Google Git
Sign in
android / platform / external / ComputeLibrary / c05589415 / . / build / android-arm64v8a / src / core / CL / cl_kernels / pooling_layer.clembed
blob: 5b0adfa4b9222415d2b9d514dfe3131eec972f6e [file] [log] [blame]
R"(
/*
* Copyright (c) 2017-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
/*
* Copyright (c) 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
/*
* Copyright (c) 2019-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef ARM_COMPUTE_REPEAT_H
#define ARM_COMPUTE_REPEAT_H
/*
* Copyright (c) 2016-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef ARM_COMPUTE_HELPER_H
#define ARM_COMPUTE_HELPER_H
/*
* Copyright (c) 2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
/** Store the 0 to (n-1)th rows of the given variables
* @name STORE_ROW_n
*
* @param[in] N0 The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
#define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
#define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
#define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
#define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
#define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
#define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
#define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
#define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
#define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
#define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
#define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
#define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
#define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
#define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
#define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
/** @} */ // end of groupd STORE_ROW_n
/** Convert and store the 0th to (n-1)th rows of the given variables
* @name CONVERT_STORE_ROW_n
*
* @param[in] N0 The size of the vectors
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
#define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
#define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
#define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
#define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
#define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
#define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
#define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
#define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
#define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
#define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
#define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
#define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
#define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
#define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
#define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE(N0) \
(CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
/** @} */ // end of groupd CONVERT_STORE_ROW_n
/** Store a block of the given size M0xN0
* @name STORE_BLOCK
*
* Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store
* @param[in] N0 The size of each vector
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
/** @} */ // end of group STORE_BLOCK
/** Convert and store a block of the given size M0xN0
* @name CONVERT_STORE_BLOCK
*
* Supported cases are M0=1,2,3,...,16 and N0=2,3,4,8,16.
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store
* @param[in] N0 The size of each vector
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
/** @} */ // end of group CONVERT_STORE_BLOCK
/** Partially store the 0 to (n-1)th rows of the given variables
* @name STORE_ROW_PARTIAL_n
* Within each row, store the lower @p STORE_N0 elements of vectors of width @p N0
*
* @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* @param[in] N0 The width of the passed in vector. Supported: 1, 2, 3, 4, 8, 16
* @param[in] STORE_N0 The **lower** size of the vectors to store. Supported: [1-16 and <= @p N0
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0));
#define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1));
#define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2));
#define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3));
#define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4));
#define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5));
#define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6));
#define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7));
#define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8));
#define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9));
#define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A));
#define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B));
#define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C));
#define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D));
#define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E));
#define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \
VSTORE_PARTIAL(N0, STORE_N0) \
(BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F));
/** @} */ // end of groupd STORE_ROW_PARTIAL_n
/** Partially store a block of the given size STORE_M0xSTORE_N0
* @name STORE_BLOCK_PARTIAL
*
* @note The vector width @p N0 is also required for correct partial storing behaviour.
* @note in case @p STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* The data to store is expected to have consecutive names for each row.
* E.g., for STORE_M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for STORE_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] STORE_M0 The number of rows to store. Supported: 1-16
* @param[in] STORE_N0 The lower number of elements of vectors to store. Supported: 1-16 and <= @p N0
* @param[in] N0 The size of each vector. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @{
*/
#define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
/** Store a block that can be partial in both x and y dimensions
*
* @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store, for non-partial blocks. Supported: 1-16
* @param[in] N0 The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
* @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
* @param[in] PARTIAL_COND_Y Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
* @param[in] PARTIAL_COND_X Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
*/
#define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y)) \
{ \
STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
} \
else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X)) \
{ \
STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
} \
else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X)) \
{ \
STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
} \
else \
{ \
STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
}
/** Store a block that can only be partial in x but not y.
*
* @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store, for non-partial blocks. Supported: 1-16
* @param[in] N0 The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported range: [1, @p N0)
* @param[in] PARTIAL_COND_X Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
*/
#define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \
if(!(PARTIAL_COND_X)) \
{ \
STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
} \
else \
{ \
STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
}
/** Store a block that can only be partial in y but not x.
*
* @note in case @p N0 or @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store, for non-partial blocks. Supported: 1-16
* @param[in] N0 The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported range: [1, @p M0)
* @param[in] PARTIAL_COND_Y Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
*/
#define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \
if(!(PARTIAL_COND_Y)) \
{ \
STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
} \
else \
{ \
STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \
}
/** @} */ // end of group STORE_BLOCK_PARTIAL
#if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
/** Boundary-aware GEMM block store
* @name STORE_BLOCK_BOUNDARY_AWARE
* This macro assumes the following schemes to achieve boundary-awareness:
* - Overlapping load in Y axis from lhs tensor. This implies lhs has no padding along y dim.
* - Non-Overlapping(normal) load from rhs tensor. This imples rhs can have paddings.
* - Overlapping load in Y axis from bias tensor. This implies rhs has no padding along y dim.
* The macro then ensures that the dst tensor can be stored without any paddings in both x and y dim.
*
* In the y dimension, we place the partial blocks **at the beginning** while in the x dimension, we place the partial
* blocks **at the end**.
* Say, the dst tensor is of shape MxN and we have M0 and N0 as the block size, this is how we define "partial blocks"/
* "boundary block" (we use the 2 terms "partial blocks" and "boundary blocks" interchangeably) and its various parameters:
*
* *--x--> x == 0 x == 1
* | |<------------------------------N-------------------------->|
* y |<--------------N0------------->|<----PARTIAL_STORE_N0----->|
* | -------------#############################################################
* * | | |...............................|...........................|
* y == 0 | PAR_..._M0 |......Boundary block in y......|.Boundary block in x and y.|
* | | |...............................|...........................|
* M --#############################################################
* | | | |...........................|
* y == 1 | M0 | Non-boundary block |....Boundary block in x....|
* | | | |...........................|
* |------------#############################################################
*
* Then @p PARTIAL_STORE_M0 = M % M0 and @p PARTIAL_STORE_N0 = N % N0
*
* @note in cases @p PARTIAL_STORE_N0 != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* It automatically detects if a giving M,N,M0,N0 combination can yield partial blocks in either X and Y dimension,
* and select corresponding store methods such that the boundary detection logic is only added when needed.
*
* The data to store is expected to have consecutive names for each row.
* E.g., for M0=3 and basename=c, the expected names are c0, c1 and c2.
* The Z offset is expected to have consecutive names.
* E.g., for M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2.
*
* @param[in] M0 The number of rows to store, for non-partial blocks. Supported: 1-16
* @param[in] N0 The size of each vector, for non-partial blocks. Supported: 1, 2, 3, 4, 8, 16
* @param[in] DATA_TYPE The data type of the vectors
* @param[in] BASENAME The basename of the variables
* @param[in] PTR The base pointer
* @param[in] STRIDE_Y The stride value in y-axis direction
* @param[in] Z The offset in z-axis direction
* @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
* @param[in] PARTIAL_STORE_N0 The partial size in x, for partial blocks. Supported: [0, @p N0)
* @param[in] PARTIAL_COND_Y Condition on the y axis to perform the partial store Y. True to use PARTIAL_STORE_M0 rather than M0.
* @param[in] PARTIAL_COND_X Condition on the x axis to perform the partial store X. True to use PARTIAL_STORE_N0 rather than N0.
* @{
*/
#if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
// Case1: No partial blocks in either x or y
#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z)
#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0
// Case2: Partial blocks in y
#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y)
#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0
// Case3: Partial blocks in x
#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X)
#else // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
// Case4: Partial blocks in both x and y
#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \
STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X)
#endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0
#endif // defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)
/** @} */ // end of group STORE_BLOCK_BOUNDARY_AWARE
#if defined(PARTIAL_STORE_M0)
/** Compute the start m0 row (LHS, BIAS and DST) in a boundary-aware way so as to avoid padding
* @name COMPUTE_M0_START_ROW
* If there're any partial blocks in y dimension, they are placed at the beginning of the rows.
* This shift amount is added to all rows such that the partial block (at the beginning) overlaps with the subsequent
* blocks in the y dimension to avoid any padding.
* EG: M0=4, PARTIAL_STORE_M0=1:
* | Non-overlapping | +M0_ROW_SHIFT (Overlapping)
* block 0 (partial)| start row = 0 | start row = 0
* block 1 (full) | start row = 4 | start row = 1
* block 2 (full) | start row = 8 | start row = 5
*
* @param[in] y Global id of current block in y.
* @param[in] M0 The number of rows to store, for non-partial blocks. Supported: 1-16
* @param[in] PARTIAL_STORE_M0 The partial size in y, for partial blocks. Supported: [0, @p M0)
* @{
*/
#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0))))
#else // defined(PARTIAL_STORE_M0)
#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \
((uint)(y * M0))
#endif // defined(PARTIAL_STORE_M0)
/** @} */ // end of group COMPUTE_M0_START_ROW
/** Store a vector that can only be partial in x.
*
* @note in case @p vec_size or @p leftover != 1, 2, 3, 4, 8, 16, extra vstore(s) will be invoked, thus incurring small performance penalty.
*
* The data to store is expected to end in a 0.
* E.g., for basename=c, the expected name is c0.
*
* @param[in] basename The name of the variable without trailing 0
* @param[in] data_type The data type of the vector
* @param[in] ptr The base pointer
* @param[in] vec_size The vector size if cond = false. Supported: 1, 2, 3, 4, 8, 16
* @param[in] leftover The vector size if cond = true. Supported range: [1, @p vec_size0)
* @param[in] cond Condition to select either vec_size0 or vec_size1
* @{
*/
#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \
STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond)
/** @} */ // end of group STORE_VECTOR_SELECT
#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable
#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
#pragma OPENCL EXTENSION cl_arm_printf : enable
#endif // defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf)
#define GPU_ARCH_MIDGARD 0x100
#define GPU_ARCH_BIFROST 0x200
/** Concatenate two inputs.
*
* @param[in] a The first input to be concatenated
* @param[in] b The second input to be concatenated
*
* @return The concatenated output
*/
#define CONCAT(a, b) a##b
/** Expand the given vector
*
* @param[in] x The vector to be expanded
*
* @return The expanded output
*/
#define EXPAND(x) x
/** Clamp the given value between an upper and lower bound.
*
* @param[in] x The value to be clamped
* @param[in] min_val The lower bound
* @param[in] max_val The upper bound
*
* @return The clamped value.
*/
#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
/** REVn reverses the given vector whose size is n.
* @name REVn
*
* @param[in] x The vector to be reversed
*
* @return The reversed vector
* @{
*/
#define REV1(x) ((x))
#define REV2(x) ((x).s10)
#define REV3(x) ((x).s210)
#define REV4(x) ((x).s3210)
#define REV8(x) ((x).s76543210)
#define REV16(x) ((x).sFEDCBA9876543210)
/** @} */ // end of group REVn
/** Reverse the given vector.
* @name REVERSE
*
* @param[in] x The vector to be reversed
* @param[in] s The size of the vector
*
* @return The reversed vector
* @{
*/
#define REVERSE_STR(x, s) REV##s((x))
#define REVERSE(x, s) REVERSE_STR(x, s)
/** @} */ // end of group REVERSE
/** Circular-right-shift (rotate-right) the vector of size s by the amount of n.
* @name ROTs_n
*
* @param[in] x The vector to be shifted
*
* @return The shifted vector
* @{
*/
#define ROT1_0(x) ((x))
#define ROT2_0(x) ((x))
#define ROT2_1(x) ((x).s10)
#define ROT3_0(x) ((x))
#define ROT3_1(x) ((x).s201)
#define ROT3_2(x) ((x).s120)
#define ROT4_0(x) ((x))
#define ROT4_1(x) ((x).s3012)
#define ROT4_2(x) ((x).s2301)
#define ROT4_3(x) ((x).s1230)
#define ROT8_0(x) ((x))
#define ROT8_1(x) ((x).s70123456)
#define ROT8_2(x) ((x).s67012345)
#define ROT8_3(x) ((x).s56701234)
#define ROT8_4(x) ((x).s45670123)
#define ROT8_5(x) ((x).s34567012)
#define ROT8_6(x) ((x).s23456701)
#define ROT8_7(x) ((x).s12345670)
#define ROT16_0(x) ((x))
#define ROT16_1(x) ((x).sF0123456789ABCDE)
#define ROT16_2(x) ((x).sEF0123456789ABCD)
#define ROT16_3(x) ((x).sDEF0123456789ABC)
#define ROT16_4(x) ((x).sCDEF0123456789AB)
#define ROT16_5(x) ((x).sBCDEF0123456789A)
#define ROT16_6(x) ((x).sABCDEF0123456789)
#define ROT16_7(x) ((x).s9ABCDEF012345678)
#define ROT16_8(x) ((x).s89ABCDEF01234567)
#define ROT16_9(x) ((x).s789ABCDEF0123456)
#define ROT16_10(x) ((x).s6789ABCDEF012345)
#define ROT16_11(x) ((x).s56789ABCDEF01234)
#define ROT16_12(x) ((x).s456789ABCDEF0123)
#define ROT16_13(x) ((x).s3456789ABCDEF012)
#define ROT16_14(x) ((x).s23456789ABCDEF01)
#define ROT16_15(x) ((x).s123456789ABCDEF0)
/** @} */ // end of group ROTs_n
/** Circular-right-shift (rotate-right) the given vector by the given amount.
* @name ROTATE
*
* @param[in] x The vector to be shifted
* @param[in] s The size of the vector
* @param[in] n The amount to be shifted
*
* @return The shifted vector
* @{
*/
#define ROTATE_STR(x, s, n) ROT##s##_##n(x)
#define ROTATE(x, s, n) ROTATE_STR(x, s, n)
/** @} */ // end of group ROTATE
/** Creates a vector of size n filled with offset values corresponding to the location of each element.
* @name V_OFFSn
*
* @param[in] dt The data type of the output vector
*
* @return The vector filled with offset values
* @{
*/
#define V_OFFS1(dt) (dt##1)(0)
#define V_OFFS2(dt) (dt##2)(0, 1)
#define V_OFFS3(dt) (dt##3)(0, 1, 2)
#define V_OFFS4(dt) (dt##4)(0, 1, 2, 3)
#define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7)
#define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
/** @} */ // end of group V_OFFSn
/** Create a vector filled with offset values corresponding to the location of each element.
* @name VEC_OFFS
*
* @param[in] dt The data type of the output vector
* @param[in] s The size of the output vector
*
* @return The vector filled with offset values
* @{
*/
#define VEC_OFFS_STR(dt, s) V_OFFS##s(dt)
#define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s)
/** @} */ // end of group VEC_OFFS
#define VLOAD_STR(size) vload##size
#define VLOAD(size) VLOAD_STR(size)
#define PIXEL_UNIT4 1
#define PIXEL_UNIT8 2
#define PIXEL_UNIT16 4
/** Utility macro to convert a vector size in pixel unit.
*
* @name CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT
*
* @param[in] vec_size Vector size. Only 4,8 and 16 is supported
*
* @return The pixel unit (number of pixels)
* @{
*/
#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size
#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size)
/** @} */ // end of group CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT
#define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord)));
#define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)));
#define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord)));
#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
#define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord)));
#define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)));
#define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord)));
#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16)
/** Utility macro to read a 2D OpenCL image object.
*
* @note Coordinates are not normalized
*
* @param[in] data_type Data type
* @param[in] n0 Number of pixel to read. Only 1,2 and 4 is supported
* @param[in] img OpenCL image object
* @param[in] x_coord The x coordinate for the top-left pixel
* @param[in] y_coord The y coordinate for the top-left pixel
*
* @return Pixels from the 2D OpenCL image object
* @{
*/
#define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord)
#define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord)
#define VSTORE_STR(size) vstore##size
#define VSTORE(size) VSTORE_STR(size)
#define float1 float
#define half1 half
#define char1 char
#define uchar1 uchar
#define short1 short
#define ushort1 ushort
#define int1 int
#define uint1 uint
#define long1 long
#define ulong1 ulong
#define double1 double
#define vload1(OFFSET, PTR) *(OFFSET + PTR)
#define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA
/** Extended partial vstore that correctly handles scalar values as well.
* Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
* @name VSTORE_PARTIAL
*
* @note With this macro, the passed data can be both a vector and a scalar
* @note @p store_size needs to be <= @p size
* eg 1: Valid
* VSTORE_PARTIAL(16, 15) ...;
* eg 2: Invalid
* VSTORE_PARTIAL(4, 7) ...;
*
* @param[in] size The width of @p DATA. Supported values: 1(scalar), 2, 3, 4, 8, 16
* @param[in] store_size The number of lower elements to store. Supported values: 1-16, but has to be <= @p size
* @{
*/
#define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size
#define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size)
#define NO_STORE(data, offs, ptr) \
{ \
}
// Size == 1 (scalar)
#define vstore_partial_1_0 NO_STORE
#define vstore_partial_1_1 vstore1
#define vstore_partial_1_2 NO_STORE
#define vstore_partial_1_3 NO_STORE
#define vstore_partial_1_4 NO_STORE
#define vstore_partial_1_5 NO_STORE
#define vstore_partial_1_6 NO_STORE
#define vstore_partial_1_7 NO_STORE
#define vstore_partial_1_8 NO_STORE
#define vstore_partial_1_9 NO_STORE
#define vstore_partial_1_10 NO_STORE
#define vstore_partial_1_11 NO_STORE
#define vstore_partial_1_12 NO_STORE
#define vstore_partial_1_13 NO_STORE
#define vstore_partial_1_14 NO_STORE
#define vstore_partial_1_15 NO_STORE
#define vstore_partial_1_16 NO_STORE
// Size == 2
#define vstore_partial_2_0 NO_STORE
#define vstore_partial_2_1 vstore_partial_1
#define vstore_partial_2_2 vstore_partial_2
#define vstore_partial_2_3 NO_STORE
#define vstore_partial_2_4 NO_STORE
#define vstore_partial_2_5 NO_STORE
#define vstore_partial_2_6 NO_STORE
#define vstore_partial_2_7 NO_STORE
#define vstore_partial_2_8 NO_STORE
#define vstore_partial_2_9 NO_STORE
#define vstore_partial_2_10 NO_STORE
#define vstore_partial_2_11 NO_STORE
#define vstore_partial_2_12 NO_STORE
#define vstore_partial_2_13 NO_STORE
#define vstore_partial_2_14 NO_STORE
#define vstore_partial_2_15 NO_STORE
#define vstore_partial_2_16 NO_STORE
// Size == 3
#define vstore_partial_3_0 NO_STORE
#define vstore_partial_3_1 vstore_partial_1
#define vstore_partial_3_2 vstore_partial_2
#define vstore_partial_3_3 vstore_partial_3
#define vstore_partial_3_4 NO_STORE
#define vstore_partial_3_5 NO_STORE
#define vstore_partial_3_6 NO_STORE
#define vstore_partial_3_7 NO_STORE
#define vstore_partial_3_8 NO_STORE
#define vstore_partial_3_9 NO_STORE
#define vstore_partial_3_10 NO_STORE
#define vstore_partial_3_11 NO_STORE
#define vstore_partial_3_12 NO_STORE
#define vstore_partial_3_13 NO_STORE
#define vstore_partial_3_14 NO_STORE
#define vstore_partial_3_15 NO_STORE
#define vstore_partial_3_16 NO_STORE
// Size == 4
#define vstore_partial_4_0 NO_STORE
#define vstore_partial_4_1 vstore_partial_1
#define vstore_partial_4_2 vstore_partial_2
#define vstore_partial_4_3 vstore_partial_3
#define vstore_partial_4_4 vstore_partial_4
#define vstore_partial_4_5 NO_STORE
#define vstore_partial_4_6 NO_STORE
#define vstore_partial_4_7 NO_STORE
#define vstore_partial_4_8 NO_STORE
#define vstore_partial_4_9 NO_STORE
#define vstore_partial_4_10 NO_STORE
#define vstore_partial_4_11 NO_STORE
#define vstore_partial_4_12 NO_STORE
#define vstore_partial_4_13 NO_STORE
#define vstore_partial_4_14 NO_STORE
#define vstore_partial_4_15 NO_STORE
#define vstore_partial_4_16 NO_STORE
// Size == 8
#define vstore_partial_8_0 NO_STORE
#define vstore_partial_8_1 vstore_partial_1
#define vstore_partial_8_2 vstore_partial_2
#define vstore_partial_8_3 vstore_partial_3
#define vstore_partial_8_4 vstore_partial_4
#define vstore_partial_8_5 vstore_partial_5
#define vstore_partial_8_6 vstore_partial_6
#define vstore_partial_8_7 vstore_partial_7
#define vstore_partial_8_8 vstore_partial_8
#define vstore_partial_8_9 NO_STORE
#define vstore_partial_8_10 NO_STORE
#define vstore_partial_8_11 NO_STORE
#define vstore_partial_8_12 NO_STORE
#define vstore_partial_8_13 NO_STORE
#define vstore_partial_8_14 NO_STORE
#define vstore_partial_8_15 NO_STORE
#define vstore_partial_8_16 NO_STORE
// Size == 16
#define vstore_partial_16_0 NO_STORE
#define vstore_partial_16_1 vstore_partial_1
#define vstore_partial_16_2 vstore_partial_2
#define vstore_partial_16_3 vstore_partial_3
#define vstore_partial_16_4 vstore_partial_4
#define vstore_partial_16_5 vstore_partial_5
#define vstore_partial_16_6 vstore_partial_6
#define vstore_partial_16_7 vstore_partial_7
#define vstore_partial_16_8 vstore_partial_8
#define vstore_partial_16_9 vstore_partial_9
#define vstore_partial_16_10 vstore_partial_10
#define vstore_partial_16_11 vstore_partial_11
#define vstore_partial_16_12 vstore_partial_12
#define vstore_partial_16_13 vstore_partial_13
#define vstore_partial_16_14 vstore_partial_14
#define vstore_partial_16_15 vstore_partial_15
#define vstore_partial_16_16 vstore_partial_16
/** Partial vstore. Store the **lower** 0 to (n-1)th elements of the given vector while minimising the amount of vstore ops
* @name vstore_partial_n
*
* @note @p DATA needs to be a vector not a scalar
* @note n needs to be <= the vector width of the input variable @p DATA
* eg 1: Valid
* vstore_partial_15(var:float16, 0, 0xabcd);
* eg 2: Invalid
* vstore_partial_7(var:float4, 0, 0xabcd);
*
* @note in cases n == 1, 2, 3, 4, 8, 16, no extra vstore is invoked, thus there's no performance penalty.
*
* @param[in] DATA The name of the variable
* @param[in] OFFSET Offset in n
* @param[in] PTR The base pointer
* @{
*/
#define vstore_partial_1(DATA, OFFSET, PTR) \
vstore1(DATA.s0, OFFSET, PTR);
#define vstore_partial_2(DATA, OFFSET, PTR) \
vstore2(DATA.s01, OFFSET, PTR);
#define vstore_partial_3(DATA, OFFSET, PTR) \
vstore3(DATA.s012, OFFSET, PTR);
#define vstore_partial_4(DATA, OFFSET, PTR) \
vstore4(DATA.s0123, OFFSET, PTR);
#define vstore_partial_5(DATA, OFFSET, PTR) \
vstore_partial_4(DATA.s0123, OFFSET, PTR); \
vstore1(DATA.s4, OFFSET, PTR + 4);
#define vstore_partial_6(DATA, OFFSET, PTR) \
vstore_partial_4(DATA.s0123, OFFSET, PTR); \
vstore_partial_2(DATA.s45, OFFSET, PTR + 4);
#define vstore_partial_7(DATA, OFFSET, PTR) \
vstore_partial_4(DATA.s0123, OFFSET, PTR); \
vstore_partial_3(DATA.s456, OFFSET, PTR + 4);
#define vstore_partial_8(DATA, OFFSET, PTR) \
vstore8(DATA.s01234567, OFFSET, PTR);
#define vstore_partial_9(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore1(DATA.s8, OFFSET, PTR + 8);
#define vstore_partial_10(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_2(DATA.s89, OFFSET, PTR + 8);
#define vstore_partial_11(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_3(DATA.s89a, OFFSET, PTR + 8);
#define vstore_partial_12(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8);
#define vstore_partial_13(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8);
#define vstore_partial_14(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8);
#define vstore_partial_15(DATA, OFFSET, PTR) \
vstore_partial_8(DATA.s01234567, OFFSET, PTR); \
vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8);
#define vstore_partial_16(DATA, OFFSET, PTR) \
vstore16(DATA, OFFSET, PTR);
/** @} */ // end of groupd vstore_partial_n
/** @} */ // end of groupd VSTORE_PARTIAL
// Convert built-in functions with _sat modifier are not supported in floating point so we create defines
// without _sat to overcome this issue
#define convert_float_sat convert_float
#define convert_float1_sat convert_float
#define convert_float2_sat convert_float2
#define convert_float3_sat convert_float3
#define convert_float4_sat convert_float4
#define convert_float8_sat convert_float8
#define convert_float16_sat convert_float16
#define convert_half_sat convert_float
#define convert_half1_sat convert_half
#define convert_half2_sat convert_half2
#define convert_half3_sat convert_half3
#define convert_half4_sat convert_half4
#define convert_half8_sat convert_half8
#define convert_half16_sat convert_half16
#define convert_float1 convert_float
#define convert_half1 convert_half
#define convert_char1 convert_char
#define convert_uchar1 convert_uchar
#define convert_short1 convert_short
#define convert_ushort1 convert_ushort
#define convert_int1 convert_int
#define convert_uint1 convert_uint
#define convert_long1 convert_long
#define convert_ulong1 convert_ulong
#define convert_double1 convert_double
#define convert_char1_sat convert_char_sat
#define convert_uchar1_sat convert_uchar_sat
#define convert_short1_sat convert_short_sat
#define convert_ushort1_sat convert_ushort_sat
#define convert_int1_sat convert_int_sat
#define convert_uint1_sat convert_uint_sat
#define convert_long1_sat convert_long_sat
#define convert_ulong1_sat convert_ulong_sat
#define convert_double1_sat convert_double_sat
#define VEC_DATA_TYPE_STR(type, size) type##size
#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
#define CONVERT_STR(x, type) (convert_##type((x)))
#define CONVERT(x, type) CONVERT_STR(x, type)
#define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x)))
#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)
#define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x)))
#define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round)
#define select_vec_dt_uchar(size) uchar##size
#define select_vec_dt_char(size) char##size
#define select_vec_dt_ushort(size) ushort##size
#define select_vec_dt_short(size) short##size
#define select_vec_dt_half(size) short##size
#define select_vec_dt_uint(size) uint##size
#define select_vec_dt_int(size) int##size
#define select_vec_dt_float(size) int##size
#define select_vec_dt_ulong(size) ulong##size
#define select_vec_dt_long(size) long##size
#define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size)
#define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size)
#define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1)
#define sum_reduce_1(x) (x)
#define sum_reduce_2(x) ((x).s0) + ((x).s1)
#define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2)
#define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23)
#define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567)
#define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF)
#define SUM_REDUCE_STR(x, size) sum_reduce_##size(x)
#define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size)
#define max_reduce_1(x) (x)
#define max_reduce_2(x) max(((x).s0), ((x).s1))
#define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2))
#define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23))
#define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567))
#define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF))
#define MAX_REDUCE_STR(x, size) max_reduce_##size(x)
#define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size)
#define VECTOR_DECLARATION(name) \
__global uchar *name##_ptr, \
uint name##_stride_x, \
uint name##_step_x, \
uint name##_offset_first_element_in_bytes
#define IMAGE_DECLARATION(name) \
__global uchar *name##_ptr, \
uint name##_stride_x, \
uint name##_step_x, \
uint name##_stride_y, \
uint name##_step_y, \
uint name##_offset_first_element_in_bytes
#define TENSOR3D_DECLARATION(name) \
__global uchar *name##_ptr, \
uint name##_stride_x, \
uint name##_step_x, \
uint name##_stride_y, \
uint name##_step_y, \
uint name##_stride_z, \
uint name##_step_z, \
uint name##_offset_first_element_in_bytes
#define TENSOR4D_DECLARATION(name) \
__global uchar *name##_ptr, \
uint name##_stride_x, \
uint name##_step_x, \
uint name##_stride_y, \
uint name##_step_y, \
uint name##_stride_z, \
uint name##_step_z, \
uint name##_stride_w, \
uint name##_step_w, \
uint name##_offset_first_element_in_bytes
#define CONVERT_TO_VECTOR_STRUCT(name) \
update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)
#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0)
#define CONVERT_TO_IMAGE_STRUCT(name) \
update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)
#define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0)
#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)
#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z)
#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)
#define CONVERT_TO_TENSOR3D_STRUCT(name) \
update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
name##_stride_z, name##_step_z)
#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0)
#define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size) \
update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size)
#define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \
update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size)
#define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name) \
tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
name##_stride_z, name##_step_z)
/** Structure to hold Vector information */
typedef struct Vector
{
__global uchar *ptr; /**< Pointer to the starting postion of the buffer */
int offset_first_element_in_bytes; /**< The offset of the first element in the source image */
int stride_x; /**< Stride of the image in X dimension (in bytes) */
} Vector;
/** Structure to hold Image information */
typedef struct Image
{
__global uchar *ptr; /**< Pointer to the starting postion of the buffer */
int offset_first_element_in_bytes; /**< The offset of the first element in the source image */
int stride_x; /**< Stride of the image in X dimension (in bytes) */
int stride_y; /**< Stride of the image in Y dimension (in bytes) */
} Image;
/** Structure to hold 3D tensor information */
typedef struct Tensor3D
{
__global uchar *ptr; /**< Pointer to the starting postion of the buffer */
int offset_first_element_in_bytes; /**< The offset of the first element in the source image */
int stride_x; /**< Stride of the image in X dimension (in bytes) */
int stride_y; /**< Stride of the image in Y dimension (in bytes) */
int stride_z; /**< Stride of the image in Z dimension (in bytes) */
} Tensor3D;
/** Structure to hold 4D tensor information */
typedef struct Tensor4D
{
__global uchar *ptr; /**< Pointer to the starting postion of the buffer */
int offset_first_element_in_bytes; /**< The offset of the first element in the source image */
int stride_x; /**< Stride of the image in X dimension (in bytes) */
int stride_y; /**< Stride of the image in Y dimension (in bytes) */
int stride_z; /**< Stride of the image in Z dimension (in bytes) */
int stride_w; /**< Stride of the image in W dimension (in bytes) */
} Tensor4D;
/** Wrap vector information into an Vector structure, and make the pointer point at this workitem's data.
*
* @param[in] ptr Pointer to the starting postion of the buffer
* @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
* @param[in] stride_x Stride of the vector in X dimension (in bytes)
* @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes)
*
* @return An image object
*/
inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x)
{
Vector vector =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
};
vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x;
return vector;
}
/** Wrap image information into an Image structure, and make the pointer point at this workitem's data.
*
* @param[in] ptr Pointer to the starting postion of the buffer
* @param[in] offset_first_element_in_bytes The offset of the first element in the source image
* @param[in] stride_x Stride of the image in X dimension (in bytes)
* @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] stride_y Stride of the image in Y dimension (in bytes)
* @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes)
*
* @return An image object
*/
inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
{
Image img =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
.stride_y = stride_y
};
img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y;
return img;
}
/** Wrap 3D tensor information into an image structure, and make the pointer point at this workitem's data.
*
* @param[in] ptr Pointer to the starting postion of the buffer
* @param[in] offset_first_element_in_bytes The offset of the first element in the source image
* @param[in] stride_x Stride of the image in X dimension (in bytes)
* @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] stride_y Stride of the image in Y dimension (in bytes)
* @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] stride_z Stride of the image in Z dimension (in bytes)
* @param[in] step_z stride_z * number of elements along Z processed per workitem(in bytes)
*
* @return A 3D tensor object
*/
inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
Image img =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
.stride_y = stride_y
};
img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
return img;
}
/** Wrap 3D tensor information into an tensor structure, and make the pointer point at this workitem's data.
*
* @param[in] ptr Pointer to the starting postion of the buffer
* @param[in] offset_first_element_in_bytes The offset of the first element in the source image
* @param[in] stride_x Stride of the image in X dimension (in bytes)
* @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] stride_y Stride of the image in Y dimension (in bytes)
* @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] stride_z Stride of the image in Z dimension (in bytes)
* @param[in] step_z stride_z * number of elements along Z processed per workitem(in bytes)
*
* @return A 3D tensor object
*/
inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
Tensor3D tensor =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
.stride_y = stride_y,
.stride_z = stride_z
};
tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z;
return tensor;
}
/** Wrap 3D tensor information into an tensor structure.
*
* @param[in] ptr Pointer to the starting postion of the buffer
* @param[in] offset_first_element_in_bytes The offset of the first element in the source image
* @param[in] stride_x Stride of the image in X dimension (in bytes)
* @param[in] step_x stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] stride_y Stride of the image in Y dimension (in bytes)
* @param[in] step_y stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] stride_z Stride of the image in Z dimension (in bytes)
* @param[in] step_z stride_z * number of elements along Z processed per workitem(in bytes)
*
* @return A 3D tensor object
*/
inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
{
Tensor3D tensor =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
.stride_y = stride_y,
.stride_z = stride_z
};
return tensor;
}
inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w,
uint step_w,
uint mod_size)
{
Tensor4D tensor =
{
.ptr = ptr,
.offset_first_element_in_bytes = offset_first_element_in_bytes,
.stride_x = stride_x,
.stride_y = stride_y,
.stride_z = stride_z,
.stride_w = stride_w
};
tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w;
return tensor;
}
/** Get the pointer position of a Vector
*
* @param[in] vec Pointer to the starting position of the buffer
* @param[in] x Relative X position
*/
inline __global const uchar *vector_offset(const Vector *vec, int x)
{
return vec->ptr + x * vec->stride_x;
}
/** Get the pointer position of a Image
*
* @param[in] img Pointer to the starting position of the buffer
* @param[in] x Relative X position
* @param[in] y Relative Y position
*/
inline __global uchar *offset(const Image *img, int x, int y)
{
return img->ptr + x * img->stride_x + y * img->stride_y;
}
/** Get the pointer position of a Tensor3D
*
* @param[in] tensor Pointer to the starting position of the buffer
* @param[in] x Relative X position
* @param[in] y Relative Y position
* @param[in] z Relative Z position
*/
inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
{
return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z;
}
/** Get the pointer position of a Tensor4D
*
* @param[in] tensor Pointer to the starting position of the buffer
* @param[in] x Relative X position
* @param[in] y Relative Y position
* @param[in] z Relative Z position
* @param[in] w Relative W position
*/
inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w)
{
return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w;
}
/** Get the offset for a given linear index of a Tensor3D
*
* @param[in] tensor Pointer to the starting position of the buffer
* @param[in] width Width of the input tensor
* @param[in] height Height of the input tensor
* @param[in] depth Depth of the input tensor
* @param[in] index Linear index
*/
inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index)
{
uint num_elements = width * height;
const uint z = index / num_elements;
index %= num_elements;
const uint y = index / width;
index %= width;
const uint x = index;
return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes;
}
#endif // _HELPER_H
/** Macros that help in loop unrolling */
//Repeat macros with 3 param, excluding the implicit ID param
#define REPEAT_3_1(P_X, P_A, P_B, P_C) P_X##_DEF(0, P_A, P_B, P_C)
#define REPEAT_3_2(P_X, P_A, P_B, P_C) \
P_X##_DEF(1, P_A, P_B, P_C); \
REPEAT_3_1(P_X, P_A, P_B, P_C)
#define REPEAT_3_3(P_X, P_A, P_B, P_C) \
P_X##_DEF(2, P_A, P_B, P_C); \
REPEAT_3_2(P_X, P_A, P_B, P_C)
#define REPEAT_3_4(P_X, P_A, P_B, P_C) \
P_X##_DEF(3, P_A, P_B, P_C); \
REPEAT_3_3(P_X, P_A, P_B, P_C)
#define REPEAT_3_5(P_X, P_A, P_B, P_C) \
P_X##_DEF(4, P_A, P_B, P_C); \
REPEAT_3_4(P_X, P_A, P_B, P_C)
#define REPEAT_3_6(P_X, P_A, P_B, P_C) \
P_X##_DEF(5, P_A, P_B, P_C); \
REPEAT_3_5(P_X, P_A, P_B, P_C)
#define REPEAT_3_7(P_X, P_A, P_B, P_C) \
P_X##_DEF(6, P_A, P_B, P_C); \
REPEAT_3_6(P_X, P_A, P_B, P_C)
#define REPEAT_3_8(P_X, P_A, P_B, P_C) \
P_X##_DEF(7, P_A, P_B, P_C); \
REPEAT_3_7(P_X, P_A, P_B, P_C)
#define REPEAT_3_9(P_X, P_A, P_B, P_C) \
P_X##_DEF(8, P_A, P_B, P_C); \
REPEAT_3_8(P_X, P_A, P_B, P_C)
#define REPEAT_3_10(P_X, P_A, P_B, P_C) \
P_X##_DEF(9, P_A, P_B, P_C); \
REPEAT_3_9(P_X, P_A, P_B, P_C)
#define REPEAT_3_11(P_X, P_A, P_B, P_C) \
P_X##_DEF(A, P_A, P_B, P_C); \
REPEAT_3_10(P_X, P_A, P_B, P_C)
#define REPEAT_3_12(P_X, P_A, P_B, P_C) \
P_X##_DEF(B, P_A, P_B, P_C); \
REPEAT_3_11(P_X, P_A, P_B, P_C)
#define REPEAT_3_13(P_X, P_A, P_B, P_C) \
P_X##_DEF(C, P_A, P_B, P_C); \
REPEAT_3_12(P_X, P_A, P_B, P_C)
#define REPEAT_3_14(P_X, P_A, P_B, P_C) \
P_X##_DEF(D, P_A, P_B, P_C); \
REPEAT_3_13(P_X, P_A, P_B, P_C)
#define REPEAT_3_15(P_X, P_A, P_B, P_C) \
P_X##_DEF(E, P_A, P_B, P_C); \
REPEAT_3_14(P_X, P_A, P_B, P_C)
#define REPEAT_3_16(P_X, P_A, P_B, P_C) \
P_X##_DEF(F, P_A, P_B, P_C); \
REPEAT_3_15(P_X, P_A, P_B, P_C)
#define REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_3_##P_NUM(P_OP, P_A, P_B, P_C) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
#define REPEAT_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C)
// Repeat macros with 4 param, excluding the implicit ID param
#define REPEAT_4_1(P_X, P_A, P_B, P_C, P_D) P_X##_DEF(0, P_A, P_B, P_C, P_D)
#define REPEAT_4_2(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(1, P_A, P_B, P_C, P_D); \
REPEAT_4_1(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_3(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(2, P_A, P_B, P_C, P_D); \
REPEAT_4_2(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_4(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(3, P_A, P_B, P_C, P_D); \
REPEAT_4_3(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_5(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(4, P_A, P_B, P_C, P_D); \
REPEAT_4_4(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_6(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(5, P_A, P_B, P_C, P_D); \
REPEAT_4_5(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_7(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(6, P_A, P_B, P_C, P_D); \
REPEAT_4_6(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_8(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(7, P_A, P_B, P_C, P_D); \
REPEAT_4_7(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_9(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(8, P_A, P_B, P_C, P_D); \
REPEAT_4_8(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_10(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(9, P_A, P_B, P_C, P_D); \
REPEAT_4_9(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_11(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(A, P_A, P_B, P_C, P_D); \
REPEAT_4_10(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_12(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(B, P_A, P_B, P_C, P_D); \
REPEAT_4_11(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_13(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(C, P_A, P_B, P_C, P_D); \
REPEAT_4_12(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_14(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(D, P_A, P_B, P_C, P_D); \
REPEAT_4_13(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_15(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(E, P_A, P_B, P_C, P_D); \
REPEAT_4_14(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_4_16(P_X, P_A, P_B, P_C, P_D) \
P_X##_DEF(F, P_A, P_B, P_C, P_D); \
REPEAT_4_15(P_X, P_A, P_B, P_C, P_D)
#define REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_4_##P_NUM(P_OP, P_A, P_B, P_C, P_D) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
#define REPEAT_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D)
// Macro for initializing N variables. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
#define VAR_INIT_TO_CONST_DEF(ID, TYPE, VAR, VAL) TYPE VAR##ID = VAL
#define REPEAT_VAR_INIT_TO_CONST(N, TYPE, VAR, VAL) REPEAT_3_N(N, VAR_INIT_TO_CONST, TYPE, VAR, VAL)
// Macro for initializing N variables by converting the data type. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
#define VAR_INIT_CONVERT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT(VAR_IN##ID, TYPE_OUT)
#define REPEAT_VAR_INIT_CONVERT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT, TYPE_OUT, VAR_IN, VAR_OUT)
// Macro for initializing N variables by converting the data type with saturation. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
#define VAR_INIT_CONVERT_SAT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT_SAT(VAR_IN##ID, TYPE_OUT)
#define REPEAT_VAR_INIT_CONVERT_SAT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT_SAT, TYPE_OUT, VAR_IN, VAR_OUT)
// Macro for adding a constant to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define ADD_CONST_TO_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID += (TYPE)VAL
#define REPEAT_ADD_CONST_TO_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, ADD_CONST_TO_VAR, TYPE, VAR, VAL)
// Macro for multiplying N variables (VAR_B) by a constant (VAL) and adding to other N variables (VAR_A). Generates N statements that defines VAR_A##N =RHS_ACCESSOR_DEF(...)
#define MLA_VAR_WITH_CONST_VEC_DEF(ID, VAR_A, VAR_B, VAL) VAR_A##ID += VAR_B##ID * VAL
#define REPEAT_MLA_VAR_WITH_CONST_VEC(N, VAR_A, VAR_B, VAL) REPEAT_3_N(N, MLA_VAR_WITH_CONST_VEC, VAR_A, VAR_B, VAL)
// Macro for adding a vector to N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define ADD_VECTOR_TO_VAR_DEF(ID, TYPE, VAR, VEC) VAR##ID += VEC
#define REPEAT_ADD_VECTOR_TO_VAR(N, VAR, VEC) REPEAT_3_N(N, ADD_VECTOR_TO_VAR, "", VAR, VEC)
// Macro for adding a two N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define ADD_TWO_VARS_DEF(ID, TYPE, VAR_A, VAR_B) VAR_A##ID += VAR_B##ID
#define REPEAT_ADD_TWO_VARS(N, VAR_A, VAR_B) REPEAT_3_N(N, ADD_TWO_VARS, "", VAR_A, VAR_B)
// Macro for performing Max between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define MAX_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = max(VAR##ID, (TYPE)VAL)
#define REPEAT_MAX_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MAX_CONST_VAR, TYPE, VAR, VAL)
// Macro for performing Min between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define MIN_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = min(VAR##ID, (TYPE)VAL)
#define REPEAT_MIN_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MIN_CONST_VAR, TYPE, VAR, VAL)
// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
// Macro for performing per-channel ASYMM_MULT_BY_QUANT_MULTIPLIER to N variables.
#define ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) \
({ \
VEC_DATA_TYPE(int, N0) \
VAR##ID_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0); \
VEC_DATA_TYPE(int, N0) \
VAR##ID_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0); \
VAR##ID = select(VAR##ID_shift_lt0, VAR##ID_shift_gt0, RES_SHIFT >= 0); \
})
#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL, SIZE, VAR, RES_MUL, RES_SHIFT)
#endif // ARM_COMPUTE_REPEAT_H
#if defined(POOL_AVG) || defined(POOL_L2)
#define POOL_OP(x, y) ((x) + (y))
#else /* defined(POOL_AVG) || defined(POOL_L2) */
#define POOL_OP(x, y) (fmax((x), (y)))
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
#define POW2_OP(x, vec_size) ((x) * (x))
#else /* defined(POOL_L2) */
#define POW2_OP(x, vec_size) (x)
#endif /* defined(POOL_L2) */
#define DIV_OP(x, y) (x * (1.f / y))
#define SQRT_OP(x) sqrt((x))
#if STRIDE_X == 1
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE1(res, input, output)
#elif STRIDE_X == 2 /* STRIDE_X == 1 */
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE2(res, input, output)
#elif STRIDE_X == 3 /* STRIDE_X not equals 1 or 2 */
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE3(res, input, output)
#endif /* STRIDE_X == 3 */
#if defined(FP_MIXED_PRECISION)
#define CONVERT_TO_ACC_DATA_TYPE(x, n) CONVERT(x, VEC_DATA_TYPE(ACC_DATA_TYPE, n))
#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) \
CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n)
#else /* defined(FP_MIXED_PRECISION) */
#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) vload##n(offset, ptr)
#endif /* defined(FP_MIXED_PRECISION) */
#define POOLING3x3_STRIDE1(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 4); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 4); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 4); \
data00 = POW2_OP(data00, 4); \
data01 = POW2_OP(data01, 2); \
data10 = POW2_OP(data10, 4); \
data11 = POW2_OP(data11, 2); \
data20 = POW2_OP(data20, 4); \
data21 = POW2_OP(data21, 2); \
\
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data01.s0, data00.s3, data01.s01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data11.s0, data10.s3, data11.s01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data21.s0, data20.s3, data21.s01); \
\
values00 = POOL_OP(values00, values10); \
values01 = POOL_OP(values01, values11); \
values00 = POOL_OP(values00, values20); \
values01 = POOL_OP(values01, values21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \
})
#define POOLING3x3_STRIDE2(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
ACC_DATA_TYPE data01 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
ACC_DATA_TYPE data11 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
ACC_DATA_TYPE data21 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8)); \
data00 = POW2_OP(data00, 8); \
data01 = POW2_OP(data01, 1); \
data10 = POW2_OP(data10, 8); \
data11 = POW2_OP(data11, 1); \
data20 = POW2_OP(data20, 8); \
data21 = POW2_OP(data21, 1); \
\
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s667, data01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data10.s667, data11); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data20.s667, data21); \
\
values00 = POOL_OP(values00, values10); \
values01 = POOL_OP(values01, values11); \
values00 = POOL_OP(values00, values20); \
values01 = POOL_OP(values01, values21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \
})
#define POOLING3x3_STRIDE3(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \
data00 = POW2_OP(data00, 8); \
data01 = POW2_OP(data01, 4); \
data10 = POW2_OP(data10, 8); \
data11 = POW2_OP(data11, 4); \
data20 = POW2_OP(data20, 8); \
data21 = POW2_OP(data21, 4); \
\
data00 = POOL_OP(data00, data10); \
data01 = POOL_OP(data01, data11); \
data00 = POOL_OP(data00, data20); \
data01 = POOL_OP(data01, data21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s036, data01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s147, data01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s25, data01.s03)); \
})
ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h,
const int pad_x, const int pad_y, const int stride_x, const int stride_y)
{
int start_x = get_global_id(0) * stride_x - pad_x;
int start_y = get_global_id(1) * stride_y - pad_y;
const int end_x = min(start_x + pool_size_x, upper_bound_w);
const int end_y = min(start_y + pool_size_y, upper_bound_h);
#if defined(EXCLUDE_PADDING)
start_x = max(0, start_x);
start_y = max(0, start_y);
#endif /* defined(EXCLUDE_PADDING) */
return ((end_y - start_y) * (end_x - start_x));
}
/** Performs a pooling function of pool size equal to 2.
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void pooling_layer_2(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
// Load data
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 = POW2_OP(data0, 2);
data1 = POW2_OP(data1, 2);
#endif /* defined(POOL_L2) */
// Perform calculations
data0 = POOL_OP(data0, data1);
ACC_DATA_TYPE res = POOL_OP(data0.s0, data0.s1);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average or l2 pooling
res = DIV_OP(res, calculate_avg_scale(2, 2, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
/** Performs a pooling function of pool size equal to 3
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void pooling_layer_3(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
// Load data
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data2 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 = POW2_OP(data0, 3);
data1 = POW2_OP(data1, 3);
data2 = POW2_OP(data2, 3);
#endif /* defined(POOL_L2) */
// Perform calculations
data0 = POOL_OP(data0, data1);
data0 = POOL_OP(data0, data2);
ACC_DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res = DIV_OP(res, calculate_avg_scale(3, 3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
#if defined(POOLING3x3)
#define CONVERT_OP(data_type) convert_##data_type##4
#define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type)
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h,
const int pad_x, const int pad_y, const int stride_x, const int stride_y)
{
int4 start_x = ((int4)get_global_id(0) * 4 + (int4)(0, 1, 2, 3)) * (int4)stride_x - (int4)pad_x;
int start_y = get_global_id(1) * stride_y - pad_y;
const int4 end_x = min(start_x + (int4)pool_size, (int4)upper_bound_w);
const int end_y = min(start_y + pool_size, upper_bound_h);
#if defined(EXCLUDE_PADDING)
start_x = max((int4)0, start_x);
start_y = max(0, start_y);
#endif /* defined(EXCLUDE_PADDING) */
return (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(ACC_DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x));
}
/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void pooling_layer_optimized_3(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
res;
// Perform pooling 3x3 for 4 output elements
POOLING3x3(res, input, output);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res *= calculate_avg_scale4(3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
vstore4(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 4)), 0, (__global DATA_TYPE *)output.ptr);
}
#endif // defined(POOLING3x3)
#if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
/** Performs a pooling function of pool size equal to N (NCHW)
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
* @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void pooling_layer_MxN_nchw(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
vdata = INITIAL_VALUE;
ACC_DATA_TYPE sdata = INITIAL_VALUE;
// Load data
for(int y = 0; y < POOL_SIZE_Y; y++)
{
int x = 0;
for(; x <= ((int)POOL_SIZE_X - 8); x += 8)
{
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif /* defined(POOL_L2) */
vdata = POOL_OP(vdata, data0);
}
// Leftover
for(; x < (int)POOL_SIZE_X; ++x)
{
ACC_DATA_TYPE data0 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif /* defined(POOL_L2) */
sdata = POOL_OP(sdata, data0);
}
}
// Reduce result
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
reduce4 = POOL_OP(vdata.s0123, vdata.s4567);
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
reduce2 = POOL_OP(reduce4.s01, reduce4.s23);
ACC_DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1);
res = POOL_OP(res, sdata);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res = DIV_OP(res, calculate_avg_scale(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
#endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
inline void offset_no_padding_nchw(const Tensor3D *input, uint *offset_top, uint *offset_bottom)
{
const int pad_horiz = PAD_TENSOR_LEFT + PAD_TENSOR_RIGHT;
const int pad_vert = PAD_TENSOR_TOP + PAD_TENSOR_BOTTOM;
const int x = get_global_id(0) * STRIDE_X;
const int y = get_global_id(1) * STRIDE_Y;
const int z = get_global_id(2);
//x axis: width, y axis: height, z axis: component
const uint padded_offset = input->offset_first_element_in_bytes
+ x * input->stride_x
+ y * input->stride_y
+ z * input->stride_z;
const uint offset_base = padded_offset
- y * pad_horiz * sizeof(DATA_TYPE) /* Horizontal padding for each row */
- PAD_TENSOR_TOP * input->stride_y /* top padding */
- z * MAX_HEIGHT * pad_horiz * sizeof(DATA_TYPE) - z * pad_vert * input->stride_y /* Z plane padding */
- PAD_TENSOR_LEFT * sizeof(DATA_TYPE);
#if defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT)
*offset_top = (uint)((offset_base / sizeof(DATA_TYPE)) % (TENSOR_CHANNEL * TENSOR_WIDTH * TENSOR_HEIGHT));
#else /* defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT) */
*offset_top = (uint)(offset_base / sizeof(DATA_TYPE));
#endif /* defined(TENSOR_CHANNEL) && defined(TENSOR_WIDTH) && defined(TENSOR_HEIGHT) */
*offset_bottom = *offset_top + input->stride_y / sizeof(DATA_TYPE) - pad_horiz;
return;
}
#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
/** Performs a MAX pooling of pool size equal to 2, and record max value indices for NCHW.
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F32
* @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13;
* @note Tensors width and height must be passed at compile time using -DMAX_WIDTH and -DMAX_HEIGHT
* @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* @note Tensor padding values must be passed at compile time using PAD_TENSOR_LEFT, PAD_TENSOR_RIGHT, PAD_TENSOR_TOP and PAD_TENSOR_BOTTOM
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F32
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] indices_ptr Pointer to the indices tensor. Supported data types: U32
* @param[in] indices_stride_x Stride of the indices tensor in X dimension (in bytes)
* @param[in] indices_step_x indices_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] indices_stride_y Stride of the indices tensor in Y dimension (in bytes)
* @param[in] indices_step_y indices_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] indices_stride_z Stride of the indices tensor in Z dimension (in bytes)
* @param[in] indices_step_z indices_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] indices_offset_first_element_in_bytes The offset of the first element in the indices tensor
*/
__kernel void pooling_layer_2_nchw_indices_fp32(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output),
TENSOR3D_DECLARATION(indices))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
Tensor3D indices = CONVERT_TO_TENSOR3D_STRUCT(indices);
// Load data
float2 data0 = VLOAD(2)(0, (__global float *)tensor3D_offset(&input, 0, 0, 0));
float2 data1 = VLOAD(2)(0, (__global float *)tensor3D_offset(&input, 0, 1, 0));
// Perform calculations
float data0_max = POOL_OP(data0.s0, data0.s1);
float data1_max = POOL_OP(data1.s0, data1.s1);
float res = POOL_OP(data0_max, data1_max);
// Store result
*(__global float *)output.ptr = res;
#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
uint offset_top = 0;
uint offset_bottom = 0;
offset_no_padding_nchw(&input, &offset_top, &offset_bottom);
uint index0 = select(offset_top + 1, offset_top, isgreaterequal(data0.s0, data0.s1));
uint index1 = select(offset_bottom + 1, offset_bottom, isgreaterequal(data1.s0, data1.s1));
uint index = select(index1, index0, isgreaterequal(data0_max, data1_max));
*(__global uint *)indices.ptr = index;
#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
}
/** Performs a MAX pooling of pool size equal to 2, and record max value indices for NCHW.
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F16
* @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13;
* @note Tensors width and height must be passed at compile time using -DMAX_WIDTH and -DMAX_HEIGHT
* @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* @note Tensor padding values must be passed at compile time using PAD_TENSOR_LEFT, PAD_TENSOR_RIGHT, PAD_TENSOR_TOP and PAD_TENSOR_BOTTOM
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F16
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] indices_ptr Pointer to the indices tensor. Supported data types: U32
* @param[in] indices_stride_x Stride of the indices tensor in X dimension (in bytes)
* @param[in] indices_step_x indices_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] indices_stride_y Stride of the indices tensor in Y dimension (in bytes)
* @param[in] indices_step_y indices_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] indices_stride_z Stride of the indices tensor in Z dimension (in bytes)
* @param[in] indices_step_z indices_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] indices_offset_first_element_in_bytes The offset of the first element in the indices tensor
*/
__kernel void pooling_layer_2_nchw_indices_fp16(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output),
TENSOR3D_DECLARATION(indices))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
Tensor3D indices = CONVERT_TO_TENSOR3D_STRUCT(indices);
// Load data
half2 data0 = VLOAD(2)(0, (__global half *)tensor3D_offset(&input, 0, 0, 0));
half2 data1 = VLOAD(2)(0, (__global half *)tensor3D_offset(&input, 0, 1, 0));
// Perform calculations
half data0_max = POOL_OP(data0.s0, data0.s1);
half data1_max = POOL_OP(data1.s0, data1.s1);
half res = POOL_OP(data0_max, data1_max);
// Store result
*(__global half *)output.ptr = res;
#if defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
uint offset_top = 0;
uint offset_bottom = 0;
offset_no_padding_nchw(&input, &offset_top, &offset_bottom);
uint index0 = select(offset_top + 1, offset_top, isgreaterequal(data0.s0, data0.s1));
uint index1 = select(offset_bottom + 1, offset_bottom, isgreaterequal(data1.s0, data1.s1));
uint index = select(index1, index0, isgreaterequal(data0_max, data1_max));
*(__global uint *)indices.ptr = index;
#endif //defined(PAD_TENSOR_LEFT) && defined(PAD_TENSOR_RIGHT) && defined(PAD_TENSOR_TOP) && defined(PAD_TENSOR_BOTTOM)
}
#if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE)
#if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
/** Performs pooling layer of size equal to MxN. This OpenCL kernel can perform the following pooling types:
* -# max, -DPOOL_MAX must be passed at compile time
* -# average, -DPOOL_AVG must be passed at compile time. If padding has to be expluded, -DEXCLUDE_PADDING should be passed at compile time
* -# l2 normalisation, -DPOOL_L2 must be passed at compile time
*
* @note Datatype must be passed at compile type using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F32/F16
* @note Accumulation data type must be passed at compile time using -DACC_DATA_TYPE e.g. -DACC_DATA_TYPE=float
* @note If -DFP_MIXED_PRECISION is passed at compile time, the kernel will use F32 for the partial result
* @note Pool size must be passed at compile time using -DPOOL_SIZE_X and -DPOOL_SIZE_Y. e.g. -DPOOL_SIZE_X=4, -DPOOL_SIZE_Y=4
* @note Input tensor width and height must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT
* @note Output tensor height, channels and batch size must be passed at compile time using -DDST_HEIGHT, -DDST_CHANNELS and -DDST_BATCH_SIZE
* @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* @note Pool pads must be passed at compile time using -DPAD_X and -DPAD_Y
* @note Vector size must be passed at compile time using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
* @note Leftover vector size must be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F32/F16
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void pooling_layer_MxN_nhwc(
TENSOR4D_DECLARATION(input),
TENSOR4D_DECLARATION(output))
{
// Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0
// Note: If C is less than VEC_SIZE, VEC_SIZE should be SHRINKED to the closest smaller VEC_SIZE. This operation is performed on the host side
int offset_c = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) * sizeof(DATA_TYPE);
int idx_out_w = get_global_id(1);
#if DST_BATCH_SIZE != 1
// If batch size != 1, the batch size dimension is collapsed over the height dimension
int idx_out_h = get_global_id(2) % DST_HEIGHT;
int idx_out_n = get_global_id(2) / DST_HEIGHT;
#else //DST_BATCH_SIZE != 1
int idx_out_h = get_global_id(2);
int idx_out_n = 0;
#endif // DST_BATCH_SIZE != 1
int idx_in_w = idx_out_w * STRIDE_X - PAD_X;
int idx_in_h = idx_out_h * STRIDE_Y - PAD_Y;
int pool_x_s = max((int)0, -idx_in_w);
int pool_x_e = min((int)POOL_SIZE_X, (int)SRC_WIDTH - idx_in_w);
int pool_y_s = max((int)0, -idx_in_h);
int pool_y_e = min((int)POOL_SIZE_Y, (int)SRC_HEIGHT - idx_in_h);
__global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes +
offset_c +
idx_out_n * input_stride_w;
__global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes +
offset_c +
idx_out_w * output_stride_y +
idx_out_h * output_stride_z +
idx_out_n * output_stride_w;
#if ((defined(POOL_AVG) || defined(POOL_L2)))
#if defined(EXCLUDE_PADDING)
int filter_size = 0;
#else // defined(EXCLUDE_PADDING)
int filter_size = POOL_SIZE_X * POOL_SIZE_Y;
#endif // defined(EXCLUDE_PADDING)
#endif // ((defined(POOL_AVG) || defined(POOL_L2)))
VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
res0 = INITIAL_VALUE;
for(int y = pool_y_s; y < pool_y_e; ++y)
{
for(int x = pool_x_s; x < pool_x_e; ++x)
{
VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) data0;
#if defined(FP_MIXED_PRECISION)
// In case of FP_MIXED_PRECISION, ACC_DATA_TYPE is != DATA_TYPE
data0 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + (x + idx_in_w) * input_stride_y + (y + idx_in_h) * input_stride_z)), VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
#else // defined(FP_MIXED_PRECISION)
data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + (x + idx_in_w) * input_stride_y + (y + idx_in_h) * input_stride_z));
#endif // defined(FP_MIXED_PRECISION)
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif // defined(POOL_L2)
res0 = POOL_OP(res0, data0);
#if ((defined(POOL_AVG) || defined(POOL_L2))) && defined(EXCLUDE_PADDING)
filter_size++;
#endif // ((defined(POOL_AVG) || defined(POOL_L2))) && defined(EXCLUDE_PADDING)
}
}
#if defined(POOL_AVG) || defined(POOL_L2)
res0 /= (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))filter_size;
#endif // defined(POOL_AVG) || defined(POOL_L2)
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res0 = SQRT_OP(res0);
#endif // defined(POOL_L2)
// Store result
#if defined(FP_MIXED_PRECISION)
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) res_converted0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
STORE_VECTOR_SELECT(res_converted, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
#else // defined(FP_MIXED_PRECISION)
STORE_VECTOR_SELECT(res, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
#endif // defined(FP_MIXED_PRECISION)
}
#endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
#define SELECT_TYPE SELECT_VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
/** Performs pooling layer of size equal to 2. This OpenCL kernel can perform the following pooling types:
* -# max, -DPOOL_MAX must be passed at compile time
* -# max extracting the max index, -DPOOL_MAX and -DEXTRACT_MAX_INDEX must be passed at compile time
* -# average, -DPOOL_AVG must be passed at compile time. If padding has to be expluded, -DEXCLUDE_PADDING should be passed at compile time
* -# l2 normalisation, -DPOOL_L2 must be passed at compile time
*
* @note Datatype must be passed at compile type using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F32/F16
* @note Accumulation data type must be passed at compile time using -DACC_DATA_TYPE e.g. -DACC_DATA_TYPE=float
* @note If -DFP_MIXED_PRECISION is passed at compile time, the kernel will use F32 for the partial result
* @note Input tensor width and height must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT
* @note Output tensor height, channels and batch size must be passed at compile time using -DDST_HEIGHT, -DDST_CHANNELS and -DDST_BATCH_SIZE
* @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* @note Pool pads must be passed at compile time using -DPAD_X and -DPAD_Y
* @note Vector size must be passed at compile time using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
* @note Leftover vector size must be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: F32/F16
* @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] indices_ptr (Optional) Pointer to the indices tensor. Supported data types: U32
* @param[in] indices_stride_x (Optional) Stride of the indices tensor in X dimension (in bytes)
* @param[in] indices_step_x (Optional) indices_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] indices_stride_y (Optional) Stride of the indices tensor in Y dimension (in bytes)
* @param[in] indices_step_y (Optional) indices_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] indices_stride_z (Optional) Stride of the indices tensor in Z dimension (in bytes)
* @param[in] indices_step_z (Optional) indices_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] indices_stride_w (Optional) Stride of the indices tensor in W dimension (in bytes)
* @param[in] indices_step_w (Optional) indices_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] indices_offset_first_element_in_bytes (Optional) The offset of the first element in the indices tensor
*/
__kernel void pooling_layer_2x2_nhwc(
TENSOR4D_DECLARATION(input),
TENSOR4D_DECLARATION(output)
#if defined(EXTRACT_MAX_INDEX) && defined(POOL_MAX)
,
TENSOR4D_DECLARATION(indices)
#endif // defined(EXTRACT_MAX_INDEX) && defined(POOL_MAX)
)
{
// Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0
// Note: If C is less than VEC_SIZE, VEC_SIZE should be SHRINKED to the closest smaller VEC_SIZE. This operation is performed on the host side
int idx_out_c = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
int idx_out_w = get_global_id(1);
#if DST_BATCH_SIZE != 1
// If batch size != 1, the batch size dimension is collapsed over the height dimension
int idx_out_h = get_global_id(2) % DST_HEIGHT;
int idx_out_n = get_global_id(2) / DST_HEIGHT;
#else //SRC_BATCH_SIZE != 1
int idx_out_h = get_global_id(2);
int idx_out_n = 0;
#endif // SRC_BATCH_SIZE != 1
int idx_in_w = idx_out_w * STRIDE_X - PAD_X;
int idx_in_h = idx_out_h * STRIDE_Y - PAD_Y;
__global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes +
idx_out_c * sizeof(DATA_TYPE) +
idx_out_n * input_stride_w;
__global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes +
idx_out_c * sizeof(DATA_TYPE) +
idx_out_w * output_stride_y +
idx_out_h * output_stride_z +
idx_out_n * output_stride_w;
int pool_x_s = max((int)0, -idx_in_w);
int pool_x_e = min((int)2, (int)SRC_WIDTH - idx_in_w);
int pool_y_s = max((int)0, -idx_in_h);
int pool_y_e = min((int)2, (int)SRC_HEIGHT - idx_in_h);
int filter_size = (pool_x_e - pool_x_s) * (pool_y_e - pool_y_s);
int x0 = pool_x_s + idx_in_w;
int y0 = pool_y_s + idx_in_h;
int x1 = pool_x_e - 1 + idx_in_w;
int y1 = pool_y_e - 1 + idx_in_h;
REPEAT_VAR_INIT_TO_CONST(4, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE), data, 0);
#if defined(FP_MIXED_PRECISION)
// In case of FP_MIXED_PRECISION, ACC_DATA_TYPE is != DATA_TYPE
data0 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x0 * input_stride_y + y0 * input_stride_z)), VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
data1 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x1 * input_stride_y + y0 * input_stride_z)), VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
data2 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x0 * input_stride_y + y1 * input_stride_z)), VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
data3 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x1 * input_stride_y + y1 * input_stride_z)), VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
#else // defined(FP_MIXED_PRECISION)
data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x0 * input_stride_y + y0 * input_stride_z));
data1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x1 * input_stride_y + y0 * input_stride_z));
data2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x0 * input_stride_y + y1 * input_stride_z));
data3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + x1 * input_stride_y + y1 * input_stride_z));
#endif // defined(FP_MIXED_PRECISION)
#if !defined(POOL_MAX)
if(filter_size != 4)
{
SELECT_TYPE cond_w_s = (SELECT_TYPE)idx_in_w < (SELECT_TYPE)0;
SELECT_TYPE cond_w_e = (SELECT_TYPE)idx_in_w >= (SELECT_TYPE)(SRC_WIDTH - 1);
SELECT_TYPE cond_h_s = (SELECT_TYPE)idx_in_h < (SELECT_TYPE)0;
SELECT_TYPE cond_h_e = (SELECT_TYPE)idx_in_h >= (SELECT_TYPE)(SRC_HEIGHT - 1);
// Make invalid the values loaded if the x or y coordinate was clamped (out-of-bound)
data0 = select(data0, (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))INITIAL_VALUE, (SELECT_TYPE)(cond_w_s | cond_h_s));
data1 = select(data1, (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))INITIAL_VALUE, (SELECT_TYPE)(cond_w_e | cond_h_s));
data2 = select(data2, (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))INITIAL_VALUE, (SELECT_TYPE)(cond_w_s | cond_h_e));
data3 = select(data3, (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))INITIAL_VALUE, (SELECT_TYPE)(cond_w_e | cond_h_e));
}
#endif // !defined(POOL_MAX)
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
data1 *= data1;
data2 *= data2;
data3 *= data3;
#endif /* defined(POOL_L2) */
VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
res0 = data0;
res0 = POOL_OP(res0, data1);
res0 = POOL_OP(res0, data2);
res0 = POOL_OP(res0, data3);
#if defined(POOL_AVG) || defined(POOL_L2)
#if defined(EXCLUDE_PADDING)
res0 /= (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))filter_size;
#else // !defined(EXCLUDE_PADDING)
res0 /= (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))4;
#endif // defined(EXCLUDE_PADDING)
#endif // defined(POOL_AVG) || defined(POOL_L2)
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res0 = SQRT_OP(res0);
#endif // defined(POOL_L2)
// Store result
#if defined(FP_MIXED_PRECISION)
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) res_converted0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
STORE_VECTOR_SELECT(res_converted, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
#else // defined(FP_MIXED_PRECISION)
STORE_VECTOR_SELECT(res, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
#endif // defined(FP_MIXED_PRECISION)
#if defined(EXTRACT_MAX_INDEX) && defined(POOL_MAX)
// This part is used to return the index of the maximum value
// Note: DST_CHANNELS and DST_BATCH_SIZE can be used for either the input and output tensor
// note: Batch dimension does not contribute in the offset contribution
VEC_DATA_TYPE(uint, VEC_SIZE) base_index = (uint)idx_out_c;
base_index += VEC_OFFS(uint, VEC_SIZE);
VEC_DATA_TYPE(uint, VEC_SIZE) index0 = base_index + (uint)x0 * DST_CHANNELS + (uint)y0 * (DST_CHANNELS * SRC_WIDTH);
VEC_DATA_TYPE(uint, VEC_SIZE) index1 = base_index + (uint)x1 * DST_CHANNELS + (uint)y0 * (DST_CHANNELS * SRC_WIDTH);
VEC_DATA_TYPE(uint, VEC_SIZE) index2 = base_index + (uint)x0 * DST_CHANNELS + (uint)y1 * (DST_CHANNELS * SRC_WIDTH);
VEC_DATA_TYPE(uint, VEC_SIZE) index3 = base_index + (uint)x1 * DST_CHANNELS + (uint)y1 * (DST_CHANNELS * SRC_WIDTH);
index0 = select(index1, index0, CONVERT(isgreaterequal(data0, data1), VEC_DATA_TYPE(int, VEC_SIZE)));
index1 = select(index3, index2, CONVERT(isgreaterequal(data2, data3), VEC_DATA_TYPE(int, VEC_SIZE)));
index0 = select(index1, index0, CONVERT(isgreaterequal(max(data0, data1), max(data2, data3)), VEC_DATA_TYPE(int, VEC_SIZE)));
__global unsigned char *idx_base_ptr = indices_ptr + indices_offset_first_element_in_bytes +
idx_out_c * sizeof(uint) +
idx_out_w * indices_stride_y +
idx_out_h * indices_stride_z +
idx_out_n * indices_stride_w;
// Store result
STORE_VECTOR_SELECT(index, uint, idx_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, ((VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0));
#endif // defined(EXTRACT_MAX_INDEX) && defined(POOL_MAX)
}
#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE)
)"