYMM_MULT(x2, x, size); \
VEC_DATA_TYPE(int, size) \
x4 = ASYMM_MULT(x2, x2, size); \
VEC_DATA_TYPE(int, size) \
x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \
VEC_DATA_TYPE(int, size) \
x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \
VEC_DATA_TYPE(int, size) \
x4_over_24_plus_x3_over_6_plus_x2_over_2 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size); \
return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \
}
/** Each bit of the result is set to the corresponding bit of either then_val or
* else_val depending on whether the corresponding bit of if_mask is set.
* Equivalent to the VBSL instruction in ARM NEON.
*
* @param[in] size Size of vector.
*
* @returns Result contaning bits from @p then_val or from @p else_val depending on corresponding bit in @p if_mask is set or not.
*/
#define ASYMM_SELECT_USING_MASK_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size(VEC_DATA_TYPE(int, size) if_mask, VEC_DATA_TYPE(int, size) then_val, VEC_DATA_TYPE(int, size) else_val) \
{ \
return (if_mask & then_val) ^ (~if_mask & else_val); \
}
/** For each element of input vector, the corresponding bits of the result item are set
* if the input item is zero.
*
* @param[in] size Size of vector.
*
* @returns Output vector with bits set when corresponding bit in @p a is zero.
*/
#define ASYMM_MASK_IF_ZERO_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \
{ \
const VEC_DATA_TYPE(int, size) all_zeros = 0; \
const VEC_DATA_TYPE(int, size) all_ones = ~0; \
return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a == 0)); \
}
/** For each element of input vector, the corresponding bits of the result item are set
* if the input item is non-zero.
*
* @param[in] size Size of vector.
*
* @returns Output vector with bits set when corresponding bit in @p a is non zero.
*/
#define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \
{ \
const VEC_DATA_TYPE(int, size) all_zeros = 0; \
const VEC_DATA_TYPE(int, size) all_ones = ~0; \
return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a != 0)); \
}
#define EXP_BARREL_SHIFTER_IMPL(size) \
inline VEC_DATA_TYPE(int, size) exp_barrel_shifter##size(VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits, int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder) \
{ \
if(k_integer_bits > exponent) \
{ \
const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \
return ASYMM_SELECT_USING_MASK( \
ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \
ASYMM_MULT(result, fp_multiplier, size), result, size); \
} \
\
return result; \
}
/** Calculates \f$ exp(x) \f$ for x < 0.
*
* @param[in] size Size of vector.
*
* @return Result in fixed-point format Q0.
*/
#define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \
{ \
const int k_fractional_bits = 31 - k_integer_bits; \
VEC_DATA_TYPE(int, size) \
k_one_quarter = 1 << (k_fractional_bits - 2); \
VEC_DATA_TYPE(int, size) \
mask = k_one_quarter - 1; \
VEC_DATA_TYPE(int, size) \
a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \
VEC_DATA_TYPE(int, size) \
a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \
VEC_DATA_TYPE(int, size) \
result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, size); \
VEC_DATA_TYPE(int, size) \
remainder = a_mod_quarter_minus_one_quarter - a; \
\
result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \
result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \
\
if(k_integer_bits > 5) \
{ \
const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \
result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \
} \
\
const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \
}
/** Calculates the product of a integer value by a power of two, with either a positive exponent
* (equivalent to an arithmetic left shift, saturating) or a negative exponent
* (equivalent to an arithmetic right shift, rounding to nearest).
*
* @param[in] size Size of vector.
*
* @return Arithmetic left or right shift.
*/
#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \
{ \
if(exponent < 0) \
{ \
return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \
} \
\
const VEC_DATA_TYPE(int, size) min = INT_MIN; \
const VEC_DATA_TYPE(int, size) max = INT_MAX; \
int threshold = ((1 << (31 - exponent)) - 1); \
VEC_DATA_TYPE(int, size) \
positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \
VEC_DATA_TYPE(int, size) \
negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \
VEC_DATA_TYPE(int, size) \
result = x << exponent; \
result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \
result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \
return result; \
}
/** Calculates (a+b)/2, rounded to the nearest integer.
* Equivalent to VRHADD in the ARM NEON instruction set.
*
* @param[in] size Size of vector.
*
* @return (a+b)/2, rounded to the nearest integer.
*/
#define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \
{ \
VEC_DATA_TYPE(long, size) \
a64 = convert_long##size(a); \
VEC_DATA_TYPE(long, size) \
b64 = convert_long##size(b); \
VEC_DATA_TYPE(long, size) \
sum = a64 + b64; \
const VEC_DATA_TYPE(long, size) one = 1; \
const VEC_DATA_TYPE(long, size) minus_one = -1; \
VEC_DATA_TYPE(long, size) \
sign = select(minus_one, one, (SELECT_VEC_DATA_TYPE(long, size))(sum >= 0)); \
return convert_int##size((sum + sign) / 2); \
}
/** Calculates \f$ 1 / (1 + x) \f$ for x in (0, 1).
*
* @param[in] size Size of vector.
*
* @return Result in fixed-point format Q0.
*/
#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \
{ \
const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \
const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \
VEC_DATA_TYPE(int, size) \
half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \
const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \
const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \
VEC_DATA_TYPE(int, size) \
x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \
for(int i = 0; i < 3; i++) \
{ \
VEC_DATA_TYPE(int, size) \
half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \
VEC_DATA_TYPE(int, size) \
one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \
VEC_DATA_TYPE(int, size) \
tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \
x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \
} \
return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \
}
/** Considering the integer value as fixed-point, change the number of integer bits and update value accordingly.
*
* @param[in] size Size of vector.
*
* @return Rescaled value.
*/
#define ASYMM_RESCALE_IMPL(size) \
inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \
{ \
int exponent = src_integer_bits - dst_integer_bits; \
return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \
}
#define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale)
#define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size)
#define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale)
#define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size)
#define ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent)
#define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size)
#define ASYMM_MULT_STR(a, b, size) asymm_mult##size(a, b)
#define ASYMM_MULT(a, b, size) ASYMM_MULT_STR(a, b, size)
#define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \
ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size)
#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \
ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size)
#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a)
#define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) asymm_select_using_mask##size(if_mask, then_val, else_val)
#define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a)
#define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a)
#define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder, size) exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder)
#define ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits)
#define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size)
#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_STR(a, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(a)
#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(a, size) ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_STR(a, size)
#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) asymm_saturating_rounding_mult_by_pow2##size(x, exponent)
#define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
#define ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
#define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size)
#define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size) \
inline VEC_DATA_TYPE(int, size) multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \
{ \
const int left_shift = shift > 0 ? shift : 0; \
const int right_shift = shift > 0 ? 0 : -shift; \
return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size), right_shift, size); \
}
#define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) multiply_by_quantized_multiplier##size(input, qmul, shift)
QUANTIZE_IMPL(uchar, 1)
QUANTIZE_IMPL(char, 1)
QUANTIZE_IMPL(uint, 1)
QUANTIZE_IMPL(int, 1)
QUANTIZE_IMPL(uchar, 4)
QUANTIZE_IMPL(ushort, 4)
QUANTIZE_IMPL(short, 4)
QUANTIZE_IMPL(uchar, 16)
QUANTIZE_IMPL(char, 16)
QUANTIZE_IMPL(ushort, 16)
QUANTIZE_IMPL(short, 16)
QUANTIZE_IMPL(uint, 16)
QUANTIZE_IMPL(int, 16)
DEQUANTIZE_IMPL(uchar, 1)
DEQUANTIZE_IMPL(char, 1)
DEQUANTIZE_IMPL(uint, 1)
DEQUANTIZE_IMPL(int, 1)
DEQUANTIZE_IMPL(uchar, 4)
DEQUANTIZE_IMPL(ushort, 4)
DEQUANTIZE_IMPL(short, 4)
DEQUANTIZE_IMPL(uchar, 16)
DEQUANTIZE_IMPL(char, 16)
DEQUANTIZE_IMPL(ushort, 16)
DEQUANTIZE_IMPL(short, 16)
DEQUANTIZE_IMPL(uint, 16)
DEQUANTIZE_IMPL(int, 16)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(3)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
ASYMM_MULT_IMPL(1)
ASYMM_MULT_IMPL(2)
ASYMM_MULT_IMPL(3)
ASYMM_MULT_IMPL(4)
ASYMM_MULT_IMPL(8)
ASYMM_MULT_IMPL(16)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(1)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(3)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16)
ASYMM_SELECT_USING_MASK_IMPL(1)
ASYMM_SELECT_USING_MASK_IMPL(2)
ASYMM_SELECT_USING_MASK_IMPL(3)
ASYMM_SELECT_USING_MASK_IMPL(4)
ASYMM_SELECT_USING_MASK_IMPL(8)
ASYMM_SELECT_USING_MASK_IMPL(16)
ASYMM_MASK_IF_ZERO_IMPL(1)
ASYMM_MASK_IF_ZERO_IMPL(2)
ASYMM_MASK_IF_ZERO_IMPL(3)
ASYMM_MASK_IF_ZERO_IMPL(4)
ASYMM_MASK_IF_ZERO_IMPL(8)
ASYMM_MASK_IF_ZERO_IMPL(16)
ASYMM_MASK_IF_NON_ZERO_IMPL(1)
ASYMM_MASK_IF_NON_ZERO_IMPL(2)
ASYMM_MASK_IF_NON_ZERO_IMPL(3)
ASYMM_MASK_IF_NON_ZERO_IMPL(4)
ASYMM_MASK_IF_NON_ZERO_IMPL(8)
ASYMM_MASK_IF_NON_ZERO_IMPL(16)
EXP_BARREL_SHIFTER_IMPL(1)
EXP_BARREL_SHIFTER_IMPL(2)
EXP_BARREL_SHIFTER_IMPL(3)
EXP_BARREL_SHIFTER_IMPL(4)
EXP_BARREL_SHIFTER_IMPL(8)
EXP_BARREL_SHIFTER_IMPL(16)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(1)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(3)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(1)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(3)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16)
ASYMM_ROUNDING_HALF_SUM_IMPL(1)
ASYMM_ROUNDING_HALF_SUM_IMPL(2)
ASYMM_ROUNDING_HALF_SUM_IMPL(3)
ASYMM_ROUNDING_HALF_SUM_IMPL(4)
ASYMM_ROUNDING_HALF_SUM_IMPL(8)
ASYMM_ROUNDING_HALF_SUM_IMPL(16)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(1)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(3)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16)
ASYMM_RESCALE_IMPL(1)
ASYMM_RESCALE_IMPL(2)
ASYMM_RESCALE_IMPL(3)
ASYMM_RESCALE_IMPL(4)
ASYMM_RESCALE_IMPL(8)
ASYMM_RESCALE_IMPL(16)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(1)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(2)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(3)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(4)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(8)
MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(16)
#endif // ARM_COMPUTE_HELPERS_ASYMM_H
#ifndef VEC_SIZE
#if defined(N0)
#define VEC_SIZE N0
#else /* defined(N0) */
#define VEC_SIZE 8
#endif /* defined(N0) */
#endif /* VEC_SIZE */
#if defined(ACTIVATION_TYPE) && defined(CONST_0)
/*
* 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.
*/
/*
* 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
#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
#if defined(S1_VAL) && !defined(S2_VAL)
#define S2_VAL S1_VAL
#endif // defined(S1_VAL) && !defined(S2_VAL)
#if defined(O1_VAL) && !defined(O2_VAL)
#define O2_VAL O1_VAL
#endif // defined(O1_VAL) && !defined(O2_VAL)
// RELU Activation
inline TYPE relu_op(TYPE x)
{
return max((TYPE)CONST_0, x);
}
// Bounded RELU Activation
inline TYPE brelu_op(TYPE x)
{
return min((TYPE)A_VAL, max((TYPE)CONST_0, x));
}
// Lower Upper Bounded RELU Activation
inline TYPE lu_brelu_op(TYPE x)
{
return min(max(x, (TYPE)B_VAL), (TYPE)A_VAL);
}
// Hard Swish Activation
inline TYPE hard_swish_op(TYPE x)
{
return (x * ((min(max((TYPE)(x + (TYPE)3.f), (TYPE)0.f), (TYPE)6.f)) * (TYPE)0.166666667f));
}
#define ACTIVATION_OP2(op, x) op##_op(x)
#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x)
#if defined(S1_VAL) && defined(S2_VAL)
#if defined(O1_VAL) && defined(O2_VAL)
#define PERFORM_ACTIVATION_QUANT(act, data) \
({ \
data = ACTIVATION_OP(act, data); \
\
VEC_DATA_TYPE(float, VEC_SIZE) \
fdata = CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)); \
\
fdata = round((fdata - (float)O1_VAL) * ((float)S1_VAL / (float)S2_VAL) + (float)O2_VAL); \
data = CONVERT_SAT(fdata, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); \
})
#else // defined(O1_VAL) && defined(O2_VAL)
#define PERFORM_ACTIVATION_QUANT(act, data) \
({ \
data = ACTIVATION_OP(act, data); \
\
VEC_DATA_TYPE(float, VEC_SIZE) \
fdata = CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)); \
\
fdata = round((fdata) * ((float)S1_VAL / (float)S2_VAL)); \
data = CONVERT_SAT(fdata, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); \
})
#endif /* defined(O1_VAL) && defined(O2_VAL) */
#else /* defined(S1_VAL) && defined(S2_VAL) */
#define PERFORM_ACTIVATION_QUANT(act, data) \
({ \
data = ACTIVATION_OP(act, data); \
})
#endif /* defined(S1_VAL) && defined(S2_VAL) */
#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
#if defined(FLOAT_DOMAIN)
// Activations performed in the float domain
/*
* 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)
/** This performs an activation function on quantized inputs with float transformations.
*
* @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
* @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
* @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
* @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively.
* @note Quantization offsets of the input/output tensors are passed in only if asymmetric with -DO1_VAL= and -DO2_VAL= respectively.
* @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128.
*
* @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM16
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
*/
__kernel void activation_layer_quant_f32(
TENSOR3D_DECLARATION(input)
#ifndef IN_PLACE
,
TENSOR3D_DECLARATION(output)
#endif /* not IN_PLACE */
)
{
uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0);
// Get pixels pointer
__global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
#ifdef IN_PLACE
__global uchar *output_addr = input_addr;
#else /* IN_PLACE */
__global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
#endif /* IN_PLACE */
// Load data
TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
VEC_FLOAT data_flt = CONVERT(data0, VEC_FLOAT);
#if defined(O1_VAL)
data_flt = round(data_flt - (float)O1_VAL) * ((float)S1_VAL);
#else // defined(O1_VAL)
data_flt = round(data_flt) * ((float)S1_VAL);
#endif // defined(O1_VAL)
data_flt = ACTIVATION(ACT, float, VEC_SIZE, data_flt, A_VAL, B_VAL);
#if defined(O2_VAL)
data0 = CONVERT_SAT(round(data_flt / ((float)S2_VAL)) + (float)O2_VAL, TYPE);
#else // defined(O2_VAL)
data0 = CONVERT_SAT(round(data_flt / ((float)S2_VAL)), TYPE);
#endif // defined(O2_VAL)
// Store result
STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
}
#else // defined(FLOAT_DOMAIN)
// Activations performed in the quantized domain
#if defined(ACT)
/** This performs an activation function on quantized inputs.
*
* @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
* @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
* @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
* @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively.
* @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively.
* @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128.
*
* @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM16
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
*/
__kernel void activation_layer_quant(
TENSOR3D_DECLARATION(input)
#ifndef IN_PLACE
,
TENSOR3D_DECLARATION(output)
#endif /* not IN_PLACE */
)
{
uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0);
// Get pixels pointer
__global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
#ifdef IN_PLACE
__global uchar *output_addr = input_addr;
#else /* IN_PLACE */
__global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
#endif /* IN_PLACE */
// Load data
TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
data0 = PERFORM_ACTIVATION_QUANT(ACT, data0);
// Store result
STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
}
#endif // defined(ACT)
#endif // defined(FLOAT_DOMAIN)
#define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QUANT(ACTIVATION_TYPE, x)
#else /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
#define ACTIVATION_FUNC(x) (x)
#endif /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
#define VEC_SHORT VEC_DATA_TYPE(short, VEC_SIZE)
#if defined(DATA_TYPE) && defined(WEIGHTS_TYPE)
#define VEC_TYPE(size) VEC_DATA_TYPE(DATA_TYPE, size)
#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
#if defined(WEIGHTS_PROMOTED_TYPE)
#define VEC_WEIGHTS_PROMOTED_TYPE(size) VEC_DATA_TYPE(WEIGHTS_PROMOTED_TYPE, size)
#if 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)
#define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), val);
#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#define ARM_DOT(x, y, val) val += arm_dot((x), (y));
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS)
#if CONV_STRIDE_X > 3
#error "Stride X not supported"
#endif /* CONV_STRIDE_X > 3 */
#if !defined(IS_DOT8)
#if DILATION_X == 1
#if CONV_STRIDE_X == 1
#define GET_VALUES(first_value, left, middle, right) \
({ \
int8 temp0 = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value)), int8); \
int2 temp1 = CONVERT(vload2(0, (__global DATA_TYPE *)(first_value + 8 * sizeof(DATA_TYPE))), int2); \
\
left = CONVERT(temp0.s01234567, int8); \
middle = CONVERT((int8)(temp0.s1234, temp0.s567, temp1.s0), int8); \
right = CONVERT((int8)(temp0.s2345, temp0.s67, temp1.s01), int8); \
})
#elif CONV_STRIDE_X == 2
#define GET_VALUES(first_value, left, middle, right) \
({ \
int16 temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value)), int16); \
int temp1 = CONVERT(*((__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))), int); \
\
left = CONVERT(temp0.s02468ace, int8); \
middle = CONVERT(temp0.s13579bdf, int8); \
right = CONVERT((int8)(temp0.s2468, temp0.sace, temp1), int8); \
})
#else /* CONV_STRIDE_X */
#define GET_VALUES(first_value, left, middle, right) \
({ \
int16 temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value)), int16); \
int8 temp1 = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))), int8); \
\
left = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \
middle = CONVERT((int8)(temp0.s147a, temp0.sd, temp1.s036), int8); \
right = CONVERT((int8)(temp0.s258b, temp0.se, temp1.s147), int8); \
})
#endif /* CONV_STRIDE_X */
#else /* DILATION_X == 1 */
#if CONV_STRIDE_X == 1
#define GET_VALUES(first_value, left, middle, right) \
({ \
left = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value)), int8); \
middle = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))), int8); \
right = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))), int8); \
})
#elif CONV_STRIDE_X == 2
#define GET_VALUES(first_value, left, middle, right) \
({ \
int16 temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value)), int16); \
left = CONVERT(temp0.s02468ace, int8); \
\
temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))), int16); \
middle = CONVERT(temp0.s02468ace, int8); \
\
temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))), int16); \
right = CONVERT(temp0.s02468ace, int8); \
})
#else /* CONV_STRIDE_X */
#define GET_VALUES(first_value, left, middle, right) \
({ \
int16 temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value)), int16); \
int8 temp1 = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))), int8); \
left = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \
\
temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))), int16); \
temp1 = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + (16 + DILATION_X) * sizeof(DATA_TYPE))), int8); \
middle = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \
\
temp0 = CONVERT(vload16(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))), int16); \
temp1 = CONVERT(vload8(0, (__global DATA_TYPE *)(first_value + (16 + 2 * DILATION_X) * sizeof(DATA_TYPE))), int8); \
right = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \
})
#endif /* CONV_STRIDE_X */
#endif /* DILATION_X==1 */
/** This function computes the depthwise convolution quantized.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
*/
__kernel void dwc_3x3_native_quantized8_nchw(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts)
#if defined(HAS_BIAS)
,
VECTOR_DECLARATION(biases)
#endif //defined(HAS_BIAS)
)
{
Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers);
Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts);
// Extract channel and linearized batch indices
const int channel = get_global_id(2) % DST_CHANNELS;
const int batch = get_global_id(2) / DST_CHANNELS;
#if defined(HAS_BIAS)
Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
int bias_value = *((__global int *)(vector_offset(&biases, channel)));
#endif //defined(HAS_BIAS)
// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)
src.ptr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
__global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
VEC_DATA_TYPE(WEIGHTS_TYPE, 3)
w0 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 0 * weights_stride_y));
VEC_DATA_TYPE(WEIGHTS_TYPE, 3)
w1 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 1 * weights_stride_y));
VEC_DATA_TYPE(WEIGHTS_TYPE, 3)
w2 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * weights_stride_y));
#if defined(PER_CHANNEL_QUANTIZATION)
const int output_multiplier = *((__global int *)vector_offset(&output_multipliers, channel));
const int output_shift = *((__global int *)vector_offset(&output_shifts, channel));
#endif // defined(PER_CHANNEL_QUANTIZATION)
int8 values0 = 0;
int8 sum0 = 0;
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
int8 values1 = 0;
int8 sum1 = 0;
#endif /* CONV_STRIDE_Y &&DILATION_Y==1 */
// Row0
int8 left, middle, right;
GET_VALUES(src.ptr + 0 * src_stride_y, left, middle, right);
values0 += left * (int8)(w0.s0);
values0 += middle * (int8)(w0.s1);
values0 += right * (int8)(w0.s2);
#if WEIGHTS_OFFSET != 0
sum0 += left + middle + right;
#endif /* WEIGHTS_OFFSET != 0 */
// Row1
GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left, middle, right);
values0 += left * (int8)(w1.s0);
values0 += middle * (int8)(w1.s1);
values0 += right * (int8)(w1.s2);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += left * (int8)(w0.s0);
values1 += middle * (int8)(w0.s1);
values1 += right * (int8)(w0.s2);
#endif /* CONV_STRIDE_Y && DILATION_Y== 1 */
#if WEIGHTS_OFFSET != 0
int8 tmp = left + middle + right;
sum0 += tmp;
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
sum1 += tmp;
#endif /* CONV_STRIDE_Y &&DILATION_Y== 1 */
#endif /* WEIGHTS_OFFSET != 0 */
// Row2
GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left, middle, right);
values0 += left * (int8)(w2.s0);
values0 += middle * (int8)(w2.s1);
values0 += right * (int8)(w2.s2);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += left * (int8)(w1.s0);
values1 += middle * (int8)(w1.s1);
values1 += right * (int8)(w1.s2);
#endif /* CONV_STRIDE_Y &&DILATION_Y == 1 */
#if WEIGHTS_OFFSET != 0
tmp = left + middle + right;
sum0 += tmp;
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
sum1 += tmp;
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
#endif /* WEIGHTS_OFFSET != 0 */
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
// Row3
GET_VALUES(src.ptr + 3 * src_stride_y, left, middle, right);
values1 += left * (int8)(w2.s0);
values1 += middle * (int8)(w2.s1);
values1 += right * (int8)(w2.s2);
#if WEIGHTS_OFFSET != 0
sum1 += left + middle + right;
#endif /* WEIGHTS_OFFSET != 0 */
#endif /* CONV_STRIDE_Y && DILATION_Y == 1 */
#if defined(HAS_BIAS)
values0 += (int8)(bias_value);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += (int8)(bias_value);
#endif /* CONV_STRIDE_Y & &DILATION_Y == 1 */
#endif //defined(HAS_BIAS)
#if WEIGHTS_OFFSET != 0
values0 += sum0 * (int8)(WEIGHTS_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += sum1 * (int8)(WEIGHTS_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
#endif /* WEIGHTS_OFFSET != 0 */
#if INPUT_OFFSET != 0
VEC_WEIGHTS_PROMOTED_TYPE(3)
tmp_we = CONVERT(w0, VEC_WEIGHTS_PROMOTED_TYPE(3)) + CONVERT(w1, VEC_WEIGHTS_PROMOTED_TYPE(3)) + CONVERT(w2, VEC_WEIGHTS_PROMOTED_TYPE(3));
WEIGHTS_PROMOTED_TYPE sum_weights = tmp_we.s0 + tmp_we.s1 + tmp_we.s2;
values0 += sum_weights * (int8)(INPUT_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += sum_weights * (int8)(INPUT_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
#endif /* INPUT_OFFSET != 0 */
#if K_OFFSET != 0
values0 += (int8)(K_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += (int8)(K_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
#endif /* K_OFFSET != 0 */
#if defined(REAL_MULTIPLIER)
values0 = CONVERT(round(CONVERT(values0, float8) * (float8)REAL_MULTIPLIER), int8);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
int8 res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8);
int8 res0_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8);
values0 = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_OFFSET < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
values0 += (int8)OUTPUT_OFFSET;
VEC_TYPE(8)
res0 = CONVERT_SAT(values0, VEC_TYPE(8));
vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
#if defined(REAL_MULTIPLIER)
values1 = CONVERT(round(CONVERT(values1, float8) * (float8)REAL_MULTIPLIER), int8);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
int8 res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8);
int8 res1_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8);
values1 = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_OFFSET < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
values1 += (int8)OUTPUT_OFFSET;
VEC_TYPE(8)
res1 = CONVERT_SAT(values1, VEC_TYPE(8));
vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
}
#else // !defined(IS_DOT8)
#if DILATION_X == 1
#if CONV_STRIDE_X == 1
#define GET_VALUES(first_value, left, middle, right) \
({ \
VEC_TYPE(8) \
temp0 = vload8(0, (__global DATA_TYPE *)(first_value)); \
VEC_TYPE(2) \
temp1 = vload2(0, (__global DATA_TYPE *)(first_value + 8 * sizeof(DATA_TYPE))); \
\
left = temp0.s01234567; \
middle = (VEC_TYPE(8))(temp0.s1234, temp0.s567, temp1.s0); \
right = (VEC_TYPE(8))(temp0.s2345, temp0.s67, temp1.s01); \
})
#elif CONV_STRIDE_X == 2
#define GET_VALUES(first_value, left, middle, right) \
({ \
VEC_TYPE(16) \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value)); \
DATA_TYPE temp1 = *((__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))); \
\
left = temp0.s02468ace; \
middle = temp0.s13579bdf; \
right = (VEC_TYPE(8))(temp0.s2468, temp0.sace, temp1); \
})
#else /* CONV_STRIDE_X */
#define GET_VALUES(first_value, left, middle, right) \
({ \
VEC_TYPE(16) \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value)); \
VEC_TYPE(8) \
temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))); \
\
left = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \
middle = (VEC_TYPE(8))(temp0.s147a, temp0.sd, temp1.s036); \
right = (VEC_TYPE(8))(temp0.s258b, temp0.se, temp1.s147); \
})
#endif /* CONV_STRIDE_X */
#else /*DILATION_X==1*/
#if CONV_STRIDE_X == 1
#define GET_VALUES(first_value, left, middle, right) \
({ \
left = vload8(0, (__global DATA_TYPE *)(first_value)); \
middle = vload8(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))); \
right = vload8(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))); \
})
#elif CONV_STRIDE_X == 2
#define GET_VALUES(first_value, left, middle, right) \
({ \
VEC_TYPE(16) \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value)); \
left = temp0.s02468ace; \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))); \
middle = temp0.s02468ace; \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))); \
right = temp0.s02468ace; \
})
#else /* CONV_STRIDE_X */
#define GET_VALUES(first_value, left, middle, right) \
({ \
VEC_TYPE(16) \
temp0 = vload16(0, (__global DATA_TYPE *)(first_value)); \
VEC_TYPE(8) \
temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE)))); \
left = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \
\
temp0 = vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE))); \
temp1 = vload8(0, (__global DATA_TYPE *)(first_value + (16 + DILATION_X) * sizeof(DATA_TYPE))); \
middle = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \
\
temp0 = vload16(0, (__global DATA_TYPE *)(first_value + 2 * DILATION_X * sizeof(DATA_TYPE))); \
temp1 = vload8(0, (__global DATA_TYPE *)(first_value + (16 + 2 * DILATION_X) * sizeof(DATA_TYPE))); \
right = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25); \
})
#endif /* CONV_STRIDE_X */
#endif /*DILATION_X==1*/
/** This function computes the depthwise convolution quantized using dot product when the data layout is NCHW.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
*/
__kernel void dwc_3x3_native_quantized8_dot8_nchw(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts)
#if defined(HAS_BIAS)
,
VECTOR_DECLARATION(biases)
#endif //defined(HAS_BIAS)
)
{
Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers);
Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts);
// Extract channel and linearized batch indices
const int channel = get_global_id(2) % DST_CHANNELS;
const int batch = get_global_id(2) / DST_CHANNELS;
#if defined(HAS_BIAS)
Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
const int bias_value = *((__global int *)(vector_offset(&biases, channel)));
#endif //defined(HAS_BIAS)
// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)
src.ptr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;
__global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;
VEC_TYPE(3)
w0 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 0 * weights_stride_y));
VEC_TYPE(3)
w1 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 1 * weights_stride_y));
VEC_TYPE(3)
w2 = vload3(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * weights_stride_y));
const int output_multiplier = *((__global int *)vector_offset(&output_multipliers, 0));
const int output_shift = *((__global int *)vector_offset(&output_shifts, 0));
VEC_TYPE(8)
left0, middle0, right0;
VEC_TYPE(8)
left1, middle1, right1;
VEC_TYPE(8)
left2, middle2, right2;
int8 values0 = 0;
int8 sum0 = 0;
GET_VALUES(src.ptr + 0 * src_stride_y, left0, middle0, right0);
GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left1, middle1, right1);
GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left2, middle2, right2);
#if WEIGHTS_OFFSET != 0
sum0 += convert_int8(left0) + convert_int8(middle0) + convert_int8(right0);
sum0 += convert_int8(left1) + convert_int8(middle1) + convert_int8(right1);
sum0 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2);
#endif /* WEIGHTS_OFFSET != 0 */
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
// If conv_stride_y is equals to 1, we compute two output rows
VEC_TYPE(8)
left3, middle3, right3;
int8 values1 = 0;
int8 sum1 = 0;
GET_VALUES(src.ptr + 3 * src_stride_y, left3, middle3, right3);
#if WEIGHTS_OFFSET != 0
sum1 += convert_int8(left1) + convert_int8(middle1) + convert_int8(right1);
sum1 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2);
sum1 += convert_int8(left3) + convert_int8(middle3) + convert_int8(right3);
#endif /* WEIGHTS_OFFSET != 0 */
#endif // CONV_STRIDE_Y == 1 && DILATION_Y==1
ARM_DOT((VEC_TYPE(4))(left0.s0, middle0.s0, right0.s0, left1.s0), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s0);
ARM_DOT((VEC_TYPE(4))(middle1.s0, right1.s0, left2.s0, middle2.s0), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s0);
values0.s0 += right2.s0 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s1, middle0.s1, right0.s1, left1.s1), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s1);
ARM_DOT((VEC_TYPE(4))(middle1.s1, right1.s1, left2.s1, middle2.s1), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s1);
values0.s1 += right2.s1 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s2, middle0.s2, right0.s2, left1.s2), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s2);
ARM_DOT((VEC_TYPE(4))(middle1.s2, right1.s2, left2.s2, middle2.s2), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s2);
values0.s2 += right2.s2 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s3, middle0.s3, right0.s3, left1.s3), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s3);
ARM_DOT((VEC_TYPE(4))(middle1.s3, right1.s3, left2.s3, middle2.s3), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s3);
values0.s3 += right2.s3 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s4, middle0.s4, right0.s4, left1.s4), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s4);
ARM_DOT((VEC_TYPE(4))(middle1.s4, right1.s4, left2.s4, middle2.s4), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s4);
values0.s4 += right2.s4 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s5, middle0.s5, right0.s5, left1.s5), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s5);
ARM_DOT((VEC_TYPE(4))(middle1.s5, right1.s5, left2.s5, middle2.s5), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s5);
values0.s5 += right2.s5 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s6, middle0.s6, right0.s6, left1.s6), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s6);
ARM_DOT((VEC_TYPE(4))(middle1.s6, right1.s6, left2.s6, middle2.s6), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s6);
values0.s6 += right2.s6 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left0.s7, middle0.s7, right0.s7, left1.s7), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values0.s7);
ARM_DOT((VEC_TYPE(4))(middle1.s7, right1.s7, left2.s7, middle2.s7), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values0.s7);
values0.s7 += right2.s7 * w2.s2;
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
ARM_DOT((VEC_TYPE(4))(left1.s0, middle1.s0, right1.s0, left2.s0), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s0);
ARM_DOT((VEC_TYPE(4))(middle2.s0, right2.s0, left3.s0, middle3.s0), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s0);
values1.s0 += right3.s0 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s1, middle1.s1, right1.s1, left2.s1), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s1);
ARM_DOT((VEC_TYPE(4))(middle2.s1, right2.s1, left3.s1, middle3.s1), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s1);
values1.s1 += right3.s1 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s2, middle1.s2, right1.s2, left2.s2), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s2);
ARM_DOT((VEC_TYPE(4))(middle2.s2, right2.s2, left3.s2, middle3.s2), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s2);
values1.s2 += right3.s2 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s3, middle1.s3, right1.s3, left2.s3), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s3);
ARM_DOT((VEC_TYPE(4))(middle2.s3, right2.s3, left3.s3, middle3.s3), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s3);
values1.s3 += right3.s3 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s4, middle1.s4, right1.s4, left2.s4), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s4);
ARM_DOT((VEC_TYPE(4))(middle2.s4, right2.s4, left3.s4, middle3.s4), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s4);
values1.s4 += right3.s4 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s5, middle1.s5, right1.s5, left2.s5), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s5);
ARM_DOT((VEC_TYPE(4))(middle2.s5, right2.s5, left3.s5, middle3.s5), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s5);
values1.s5 += right3.s5 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s6, middle1.s6, right1.s6, left2.s6), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s6);
ARM_DOT((VEC_TYPE(4))(middle2.s6, right2.s6, left3.s6, middle3.s6), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s6);
values1.s6 += right3.s6 * w2.s2;
ARM_DOT((VEC_TYPE(4))(left1.s7, middle1.s7, right1.s7, left2.s7), (VEC_TYPE(4))(w0.s0, w0.s1, w0.s2, w1.s0), values1.s7);
ARM_DOT((VEC_TYPE(4))(middle2.s7, right2.s7, left3.s7, middle3.s7), (VEC_TYPE(4))(w1.s1, w1.s2, w2.s0, w2.s1), values1.s7);
values1.s7 += right3.s7 * w2.s2;
#endif // CONV_STRIDE_Y == 1 && DILATION_Y==1
#if defined(HAS_BIAS)
values0 += (int8)(bias_value);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += (int8)(bias_value);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
#endif //defined(HAS_BIAS)
#if WEIGHTS_OFFSET != 0
values0 += sum0 * (int8)(WEIGHTS_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += sum1 * (int8)(WEIGHTS_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1 */
#endif /* WEIGHTS_OFFSET != 0 */
#if INPUT_OFFSET != 0
WEIGHTS_PROMOTED_TYPE sum_weights = 0;
VEC_WEIGHTS_PROMOTED_TYPE(3)
tmp_we = CONVERT(w0, VEC_WEIGHTS_PROMOTED_TYPE(3)) + CONVERT(w1, VEC_WEIGHTS_PROMOTED_TYPE(3)) + CONVERT(w2, VEC_WEIGHTS_PROMOTED_TYPE(3));
sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2;
values0 += sum_weights * (int8)(INPUT_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += sum_weights * (int8)(INPUT_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
#endif /* INPUT_OFFSET != 0 */
#if K_OFFSET != 0
values0 += (int8)(K_OFFSET);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
values1 += (int8)(K_OFFSET);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
#endif /* K_OFFSET != 0 */
#if defined(REAL_MULTIPLIER)
values0 = CONVERT(round(CONVERT(values0, float8) * (float8)REAL_MULTIPLIER), int8);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
int8 res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8);
int8 res0_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8);
values0 = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_OFFSET < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
values0 += (int8)OUTPUT_OFFSET;
VEC_TYPE(8)
res0 = CONVERT_SAT(values0, VEC_TYPE(8));
vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr);
#if CONV_STRIDE_Y == 1 && DILATION_Y == 1
#if defined(REAL_MULTIPLIER)
values1 = CONVERT(round(CONVERT(values1, float8) * (float8)REAL_MULTIPLIER), int8);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
int8 res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8);
int8 res1_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8);
values1 = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_OFFSET < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
values1 += (int8)OUTPUT_OFFSET;
VEC_TYPE(8)
res1 = CONVERT_SAT(values1, VEC_TYPE(8));
vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y);
#endif /* CONV_STRIDE_Y == 1 && DILATION_Y==1*/
}
#endif // !defined(IS_DOT8)
#endif /* defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) */
#if defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
#define asymm_mult_by_quant_multiplier_less_than_one(x, y, z) ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, y, z, VEC_SIZE)
#define MULTIPLY_ADD(x, y, acc) acc += CONVERT(CONVERT(x, VEC_WEIGHTS_PROMOTED_TYPE(VEC_SIZE)) * CONVERT(y, VEC_WEIGHTS_PROMOTED_TYPE(VEC_SIZE)), VEC_INT)
#if WEIGHTS_OFFSET != 0
#define MULTIPLY_ADD_ACCUMULATE(x, y, acc, sum) \
({ \
sum += CONVERT(x, VEC_INT); \
MULTIPLY_ADD(x, y, acc); \
})
#else /* WEIGHTS_OFFSET != 0 */
#define MULTIPLY_ADD_ACCUMULATE(x, y, acc, sum) MULTIPLY_ADD(x, y, acc)
#endif /* WEIGHTS_OFFSET != 0 */
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#define DOT_PRODUCT(acc, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1) \
({ \
ARM_DOT((VEC_TYPE(4))(val0, val1, val2, val3), w0.s0123, acc); \
ARM_DOT((VEC_TYPE(4))(val4, val5, val6, val7), w0.s4567, acc); \
acc += val8 * w1; \
})
#define DOT_PRODUCT_REDUCTION(sum, val0, val1, val2, val3, val4, val5, val6, val7, val8) \
({ \
sum = val0; \
ARM_DOT((VEC_TYPE(4))(val1, val2, val3, val4), (VEC_TYPE(4))1, sum); \
ARM_DOT((VEC_TYPE(4))(val5, val6, val7, val8), (VEC_TYPE(4))1, sum); \
})
#define DOT_PRODUCT_REDUCTION_WEIGHTS(sum, w0, w1) \
({ \
sum = w1; \
ARM_DOT(w0.s0123, (VEC_TYPE(4))1, sum); \
ARM_DOT(w0.s4567, (VEC_TYPE(4))1, sum); \
})
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && VEC_SIZE == 4
/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width or height is not 1.
*
* @note This kernel assumes VEC_SIZE is 4.
* @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
* @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
* @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
* @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
* @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor reshaped. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
* @param[in] max_offset Max offset for the input tensor
*/
__kernel void dwc_3x3_reshaped_quantized8_nhwc(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst),
IMAGE_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts),
#if defined(HAS_BIAS)
VECTOR_DECLARATION(biases),
#endif /* defined(HAS_BIAS) */
int max_offset)
{
const int x = get_global_id(0); // channels
const int y = get_global_id(1); // spatial coordinate x
#if defined(DST_DEPTH)
int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
int b = get_global_id(2) / (int)DST_DEPTH; // batch
#else // defined(DST_DEPTH)
int z = get_global_id(2); // spatial coordinate y
#endif // defined(DST_DEPTH)
__global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
#if defined(DST_DEPTH)
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
#endif /* defined(DST_DEPTH) */
int z_coord = 0;
int4 offset = 0;
int4 y_coord = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3)) - (int)CONV_PAD_LEFT;
// Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
// We compute VEC_SIZEx1x1 [C,W,H] elements
VEC_INT acc = 0, sum = 0;
// Load weights
VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16));
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w0 = w0_tmp.s0123;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w1 = w0_tmp.s4567;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w2 = w0_tmp.s89AB;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w3 = w0_tmp.sCDEF;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w4 = w1_tmp.s0123;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w5 = w1_tmp.s4567;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w6 = w1_tmp.s89AB;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w7 = w1_tmp.sCDEF;
#if INPUT_OFFSET != 0
VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT)
+ CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT)
+ CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT);
#endif /* INPUT_OFFSET != 0 */
// Load input values
// z == 0
// Clamp z_coord as for z = 0, it can be negative
// z_coord is casted to unsigned int in order to use just a min() operation
// A "-1" 32 bit signed variable converted to unsigned gives 4294967295
z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP;
z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
offset = y_offset + (int4)(z_coord * src_stride_z);
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
// z == 1
// z_coord can be only negative for z = 0 so we do not need to clamp it
// Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y;
offset = y_offset + (int4)(z_coord * src_stride_z);
VEC_TYPE(VEC_SIZE)
values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
// z == 2
// Offset can be out-of-bound so we need to check if it is greater than max_offset
z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2;
offset = y_offset + (int4)(z_coord * src_stride_z);
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
MULTIPLY_ADD_ACCUMULATE(values0, w0, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values1, w1, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values2, w2, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values3, w3, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values4, w4, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values5, w5, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values6, w6, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values7, w7, acc, sum);
MULTIPLY_ADD_ACCUMULATE(values8, w8, acc, sum);
#if defined(HAS_BIAS)
Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
acc += bias_values;
#endif // defined(HAS_BIAS)
#if WEIGHTS_OFFSET != 0
acc += WEIGHTS_OFFSET * sum;
#endif /* WEIGHTS_OFFSET != 0 */
#if INPUT_OFFSET != 0
acc += INPUT_OFFSET * sum_we;
#endif /* INPUT_OFFSET != 0 */
#if K_OFFSET != 0
acc += (VEC_INT)K_OFFSET;
#endif /* K_OFFSET != 0 */
#if defined(REAL_MULTIPLIER)
acc = CONVERT(round(CONVERT(acc, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers);
Vector output_shifts = CONVERT_TO_VECTOR_STRUCT(output_shifts);
VEC_INT output_multiplier = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr);
VEC_INT output_shift = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr);
VEC_INT res_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, output_multiplier, output_shift, VEC_SIZE);
VEC_INT res_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc, output_multiplier, output_shift);
acc = select(res_shift_lt0, res_shift_gt0, output_shift >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
acc = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
#else // OUTPUT_SHIFT < 0
acc = asymm_mult_by_quant_multiplier_less_than_one(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
#endif // OUTPUT_SHIFT < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
acc += (VEC_INT)OUTPUT_OFFSET;
VEC_TYPE(VEC_SIZE)
res = CONVERT_SAT(acc, VEC_TYPE(VEC_SIZE));
#if defined(DST_DEPTH)
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z;
#endif /* defined(DST_DEPTH) */
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res), 0, (__global DATA_TYPE *)(dst_addr));
}
#endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y)
#if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) && VEC_SIZE == 4
/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1.
*
* @note This kernel assumes VEC_SIZE is 4.
* @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
* @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
* @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
* @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2)
* @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1).
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
* @param[in] max_offset Max offset for the input tensor
*/
__kernel void dwc_3x3_reshaped_quantized8_stride1_nhwc(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst),
IMAGE_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts),
#if defined(HAS_BIAS)
VECTOR_DECLARATION(biases),
#endif /* defined(HAS_BIAS) */
int max_offset)
{
int x = get_global_id(0);
int y = get_global_id(1);
#if defined(DST_DEPTH)
int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
int b = get_global_id(2) / (int)DST_DEPTH; // batch
#else // defined(DST_DEPTH)
int z = get_global_id(2); // spatial coordinate y
#endif // defined(DST_DEPTH)
__global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
#if defined(DST_DEPTH)
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
#endif /* defined(DST_DEPTH) */
int z_coord = 0;
int4 offset = 0;
int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT;
// Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
// We compute 4x2x2 [C,W,H] elements
VEC_INT acc0 = 0, sum0 = 0;
VEC_INT acc1 = 0, sum1 = 0;
VEC_INT acc2 = 0, sum2 = 0;
VEC_INT acc3 = 0, sum3 = 0;
// Load weights
VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16));
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w0 = w0_tmp.s0123;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w1 = w0_tmp.s4567;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w2 = w0_tmp.s89AB;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w3 = w0_tmp.sCDEF;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w4 = w1_tmp.s0123;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w5 = w1_tmp.s4567;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w6 = w1_tmp.s89AB;
VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
w7 = w1_tmp.sCDEF;
#if INPUT_OFFSET != 0
VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT)
+ CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT)
+ CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT);
#endif /* INPUT_OFFSET != 0 */
// Load input values
// z == 0
// Clamp z_coord as for z = 0, it can be negative
// z_coord is casted to unsigned int in order to use just a min() operation
// A "-1" 32 bit signed variable converted to unsigned gives 4294967295
z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP;
z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
offset = y_offset + (int4)(z_coord * src_stride_z);
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
// z == 1
// z_coord can be only negative for z = 0 so we do not need to clamp it
// Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1;
offset = y_offset + (int4)(z_coord * src_stride_z);
VEC_TYPE(VEC_SIZE)
values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
// z == 2
// After z = 1 we can simply add src_stride_z to offset without updating z_coord
// However offset can be out-of-bound so we need to check if it is greater than max_offset
offset += (int4)src_stride_z;
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
// z == 3
// After z = 1 we can simply add src_stride_z to offset without updating z_coord
// However offset can be out-of-bound so we need to check if it is greater than max_offset
offset += (int4)(src_stride_z);
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
MULTIPLY_ADD_ACCUMULATE(values0, w0, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values1, w1, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values2, w2, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values1, w0, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values2, w1, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values3, w2, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values4, w3, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values5, w4, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values6, w5, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values5, w3, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values6, w4, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values7, w5, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values8, w6, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values9, w7, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values10, w8, acc0, sum0);
MULTIPLY_ADD_ACCUMULATE(values9, w6, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values10, w7, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values11, w8, acc1, sum1);
MULTIPLY_ADD_ACCUMULATE(values4, w0, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values5, w1, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values6, w2, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values5, w0, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values6, w1, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values7, w2, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values8, w3, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values9, w4, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values10, w5, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values9, w3, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values10, w4, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values11, w5, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values12, w6, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values13, w7, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values14, w8, acc2, sum2);
MULTIPLY_ADD_ACCUMULATE(values13, w6, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values14, w7, acc3, sum3);
MULTIPLY_ADD_ACCUMULATE(values15, w8, acc3, sum3);
#if defined(HAS_BIAS)
Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
acc0 += bias_values;
acc1 += bias_values;
acc2 += bias_values;
acc3 += bias_values;
#endif /* defined(HAS_BIAS) */
#if WEIGHTS_OFFSET != 0
acc0 += WEIGHTS_OFFSET * sum0;
acc1 += WEIGHTS_OFFSET * sum1;
acc2 += WEIGHTS_OFFSET * sum2;
acc3 += WEIGHTS_OFFSET * sum3;
#endif /* WEIGHTS_OFFSET != 0 */
#if INPUT_OFFSET != 0
VEC_INT offs = INPUT_OFFSET * sum_we;
acc0 += offs;
acc1 += offs;
acc2 += offs;
acc3 += offs;
#endif /* INPUT_OFFSET != 0 */
#if K_OFFSET != 0
acc0 += (VEC_INT)K_OFFSET;
acc1 += (VEC_INT)K_OFFSET;
acc2 += (VEC_INT)K_OFFSET;
acc3 += (VEC_INT)K_OFFSET;
#endif /* K_OFFSET != 0 */
#if defined(REAL_MULTIPLIER)
acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
acc2 = CONVERT(round(CONVERT(acc2, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
acc3 = CONVERT(round(CONVERT(acc3, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
#else // defined(REAL_MULTIPLIER)
#if defined(PER_CHANNEL_QUANTIZATION)
Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers);
Vector output_shifts = CONVERT_TO_VECTOR_STRUCT(output_shifts);
VEC_INT output_multiplier = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr);
VEC_INT output_shift = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr);
VEC_INT res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, output_multiplier, output_shift, VEC_SIZE);
VEC_INT res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, output_multiplier, output_shift, VEC_SIZE);
VEC_INT res2_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, output_multiplier, output_shift, VEC_SIZE);
VEC_INT res3_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, output_multiplier, output_shift, VEC_SIZE);
VEC_INT res0_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, output_multiplier, output_shift);
VEC_INT res1_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc1, output_multiplier, output_shift);
VEC_INT res2_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc2, output_multiplier, output_shift);
VEC_INT res3_shift_gt0 = asymm_mult_by_quant_multiplier_less_than_one(acc3, output_multiplier, output_shift);
acc0 = select(res0_shift_lt0, res0_shift_gt0, output_shift >= 0);
acc1 = select(res1_shift_lt0, res1_shift_gt0, output_shift >= 0);
acc2 = select(res2_shift_lt0, res2_shift_gt0, output_shift >= 0);
acc3 = select(res3_shift_lt0, res3_shift_gt0, output_shift >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
acc0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
acc1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
acc2 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
acc3 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
#else // OUTPUT_SHIFT < 0
acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
acc3 = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
#endif // OUTPUT_SHIFT < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
#endif // defined(REAL_MULTIPLIER)
acc0 += (VEC_INT)OUTPUT_OFFSET;
acc1 += (VEC_INT)OUTPUT_OFFSET;
acc2 += (VEC_INT)OUTPUT_OFFSET;
acc3 += (VEC_INT)OUTPUT_OFFSET;
VEC_TYPE(VEC_SIZE)
res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE));
VEC_TYPE(VEC_SIZE)
res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE));
VEC_TYPE(VEC_SIZE)
res2 = CONVERT_SAT(acc2, VEC_TYPE(VEC_SIZE));
VEC_TYPE(VEC_SIZE)
res3 = CONVERT_SAT(acc3, VEC_TYPE(VEC_SIZE));
#if defined(DST_DEPTH)
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z + b * dst_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z;
#endif /* defined(DST_DEPTH) */
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res0), 0, dst_addr + 0 * dst_stride_y);
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res1), 0, dst_addr + 1 * dst_stride_y);
#if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2)
#endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
{
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res2), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z));
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res3), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z));
}
}
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE == 4
/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1 using dot product.
*
* @note Per-channel quantization is not supported by this kernel.
* @note This kernel assumes VEC_SIZE is 4.
* @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
* @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
* @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
* @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2)
* @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1).
* @note If REAL_MULTIPLIER is passed at compile time (i.e. -DREAL_MULTIPLIER=1.355f), the final quantization is performed using a floating point multiplication.
* If not, the quantization will be performed using a fixed point multiplication
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
* @param[in] max_offset The maximum allowed offset for the input tensor
*/
__kernel void dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst),
IMAGE_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts),
#if defined(HAS_BIAS)
VECTOR_DECLARATION(biases),
#endif // defined(HAS_BIAS)
int max_offset)
{
int x = get_global_id(0);
int y = get_global_id(1);
#if defined(DST_DEPTH)
int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
int b = get_global_id(2) / (int)DST_DEPTH; // batch
#else // defined(DST_DEPTH)
int z = get_global_id(2); // spatial coordinate y
#endif // defined(DST_DEPTH)
__global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
#if defined(DST_DEPTH)
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
#endif /* defined(DST_DEPTH) */
int z_coord = 0;
int4 offset = 0;
int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT;
// Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
// We compute 4x2x1 [C,W,H] elements
VEC_INT acc0 = 0;
VEC_INT acc1 = 0;
VEC_INT sum0 = 0;
VEC_INT sum1 = 0;
// Load weights
VEC_TYPE(16)
w0 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
VEC_TYPE(16)
w1 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
VEC_TYPE(4)
w2 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 32));
#if INPUT_OFFSET != 0
// Initilize the final result with the weights reduction multiplied by INPUT_OFFSET
DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s0, w0.s01234567, w0.s8);
DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s2, w1.s23456789, w1.sA);
DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
// Multiply the weights reduction with INPUT_OFFSET
acc0 = INPUT_OFFSET * acc0;
acc1 = acc0;
#endif // INPUT_OFFSET != 0
// Load input values
// z == 0
// Clamp z_coord as for z = 0, it can be negative
// z_coord is casted to unsigned int in order to use just a min() operation
// A "-1" 32 bit signed variable converted to unsigned gives 4294967295
z_coord = z - (int)CONV_PAD_TOP;
z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
offset = y_offset + (int4)(z_coord * src_stride_z);
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
// z == 1
// z_coord can be only negative for z = 0 so we do not need to clamp it
// Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
z_coord = z - (int)CONV_PAD_TOP + 1;
offset = y_offset + (int4)(z_coord * src_stride_z);
VEC_TYPE(VEC_SIZE)
values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
// z == 2
// After z = 1 we can simply add src_stride_z to offset without updating z_coord
// However offset can be out-of-bound so we need to check if it is greater than max_offset
offset += (int4)src_stride_z;
offset = min(offset, (int4)max_offset);
VEC_TYPE(VEC_SIZE)
values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
VEC_TYPE(VEC_SIZE)
values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
VEC_TYPE(VEC_SIZE)
values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
VEC_TYPE(VEC_SIZE)
values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
DOT_PRODUCT_REDUCTION(sum0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0);
DOT_PRODUCT_REDUCTION(sum1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0);
DOT_PRODUCT(acc0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0, w0.s01234567, w0.s8);
DOT_PRODUCT(acc1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0, w0.s01234567, w0.s8);
DOT_PRODUCT_REDUCTION(sum0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1);
DOT_PRODUCT_REDUCTION(sum1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1);
DOT_PRODUCT(acc0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
DOT_PRODUCT(acc1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
DOT_PRODUCT_REDUCTION(sum0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2);
DOT_PRODUCT_REDUCTION(sum1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2);
DOT_PRODUCT(acc0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2, w1.s23456789, w1.sA);
DOT_PRODUCT(acc1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2, w1.s23456789, w1.sA);
DOT_PRODUCT_REDUCTION(sum0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3);
DOT_PRODUCT_REDUCTION(sum1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3);
DOT_PRODUCT(acc0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
DOT_PRODUCT(acc1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
#if defined(HAS_BIAS)
Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
acc0 += bias_values;
acc1 += bias_values;
#endif // defined(HAS_BIAS)
#if WEIGHTS_OFFSET != 0
acc0 += WEIGHTS_OFFSET * sum0;
acc1 += WEIGHTS_OFFSET * sum1;
#endif // WEIGHTS_OFFSET != 0
#if K_OFFSET != 0
acc0 += (VEC_INT)K_OFFSET;
acc1 += (VEC_INT)K_OFFSET;
#endif // K_OFFSET != 0
#if defined(REAL_MULTIPLIER)
acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
#else // defined(REAL_MULTIPLIER)
#if OUTPUT_SHIFT < 0
acc0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
acc1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
#else // OUTPUT_SHIFT < 0
acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
#endif // OUTPUT_SHIFT < 0
#endif // defined(REAL_MULTIPLIER)
acc0 += (VEC_INT)OUTPUT_OFFSET;
acc1 += (VEC_INT)OUTPUT_OFFSET;
VEC_TYPE(VEC_SIZE)
res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE));
VEC_TYPE(VEC_SIZE)
res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE));
#if defined(DST_DEPTH)
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w;
#else /* defined(DST_DEPTH) */
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z;
#endif /* defined(DST_DEPTH) */
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res0), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
VSTORE(VEC_SIZE)
(ACTIVATION_FUNC(res1), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
}
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE==4
#endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED)
#endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
#endif // defined(WEIGHTS_PROMOTED_TYPE)
#endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
#if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_SHIFT) && defined(OUTPUT_MULTIPLIER) && defined(VEC_SIZE_LEFTOVER)
/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped
*
* @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2)
* @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1)
* @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112)
* @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80)
* @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5)
* @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
* @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
* @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
* @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
* @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source 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 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 tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor 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 tensor 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 Y processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] output_multipliers_ptr Pointer to the output multipliers vector. Supported data types: S32
* @param[in] output_multipliers_stride_x Stride of the output multipliers vector in X dimension (in bytes)
* @param[in] output_multipliers_step_x output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
* @param[in] output_shifts_ptr Pointer to the output shifts vector. Supported data types: S32
* @param[in] output_shifts_stride_x Stride of the output shifts vector in X dimension (in bytes)
* @param[in] output_shifts_step_x output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_shifts_offset_first_element_in_bytes The offset of the first element in the output shifts vector
* @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: S32
* @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
*/
__kernel void dwc_MxN_native_quantized8_nhwc(
TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
VECTOR_DECLARATION(output_multipliers),
VECTOR_DECLARATION(output_shifts)
#if defined(HAS_BIAS)
,
VECTOR_DECLARATION(biases)
#endif // defined(HAS_BIAS)
)
{
int x_offs = max((int)(get_global_id(0) * N0 - (N0 - VEC_SIZE_LEFTOVER) % N0), 0);
int y = get_global_id(1); // spatial coordinate x
#if defined(DST_DEPTH)
int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
int b = get_global_id(2) / (int)DST_DEPTH; // batch
#else // defined(DST_DEPTH)
int z = get_global_id(2); // spatial coordinate y
#endif // defined(DST_DEPTH)
__global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE);
__global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER + y * dst_stride_y + z * dst_stride_z;
__global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x_offs * sizeof(WEIGHTS_TYPE) * (int)DEPTH_MULTIPLIER;
#if defined(HAS_BIAS)
__global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x_offs * sizeof(int) * (int)DEPTH_MULTIPLIER;
#endif // defined(HAS_BIAS)
#if defined(PER_CHANNEL_QUANTIZATION)
__global uchar *out_mul_addr = output_multipliers_ptr + output_multipliers_offset_first_element_in_bytes + x_offs * sizeof(int) * (int)DEPTH_MULTIPLIER;
__global uchar *out_shift_addr = output_shifts_ptr + output_shifts_offset_first_element_in_bytes + x_offs * sizeof(int) * (int)DEPTH_MULTIPLIER;
#endif // defined(PER_CHANNEL_QUANTIZATION)
#if defined(DST_DEPTH)
s_addr += b * src_stride_w;
d_addr += b * dst_stride_w;
#endif // defined(DST_DEPTH)
#if DEPTH_MULTIPLIER > 1
for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d)
{
#endif // DEPTH_MULTIPLIER > 1
// Each work-item computes N0x1x1 elements
VEC_INT res = 0;
int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT;
int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;
for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
{
if(y_coord >= 0 && y_coord < SRC_DIM2)
{
int x_coord_tmp = x_coord;
for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
{
if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1)
{
int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z;
int w_offset = xk * weights_stride_y + yk * weights_stride_z;
// Load input and weights values
VEC_INT i = CONVERT(VLOAD(N0)(0, (__global DATA_TYPE *)(s_addr + s_offset)), VEC_INT);
VEC_INT w = CONVERT(VLOAD(N0)(0, (__global WEIGHTS_TYPE *)(w_addr + w_offset)), VEC_INT);
res += (i + (VEC_INT)INPUT_OFFSET) * (w + (VEC_INT)WEIGHTS_OFFSET);
}
x_coord_tmp += DILATION_X;
}
}
y_coord += DILATION_Y;
}
#if defined(HAS_BIAS)
VEC_INT bias = VLOAD(N0)(0, (__global int *)(b_addr));
res += bias;
#endif // defined(HAS_BIAS)
#if defined(PER_CHANNEL_QUANTIZATION)
VEC_INT output_multiplier = VLOAD(N0)(0, (__global int *)(out_mul_addr));
VEC_INT output_shift = VLOAD(N0)(0, (__global int *)(out_shift_addr));
VEC_INT res_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(res, output_multiplier, output_shift, N0);
VEC_INT res_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(res, output_multiplier, output_shift, N0);
res = select(res_shift_lt0, res_shift_gt0, (VEC_INT)(output_shift) >= 0);
#else // defined(PER_CHANNEL_QUANTIZATION)
#if OUTPUT_SHIFT < 0
res = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, N0);
#else // OUTPUT_SHIFT < 0
res = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, N0);
#endif // OUTPUT_OFFSET < 0
#endif // defined(PER_CHANNEL_QUANTIZATION)
res += (VEC_INT)OUTPUT_OFFSET;
VEC_TYPE(VEC_SIZE)
res0 = CONVERT_SAT(res, VEC_TYPE(VEC_SIZE));
res0 = ACTIVATION_FUNC(res0);
STORE_VECTOR_SELECT(res, DATA_TYPE, d_addr, N0, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
#if DEPTH_MULTIPLIER > 1
w_addr += sizeof(WEIGHTS_TYPE);
d_addr += sizeof(DATA_TYPE);
#if defined(PER_CHANNEL_QUANTIZATION)
out_mul_addr += sizeof(int);
out_shift_addr += sizeof(int);
#endif // defined(PER_CHANNEL_QUANTIZATION)
#if defined(HAS_BIAS)
b_addr += sizeof(int);
#endif // defined(HAS_BIAS)
}
#endif // DEPTH_MULTIPLIER > 1
}
#endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_SHIFT) && defined(OUTPUT_MULTIPLIER) && defined(VEC_SIZE_LEFTOVER)
#endif // defined(DATA_TYPE) && defined(WEIGHTS_TYPE)
)"