From 50335fd3d0734157382741fcf1bfdaf630c60c4b Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Thu, 26 Aug 2021 14:47:55 +0100 Subject: Remove padding from ClGemmMatrixMultiplyReshapedKernel Create new macros for loading values from memory while being aware of boundaries of the tensor to not generate page faults. Resolves: COMPMID-4447 Change-Id: If9a455291e395ebd9070ebe5e120b3064d8fab29 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6168 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins --- src/core/CL/cl_kernels/gemm_helpers.h | 300 ++++++++++++++++++++++++++++++++++ 1 file changed, 300 insertions(+) (limited to 'src/core/CL/cl_kernels/gemm_helpers.h') diff --git a/src/core/CL/cl_kernels/gemm_helpers.h b/src/core/CL/cl_kernels/gemm_helpers.h index 3bbd243ff5..e4ed47e7e6 100644 --- a/src/core/CL/cl_kernels/gemm_helpers.h +++ b/src/core/CL/cl_kernels/gemm_helpers.h @@ -398,6 +398,306 @@ #define LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) /** @} */ // end of group LOAD_BLOCK +/** Partially load the 0 to (n-1)th rows of the given variables + * @name LOAD_ROW_PARTIAL_n + * Within each row, load the lower @p LOAD_N0 elements of vectors of width @p N0 + * + * @note in case @p LOAD_N0 != 1, 2, 3, 4, 8, 16, extra vload(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] LOAD_N0 The **lower** size of the vectors to load. 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] OFFSET The offset within a row + * @param[in] STRIDE_Y The stride value in y-axis direction + * @param[in] Z The offset in z-axis direction + * @{ + */ +#define LOAD_ROW_PARTIAL_1(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y + Z##0)); + +#define LOAD_ROW_PARTIAL_2(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_1(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y + Z##1)); + +#define LOAD_ROW_PARTIAL_3(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_2(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y + Z##2)); + +#define LOAD_ROW_PARTIAL_4(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_3(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y + Z##3)); + +#define LOAD_ROW_PARTIAL_5(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_4(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y + Z##4)); + +#define LOAD_ROW_PARTIAL_6(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_5(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y + Z##5)); + +#define LOAD_ROW_PARTIAL_7(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_6(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y + Z##6)); + +#define LOAD_ROW_PARTIAL_8(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_7(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y + Z##7)); + +#define LOAD_ROW_PARTIAL_9(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_8(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y + Z##8)); + +#define LOAD_ROW_PARTIAL_10(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_9(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y + Z##9)); + +#define LOAD_ROW_PARTIAL_11(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_10(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y + Z##A)); + +#define LOAD_ROW_PARTIAL_12(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_11(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y + Z##B)); + +#define LOAD_ROW_PARTIAL_13(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_12(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y + Z##C)); + +#define LOAD_ROW_PARTIAL_14(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_13(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y + Z##D)); + +#define LOAD_ROW_PARTIAL_15(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_14(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y + Z##E)); + +#define LOAD_ROW_PARTIAL_16(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + LOAD_ROW_PARTIAL_15(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) \ + VLOAD_PARTIAL(N0, LOAD_N0) \ + (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y + Z##F)); +/** @} */ // end of groupd LOAD_ROW_PARTIAL_n + +/** Partially load a block of the given size LOAD_M0xLOAD_N0 + * @name LOAD_BLOCK_PARTIAL + * + * @note The vector width @p N0 is also required for correct partial storing behaviour. + * @note in case @p LOAD_N0 != 1, 2, 3, 4, 8, 16, extra vload(s) will be invoked, thus incurring small performance penalty. + * + * The data to load is expected to have consecutive names for each row. + * E.g., for LOAD_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 LOAD_M0=3 and Z=zin, the expected z offset names are zin0, zin1 and zin2. + * + * @param[in] LOAD_M0 The number of rows to load. Supported: 1-16 + * @param[in] LOAD_N0 The lower number of elements of vectors to load. 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] OFFSET The offset within a row + * @param[in] STRIDE_Y The stride value in y-axis direction + * @param[in] Z The offset in z-axis direction + * @{ + */ +#define LOAD_BLOCK_PARTIAL_STR(LOAD_M0, LOAD_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_ROW_PARTIAL_##LOAD_M0(N0, LOAD_N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) +#define LOAD_BLOCK_PARTIAL(LOAD_M0, LOAD_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_PARTIAL_STR(LOAD_M0, LOAD_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) +/** Load 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 vload(s) will be invoked, thus incurring small performance penalty. + * + * The data to load 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 load, 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] OFFSET The offset within a row + * @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 load Y. True to use PARTIAL_STORE_M0 rather than M0. + * @param[in] PARTIAL_COND_X Condition on the x axis to perform the partial load X. True to use PARTIAL_STORE_N0 rather than N0. + */ +#define LOAD_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y)) \ + { \ + LOAD_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } \ + else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X)) \ + { \ + LOAD_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } \ + else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X)) \ + { \ + LOAD_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } \ + else \ + { \ + LOAD_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } +/** Load 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 vload(s) will be invoked, thus incurring small performance penalty. + * + * The data to load 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 load, 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] OFFSET The offset within a row + * @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 load X. True to use PARTIAL_STORE_N0 rather than N0. + */ +#define LOAD_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \ + if(!(PARTIAL_COND_X)) \ + { \ + LOAD_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } \ + else \ + { \ + LOAD_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } +/** Load 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 vload(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] OFFSET The offset within a row + * @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 LOAD_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \ + if(!(PARTIAL_COND_Y)) \ + { \ + LOAD_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } \ + else \ + { \ + LOAD_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z); \ + } +/** @} */ // end of group LOAD_BLOCK_PARTIAL +/** Boundary-aware GeMM block load + * @name LOAD_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 src tensor can be loaded 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 src 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 vload(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 load methods such that the boundary detection logic is only added when needed. + * + * The data to load 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. + * + * The macro will result in a declaration of @p M0 vectors of size @p N0 with data + * type @p DATA_TYPE containing values partially loaded from the specified + * address in memory. The remaining (N0 - PARTIAL_STORE_N0) elements will be + * filled with zeros. + * + * @param[in] M0 The number of rows to load, 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] OFFSET The offset within a row + * @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 load Y. True to use PARTIAL_STORE_M0 rather than M0. + * @param[in] PARTIAL_COND_X Condition on the x axis to perform the partial load 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 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) + +#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0 +// Case2: Partial blocks in y +#define LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), BASENAME, 0); \ + LOAD_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) + +#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0 +// Case3: Partial blocks in x +#define LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), BASENAME, 0); \ + LOAD_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, 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 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), BASENAME, 0); \ + LOAD_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) + +#endif // PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0 + /** Loads the rows from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1). * @name LOAD_TEXTURE2D_ROW_n * -- cgit v1.2.1