R"( #ifndef ARM_COMPUTE_HELPER_H #define ARM_COMPUTE_HELPER_H #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)); #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)); #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) #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) #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)); #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) #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); \ } #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); \ } #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); \ } #if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) #if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0 #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 #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 #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 #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 #endif #if defined(PARTIAL_STORE_M0) #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \ ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0)))) #else #define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \ ((uint)(y * M0)) #endif #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) #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) #pragma OPENCL EXTENSION cl_khr_fp16 : enable #endif #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 #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 #if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf) #pragma OPENCL EXTENSION cl_arm_printf : enable #endif #define GPU_ARCH_MIDGARD 0x100 #define GPU_ARCH_BIFROST 0x200 #define GPU_ARCH_VALHALL 0x300 #define CONCAT(a, b) a##b #define EXPAND(x) x #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) #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) #define REVERSE_STR(x, s) REV##s((x)) #define REVERSE(x, s) REVERSE_STR(x, s) #define ROT1_0(x) ((x)) #define ROT1_1(x) ((x)) #define ROT2_0(x) ((x)) #define ROT2_1(x) ((x).s10) #define ROT2_2(x) ((x)) #define ROT3_0(x) ((x)) #define ROT3_1(x) ((x).s201) #define ROT3_2(x) ((x).s120) #define ROT3_3(x) ((x)) #define ROT4_0(x) ((x)) #define ROT4_1(x) ((x).s3012) #define ROT4_2(x) ((x).s2301) #define ROT4_3(x) ((x).s1230) #define ROT4_4(x) ((x)) #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 ROT8_8(x) ((x)) #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) #define ROT16_16(x) ((x)) #define ROTATE_STR(x, s, n) ROT##s##_##n(x) #define ROTATE(x, s, n) ROTATE_STR(x, s, n) #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) #define VEC_OFFS_STR(dt, s) V_OFFS##s(dt) #define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s) #define VLOAD_STR(size) vload##size #define VLOAD(size) VLOAD_STR(size) #define VLOAD_PARTIAL_STR(size, load_size) vload_partial_##size##_##load_size #define VLOAD_PARTIAL(size, load_size) VLOAD_PARTIAL_STR(size, load_size) #define NO_LOAD(data, offs, ptr) \ { \ } #define vload_partial_1_0 NO_LOAD #define vload_partial_1_1 vload1 #define vload_partial_1_2 NO_LOAD #define vload_partial_1_3 NO_LOAD #define vload_partial_1_4 NO_LOAD #define vload_partial_1_5 NO_LOAD #define vload_partial_1_6 NO_LOAD #define vload_partial_1_7 NO_LOAD #define vload_partial_1_8 NO_LOAD #define vload_partial_1_9 NO_LOAD #define vload_partial_1_10 NO_LOAD #define vload_partial_1_11 NO_LOAD #define vload_partial_1_12 NO_LOAD #define vload_partial_1_13 NO_LOAD #define vload_partial_1_14 NO_LOAD #define vload_partial_1_15 NO_LOAD #define vload_partial_1_16 NO_LOAD #define vload_partial_2_0 NO_LOAD #define vload_partial_2_1 vload_partial_1 #define vload_partial_2_2 vload_partial_2 #define vload_partial_2_3 NO_LOAD #define vload_partial_2_4 NO_LOAD #define vload_partial_2_5 NO_LOAD #define vload_partial_2_6 NO_LOAD #define vload_partial_2_7 NO_LOAD #define vload_partial_2_8 NO_LOAD #define vload_partial_2_9 NO_LOAD #define vload_partial_2_10 NO_LOAD #define vload_partial_2_11 NO_LOAD #define vload_partial_2_12 NO_LOAD #define vload_partial_2_13 NO_LOAD #define vload_partial_2_14 NO_LOAD #define vload_partial_2_15 NO_LOAD #define vload_partial_2_16 NO_LOAD #define vload_partial_3_0 NO_LOAD #define vload_partial_3_1 vload_partial_1 #define vload_partial_3_2 vload_partial_2 #define vload_partial_3_3 vload_partial_3 #define vload_partial_3_4 NO_LOAD #define vload_partial_3_5 NO_LOAD #define vload_partial_3_6 NO_LOAD #define vload_partial_3_7 NO_LOAD #define vload_partial_3_8 NO_LOAD #define vload_partial_3_9 NO_LOAD #define vload_partial_3_10 NO_LOAD #define vload_partial_3_11 NO_LOAD #define vload_partial_3_12 NO_LOAD #define vload_partial_3_13 NO_LOAD #define vload_partial_3_14 NO_LOAD #define vload_partial_3_15 NO_LOAD #define vload_partial_3_16 NO_LOAD #define vload_partial_4_0 NO_LOAD #define vload_partial_4_1 vload_partial_1 #define vload_partial_4_2 vload_partial_2 #define vload_partial_4_3 vload_partial_3 #define vload_partial_4_4 vload_partial_4 #define vload_partial_4_5 NO_LOAD #define vload_partial_4_6 NO_LOAD #define vload_partial_4_7 NO_LOAD #define vload_partial_4_8 NO_LOAD #define vload_partial_4_9 NO_LOAD #define vload_partial_4_10 NO_LOAD #define vload_partial_4_11 NO_LOAD #define vload_partial_4_12 NO_LOAD #define vload_partial_4_13 NO_LOAD #define vload_partial_4_14 NO_LOAD #define vload_partial_4_15 NO_LOAD #define vload_partial_4_16 NO_LOAD #define vload_partial_8_0 NO_LOAD #define vload_partial_8_1 vload_partial_1 #define vload_partial_8_2 vload_partial_2 #define vload_partial_8_3 vload_partial_3 #define vload_partial_8_4 vload_partial_4 #define vload_partial_8_5 vload_partial_5 #define vload_partial_8_6 vload_partial_6 #define vload_partial_8_7 vload_partial_7 #define vload_partial_8_8 vload_partial_8 #define vload_partial_8_9 NO_LOAD #define vload_partial_8_10 NO_LOAD #define vload_partial_8_11 NO_LOAD #define vload_partial_8_12 NO_LOAD #define vload_partial_8_13 NO_LOAD #define vload_partial_8_14 NO_LOAD #define vload_partial_8_15 NO_LOAD #define vload_partial_8_16 NO_LOAD #define vload_partial_16_0 NO_LOAD #define vload_partial_16_1 vload_partial_1 #define vload_partial_16_2 vload_partial_2 #define vload_partial_16_3 vload_partial_3 #define vload_partial_16_4 vload_partial_4 #define vload_partial_16_5 vload_partial_5 #define vload_partial_16_6 vload_partial_6 #define vload_partial_16_7 vload_partial_7 #define vload_partial_16_8 vload_partial_8 #define vload_partial_16_9 vload_partial_9 #define vload_partial_16_10 vload_partial_10 #define vload_partial_16_11 vload_partial_11 #define vload_partial_16_12 vload_partial_12 #define vload_partial_16_13 vload_partial_13 #define vload_partial_16_14 vload_partial_14 #define vload_partial_16_15 vload_partial_15 #define vload_partial_16_16 vload_partial_16 #define vload_partial_1(DATA, OFFSET, PTR) \ DATA.s0 = vload1(OFFSET, PTR); #define vload_partial_2(DATA, OFFSET, PTR) \ DATA.s01 = vload2(OFFSET, PTR); #define vload_partial_3(DATA, OFFSET, PTR) \ DATA.s012 = vload3(OFFSET, PTR); #define vload_partial_4(DATA, OFFSET, PTR) \ DATA.s0123 = vload4(OFFSET, PTR); #define vload_partial_5(DATA, OFFSET, PTR) \ vload_partial_4(DATA.s0123, OFFSET, PTR); \ DATA.s4 = vload1(OFFSET, PTR + 4); #define vload_partial_6(DATA, OFFSET, PTR) \ vload_partial_4(DATA.s0123, OFFSET, PTR); \ vload_partial_2(DATA.s45, OFFSET, PTR + 4); #define vload_partial_7(DATA, OFFSET, PTR) \ vload_partial_4(DATA.s0123, OFFSET, PTR); \ vload_partial_3(DATA.s456, OFFSET, PTR + 4); #define vload_partial_8(DATA, OFFSET, PTR) \ DATA.s01234567 = vload8(OFFSET, PTR); #define vload_partial_9(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ DATA.s8 = vload1(OFFSET, PTR + 8); #define vload_partial_10(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_2(DATA.s89, OFFSET, PTR + 8); #define vload_partial_11(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_3(DATA.s89A, OFFSET, PTR + 8); #define vload_partial_12(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_4(DATA.s89AB, OFFSET, PTR + 8); #define vload_partial_13(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_5(DATA.s89ABCDEF, OFFSET, PTR + 8); #define vload_partial_14(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_6(DATA.s89ABCDEF, OFFSET, PTR + 8); #define vload_partial_15(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_7(DATA.s89ABCDEF, OFFSET, PTR + 8); #define vload_partial_16(DATA, OFFSET, PTR) \ DATA = vload16(OFFSET, PTR); #define PIXEL_UNIT4 1 #define PIXEL_UNIT8 2 #define PIXEL_UNIT16 4 #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) #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 #define write_image2d_floatx1(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values)); #define write_image2d_floatx2(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567)); #define write_image2d_floatx4(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imagef(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imagef(img, (int2)(x_coord + 3, y_coord), values.sCDEF)); #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) #define write_image2d_halfx1(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values)); #define write_image2d_halfx2(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567)); #define write_image2d_halfx4(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imageh(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imageh(img, (int2)(x_coord + 3, y_coord), values.sCDEF)); #endif #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 WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values) write_image2d_##data_type##x##n0(img, x_coord, y_coord, values) #define WRITE_IMAGE2D(data_type, n0, img, x_coord, y_coord, values) WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values) #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 #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) \ { \ } #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 #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 #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 #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 #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 #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 #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); #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_uchar2_sat convert_uchar2_sat #define convert_uchar3_sat convert_uchar3_sat #define convert_uchar4_sat convert_uchar4_sat #define convert_uchar8_sat convert_uchar8_sat #define convert_uchar16_sat convert_uchar16_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 signed_int_vec_dt_uchar(size) char##size #define signed_int_vec_dt_char(size) char##size #define signed_int_vec_dt_ushort(size) short##size #define signed_int_vec_dt_short(size) short##size #define signed_int_vec_dt_half(size) short##size #define signed_int_vec_dt_uint(size) int##size #define signed_int_vec_dt_int(size) int##size #define signed_int_vec_dt_float(size) int##size #define signed_int_vec_dt_ulong(size) long##size #define signed_int_vec_dt_long(size) long##size #define SIGNED_INT_VEC_DATA_TYPE_STR(type, size) signed_int_vec_dt_##type(size) #define SIGNED_INT_VEC_DATA_TYPE(type, size) SIGNED_INT_VEC_DATA_TYPE_STR(type, size) #define SIGNED_INT_DATA_TYPE(type) SIGNED_INT_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 prod_reduce_1(x) (x) #define prod_reduce_2(x) ((x).s0) * ((x).s1) #define prod_reduce_3(x) prod_reduce_2((x).s01) * ((x).s2) #define prod_reduce_4(x) prod_reduce_2((x).s01) * prod_reduce_2((x).s23) #define prod_reduce_8(x) prod_reduce_4((x).s0123) * prod_reduce_4((x).s4567) #define prod_reduce_16(x) prod_reduce_8((x).s01234567) * prod_reduce_8((x).s89ABCDEF) #define PROD_REDUCE_STR(x, size) prod_reduce_##size(x) #define PROD_REDUCE(x, size) PROD_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 TENSOR5D_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##_stride_v, \ uint name##_step_v, \ 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) typedef struct Vector { __global uchar *ptr; int offset_first_element_in_bytes; int stride_x; } Vector; typedef struct Image { __global uchar *ptr; int offset_first_element_in_bytes; int stride_x; int stride_y; } Image; typedef struct Tensor3D { __global uchar *ptr; int offset_first_element_in_bytes; int stride_x; int stride_y; int stride_z; } Tensor3D; typedef struct Tensor4D { __global uchar *ptr; int offset_first_element_in_bytes; int stride_x; int stride_y; int stride_z; int stride_w; } Tensor4D; 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; } 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; } 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; } 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; } 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; } inline __global const uchar *vector_offset(const Vector *vec, int x) { return vec->ptr + x * vec->stride_x; } inline __global uchar *offset(const Image *img, int x, int y) { return img->ptr + x * img->stride_x + y * img->stride_y; } 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; } 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; } 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 #ifndef SRC_CORE_CL_CL_KERNELS_TILE_HELPERS #define SRC_CORE_CL_CL_KERNELS_TILE_HELPERS #define TILE_VECTOR_SIZE1 1 #define TILE_VECTOR_SIZE2 2 #define TILE_VECTOR_SIZE3 3 #define TILE_VECTOR_SIZE4 4 #define TILE_VECTOR_SIZE5 8 #define TILE_VECTOR_SIZE6 8 #define TILE_VECTOR_SIZE7 8 #define TILE_VECTOR_SIZE8 8 #define TILE_VECTOR_SIZE9 16 #define TILE_VECTOR_SIZE10 16 #define TILE_VECTOR_SIZE11 16 #define TILE_VECTOR_SIZE12 16 #define TILE_VECTOR_SIZE13 16 #define TILE_VECTOR_SIZE14 16 #define TILE_VECTOR_SIZE15 16 #define TILE_VECTOR_SIZE16 16 #define TILE_VECTOR_TYPE1(DATA_TYPE) DATA_TYPE##1 #define TILE_VECTOR_TYPE2(DATA_TYPE) DATA_TYPE##2 #define TILE_VECTOR_TYPE3(DATA_TYPE) DATA_TYPE##3 #define TILE_VECTOR_TYPE4(DATA_TYPE) DATA_TYPE##4 #define TILE_VECTOR_TYPE5(DATA_TYPE) DATA_TYPE##8 #define TILE_VECTOR_TYPE6(DATA_TYPE) DATA_TYPE##8 #define TILE_VECTOR_TYPE7(DATA_TYPE) DATA_TYPE##8 #define TILE_VECTOR_TYPE8(DATA_TYPE) DATA_TYPE##8 #define TILE_VECTOR_TYPE9(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE10(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE11(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE12(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE13(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE14(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE15(DATA_TYPE) DATA_TYPE##16 #define TILE_VECTOR_TYPE16(DATA_TYPE) DATA_TYPE##16 #define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME) #define TILE_STR(DATA_TYPE, H, W, BASENAME) \ union { \ DATA_TYPE s[TILE_VECTOR_SIZE##W]; \ TILE_VECTOR_TYPE##W(DATA_TYPE) v; \ } BASENAME[H] #define TENSOR4D_IMAGE(name) \ __read_only image2d_t name##_img, \ __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 TENSOR4D_BUFFER(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 TENSOR4D_STR(name, type) TENSOR4D_##type(name) #define TENSOR4D(name, type) TENSOR4D_STR(name, type) #define TENSOR4D_T_IMAGE(name) \ __read_only image2d_t name##_img, \ __global uchar *name##_ptr, \ uint name##_stride_y, \ uint name##_stride_z, \ uint name##_stride_w, \ uint name##_c, \ uint name##_w, \ uint name##_h, \ uint name##_n, \ uint name##_offset_first_element_in_bytes #define TENSOR4D_T_BUFFER(name) \ __global uchar *name##_ptr, \ uint name##_stride_y, \ uint name##_stride_z, \ uint name##_stride_w, \ uint name##_c, \ uint name##_w, \ uint name##_h, \ uint name##_n, \ uint name##_offset_first_element_in_bytes #define TENSOR4D_T_STR(name, type) TENSOR4D_T_##type(name) #define TENSOR4D_T(name, type) TENSOR4D_T_STR(name, type) #define TENSOR4D_RO_T_IMAGE(name) \ __read_only image2d_t name##_img, \ TENSOR4D_T_BUFFER(name) #define TENSOR4D_RO_T_BUFFER(name) TENSOR4D_T_BUFFER(name) #define TENSOR4D_RO_T_STR(name, type) TENSOR4D_RO_T_##type(name) #define TENSOR4D_RO_T(name, type) TENSOR4D_RO_T_STR(name, type) #define TENSOR4D_WO_T_IMAGE(name) \ __write_only image2d_t name##_img, \ TENSOR4D_T_BUFFER(name) #define TENSOR4D_WO_T_BUFFER(name) TENSOR4D_T_BUFFER(name) #define TENSOR4D_WO_T_STR(name, type) TENSOR4D_WO_T_##type(name) #define TENSOR4D_WO_T(name, type) TENSOR4D_WO_T_STR(name, type) #define TENSOR3D_T_IMAGE(name) \ __read_only image2d_t name##_img, \ __global uchar *name##_ptr, \ uint name##_stride_y, \ uint name##_stride_z, \ uint name##_w, \ uint name##_h, \ uint name##_n, \ uint name##_offset_first_element_in_bytes #define TENSOR3D_T_BUFFER(name) \ __global uchar *name##_ptr, \ uint name##_stride_y, \ uint name##_stride_z, \ uint name##_w, \ uint name##_h, \ uint name##_n, \ uint name##_offset_first_element_in_bytes #define TENSOR3D_T_STR(name, type) TENSOR3D_T_##type(name) #define TENSOR3D_T(name, type) TENSOR3D_T_STR(name, type) #if !defined(UNROLL_WITH_PRAGMA) #define UNROLL_INCR(idx, step, macro) idx += (step); (macro) #define LOOP_UNROLLING_1(idx, step, macro) (macro) #define LOOP_UNROLLING_2(idx, step, macro) LOOP_UNROLLING_1(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_3(idx, step, macro) LOOP_UNROLLING_2(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_4(idx, step, macro) LOOP_UNROLLING_3(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_5(idx, step, macro) LOOP_UNROLLING_4(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_6(idx, step, macro) LOOP_UNROLLING_5(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_7(idx, step, macro) LOOP_UNROLLING_6(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_8(idx, step, macro) LOOP_UNROLLING_7(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_9(idx, step, macro) LOOP_UNROLLING_8(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_10(idx, step, macro) LOOP_UNROLLING_9(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_11(idx, step, macro) LOOP_UNROLLING_10(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_12(idx, step, macro) LOOP_UNROLLING_11(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_13(idx, step, macro) LOOP_UNROLLING_12(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_14(idx, step, macro) LOOP_UNROLLING_13(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_15(idx, step, macro) LOOP_UNROLLING_14(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_16(idx, step, macro) LOOP_UNROLLING_15(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_17(idx, step, macro) LOOP_UNROLLING_16(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_18(idx, step, macro) LOOP_UNROLLING_17(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_19(idx, step, macro) LOOP_UNROLLING_18(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_20(idx, step, macro) LOOP_UNROLLING_19(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_21(idx, step, macro) LOOP_UNROLLING_20(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_22(idx, step, macro) LOOP_UNROLLING_21(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_23(idx, step, macro) LOOP_UNROLLING_22(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_24(idx, step, macro) LOOP_UNROLLING_23(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_25(idx, step, macro) LOOP_UNROLLING_24(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_26(idx, step, macro) LOOP_UNROLLING_25(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_27(idx, step, macro) LOOP_UNROLLING_26(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_28(idx, step, macro) LOOP_UNROLLING_27(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_29(idx, step, macro) LOOP_UNROLLING_28(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_30(idx, step, macro) LOOP_UNROLLING_29(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_31(idx, step, macro) LOOP_UNROLLING_30(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_32(idx, step, macro) LOOP_UNROLLING_31(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_33(idx, step, macro) LOOP_UNROLLING_32(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_34(idx, step, macro) LOOP_UNROLLING_33(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_35(idx, step, macro) LOOP_UNROLLING_34(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_36(idx, step, macro) LOOP_UNROLLING_35(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_37(idx, step, macro) LOOP_UNROLLING_36(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_38(idx, step, macro) LOOP_UNROLLING_37(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_39(idx, step, macro) LOOP_UNROLLING_38(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_40(idx, step, macro) LOOP_UNROLLING_39(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_41(idx, step, macro) LOOP_UNROLLING_40(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_42(idx, step, macro) LOOP_UNROLLING_41(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_43(idx, step, macro) LOOP_UNROLLING_42(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_44(idx, step, macro) LOOP_UNROLLING_43(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_45(idx, step, macro) LOOP_UNROLLING_44(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_46(idx, step, macro) LOOP_UNROLLING_45(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_47(idx, step, macro) LOOP_UNROLLING_46(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_48(idx, step, macro) LOOP_UNROLLING_47(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_49(idx, step, macro) LOOP_UNROLLING_48(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_50(idx, step, macro) LOOP_UNROLLING_49(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_51(idx, step, macro) LOOP_UNROLLING_50(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_52(idx, step, macro) LOOP_UNROLLING_51(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_53(idx, step, macro) LOOP_UNROLLING_52(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_54(idx, step, macro) LOOP_UNROLLING_53(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_55(idx, step, macro) LOOP_UNROLLING_54(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_56(idx, step, macro) LOOP_UNROLLING_55(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_57(idx, step, macro) LOOP_UNROLLING_56(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_58(idx, step, macro) LOOP_UNROLLING_57(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_59(idx, step, macro) LOOP_UNROLLING_58(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_60(idx, step, macro) LOOP_UNROLLING_59(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_61(idx, step, macro) LOOP_UNROLLING_60(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_62(idx, step, macro) LOOP_UNROLLING_61(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_63(idx, step, macro) LOOP_UNROLLING_62(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_64(idx, step, macro) LOOP_UNROLLING_63(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_65(idx, step, macro) LOOP_UNROLLING_64(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_66(idx, step, macro) LOOP_UNROLLING_65(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_67(idx, step, macro) LOOP_UNROLLING_66(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_68(idx, step, macro) LOOP_UNROLLING_67(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_69(idx, step, macro) LOOP_UNROLLING_68(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_70(idx, step, macro) LOOP_UNROLLING_69(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_71(idx, step, macro) LOOP_UNROLLING_70(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_72(idx, step, macro) LOOP_UNROLLING_71(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_73(idx, step, macro) LOOP_UNROLLING_72(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_74(idx, step, macro) LOOP_UNROLLING_73(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_75(idx, step, macro) LOOP_UNROLLING_74(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_76(idx, step, macro) LOOP_UNROLLING_75(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_77(idx, step, macro) LOOP_UNROLLING_76(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_78(idx, step, macro) LOOP_UNROLLING_77(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_79(idx, step, macro) LOOP_UNROLLING_78(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_80(idx, step, macro) LOOP_UNROLLING_79(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_81(idx, step, macro) LOOP_UNROLLING_80(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_82(idx, step, macro) LOOP_UNROLLING_81(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_83(idx, step, macro) LOOP_UNROLLING_82(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_84(idx, step, macro) LOOP_UNROLLING_83(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_85(idx, step, macro) LOOP_UNROLLING_84(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_86(idx, step, macro) LOOP_UNROLLING_85(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_87(idx, step, macro) LOOP_UNROLLING_86(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_88(idx, step, macro) LOOP_UNROLLING_87(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_89(idx, step, macro) LOOP_UNROLLING_88(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_90(idx, step, macro) LOOP_UNROLLING_89(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_91(idx, step, macro) LOOP_UNROLLING_90(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_92(idx, step, macro) LOOP_UNROLLING_91(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_93(idx, step, macro) LOOP_UNROLLING_92(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_94(idx, step, macro) LOOP_UNROLLING_93(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_95(idx, step, macro) LOOP_UNROLLING_94(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_96(idx, step, macro) LOOP_UNROLLING_95(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_97(idx, step, macro) LOOP_UNROLLING_96(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_98(idx, step, macro) LOOP_UNROLLING_97(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_99(idx, step, macro) LOOP_UNROLLING_98(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_100(idx, step, macro) LOOP_UNROLLING_99(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_101(idx, step, macro) LOOP_UNROLLING_100(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_102(idx, step, macro) LOOP_UNROLLING_101(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_103(idx, step, macro) LOOP_UNROLLING_102(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_104(idx, step, macro) LOOP_UNROLLING_103(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_105(idx, step, macro) LOOP_UNROLLING_104(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_106(idx, step, macro) LOOP_UNROLLING_105(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_107(idx, step, macro) LOOP_UNROLLING_106(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_108(idx, step, macro) LOOP_UNROLLING_107(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_109(idx, step, macro) LOOP_UNROLLING_108(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_110(idx, step, macro) LOOP_UNROLLING_109(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_111(idx, step, macro) LOOP_UNROLLING_110(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_112(idx, step, macro) LOOP_UNROLLING_111(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_113(idx, step, macro) LOOP_UNROLLING_112(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_114(idx, step, macro) LOOP_UNROLLING_113(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_115(idx, step, macro) LOOP_UNROLLING_114(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_116(idx, step, macro) LOOP_UNROLLING_115(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_117(idx, step, macro) LOOP_UNROLLING_116(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_118(idx, step, macro) LOOP_UNROLLING_117(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_119(idx, step, macro) LOOP_UNROLLING_118(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_120(idx, step, macro) LOOP_UNROLLING_119(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_121(idx, step, macro) LOOP_UNROLLING_120(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_122(idx, step, macro) LOOP_UNROLLING_121(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_123(idx, step, macro) LOOP_UNROLLING_122(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_124(idx, step, macro) LOOP_UNROLLING_123(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_125(idx, step, macro) LOOP_UNROLLING_124(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_126(idx, step, macro) LOOP_UNROLLING_125(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro) #define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \ { \ type idx = start; \ LOOP_UNROLLING_##num(idx, step, macro); \ } #else #define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \ { \ _Pragma("unroll") \ for(type idx = start; idx < (num * step); idx += step) \ { \ (macro); \ } \ } #endif #define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro) #define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0)) #define DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) #define DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT##K0##_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) #define DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ c += (C_DATA_TYPE)(a) * (C_DATA_TYPE)(b); \ }) #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product) #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0))); #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0)); #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((a), (b)); #elif defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)), (c)); #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0), (c)); #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((a), (b), (c)); #elif defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0))); #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0)); #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((a), (b)); #else #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ c += (C_DATA_TYPE)(a).s0 * (C_DATA_TYPE)(b).s0; \ c += (C_DATA_TYPE)(a).s1 * (C_DATA_TYPE)(b).s1; \ }) #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c); \ c += (C_DATA_TYPE)(a).s2 * (C_DATA_TYPE)(b).s2; \ }) #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, x, y, val) \ ({ \ val += (C_DATA_TYPE)(x).s0 * (C_DATA_TYPE)(y).s0; \ val += (C_DATA_TYPE)(x).s1 * (C_DATA_TYPE)(y).s1; \ val += (C_DATA_TYPE)(x).s2 * (C_DATA_TYPE)(y).s2; \ val += (C_DATA_TYPE)(x).s3 * (C_DATA_TYPE)(y).s3; \ }) #endif #define DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s4), ((b).s4), c); \ }) #define DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s45), ((b).s45), c); \ }) #define DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s456), ((b).s456), c); \ }) #define DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \ }) #define DOT_PRODUCT9_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s8), ((b).s8), c); \ }) #define DOT_PRODUCT10_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89), ((b).s89), c); \ }) #define DOT_PRODUCT11_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89A), ((b).s89A), c); \ }) #define DOT_PRODUCT12_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89AB), ((b).s89AB), c); \ }) #define DOT_PRODUCT13_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABC), ((b).s89ABC), c); \ }) #define DOT_PRODUCT14_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCD), ((b).s89ABCD), c); \ }) #define DOT_PRODUCT15_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCDE), ((b).s89ABCDE), c); \ }) #define DOT_PRODUCT16_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ ({ \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \ DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \ }) #define REDUCE_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) #define REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, (TILE_VECTOR_TYPE##K0(B_DATA_TYPE))1, c) #define V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) #define V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) #define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \ VLOAD(WIDTH) \ (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y))) #define V_LOAD_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) READ_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y)) #define V_STORE(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) #define V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) #define V_STORE_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) \ VSTORE(WIDTH) \ (VALUES, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y))) #define V_STORE_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) WRITE_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y), VALUES) #define T_LOAD(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \ }) \ }) #define T_LOAD_INDIRECT(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, STRIDE_Y, indirect_y, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, (indirect_y[_i].v), STRIDE_Y); \ }) \ }) #define T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, dst, indirect_y) \ ({ \ if(WIDTH1_CONDITION) \ { \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ VLOAD_PARTIAL(WIDTH0, WIDTH1) \ (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ }) \ } \ else \ { \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, (indirect_y[HEIGHT - 1 - _i].v), STRIDE_Y); \ }) \ } \ }) #define T_LOAD_NHWC(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, dst) \ ({ \ LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \ { \ LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \ { \ int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH); \ _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \ int _src_valid_y = (((X) + _xk) >= 0 && ((X) + _xk) < (int)(TENSOR_WIDTH) && ((Y) + _yk) >= 0 && ((Y) + _yk) < (int)(TENSOR_HEIGHT)); \ if(_src_valid_y != 0) \ { \ dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ } \ }) \ }) \ }) #define T_LOAD_NHWC_WITH_DILATION(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, DILATION_X, DILATION_Y, BOUNDARY_CHECK, dst) \ ({ \ LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \ { \ LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \ { \ int _src_y = (X) + _xk * (DILATION_X); \ int _src_z = ((Y) + _yk * (DILATION_Y)); \ int _src_w = (B); \ bool _src_valid_y = (((X) + _xk * (DILATION_X)) >= 0) && (((X) + _xk * (DILATION_X)) < (int)(TENSOR_WIDTH)) && (((Y) + _yk * (DILATION_Y)) >= 0) && (((Y) + _yk * (DILATION_Y)) < (int)(TENSOR_HEIGHT)); \ if(!(BOUNDARY_CHECK)) \ { \ dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \ (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \ } \ else \ { \ if(_src_valid_y) \ { \ dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \ (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \ } \ } \ }) \ }) \ }) #define T_LOAD_NHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, xi, yi, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ { \ int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH); \ _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \ int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)); \ if(_src_valid_y != 0) \ { \ dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ } \ }) \ }) #define T_LOAD2D_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) #define T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_##TENSOR_TYPE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) #define T_LOAD2D_INDIRECT_BUFFER(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ { \ if(yi[0].s[_i] >= 0) \ { \ dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \ } \ }) \ }) #define T_LOAD2D_INDIRECT_IMAGE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ { \ dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \ }) \ }) #define T_LOAD_NDHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Z, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, TENSOR_DEPTH, STRIDE_Y, xi, yi, zi, dst) \ ({ \ LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ { \ int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH) + ((Z) + zi[_i].v) * (TENSOR_WIDTH * TENSOR_HEIGHT); \ _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT) * (int)(TENSOR_DEPTH); \ int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT) \ && ((Z) + zi[_i].v) >= 0 && ((Z) + zi[_i].v) < (int)(TENSOR_DEPTH)); \ if(_src_valid_y != 0) \ { \ dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ } \ }) \ }) #define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y) \ ({ \ if(WIDTH1_CONDITION) \ { \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ VSTORE_PARTIAL(WIDTH0, WIDTH1) \ (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ }) \ } \ else \ { \ LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ { \ VSTORE(WIDTH0) \ (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ }) \ } \ }) #define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ ACC_DATA_TYPE _tm = 0; \ LOOP_UNROLLING(int, _k0, 0, 1, K0, \ { \ _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET); \ }) \ LOOP_UNROLLING(int, _n0, 0, 1, N0, \ { \ dst[_m0].s[_n0] += _tm; \ LOOP_UNROLLING(int, _k0, 0, 1, K0, \ { \ dst[_m0].s[_n0] += ((ACC_DATA_TYPE)rhs[_n0].s[_k0] * (ACC_DATA_TYPE)SRC_OFFSET); \ }) \ }) \ }) \ }) #define T_QUANTIZE8(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) #define T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_##QUANTIZATION_TYPE(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) #define T_QUANTIZE8_PER_TENSOR(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n0, 0, 1, N0, \ { \ SRC_DATA_TYPE _tmp = 0; \ SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \ SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \ long a_64 = (long)(_src); \ long b_64 = (long)(DST_MULTIPLIER); \ long ab_64 = a_64 * b_64; \ long mask1 = 1 << 30; \ long mask2 = 1 - (1 << 30); \ long is_positive_or_zero = ab_64 >= 0; \ long nudge = select(mask2, mask1, is_positive_or_zero); \ SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ if(DST_SHIFT >= 0) \ { \ long mask = ((((int)1) << DST_SHIFT) - (long)1); \ long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \ _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \ } \ _tmp += DST_OFFSET; \ dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ }) \ }) \ }) #define T_QUANTIZE8_PER_CHANNEL(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n0, 0, 1, N0, \ { \ SRC_DATA_TYPE _tmp = 0; \ SRC_DATA_TYPE _tmp2 = 0; \ SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ SRC_DATA_TYPE _dst_multiplier = dst_multipliers[0].s[_n0]; \ SRC_DATA_TYPE _dst_shift = dst_shifts[0].s[_n0]; \ _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-_dst_shift)), ((SRC_DATA_TYPE)_dst_shift < (SRC_DATA_TYPE)0)); \ SRC_DATA_TYPE overflow = _src == _dst_multiplier && _src == INT_MIN; \ long a_64 = (long)(_src); \ long b_64 = (long)(_dst_multiplier); \ long ab_64 = a_64 * b_64; \ long mask1 = 1 << 30; \ long mask2 = 1 - (1 << 30); \ long is_positive_or_zero = ab_64 >= 0; \ long nudge = select(mask2, mask1, is_positive_or_zero); \ SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ long mask = ((((int)1) << _dst_shift) - (int)1); \ long threshold = (mask >> 1) + any(_tmp); \ _tmp2 = _tmp >> _dst_shift; \ _tmp2 += select(0, 1, (_tmp & mask) > threshold); \ _tmp = select(_tmp, _tmp2, _dst_shift >= 0); \ _tmp += DST_OFFSET; \ dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ }) \ }) \ }) #define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n0, 0, 1, N0, \ { \ SRC_DATA_TYPE _tmp = 0; \ SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \ SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \ long a_64 = (long)(_src); \ long b_64 = (long)(DST_MULTIPLIER); \ long ab_64 = a_64 * b_64; \ long mask1 = 1 << 30; \ long mask2 = 1 - (1 << 30); \ long is_positive_or_zero = ab_64 >= 0; \ long nudge = select(mask2, mask1, is_positive_or_zero); \ SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ if(DST_SHIFT >= 0) \ { \ long mask = ((((int)1) << DST_SHIFT) - (int)1); \ long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \ _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \ } \ _tmp += DST_OFFSET; \ dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ }) \ }) \ }) #define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n0, 0, 1, N0, \ { \ a[_m0].s[_n0] = select((DATA_TYPE)(a[_m0].s[_n0]), (DATA_TYPE)(VALUE_TO_SET), (SELECT_DATA_TYPE(DATA_TYPE))(mask[_m0].v == (DATA_TYPE)0)); \ }) \ }) \ }) #define T_ACTIVATION(DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, src, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, N0, src[_m0].v, A_VAL, B_VAL); \ }) \ }) #define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x)) #define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_VALUE, x))) #define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL)) #define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f)) #define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x) #define ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) #define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) #define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL)) #define V_SUB(A_VAL, B_VAL) ((A_VAL) - (B_VAL)) #define V_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL)) #define V_MUL(A_VAL, B_VAL) ((A_VAL) * (B_VAL)) #define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \ }) \ }) #define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \ }) \ }) #define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant; \ }) \ }) #define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_LHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_RHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_LHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_RHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_LHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_BROADCAST_RHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \ }) \ }) #define T_ELTWISE_BROADCAST_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ }) \ }) #define T_ELTWISE_BROADCAST_LHS_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ }) \ }) #define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) #define T_ELTWISE(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ }) \ }) #define T_FLOOR(DST_DATA_TYPE, M0, N0, src, dst) \ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ dst[_m0].v = floor(CONVERT(src[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ }) \ }) #define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_char_char_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_uchar_uchar_uint(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_uchar_uchar_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) #define T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ { \ LOOP_UNROLLING(int, _m, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n, 0, 1, N0, \ { \ LOOP_UNROLLING(int, _k, 0, 1, K0, \ { \ dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_m].s[_k]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \ }) \ }) \ }) \ } #define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m, 0, 1, M0, \ { \ LOOP_UNROLLING(int, _n, 0, 1, N0, \ { \ DOT_PRODUCT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \ }) \ }) \ }) #endif #define OUTPUT_ROW_4x4_5x5(out, tmp, comm_fact) \ ({ \ comm_fact.s0 = tmp.s2 - 4.25f * tmp.s4 + tmp.s6; \ comm_fact.s1 = tmp.s1 - 4.25f * tmp.s3 + tmp.s5; \ comm_fact.s2 = 2.5f * tmp.s3; \ comm_fact.s3 = 0.5f * tmp.s1 + 2.f * tmp.s5 - comm_fact.s2; \ comm_fact.s4 = 0.25f * tmp.s2 - 1.25f * tmp.s4 + tmp.s6; \ comm_fact.s5 = 4.f * tmp.s2 + tmp.s6 - 5.f * tmp.s4; \ comm_fact.s6 = 2.f * tmp.s1 + 0.5f * tmp.s5 - comm_fact.s2; \ \ out.s0 = tmp.s0 - tmp.s6 + 5.25f * tmp.s4 - 5.25f * tmp.s2; \ out.s1 = comm_fact.s0 + comm_fact.s1; \ out.s2 = comm_fact.s0 - comm_fact.s1; \ out.s3 = comm_fact.s3 + comm_fact.s4; \ out.s4 = comm_fact.s4 - comm_fact.s3; \ out.s5 = comm_fact.s5 + comm_fact.s6; \ out.s6 = comm_fact.s5 - comm_fact.s6; \ out.s7 = tmp.s7 - tmp.s1 + 5.25f * tmp.s3 - 5.25f * tmp.s5; \ }) #define OUTPUT_ROW_2x2_7x7(out, tmp, comm_fact) \ ({ \ comm_fact.s0 = 36.0f * tmp.s2 - 13.0f * tmp.s4 + tmp.s6; \ comm_fact.s1 = 36.0f * tmp.s1 - 13.0f * tmp.s3 + 1.0f * tmp.s5; \ comm_fact.s2 = 9.0f * tmp.s2 - 10.0f * tmp.s4 + tmp.s6; \ comm_fact.s3 = 18.0f * tmp.s1 - 20.0f * tmp.s3 + 2.0f * tmp.s5; \ comm_fact.s4 = 4.0f * tmp.s2 - 5.0f * tmp.s4 + tmp.s6; \ comm_fact.s5 = 12.0f * tmp.s1 - 15.0f * tmp.s3 + 3.0f * tmp.s5; \ out.s0 = -36.0f * tmp.s0 + 49.0f * tmp.s2 + -14.0f * tmp.s4 + tmp.s6; \ out.s1 = comm_fact.s0 - comm_fact.s1; \ out.s2 = comm_fact.s0 + comm_fact.s1; \ out.s3 = comm_fact.s2 - comm_fact.s3; \ out.s4 = comm_fact.s2 + comm_fact.s3; \ out.s5 = comm_fact.s4 - comm_fact.s5; \ out.s6 = comm_fact.s4 + comm_fact.s5; \ out.s7 = -36.0f * tmp.s1 + 0.0f * tmp.s2 + 49.0f * tmp.s3 - 14.0f * tmp.s5 + tmp.s7; \ }) #if defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(OUTPUT_TILE_W) && defined(OUTPUT_TILE_H) __kernel void winograd_input_transform_2x2_3x3_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { const int x = get_global_id(0); const int y = get_global_id(1); #if defined(SRC_DEPTH) const int z = get_global_id(2) % SRC_DEPTH; const int b = get_global_id(2) / SRC_DEPTH; #else const int z = get_global_id(2); #endif #if defined(SRC_DEPTH) __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z + b * src_stride_w; #else __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; #endif src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y); #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr)); #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y))); #else VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row1 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row2 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row3 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); #endif VEC_DATA_TYPE(DATA_TYPE, 4) tmp0 = in_row0; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) tmp0 -= in_row2; #endif DATA_TYPE out00 = tmp0.s0 - tmp0.s2; DATA_TYPE out01 = tmp0.s1 + tmp0.s2; DATA_TYPE out02 = tmp0.s2 - tmp0.s1; DATA_TYPE out03 = tmp0.s1 - tmp0.s3; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) tmp1 = in_row1 + in_row2; VEC_DATA_TYPE(DATA_TYPE, 4) tmp2 = in_row2 - in_row1; VEC_DATA_TYPE(DATA_TYPE, 4) tmp3 = in_row1 - in_row3; DATA_TYPE out10 = tmp1.s0 - tmp1.s2; DATA_TYPE out11 = tmp1.s1 + tmp1.s2; DATA_TYPE out12 = tmp1.s2 - tmp1.s1; DATA_TYPE out13 = tmp1.s1 - tmp1.s3; DATA_TYPE out20 = tmp2.s0 - tmp2.s2; DATA_TYPE out21 = tmp2.s1 + tmp2.s2; DATA_TYPE out22 = tmp2.s2 - tmp2.s1; DATA_TYPE out23 = tmp2.s1 - tmp2.s3; DATA_TYPE out30 = tmp3.s0 - tmp3.s2; DATA_TYPE out31 = tmp3.s1 + tmp3.s2; DATA_TYPE out32 = tmp3.s2 - tmp3.s1; DATA_TYPE out33 = tmp3.s1 - tmp3.s3; #endif #if defined(SRC_DEPTH) __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y + b * dst_stride_w; #else __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y; #endif *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)) = out00; *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)) = out01; *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)) = out02; *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)) = out03; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) *((__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)) = out10; *((__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)) = out11; *((__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)) = out12; *((__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)) = out13; *((__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)) = out20; *((__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)) = out21; *((__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)) = out22; *((__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)) = out23; *((__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)) = out30; *((__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)) = out31; *((__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)) = out32; *((__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)) = out33; #endif } __kernel void winograd_input_transform_2x2_3x3_stepz2_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { const int x = get_global_id(0); const int y = get_global_id(1); #if defined(SRC_DEPTH) const int z = (get_global_id(2) * 2) % SRC_DEPTH; const int b = (get_global_id(2) * 2) / SRC_DEPTH; #else const int z = get_global_id(2) * 2; #endif #if defined(SRC_DEPTH) __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z + b * src_stride_w; #else __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; #endif src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y); #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr)); #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y))); #else VEC_DATA_TYPE(DATA_TYPE, 4) in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row1 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row2 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row3 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); #endif src_addr += src_stride_z; #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row4 = vload4(0, (__global DATA_TYPE *)(src_addr)); #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) in_row4 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y))); #else VEC_DATA_TYPE(DATA_TYPE, 4) in_row4 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row5 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row6 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) in_row7 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); #endif VEC_DATA_TYPE(DATA_TYPE, 4) tmp0 = in_row0; VEC_DATA_TYPE(DATA_TYPE, 4) tmp4 = in_row4; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) tmp0 -= in_row2; tmp4 -= in_row6; #endif VEC_DATA_TYPE(DATA_TYPE, 2) out00 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s0 - tmp0.s2, tmp4.s0 - tmp4.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out01 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s1 + tmp0.s2, tmp4.s1 + tmp4.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out02 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s2 - tmp0.s1, tmp4.s2 - tmp4.s1); VEC_DATA_TYPE(DATA_TYPE, 2) out03 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s1 - tmp0.s3, tmp4.s1 - tmp4.s3); #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) tmp1 = in_row1 + in_row2; VEC_DATA_TYPE(DATA_TYPE, 4) tmp2 = in_row2 - in_row1; VEC_DATA_TYPE(DATA_TYPE, 4) tmp3 = in_row1 - in_row3; VEC_DATA_TYPE(DATA_TYPE, 4) tmp5 = in_row5 + in_row6; VEC_DATA_TYPE(DATA_TYPE, 4) tmp6 = in_row6 - in_row5; VEC_DATA_TYPE(DATA_TYPE, 4) tmp7 = in_row5 - in_row7; VEC_DATA_TYPE(DATA_TYPE, 2) out10 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s0 - tmp1.s2, tmp5.s0 - tmp5.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out11 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s1 + tmp1.s2, tmp5.s1 + tmp5.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out12 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s2 - tmp1.s1, tmp5.s2 - tmp5.s1); VEC_DATA_TYPE(DATA_TYPE, 2) out13 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s1 - tmp1.s3, tmp5.s1 - tmp5.s3); VEC_DATA_TYPE(DATA_TYPE, 2) out20 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s0 - tmp2.s2, tmp6.s0 - tmp6.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out21 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s1 + tmp2.s2, tmp6.s1 + tmp6.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out22 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s2 - tmp2.s1, tmp6.s2 - tmp6.s1); VEC_DATA_TYPE(DATA_TYPE, 2) out23 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s1 - tmp2.s3, tmp6.s1 - tmp6.s3); VEC_DATA_TYPE(DATA_TYPE, 2) out30 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s0 - tmp3.s2, tmp7.s0 - tmp7.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out31 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s1 + tmp3.s2, tmp7.s1 + tmp7.s2); VEC_DATA_TYPE(DATA_TYPE, 2) out32 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s2 - tmp3.s1, tmp7.s2 - tmp7.s1); VEC_DATA_TYPE(DATA_TYPE, 2) out33 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s1 - tmp3.s3, tmp7.s1 - tmp7.s3); #endif #if defined(SRC_DEPTH) __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y + b * dst_stride_w; #else __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y; #endif vstore2(out00, 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)); vstore2(out01, 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)); vstore2(out02, 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)); vstore2(out03, 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)); #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) vstore2(out10, 0, (__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)); vstore2(out11, 0, (__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)); vstore2(out12, 0, (__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)); vstore2(out13, 0, (__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)); vstore2(out20, 0, (__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)); vstore2(out21, 0, (__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)); vstore2(out22, 0, (__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)); vstore2(out23, 0, (__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)); vstore2(out30, 0, (__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)); vstore2(out31, 0, (__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)); vstore2(out32, 0, (__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)); vstore2(out33, 0, (__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)); #endif } __kernel void winograd_input_transform_4x4_3x3_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { const int x = get_global_id(0); const int y = get_global_id(1); #if defined(SRC_DEPTH) const int z = get_global_id(2) % SRC_DEPTH; const int b = get_global_id(2) / SRC_DEPTH; #else const int z = get_global_id(2); #endif #if defined(SRC_DEPTH) __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z + b * src_stride_w; #else __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; #endif src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y); #if defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) d00 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y))); VEC_DATA_TYPE(DATA_TYPE, 2) d01 = (VEC_DATA_TYPE(DATA_TYPE, 2))(*((__global DATA_TYPE *)(src_addr + 4 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 5 * src_stride_y))); #else VEC_DATA_TYPE(DATA_TYPE, 4) d00 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d01 = vload2(2, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); #endif DATA_TYPE out0 = 0.0f; DATA_TYPE out1 = 0.0f; DATA_TYPE out2 = 0.0f; DATA_TYPE out3 = 0.0f; DATA_TYPE out4 = 0.0f; DATA_TYPE out5 = 0.0f; out0 += 16.0f * d00.s0 - 20.0f * d00.s2 + 4.0f * d01.s0; out1 += -16.0f * d00.s1 - 16.0f * d00.s2 + 4.0f * d00.s3 + 4.0f * d01.s0; out2 += 16.0f * d00.s1 - 16.0f * d00.s2 - 4.0f * d00.s3 + 4.0f * d01.s0; out3 += -8.0f * d00.s1 - 4.0f * d00.s2 + 8.0f * d00.s3 + 4.0f * d01.s0; out4 += 8.0f * d00.s1 - 4.0f * d00.s2 - 8.0f * d00.s3 + 4.0f * d01.s0; out5 += 16.0f * d00.s1 - 20.0f * d00.s3 + 4.0f * d01.s1; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 4) d40 = vload4(0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d41 = vload2(2, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y)); DATA_TYPE k0 = d41.s0; DATA_TYPE k1 = d41.s0; DATA_TYPE k2 = d41.s0; DATA_TYPE k3 = d41.s0; DATA_TYPE k4 = d41.s0; DATA_TYPE k5 = 0.0f; k0 += 4.0f * d40.s0 - 5.0f * d40.s2; k1 += -4.0f * d40.s1 - 4.0f * d40.s2 + d40.s3; k2 += 4.0f * d40.s1 - 4.0f * d40.s2 - d40.s3; k3 += -2.0f * d40.s1 + 2.0f * d40.s3 - d40.s2; k4 += 2.0f * d40.s1 - 2.0f * d40.s3 - d40.s2; k5 += 4.0f * d40.s1 - 5.0f * d40.s3 + d41.s1; out0 += k0; out1 += k1; out2 += k2; out3 += k3; out4 += k4; out5 += k5; VEC_DATA_TYPE(DATA_TYPE, 4) d20 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d21 = vload2(2, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); out0 += -20.0f * d20.s0 + 25.0f * d20.s2 - 5.0f * d21.s0; out1 += +20.0f * d20.s1 + 20.0f * d20.s2 - 5.0f * d20.s3 - 5.0f * d21.s0; out2 += -20.0f * d20.s1 + 20.0f * d20.s2 + 5.0f * d20.s3 - 5.0f * d21.s0; out3 += +10.0f * d20.s1 + 5.0f * d20.s2 - 10.0f * d20.s3 - 5.0f * d21.s0; out4 += -10.0f * d20.s1 + 5.0f * d20.s2 + 10.0f * d20.s3 - 5.0f * d21.s0; out5 += -20.0f * d20.s1 + 25.0f * d20.s3 - 5.0f * d21.s1; #endif #if defined(SRC_DEPTH) __global DATA_TYPE *dst_addr = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y + b * dst_stride_w); #else __global DATA_TYPE *dst_addr = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y); #endif uint dst_plane_stride = dst_stride_z / sizeof(DATA_TYPE); *(dst_addr) = out0; dst_addr += dst_plane_stride; *(dst_addr) = out1; dst_addr += dst_plane_stride; *(dst_addr) = out2; dst_addr += dst_plane_stride; *(dst_addr) = out3; dst_addr += dst_plane_stride; *(dst_addr) = out4; dst_addr += dst_plane_stride; *(dst_addr) = out5; dst_addr += dst_plane_stride; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) DATA_TYPE out6 = k0; DATA_TYPE out7 = k1; DATA_TYPE out8 = k2; DATA_TYPE out9 = k3; DATA_TYPE out10 = k4; DATA_TYPE out11 = k5; DATA_TYPE out12 = k0; DATA_TYPE out13 = k1; DATA_TYPE out14 = k2; DATA_TYPE out15 = k3; DATA_TYPE out16 = k4; DATA_TYPE out17 = k5; DATA_TYPE out18 = k0; DATA_TYPE out19 = k1; DATA_TYPE out20 = k2; DATA_TYPE out21 = k3; DATA_TYPE out22 = k4; DATA_TYPE out23 = k5; DATA_TYPE out24 = k0; DATA_TYPE out25 = k1; DATA_TYPE out26 = k2; DATA_TYPE out27 = k3; DATA_TYPE out28 = k4; DATA_TYPE out29 = k5; VEC_DATA_TYPE(DATA_TYPE, 4) d10 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d11 = vload2(2, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 4) d30 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d31 = vload2(2, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); DATA_TYPE part0 = -16.0f * d20.s0 + 20.0f * d20.s2 - 4.0f * d21.s0; DATA_TYPE part1 = 16.0f * d10.s0 - 20.0f * d10.s2 + 4.0f * d11.s0 - 4.0f * d30.s0 + 5.0f * d30.s2 - d31.s0; DATA_TYPE part2 = 16.0f * d20.s2 - 4.0f * d21.s0; DATA_TYPE part3 = 16.0f * d20.s1 - 4.0f * d20.s3; DATA_TYPE part4 = 16.0f * d10.s2 - 4.0f * d11.s0 - 4.0f * d30.s2 + d31.s0; DATA_TYPE part5 = 16.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + d30.s3; DATA_TYPE part6 = 4.0f * d20.s2 - 4.0f * d21.s0; DATA_TYPE part7 = 8.0f * d10.s1 - 8.0f * d10.s3 - 2.0f * d30.s1 + 2.0f * d30.s3; DATA_TYPE part8 = 4.0f * d10.s2 - 4.0f * d11.s0 - d30.s2 + d31.s0; DATA_TYPE part9 = 8.0f * d20.s1 - 8.0f * d20.s3; DATA_TYPE part10 = -16.0f * d20.s1 + 20.0f * d20.s3 - 4.0f * d21.s1; DATA_TYPE part11 = -16.0f * d10.s1 + 20.0f * d10.s3 - 4.0f * d11.s1 + 4.0f * d30.s1 - 5.0f * d30.s3 + d31.s1; DATA_TYPE part12 = 8.0f * d10.s0 - 10.0f * d10.s2 + 2.0f * d11.s0 - 8.0f * d30.s0 + 10.0f * d30.s2 - 2.0f * d31.s0; DATA_TYPE part13 = part0 * 0.25f; DATA_TYPE part14 = part2 * 0.25f; DATA_TYPE part15 = 8.0f * d10.s1 - 2.0f * d10.s3 - 8.0f * d30.s1 + 2.0f * d30.s3; DATA_TYPE part16 = 8.0f * d10.s2 - 2.0f * d11.s0 - 8.0f * d30.s2 + 2.0f * d31.s0; DATA_TYPE part17 = part3 * 0.25f; DATA_TYPE part18 = part6 * 0.25f; DATA_TYPE part19 = 4.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + 4.0f * d30.s3; DATA_TYPE part20 = 2.0f * d10.s2 - 2.0f * d11.s0 - 2.0f * d30.s2 + 2.0f * d31.s0; DATA_TYPE part21 = part9 * 0.25f; DATA_TYPE part22 = part10 * 0.25f; DATA_TYPE part23 = part11 * 0.5f + 6.0f * d30.s1 - 7.5f * d30.s3 + 1.5f * d31.s1; out6 += part0 - part1; out12 += part0 + part1; out7 += part2 + part3 + part4 + part5; out8 += part2 - part3 + part4 - part5; out13 += part2 + part3 - part4 - part5; out14 += part2 - part3 - part4 + part5; out9 += part6 + part7 + part8 + part9; out10 += part6 - part7 + part8 - part9; out15 += part6 - part7 - part8 + part9; out16 += part6 + part7 - part8 - part9; out11 += part10 + part11; out17 += part10 - part11; out18 += part13 - part12; out24 += part13 + part12; out19 += part14 + part15 + part16 + part17; out20 += part14 - part15 + part16 - part17; out25 += part14 - part15 - part16 + part17; out26 += part14 + part15 - part16 - part17; out21 += part18 + part19 + part20 + part21; out22 += part18 - part19 + part20 - part21; out27 += part18 - part19 - part20 + part21; out28 += part18 + part19 - part20 - part21; out23 += part22 + part23; out29 += part22 - part23; *(dst_addr) = out6; dst_addr += dst_plane_stride; *(dst_addr) = out7; dst_addr += dst_plane_stride; *(dst_addr) = out8; dst_addr += dst_plane_stride; *(dst_addr) = out9; dst_addr += dst_plane_stride; *(dst_addr) = out10; dst_addr += dst_plane_stride; *(dst_addr) = out11; dst_addr += dst_plane_stride; *(dst_addr) = out12; dst_addr += dst_plane_stride; *(dst_addr) = out13; dst_addr += dst_plane_stride; *(dst_addr) = out14; dst_addr += dst_plane_stride; *(dst_addr) = out15; dst_addr += dst_plane_stride; *(dst_addr) = out16; dst_addr += dst_plane_stride; *(dst_addr) = out17; dst_addr += dst_plane_stride; *(dst_addr) = out18; dst_addr += dst_plane_stride; *(dst_addr) = out19; dst_addr += dst_plane_stride; *(dst_addr) = out20; dst_addr += dst_plane_stride; *(dst_addr) = out21; dst_addr += dst_plane_stride; *(dst_addr) = out22; dst_addr += dst_plane_stride; *(dst_addr) = out23; dst_addr += dst_plane_stride; *(dst_addr) = out24; dst_addr += dst_plane_stride; *(dst_addr) = out25; dst_addr += dst_plane_stride; *(dst_addr) = out26; dst_addr += dst_plane_stride; *(dst_addr) = out27; dst_addr += dst_plane_stride; *(dst_addr) = out28; dst_addr += dst_plane_stride; *(dst_addr) = out29; dst_addr += dst_plane_stride; VEC_DATA_TYPE(DATA_TYPE, 4) d50 = vload4(0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y)); VEC_DATA_TYPE(DATA_TYPE, 2) d51 = vload2(2, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y)); out0 = 16.0f * d10.s0 - 20.0f * d10.s2 - 20.0f * d30.s0 + 25.0f * d30.s2 + 4.0f * d50.s0 - 5.0f * d50.s2 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0; out1 = -16.0f * d10.s1 - 16.0f * d10.s2 + 4.0f * d10.s3 + 20.0f * d30.s1 + 20.0f * d30.s2 - 5.0f * d30.s3 - 4.0f * d50.s1 - 4.0f * d50.s2 + d50.s3 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0; out2 = 16.0f * d10.s1 - 16.0f * d10.s2 - 4.0f * d10.s3 - 20.0f * d30.s1 + 20.0f * d30.s2 + 5.0f * d30.s3 + 4.0f * d50.s1 - 4.0f * d50.s2 - d50.s3 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0; out3 = -8.0f * d10.s1 - 4.0f * d10.s2 + 8.0f * d10.s3 + 10.0f * d30.s1 - 10.0f * d30.s3 + 5.0f * d30.s2 - 2.0f * d50.s1 + 2.0f * d50.s3 - d50.s2 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0; out4 = 8.0f * d10.s1 - 4.0f * d10.s2 - 8.0f * d10.s3 - 10.0f * d30.s1 + 5.0f * d30.s2 + 10.0f * d30.s3 + 2.0f * d50.s1 - 2.0f * d50.s3 - d50.s2 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0; out5 = 16.0f * d10.s1 - 20.0f * d10.s3 + 4.0f * d11.s1 - 20.0f * d30.s1 + 25.0f * d30.s3 - 5.0f * d31.s1 + 4.0f * d50.s1 - 5.0f * d50.s3 + d51.s1; *(dst_addr) = out0; dst_addr += dst_plane_stride; *(dst_addr) = out1; dst_addr += dst_plane_stride; *(dst_addr) = out2; dst_addr += dst_plane_stride; *(dst_addr) = out3; dst_addr += dst_plane_stride; *(dst_addr) = out4; dst_addr += dst_plane_stride; *(dst_addr) = out5; dst_addr += dst_plane_stride; #endif } __kernel void winograd_input_transform_4x4_5x5_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { const int x = get_global_id(0); const int y = get_global_id(1); #if defined(SRC_DEPTH) const int z = get_global_id(2) % SRC_DEPTH; const int b = get_global_id(2) / SRC_DEPTH; #else const int z = get_global_id(2); #endif #if defined(SRC_DEPTH) __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z + b * src_stride_w; #else __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z; #endif src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y); #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = vload8(0, (__global DATA_TYPE *)(src_addr)); #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 5 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 6 * src_stride_y)), *((__global DATA_TYPE *)(src_addr + 7 * src_stride_y))); #else const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = vload8(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row1 = vload8(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row2 = vload8(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row3 = vload8(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row4 = vload8(0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row5 = vload8(0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row6 = vload8(0, (__global DATA_TYPE *)(src_addr + 6 * src_stride_y)); const VEC_DATA_TYPE(DATA_TYPE, 8) in_row7 = vload8(0, (__global DATA_TYPE *)(src_addr + 7 * src_stride_y)); #endif VEC_DATA_TYPE(DATA_TYPE, 8) tmp0 = in_row0; VEC_DATA_TYPE(DATA_TYPE, 8) comm_fact0 = 0.0f; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) comm_fact0 += in_row2 + in_row6 - (DATA_TYPE)4.25f * in_row4; tmp0 += -in_row6 + (DATA_TYPE)5.25f * in_row4 - (DATA_TYPE)5.25f * in_row2; VEC_DATA_TYPE(DATA_TYPE, 8) comm_fact1 = in_row1 + in_row5 - (DATA_TYPE)4.25f * in_row3; VEC_DATA_TYPE(DATA_TYPE, 8) comm_fact2 = (DATA_TYPE)0.25f * in_row2 - (DATA_TYPE)1.25f * in_row4 + in_row6; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp1 = comm_fact0 + comm_fact1; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp2 = comm_fact0 - comm_fact1; comm_fact0 = (DATA_TYPE)2.5f * in_row3; comm_fact1 = (DATA_TYPE)0.5f * in_row1 - comm_fact0 + (DATA_TYPE)2.0f * in_row5; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp3 = comm_fact1 + comm_fact2; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp4 = comm_fact2 - comm_fact1; comm_fact1 = (DATA_TYPE)2.0f * in_row1 - comm_fact0 + (DATA_TYPE)0.5f * in_row5; comm_fact2 = (DATA_TYPE)4.0f * in_row2 - (DATA_TYPE)5.0f * in_row4 + in_row6; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp5 = comm_fact1 + comm_fact2; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp6 = comm_fact2 - comm_fact1; const VEC_DATA_TYPE(DATA_TYPE, 8) tmp7 = in_row7 - in_row1 + (DATA_TYPE)5.25f * in_row3 - (DATA_TYPE)5.25f * in_row5; #endif VEC_DATA_TYPE(DATA_TYPE, 8) out0; OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0); #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) VEC_DATA_TYPE(DATA_TYPE, 8) out1, out2, out3, out4, out5, out6, out7; OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0); OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0); OUTPUT_ROW_4x4_5x5(out3, tmp3, comm_fact0); OUTPUT_ROW_4x4_5x5(out4, tmp4, comm_fact0); OUTPUT_ROW_4x4_5x5(out5, tmp5, comm_fact0); OUTPUT_ROW_4x4_5x5(out6, tmp6, comm_fact0); OUTPUT_ROW_4x4_5x5(out7, tmp7, comm_fact0); #endif #if defined(SRC_DEPTH) __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y + b * dst_stride_w; #else __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y; #endif *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)) = out0.s0; *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)) = out0.s1; *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)) = out0.s2; *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)) = out0.s3; *((__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)) = out0.s4; *((__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)) = out0.s5; *((__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)) = out0.s6; *((__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)) = out0.s7; #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) *((__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)) = out1.s0; *((__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)) = out1.s1; *((__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)) = out1.s2; *((__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)) = out1.s3; *((__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)) = out1.s4; *((__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)) = out1.s5; *((__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)) = out1.s6; *((__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)) = out1.s7; *((__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z)) = out2.s0; *((__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z)) = out2.s1; *((__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z)) = out2.s2; *((__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z)) = out2.s3; *((__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z)) = out2.s4; *((__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z)) = out2.s5; *((__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z)) = out2.s6; *((__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z)) = out2.s7; *((__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z)) = out3.s0; *((__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z)) = out3.s1; *((__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z)) = out3.s2; *((__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z)) = out3.s3; *((__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z)) = out3.s4; *((__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z)) = out3.s5; *((__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z)) = out3.s6; *((__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z)) = out3.s7; *((__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z)) = out4.s0; *((__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z)) = out4.s1; *((__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z)) = out4.s2; *((__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z)) = out4.s3; *((__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z)) = out4.s4; *((__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z)) = out4.s5; *((__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z)) = out4.s6; *((__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z)) = out4.s7; *((__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z)) = out5.s0; *((__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z)) = out5.s1; *((__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z)) = out5.s2; *((__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z)) = out5.s3; *((__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z)) = out5.s4; *((__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z)) = out5.s5; *((__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z)) = out5.s6; *((__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z)) = out5.s7; *((__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z)) = out6.s0; *((__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z)) = out6.s1; *((__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z)) = out6.s2; *((__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z)) = out6.s3; *((__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z)) = out6.s4; *((__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z)) = out6.s5; *((__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z)) = out6.s6; *((__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z)) = out6.s7; *((__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z)) = out7.s0; *((__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z)) = out7.s1; *((__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z)) = out7.s2; *((__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z)) = out7.s3; *((__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z)) = out7.s4; *((__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z)) = out7.s5; *((__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z)) = out7.s6; *((__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z)) = out7.s7; #endif } #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) __kernel void winograd_input_transform_2x1_3x1_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_2x2_3x3_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_2x1_3x1_stepz2_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_2x2_3x3_stepz2_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_4x1_3x1_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_4x4_3x3_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_4x1_5x1_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_4x4_5x5_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } #endif #if defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) __kernel void winograd_input_transform_1x2_1x3_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_2x2_3x3_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_1x2_1x3_stepz2_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_2x2_3x3_stepz2_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_1x4_1x3_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_4x4_3x3_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } __kernel void winograd_input_transform_1x4_1x5_stepz1_nchw( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), uint src_stride_w, uint dst_stride_w) { winograd_input_transform_4x4_5x5_stepz1_nchw(src_ptr, src_stride_x, src_step_x, src_stride_y, src_step_y, src_stride_z, src_step_z, src_offset_first_element_in_bytes, dst_ptr, dst_stride_x, dst_step_x, dst_stride_y, dst_step_y, dst_stride_z, dst_step_z, dst_offset_first_element_in_bytes, src_stride_w, dst_stride_w); } #endif #endif )"