| 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 |
| |
| #undef CONVERT_SAT_STR |
| #undef CONVERT_SAT |
| |
| #if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) |
| |
| #define CONVERT_SAT_STR(x, type) (convert_##type##8_sat((x))) |
| #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type) |
| |
| #if defined(DATA_LAYOUT_NHWC) |
| |
| #if KERNEL_SIZE == 9 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x9(acc, src_ptr, weights_ptr) CONVOLUTION1x9_STRIDE1(acc, src_ptr, weights_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x9(acc, src_ptr, weights_ptr) CONVOLUTION1x9_STRIDE2(acc, src_ptr, weights_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x9_STRIDE1(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int8 weights_values0 = 0; \ |
| int weights_value1 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \ |
| weights_values0.s4 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \ |
| weights_values0.s5 = convert_int(*(weights_ptr + 5 * weights_stride_y)); \ |
| weights_values0.s6 = convert_int(*(weights_ptr + 6 * weights_stride_y)); \ |
| weights_values0.s7 = convert_int(*(weights_ptr + 7 * weights_stride_y)); \ |
| weights_value1 = convert_int(*(weights_ptr + 8 * weights_stride_y)); \ |
| \ |
| int8 src0 = 0; \ |
| int8 src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \ |
| src1.s0 = convert_int(*(src_ptr + 8 * weights_stride_y)); \ |
| src1.s1 = convert_int(*(src_ptr + 9 * weights_stride_y)); \ |
| src1.s2 = convert_int(*(src_ptr + 10 * weights_stride_y)); \ |
| src1.s3 = convert_int(*(src_ptr + 11 * weights_stride_y)); \ |
| src1.s4 = convert_int(*(src_ptr + 12 * weights_stride_y)); \ |
| src1.s5 = convert_int(*(src_ptr + 13 * weights_stride_y)); \ |
| src1.s6 = convert_int(*(src_ptr + 14 * weights_stride_y)); \ |
| src1.s7 = convert_int(*(src_ptr + 15 * weights_stride_y)); \ |
| \ |
| acc += src0 * (int8)weights_values0.s0; \ |
| acc += (int8)(src0.s1234, src0.s567, src1.s0) * (int8)weights_values0.s1; \ |
| acc += (int8)(src0.s234, src0.s567, src1.s01) * (int8)weights_values0.s2; \ |
| acc += (int8)(src0.s345, src0.s67, src1.s012) * (int8)weights_values0.s3; \ |
| acc += (int8)(src0.s4567, src1.s0123) * (int8)weights_values0.s4; \ |
| acc += (int8)(src0.s567, src1.s0123, src1.s4) * (int8)weights_values0.s5; \ |
| acc += (int8)(src0.s67, src1.s012, src1.s345) * (int8)weights_values0.s6; \ |
| acc += (int8)(src0.s7, src1.s0123, src1.s456) * (int8)weights_values0.s7; \ |
| acc += src1 * (int8)weights_value1; \ |
| }) |
| |
| #define CONVOLUTION1x9_STRIDE2(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int8 weights_values0 = 0; \ |
| int weights_value1 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \ |
| weights_values0.s4 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \ |
| weights_values0.s5 = convert_int(*(weights_ptr + 5 * weights_stride_y)); \ |
| weights_values0.s6 = convert_int(*(weights_ptr + 6 * weights_stride_y)); \ |
| weights_values0.s7 = convert_int(*(weights_ptr + 7 * weights_stride_y)); \ |
| weights_value1 = convert_int(*(weights_ptr + 8 * weights_stride_y)); \ |
| \ |
| int16 src0 = 0; \ |
| int8 src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \ |
| src0.s8 = convert_int(*(src_ptr + 8 * weights_stride_y)); \ |
| src0.s9 = convert_int(*(src_ptr + 9 * weights_stride_y)); \ |
| src0.sA = convert_int(*(src_ptr + 10 * weights_stride_y)); \ |
| src0.sB = convert_int(*(src_ptr + 11 * weights_stride_y)); \ |
| src0.sC = convert_int(*(src_ptr + 12 * weights_stride_y)); \ |
| src0.sD = convert_int(*(src_ptr + 13 * weights_stride_y)); \ |
| src0.sE = convert_int(*(src_ptr + 14 * weights_stride_y)); \ |
| src0.sF = convert_int(*(src_ptr + 15 * weights_stride_y)); \ |
| src1.s0 = convert_int(*(src_ptr + 16 * weights_stride_y)); \ |
| src1.s1 = convert_int(*(src_ptr + 17 * weights_stride_y)); \ |
| src1.s2 = convert_int(*(src_ptr + 18 * weights_stride_y)); \ |
| src1.s3 = convert_int(*(src_ptr + 19 * weights_stride_y)); \ |
| src1.s4 = convert_int(*(src_ptr + 20 * weights_stride_y)); \ |
| src1.s5 = convert_int(*(src_ptr + 21 * weights_stride_y)); \ |
| src1.s6 = convert_int(*(src_ptr + 22 * weights_stride_y)); \ |
| src1.s7 = convert_int(*(src_ptr + 23 * weights_stride_y)); \ |
| \ |
| acc += src0.s02468ACE * (int8)weights_values0.s0; \ |
| acc += (int8)(src0.s1357, src0.s9BDF) * (int8)weights_values0.s1; \ |
| acc += (int8)(src0.s2468, src0.sACE, src1.s0) * (int8)weights_values0.s2; \ |
| acc += (int8)(src0.s3579, src0.sBDF, src1.s1) * (int8)weights_values0.s3; \ |
| acc += (int8)(src0.s468A, src0.sCE, src1.s02) * (int8)weights_values0.s4; \ |
| acc += (int8)(src0.s579, src0.sBDF, src1.s13) * (int8)weights_values0.s5; \ |
| acc += (int8)(src0.s68A, src0.sCE, src1.s024) * (int8)weights_values0.s6; \ |
| acc += (int8)(src0.s79B, src0.sDF, src1.s135) * (int8)weights_values0.s7; \ |
| acc += (int8)(src0.s8AC, src0.sE, src1.s0246) * (int8)weights_value1; \ |
| }) |
| |
| #elif KERNEL_SIZE == 5 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x5(acc, src_ptr, weights_ptr) CONVOLUTION1x5_STRIDE1(acc, src_ptr, weights_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x5(acc, src_ptr, weights_ptr) CONVOLUTION1x5_STRIDE2(acc, src_ptr, weights_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x5_STRIDE1(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int4 weights_values0 = 0; \ |
| int weights_value1 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \ |
| weights_value1 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \ |
| \ |
| int8 src0 = 0; \ |
| int4 src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \ |
| src1.s0 = convert_int(*(src_ptr + 8 * weights_stride_y)); \ |
| src1.s1 = convert_int(*(src_ptr + 9 * weights_stride_y)); \ |
| src1.s2 = convert_int(*(src_ptr + 10 * weights_stride_y)); \ |
| src1.s3 = convert_int(*(src_ptr + 11 * weights_stride_y)); \ |
| \ |
| acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s345, src0.s67, src1.s012) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s45, src0.s67, src1.s0123) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #define CONVOLUTION1x5_STRIDE2(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int4 weights_values0 = 0; \ |
| int weights_value1 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \ |
| weights_value1 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \ |
| \ |
| int16 src0 = 0; \ |
| int4 src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \ |
| src0.s8 = convert_int(*(src_ptr + 8 * weights_stride_y)); \ |
| src0.s9 = convert_int(*(src_ptr + 9 * weights_stride_y)); \ |
| src0.sa = convert_int(*(src_ptr + 10 * weights_stride_y)); \ |
| src0.sb = convert_int(*(src_ptr + 11 * weights_stride_y)); \ |
| src0.sc = convert_int(*(src_ptr + 12 * weights_stride_y)); \ |
| src0.sd = convert_int(*(src_ptr + 13 * weights_stride_y)); \ |
| src0.se = convert_int(*(src_ptr + 14 * weights_stride_y)); \ |
| src0.sf = convert_int(*(src_ptr + 15 * weights_stride_y)); \ |
| src1.s0 = convert_int(*(src_ptr + 16 * weights_stride_y)); \ |
| src1.s1 = convert_int(*(src_ptr + 17 * weights_stride_y)); \ |
| src1.s2 = convert_int(*(src_ptr + 18 * weights_stride_y)); \ |
| src1.s3 = convert_int(*(src_ptr + 19 * weights_stride_y)); \ |
| \ |
| acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #elif KERNEL_SIZE == 3 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x3(acc, src_ptr, weights_ptr) CONVOLUTION1x3_STRIDE1(acc, src_ptr, weights_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x3(acc, src_ptr, weights_ptr) CONVOLUTION1x3_STRIDE2(acc, src_ptr, weights_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x3_STRIDE1(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int3 weights_values0 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| \ |
| int8 src0 = 0; \ |
| int2 src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \ |
| src1.s0 = convert_int(*(src_ptr + 8 * weights_stride_y)); \ |
| src1.s1 = convert_int(*(src_ptr + 9 * weights_stride_y)); \ |
| \ |
| acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| }) |
| |
| #define CONVOLUTION1x3_STRIDE2(acc, src_ptr, weights_ptr) \ |
| ({ \ |
| int3 weights_values0 = 0; \ |
| weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \ |
| weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \ |
| weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \ |
| \ |
| int16 src0 = 0; \ |
| int src1 = 0; \ |
| src0.s0 = convert_int(*(src_ptr + 0 * src_stride_y)); \ |
| src0.s1 = convert_int(*(src_ptr + 1 * src_stride_y)); \ |
| src0.s2 = convert_int(*(src_ptr + 2 * src_stride_y)); \ |
| src0.s3 = convert_int(*(src_ptr + 3 * src_stride_y)); \ |
| src0.s4 = convert_int(*(src_ptr + 4 * src_stride_y)); \ |
| src0.s5 = convert_int(*(src_ptr + 5 * src_stride_y)); \ |
| src0.s6 = convert_int(*(src_ptr + 6 * src_stride_y)); \ |
| src0.s7 = convert_int(*(src_ptr + 7 * src_stride_y)); \ |
| src0.s8 = convert_int(*(src_ptr + 8 * src_stride_y)); \ |
| src0.s9 = convert_int(*(src_ptr + 9 * src_stride_y)); \ |
| src0.sa = convert_int(*(src_ptr + 10 * src_stride_y)); \ |
| src0.sb = convert_int(*(src_ptr + 11 * src_stride_y)); \ |
| src0.sc = convert_int(*(src_ptr + 12 * src_stride_y)); \ |
| src0.sd = convert_int(*(src_ptr + 13 * src_stride_y)); \ |
| src0.se = convert_int(*(src_ptr + 14 * src_stride_y)); \ |
| src0.sf = convert_int(*(src_ptr + 15 * src_stride_y)); \ |
| src1 = convert_int(*(src_ptr + 16 * src_stride_y)); \ |
| acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2468, src0.sACE, src1) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| }) |
| |
| #elif KERNEL_SIZE == 1 |
| |
| #if STRIDE_X == 3 |
| #define INPUT_VALUE extract_input_stride3 |
| #elif STRIDE_X == 2 |
| #define INPUT_VALUE extract_input_stride2 |
| #elif STRIDE_X == 1 |
| #define INPUT_VALUE extract_input_stride1 |
| |
| #else /* STRIDE_X not equals 1, 2 or 3 */ |
| #error "Only support strides 1, 2 and 3" |
| #endif /* STRIDE_X */ |
| |
| #endif // KERNEL_SIZE == 1 |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 1. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_value, const uchar stride_y) |
| { |
| VEC_DATA_TYPE(DATA_TYPE, 8) |
| vals; |
| vals.s0 = *(input_value + 0 * stride_y); |
| vals.s1 = *(input_value + 1 * stride_y); |
| vals.s2 = *(input_value + 2 * stride_y); |
| vals.s3 = *(input_value + 3 * stride_y); |
| vals.s4 = *(input_value + 4 * stride_y); |
| vals.s5 = *(input_value + 5 * stride_y); |
| vals.s6 = *(input_value + 6 * stride_y); |
| vals.s7 = *(input_value + 7 * stride_y); |
| |
| return vals; |
| } |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 2. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_value, const uchar stride_y) |
| { |
| VEC_DATA_TYPE(DATA_TYPE, 8) |
| vals; |
| vals.s0 = *(input_value + 0 * stride_y); |
| vals.s1 = *(input_value + 2 * stride_y); |
| vals.s2 = *(input_value + 4 * stride_y); |
| vals.s3 = *(input_value + 6 * stride_y); |
| vals.s4 = *(input_value + 8 * stride_y); |
| vals.s5 = *(input_value + 10 * stride_y); |
| vals.s6 = *(input_value + 12 * stride_y); |
| vals.s7 = *(input_value + 14 * stride_y); |
| |
| return vals; |
| } |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3(__global const DATA_TYPE *input_value, const uchar stride_y) |
| { |
| VEC_DATA_TYPE(DATA_TYPE, 8) |
| vals; |
| vals.s0 = *(input_value + 0 * stride_y); |
| vals.s1 = *(input_value + 3 * stride_y); |
| vals.s2 = *(input_value + 6 * stride_y); |
| vals.s3 = *(input_value + 9 * stride_y); |
| vals.s4 = *(input_value + 12 * stride_y); |
| vals.s5 = *(input_value + 15 * stride_y); |
| vals.s6 = *(input_value + 18 * stride_y); |
| vals.s7 = *(input_value + 21 * stride_y); |
| |
| return vals; |
| } |
| |
| /** This kernel performs a direct convolution to convolve the low three dimensions. |
| * |
| * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 |
| * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
| * @note If biases are used then -DHAS_BIAS has to be passed at compile time |
| * @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234 |
| * @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4 |
| * |
| * @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 Z processed per workitem(in bytes) |
| * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| * @param[out] 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 Z processed per workitem(in bytes) |
| * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination 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_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| * @param[in] weights_step_z weights_stride_z * number of elements along Z 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] biases_ptr Pointer to the biases tensor. Supported data types: S32 |
| * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| * @param[in] input_offset Input offset quantization parameter |
| * @param[in] weight_offset Weights offset quantization parameter |
| * @param[in] output_offset Output offset quantization parameter |
| */ |
| __kernel void direct_convolution_quantized( |
| TENSOR3D_DECLARATION(src), |
| TENSOR3D_DECLARATION(dst), |
| TENSOR3D_DECLARATION(weights), |
| #ifdef HAS_BIAS |
| VECTOR_DECLARATION(biases), |
| #endif /* defined(HAS_BIAS) */ |
| unsigned int weights_stride_w, |
| int input_offset, |
| int weight_offset, |
| int output_offset) |
| { |
| Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| |
| int8 values0 = 0; |
| |
| const int id0 = get_global_id(0); |
| const int y_coord = (get_global_id(2) * STRIDE_Y) - PAD_TOP; |
| |
| __global DATA_TYPE *weights_addr = (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0); |
| __global DATA_TYPE *src_addr = (__global DATA_TYPE *)offset(&src, 0, 0) - src_stride_x * id0 + y_coord * (int)src_stride_z; |
| |
| weights_addr += id0 * weights_stride_w; |
| |
| for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) |
| { |
| #if KERNEL_SIZE == 9 |
| if(y_coord < 0) |
| { |
| const int start_z = -y_coord; |
| for(int i = start_z; i < 9; ++i) |
| { |
| CONVOLUTION1x9(values0, (src_addr + i * (int)src_stride_z), (weights_addr + i * (int)weights_stride_z)); |
| } |
| } |
| else if(y_coord > (SRC_HEIGHT - 9)) |
| { |
| // Avoid loading rows beyond the input height |
| const int end_z = SRC_HEIGHT - y_coord; |
| for(int i = 0; i < end_z; ++i) |
| { |
| CONVOLUTION1x9(values0, (src_addr + i * (int)src_stride_z), (weights_addr + i * (int)weights_stride_z)); |
| } |
| } |
| else |
| { |
| CONVOLUTION1x9(values0, src_addr, weights_addr); |
| CONVOLUTION1x9(values0, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 5 * (int)src_stride_z), (weights_addr + 5 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 6 * (int)src_stride_z), (weights_addr + 6 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 7 * (int)src_stride_z), (weights_addr + 7 * (int)weights_stride_z)); |
| CONVOLUTION1x9(values0, (src_addr + 8 * (int)src_stride_z), (weights_addr + 8 * (int)weights_stride_z)); |
| } |
| #elif KERNEL_SIZE == 5 |
| #if(PAD_TOP == 1) || (PAD_BOTTM == 1) |
| if(y_coord < 0) // special case Z = -1 doesn't exists |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| } |
| else if(get_global_id(2) == (DST_HEIGHT - 1)) |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| } |
| else |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| } |
| #elif(PAD_TOP == 2) || (PAD_BOTTM == 2) |
| if(y_coord < -1) |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| } |
| else if(y_coord == -1) |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| } |
| else if(y_coord == (SRC_HEIGHT - 3)) |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| } |
| else if(y_coord >= (SRC_HEIGHT - 4)) |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| } |
| else |
| { |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| } |
| #else /* PAD_TOP == 2 || || PAD_BOTTM == 2 */ |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z)); |
| #endif /* PAD_TOP == 1 || || PAD_BOTTM == 1 */ |
| #elif KERNEL_SIZE == 3 |
| #if(PAD_TOP > 0) || (PAD_BOTTOM > 0) |
| if(y_coord < 0) // special case Z = -1 doesn't exists |
| { |
| //skip first row and load the two next ones |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| } |
| else if(y_coord == (SRC_HEIGHT - PAD_BOTTOM - 1)) |
| { |
| // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the |
| // Z axis has no padding at all. |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| } |
| else |
| { |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| } |
| #else // PAD_TOP > 0 || PAD_BOTTOM > 0 |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z)); |
| #endif // PAD_TOP > 0 || PAD_BOTTOM > 0 |
| #elif KERNEL_SIZE == 1 |
| int weight = convert_int(*(__global DATA_TYPE *)weights_addr); |
| int8 input_value = convert_int8(INPUT_VALUE((__global DATA_TYPE *)src_addr, src_stride_y)); |
| values0 += (input_value + input_offset) * ((int8)weight + weight_offset); |
| #endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */ |
| |
| src_addr += src_stride_x; |
| weights_addr += weights_stride_x; |
| } |
| |
| #ifdef HAS_BIAS |
| Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| __global int *bias_addr = ((__global int *)(vector_offset(&biases, id0))); |
| values0 += (int8)(*bias_addr); |
| #endif /* defined(HAS_BIAS) */ |
| |
| #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_SHIFT < 0 |
| values0 = values0 + output_offset; |
| |
| VEC_DATA_TYPE(DATA_TYPE, 8) |
| values = CONVERT_SAT(values0, DATA_TYPE); |
| *(dst.ptr + 0 * dst_stride_y) = values.s0; |
| *(dst.ptr + 1 * dst_stride_y) = values.s1; |
| *(dst.ptr + 2 * dst_stride_y) = values.s2; |
| *(dst.ptr + 3 * dst_stride_y) = values.s3; |
| *(dst.ptr + 4 * dst_stride_y) = values.s4; |
| *(dst.ptr + 5 * dst_stride_y) = values.s5; |
| *(dst.ptr + 6 * dst_stride_y) = values.s6; |
| *(dst.ptr + 7 * dst_stride_y) = values.s7; |
| } |
| |
| #else // defined(DATA_LAYOUT_NHWC) |
| |
| #if KERNEL_SIZE == 9 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \ |
| int weights_value1 = convert_int(*(weights_row_ptr + 8)); \ |
| int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ |
| acc += (src0.lo + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1234, src0.s5678) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2345, src0.s6789) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s3456, src0.s789A) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s4567, src0.s89AB) + input_offset) * ((int8)weights_values0.s4 + weight_offset); \ |
| acc += ((int8)(src0.s5678, src0.s9ABC) + input_offset) * ((int8)weights_values0.s5 + weight_offset); \ |
| acc += ((int8)(src0.s6789, src0.sABCD) + input_offset) * ((int8)weights_values0.s6 + weight_offset); \ |
| acc += ((int8)(src0.s789A, src0.sBCDE) + input_offset) * ((int8)weights_values0.s7 + weight_offset); \ |
| acc += ((int8)(src0.s89AB, src0.sCDEF) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #define CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \ |
| int weights_value1 = convert_int(*(weights_row_ptr + 8)); \ |
| int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ |
| int8 src1 = convert_int8(vload8(0, src_row_ptr + 16)); \ |
| acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s468A, src0.sCE, src1.s02) + input_offset) * ((int8)weights_values0.s4 + weight_offset); \ |
| acc += ((int8)(src0.s579B, src0.sDF, src1.s13) + input_offset) * ((int8)weights_values0.s5 + weight_offset); \ |
| acc += ((int8)(src0.s68AC, src0.sE, src1.s024) + input_offset) * ((int8)weights_values0.s6 + weight_offset); \ |
| acc += ((int8)(src0.s79BD, src0.sF, src1.s135) + input_offset) * ((int8)weights_values0.s7 + weight_offset); \ |
| acc += ((int8)(src0.s8ACE, src1.s0246) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #elif KERNEL_SIZE == 5 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \ |
| int weights_value1 = convert_int(*(weights_row_ptr + 4)); \ |
| int8 src0 = convert_int8(vload8(0, src_row_ptr)); \ |
| int4 src1 = convert_int4(vload4(0, src_row_ptr + 8)); \ |
| acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s345, src0.s67, src1.s012) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s45, src0.s67, src1.s0123) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \ |
| int weights_value1 = convert_int(*(weights_row_ptr + 4)); \ |
| int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ |
| int4 src1 = convert_int4(vload4(0, src_row_ptr + 16)); \ |
| acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \ |
| acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + input_offset) * ((int8)weights_value1 + weight_offset); \ |
| }) |
| |
| #elif KERNEL_SIZE == 3 |
| |
| #if STRIDE_X == 1 |
| #define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) |
| #elif STRIDE_X == 2 |
| #define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) |
| #else /* STRIDE_X not equals 1 or 2 */ |
| #error "STRIDE_X larger than 2 is not supported" |
| #endif /* STRIDE_X */ |
| |
| #define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \ |
| int8 src0 = convert_int8(vload8(0, src_row_ptr)); \ |
| int2 src1 = convert_int2(vload2(0, src_row_ptr + 8)); \ |
| acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| }) |
| |
| #define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ |
| ({ \ |
| int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \ |
| int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ |
| int src1 = convert_int(*(src_row_ptr + 16)); \ |
| acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \ |
| acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \ |
| acc += ((int8)(src0.s2468, src0.sACE, src1) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \ |
| }) |
| |
| #elif KERNEL_SIZE == 1 |
| |
| #if STRIDE_X == 3 |
| #define INPUT_VALUE extract_input_stride3 |
| #elif STRIDE_X == 2 |
| #define INPUT_VALUE extract_input_stride2 |
| #elif STRIDE_X == 1 |
| #define INPUT_VALUE extract_input_stride1 |
| |
| #else /* STRIDE_X not equals 1, 2 or 3 */ |
| #error "Only support strides 1, 2 and 3" |
| #endif /* STRIDE_X */ |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 1. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_value) |
| { |
| return vload8(0, input_value); |
| } |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 2. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_value) |
| { |
| VEC_DATA_TYPE(DATA_TYPE, 16) |
| temp = vload16(0, input_value); |
| return temp.s02468ace; |
| } |
| |
| /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. |
| * |
| * @param[in] input_value Pointer to the first value. |
| * |
| * @return extracted input values. |
| */ |
| inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3(__global const DATA_TYPE *input_value) |
| { |
| VEC_DATA_TYPE(DATA_TYPE, 16) |
| temp1 = vload16(0, input_value); |
| VEC_DATA_TYPE(DATA_TYPE, 16) |
| temp2 = vload16(0, input_value + 12); |
| return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369); |
| } |
| |
| #else /* KERNEL_SIZE not equals 1, 3 , 5, 9 */ |
| #error "Only kernel sizes 1, 3, 5 and 9 are supported" |
| #endif /* KERNEL_SIZE */ |
| |
| /** This kernel performs a direct convolution to convolve the low three dimensions. |
| * |
| * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 |
| * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
| * @note If biases are used then -DHAS_BIAS has to be passed at compile time |
| * @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234 |
| * @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4 |
| * |
| * @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 Z processed per workitem(in bytes) |
| * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| * @param[out] 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 Z processed per workitem(in bytes) |
| * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination 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_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| * @param[in] weights_step_z weights_stride_z * number of elements along Z 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] biases_ptr Pointer to the biases tensor. Supported data types: S32 |
| * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| * @param[in] input_offset Input offset quantization parameter |
| * @param[in] weight_offset Weights offset quantization parameter |
| * @param[in] output_offset Output offset quantization parameter |
| */ |
| __kernel void direct_convolution_quantized( |
| TENSOR3D_DECLARATION(src), |
| TENSOR3D_DECLARATION(dst), |
| TENSOR3D_DECLARATION(weights), |
| #ifdef HAS_BIAS |
| VECTOR_DECLARATION(biases), |
| #endif /* defined(HAS_BIAS) */ |
| unsigned int weights_stride_w, |
| int input_offset, |
| int weight_offset, |
| int output_offset) |
| { |
| Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| |
| int8 values0 = 0; |
| |
| __global DATA_TYPE *weights_addr = (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0); |
| __global DATA_TYPE *src_addr = (__global DATA_TYPE *)offset(&src, 0, 0); |
| |
| const int kernel_index = get_global_id(2); |
| weights_addr += kernel_index * weights_stride_w; |
| |
| for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) |
| { |
| #if KERNEL_SIZE == 9 |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 5 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 6 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 6 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 7 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 7 * weights_stride_y)); |
| CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 8 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 8 * weights_stride_y)); |
| #elif KERNEL_SIZE == 5 |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y)); |
| CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y)); |
| #elif KERNEL_SIZE == 3 |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); |
| CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); |
| #elif KERNEL_SIZE == 1 |
| int weight = convert_int(*(__global DATA_TYPE *)weights_addr); |
| int8 input_value = convert_int8(INPUT_VALUE((__global DATA_TYPE *)src_addr)); |
| values0 += (input_value + input_offset) * ((int8)weight + weight_offset); |
| #endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */ |
| |
| src_addr += src_stride_z; |
| weights_addr += weights_stride_z; |
| } |
| |
| #ifdef HAS_BIAS |
| Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| __global int *bias_addr = ((__global int *)(vector_offset(&biases, kernel_index))); |
| values0 += (int8)(*bias_addr); |
| #endif /* defined(HAS_BIAS) */ |
| |
| #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_SHIFT < 0 |
| values0 = values0 + output_offset; |
| |
| vstore8(CONVERT_SAT(values0, DATA_TYPE), 0, (__global DATA_TYPE *)dst.ptr); |
| } |
| |
| #endif // defined(DATA_LAYOUT_NHWC) |
| #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) |
| |
| )" |