diff options
Diffstat (limited to 'src/core/CL/cl_kernels')
19 files changed, 1128 insertions, 2766 deletions
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl index 4665d612f..d8453ed80 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl @@ -27,7 +27,7 @@ #include "repeat.h" /** (EXPERIMENTAL_POST_OPS) gemm_mm_native kernel */ -#if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) +#if defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) #if defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) #define VFMA(a, b, c) \ @@ -107,6 +107,7 @@ #error "M0 not supported" #endif // M0 not supported +#if defined(GEMM_MM_NATIVE_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: * Post op 1: activation (optional) * Post op 2: elementwise op @@ -140,8 +141,11 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), #if defined(BETA) uint bias_stride_z, #endif //defined(BETA) - uint dst_stride_z, - uint eltwise_operand_stride_z + uint dst_stride_z, + uint eltwise_operand_stride_z, + const int M, + const int N, + const int K #if defined(REINTERPRET_INPUT_AS_3D) , uint lhs_cross_plane_pad @@ -360,5 +364,6 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), // Store output block STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } +#endif // defined(GEMM_MM_NATIVE_POST_ACT_ELTWISE_OP_ACT) #endif // defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) -#endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) +#endif // defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl index 32186c359..89577e9eb 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl @@ -27,7 +27,7 @@ /** (EXPERIMENTAL_POST_OPS) gemm_mm_reshaped kernel */ -#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N) +#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) #if defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) #if defined(MIXED_PRECISION) @@ -207,6 +207,7 @@ #error "N0 value not supported" #endif // N0 conditions +#if defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: * Post op 1: activation (optional) * Post op 2: elementwise op @@ -235,7 +236,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR IMAGE_DECLARATION(dst), // Post Op arguments IMAGE_DECLARATION(eltwise_operand), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -247,7 +247,10 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define LHS_BLOCK_SIZE ((K0) * (M0)) @@ -303,7 +306,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); - for(int i = 0; i < k; i += K0) + for(int i = 0; i < K; i += K0) { // Supported cases (M0, K0): // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 @@ -425,8 +428,9 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } +#endif // defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_POST_ACT_ELTWISE_OP_ACT) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object. * Post op 1: activation (optional) * Post op 2: elementwise op @@ -455,7 +459,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG IMAGE_DECLARATION(dst), // Post Op arguments IMAGE_DECLARATION(eltwise_operand), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -467,7 +470,10 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) @@ -643,7 +649,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT) #if defined(LHS_TRANSPOSE) @@ -755,6 +761,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG CONCAT(ARM_MM_T_NT_M0xN0x, K0) \ (M0, N0, TYPE, A, B, C) +#if defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: * Post op 1: activation (optional) * Post op 2: elementwise op @@ -774,6 +781,9 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes) * @param[in] eltwise_operand_step_y eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -783,7 +793,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR IMAGE_DECLARATION(dst), // Post Op arguments IMAGE_DECLARATION(eltwise_operand), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -795,7 +804,10 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define LHS_BLOCK_SIZE ((K0) * (M0)) @@ -858,7 +870,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr); __global DATA_TYPE *rhs = (__global DATA_TYPE *)(rhs_addr); - for(int i = 0; i < k; i += K0) + for(int i = 0; i < K; i += K0) { VEC_DATA_TYPE(DATA_TYPE, M0) a0; @@ -1083,7 +1095,9 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR #undef RHS_OFFSET_X #undef RHS_STEP_X } -#if defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_POST_ACT_ELTWISE_OP_ACT) + +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object. * Post op 1: activation (optional) * Post op 2: elementwise op @@ -1112,7 +1126,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG IMAGE_DECLARATION(dst), // Post Op arguments IMAGE_DECLARATION(eltwise_operand), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -1124,7 +1137,10 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) @@ -1401,8 +1417,8 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT) #endif // defined(LHS_TRANSPOSE) #endif // defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) -#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N)
\ No newline at end of file +#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl index e96aba613..09ddcde04 100644 --- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,7 +26,7 @@ #include "repeat.h" /** (EXPERIMENTAL_POST_OPS) gemm_mm_reshaped_only_rhs kernel */ -#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K) +#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) #if defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) #define CONCAT(a, b) a##b @@ -151,6 +151,7 @@ #error "N0 value not supported" #endif // N0 conditions +#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: * Post op 1: activation (optional) * Post op 2: elementwise op @@ -194,7 +195,10 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define RHS_BLOCK_SIZE ((K0) * (N0)) @@ -409,8 +413,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT #undef RHS_OFFSET_X #undef RHS_STEP_X } +#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_T_POST_ACT_ELTWISE_OP_ACT) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object. * Post op 1: activation (optional) * Post op 2: elementwise op @@ -430,6 +435,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARAT * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes) * @param[in] eltwise_operand_step_y eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -454,12 +462,15 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) -#define LEFTOVER_K (K % K0) + const uint LEFTOVER_K = K % K0; // Block size #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0)) @@ -562,99 +573,99 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP; } -#if LEFTOVER_K != 0 - // Note: We cannot read out-of-bound elements from the RHS matrix because - // the RHS width is always multiple of K0. This is not be true for the LHS matrix - - union UNION_VEC_TYPE + if(LEFTOVER_K != 0) { - DATA_TYPE s[K0]; - VEC_DATA_TYPE(DATA_TYPE, K0) - v; - }; + // Note: We cannot read out-of-bound elements from the RHS matrix because + // the RHS width is always multiple of K0. This is not be true for the LHS matrix + + union UNION_VEC_TYPE + { + DATA_TYPE s[K0]; + VEC_DATA_TYPE(DATA_TYPE, K0) + v; + }; - union UNION_VEC_TYPE a0 = {.v = 0 }; + union UNION_VEC_TYPE a0 = {.v = 0 }; #if M0 > 1 - union UNION_VEC_TYPE a1 = {.v = 0 }; + union UNION_VEC_TYPE a1 = {.v = 0 }; #endif // M0 > 1 #if M0 > 2 - union UNION_VEC_TYPE a2 = {.v = 0 }; + union UNION_VEC_TYPE a2 = {.v = 0 }; #endif // M0 > 2 #if M0 > 3 - union UNION_VEC_TYPE a3 = {.v = 0 }; + union UNION_VEC_TYPE a3 = {.v = 0 }; #endif // M0 > 3 #if M0 > 4 - union UNION_VEC_TYPE a4 = {.v = 0 }; + union UNION_VEC_TYPE a4 = {.v = 0 }; #endif // M0 > 4 #if M0 > 5 - union UNION_VEC_TYPE a5 = {.v = 0 }; + union UNION_VEC_TYPE a5 = {.v = 0 }; #endif // M0 > 5 #if M0 > 6 - union UNION_VEC_TYPE a6 = {.v = 0 }; + union UNION_VEC_TYPE a6 = {.v = 0 }; #endif // M0 > 6 #if M0 > 7 - union UNION_VEC_TYPE a7 = {.v = 0 }; + union UNION_VEC_TYPE a7 = {.v = 0 }; #endif // M0 > 7 - REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); + REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); - // Load from RHS matrix - LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); + // Load from RHS matrix + LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); - // Load from LHS matrix - for(int k = 0; k < LEFTOVER_K; ++k) - { - a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0); + // Load from LHS matrix + for(int k = 0; k < LEFTOVER_K; ++k) + { + a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0); #if M0 > 1 - a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1); + a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1); #endif // M0 > 1 #if M0 > 2 - a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2); + a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2); #endif // M0 > 2 #if M0 > 3 - a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3); + a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3); #endif // M0 > 3 #if M0 > 4 - a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4); + a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4); #endif // M0 > 4 #if M0 > 5 - a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5); + a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5); #endif // M0 > 5 #if M0 > 6 - a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6); + a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6); #endif // M0 > 6 #if M0 > 7 - a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7); + a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7); #endif // M0 > 7 - lhs_offset += sizeof(DATA_TYPE); - } + lhs_offset += sizeof(DATA_TYPE); + } - // Accumulate - ARM_DOT_K0XN0(K0, a0.v, b, c0); + // Accumulate + ARM_DOT_K0XN0(K0, a0.v, b, c0); #if M0 > 1 - ARM_DOT_K0XN0(K0, a1.v, b, c1); + ARM_DOT_K0XN0(K0, a1.v, b, c1); #endif // M0 > 1 #if M0 > 2 - ARM_DOT_K0XN0(K0, a2.v, b, c2); + ARM_DOT_K0XN0(K0, a2.v, b, c2); #endif // M0 > 2 #if M0 > 3 - ARM_DOT_K0XN0(K0, a3.v, b, c3); + ARM_DOT_K0XN0(K0, a3.v, b, c3); #endif // M0 > 3 #if M0 > 4 - ARM_DOT_K0XN0(K0, a4.v, b, c4); + ARM_DOT_K0XN0(K0, a4.v, b, c4); #endif // M0 > 4 #if M0 > 5 - ARM_DOT_K0XN0(K0, a5.v, b, c5); + ARM_DOT_K0XN0(K0, a5.v, b, c5); #endif // M0 > 5 #if M0 > 6 - ARM_DOT_K0XN0(K0, a6.v, b, c6); + ARM_DOT_K0XN0(K0, a6.v, b, c6); #endif // M0 > 6 #if M0 > 7 - ARM_DOT_K0XN0(K0, a7.v, b, c7); + ARM_DOT_K0XN0(K0, a7.v, b, c7); #endif // M0 > 7 - -#endif // LEFTOVER_K != 0 + } __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); @@ -723,10 +734,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X -#undef LEFTOVER_K #undef PIXEL_UNIT } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT) #define VFMA(a, b, c) \ ({ \ @@ -805,6 +815,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ #error "M0 not supported" #endif // M0 not supported +#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: * Post op 1: activation (optional) * Post op 2: elementwise op @@ -824,6 +835,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_ * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes) * @param[in] eltwise_operand_step_y eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -848,7 +862,10 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define RHS_BLOCK_SIZE ((K0) * (N0)) @@ -1087,9 +1104,11 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X +#undef RHS_STEP_LOOP } +#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_POST_ACT_ELTWISE_OP_ACT) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object. * Post op 1: activation (optional) * Post op 2: elementwise op @@ -1109,6 +1128,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARA * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes) * @param[in] eltwise_operand_step_y eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -1133,7 +1155,10 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) @@ -1145,9 +1170,11 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE #if defined(RHS_INTERLEAVE) #define RHS_OFFSET_X (PIXEL_UNIT) #define RHS_STEP_X ((PIXEL_UNIT) * (H0)) +#define RHS_STEP_LOOP (1) #else // defined(RHS_INTERLEAVE) #define RHS_OFFSET_X (RHS_BLOCK_SIZE) #define RHS_STEP_X (PIXEL_UNIT) +#define RHS_STEP_LOOP (H0) #endif // defined(RHS_INTERLEAVE) uint x = get_global_id(0); @@ -1365,7 +1392,8 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X +#undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT) #endif // defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) -#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K)
\ No newline at end of file +#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl index a76ad458a..33ab25cad 100644 --- a/src/core/CL/cl_kernels/common/gemm.cl +++ b/src/core/CL/cl_kernels/common/gemm.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,856 +24,7 @@ #include "gemm_helpers.h" #include "repeat.h" -#if defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0) -#define INC2 (VEC_DATA_TYPE(uint, 2))(0, 1) -#define INC3 (VEC_DATA_TYPE(uint, 3))(0, 1, 2) -#define INC4 (VEC_DATA_TYPE(uint, 4))(0, 1, 2, 3) -#define INC8 (VEC_DATA_TYPE(uint, 8))(0, 1, 2, 3, 4, 5, 6, 7) -#define INC16 (VEC_DATA_TYPE(uint, 16))(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) -#define CONCAT_INC(K0) INC##K0 -#define INC(K0) CONCAT_INC(K0) - -#if(SRC_WIDTH % K0) -#define BOUNDARY_CONDITION_X(x, a) \ - ({ \ - a = select(0, a, CONVERT(((x * (VEC_DATA_TYPE(uint, K0))K0 + INC(K0)) < (VEC_DATA_TYPE(uint, K0))SRC_WIDTH), VEC_DATA_TYPE(DATA_TYPE, K0))); \ - }) -#else // (SRC_WIDTH % K0) -#define BOUNDARY_CONDITION_X(x, a) \ - ({}) -#endif // (SRC_WIDTH % K0) - -#define LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin) \ - ({ \ - if(y * M0 + M0 >= SRC_HEIGHT && PARTIAL_LOAD_M0 != 0) \ - { \ - if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0)) \ - { \ - LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ - } \ - else \ - { \ - LOAD_TENSOR_M0XN0(PARTIAL_LOAD_M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ - } \ - } \ - else \ - { \ - if(x * K0 + K0 >= SRC_WIDTH && (PARTIAL_LOAD_K0 != 0)) \ - { \ - LOAD_TENSOR_M0XN0(M0, PARTIAL_LOAD_K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ - } \ - else \ - { \ - LOAD_TENSOR_M0XN0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); \ - } \ - } \ - }) - -/** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (not transposed) in - * the output matrix unrolling the values. - * - * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) - * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) - * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) - * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). - * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2) - * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1) - * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1) - * @note Only the following values for M0, K0 and V0 are supported: - * M0: 2,3,4,5,6,7,8 - * K0: 2,3,4,8,16 - * V0: greater than 0 - * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time: - * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D - * -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor. - * -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor - * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped - * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. - * - * @param[in] src_ptr Pointer to the source LHS tensor. Supported data types: All - * @param[in] src_stride_x Stride of the source LHS tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source LHS tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source LHS tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) - */ -__kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst) -#if defined(REINTERPRET_INPUT_AS_3D) - , - uint cross_plane_pad -#endif // REINTERPRET_INPUT_AS_3D - ) -{ - // Block size -#define BLOCK_SIZE ((M0) * (K0)) - - // Output offset X -#if defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (K0) -#else // defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (BLOCK_SIZE) -#endif // defined(INTERLEAVE) - - // Output step X -#if defined(INTERLEAVE) -#define OUTPUT_STEP_X (K0) * (V0) -#else // Do not interleave -#define OUTPUT_STEP_X (K0) -#endif // defined(INTERLEAVE) - - // Compute source and destination addresses - uint x = get_global_id(0); - uint y = get_global_id(1); - uint z = get_global_id(2); - - // ------------------ Compute input/output addresses --------------------------- - - // Compute the input address - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y; - - // Compute the output address - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) * - (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)); - - // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0; - REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0); - -#if defined(REINTERPRET_INPUT_AS_3D) - // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we - // multiply src_stride_z by DEPTH_GEMM3D - - input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D; - - // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D - CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y); - -#else // defined(REINTERPRET_INPUT_AS_3D) - - input_ptr += z * (uint)src_stride_z; - -#endif // defined(REINTERPRET_INPUT_AS_3D) - - // Add offset for batched GEMM - output_ptr += z * (uint)dst_stride_z; - - // ---------------------------Load input values -------------------------------- - // Load values from the LHS matrix - REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0); - - LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); - - // ---------------------------Store output values ------------------------------ - REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); - STORE_BLOCK(M0, K0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); - -#undef BLOCK_SIZE -#undef OUTPUT_OFFSET_X -#undef OUTPUT_STEP_X -} - -#if M0 == 2 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, M0) \ - res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i); \ - VSTORE(M0) \ - (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - }) -#elif M0 == 3 // M0 == 3 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, M0) \ - res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i); \ - VSTORE(M0) \ - (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - }) -#elif M0 == 4 // M0 == 4 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, M0) \ - res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ - VSTORE(M0) \ - (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - }) -#elif M0 == 5 // M0 == 5 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ - DATA_TYPE res1 = a4.s##i; \ - VSTORE(4) \ - (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - *((__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4) = res1; \ - }) -#elif M0 == 6 // M0 == 6 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ - VEC_DATA_TYPE(DATA_TYPE, 2) \ - res1 = (VEC_DATA_TYPE(DATA_TYPE, 2))(a4.s##i, a5.s##i); \ - VSTORE(4) \ - (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - VSTORE(2) \ - (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \ - }) -#elif M0 == 7 // M0 == 7 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - res0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(a0.s##i, a1.s##i, a2.s##i, a3.s##i); \ - VEC_DATA_TYPE(DATA_TYPE, 3) \ - res1 = (VEC_DATA_TYPE(DATA_TYPE, 3))(a4.s##i, a5.s##i, a6.s##i); \ - VSTORE(4) \ - (res0, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - VSTORE(3) \ - (res1, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE)) + 4); \ - }) -#elif M0 == 8 // M0 == 8 -#define TRANSPOSE_COLUMN_AND_STORE(output_ptr, output_step_x, i) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, M0) \ - res = (VEC_DATA_TYPE(DATA_TYPE, M0))(a0.s##i, a1.s##i, a2.s##i, a3.s##i, a4.s##i, a5.s##i, a6.s##i, a7.s##i); \ - VSTORE(M0) \ - (res, 0, (__global DATA_TYPE *)(output_ptr + 0x##i * output_step_x * sizeof(DATA_TYPE))); \ - }) -#else // M0 not supported -#error "M0 value not supported" -#endif // N0 conditions - -/** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (transposed) in - * the output matrix unrolling the values. - * - * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) - * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) - * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) - * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). - * @note The number of M0xK0 vertical blocks to store on the same output row must be passed at compile time using -DV0 (e.g. -DV0=2) - * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_LOAD_M0 (e.g. -DPARTIAL_LOAD_M0=1) - * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_LOAD_K0 (e.g. -DPARTIAL_LOAD_K0=1) - * @note Only the following values for M0, K0 and V0 are supported: - * M0: 2,3,4,5,6,7,8 - * K0: 2,3,4,8,16 - * V0: greater than 0 - * @note In case the input has to be reinterpreted as a 3D tensor (e.g. input of convolution layer 1x1), the following information must be passed at compile time: - * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D - * -# HEIGHT_GEMM3D: The height of the input in case it has to be reinterpreted as a 3D tensor. - * -# DEPTH_GEMM3D: The depth of the input in case it has to be reinterpreted as a 3D tensor - * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped - * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. - * - * @param[in] src_ptr Pointer to the source LHS tensor. Supported data types: All - * @param[in] src_stride_x Stride of the source LHS tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source LHS tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source LHS tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source LHS tensor - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) - */ -__kernel void gemm_reshape_lhs_matrix_t(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst) -#if defined(REINTERPRET_INPUT_AS_3D) - , - uint cross_plane_pad -#endif // REINTERPRET_INPUT_AS_3D - ) -{ - // Block size -#define BLOCK_SIZE ((M0) * (K0)) - - // Output offset X -#if defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (M0) -#else // defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (BLOCK_SIZE) -#endif // defined(INTERLEAVE) - - // Output step X -#if defined(INTERLEAVE) -#define OUTPUT_STEP_X (M0) * (V0) -#else // Do not interleave -#define OUTPUT_STEP_X (M0) -#endif // defined(INTERLEAVE) - - // Compute source and destination addresses - uint x = get_global_id(0); - uint y = get_global_id(1); - uint z = get_global_id(2); - - // ------------------ Compute input/output addresses --------------------------- - - // Compute the input address - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)K0 * sizeof(DATA_TYPE) + y * (uint)M0 * src_stride_y; - - // Compute the output address - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)BLOCK_SIZE * (uint)V0 * sizeof(DATA_TYPE)) + ((y / (uint)V0) * (uint)dst_stride_y) + ((y % V0) * - (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)); - - // Create variables: uint zin0=0, zin1=0, zin2=0...zin(M0-1)=0; - REPEAT_VAR_INIT_TO_CONST(M0, uint, zin, 0); - -#if defined(REINTERPRET_INPUT_AS_3D) - // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we - // multiply src_stride_z by DEPTH_GEMM3D - - input_ptr += z * (uint)src_stride_z * DEPTH_GEMM3D; - - // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D - CALCULATE_Z_OFFSET(M0, uint, zin, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, cross_plane_pad, src_stride_y); - -#else // defined(REINTERPRET_INPUT_AS_3D) - - input_ptr += z * (uint)src_stride_z; - -#endif // defined(REINTERPRET_INPUT_AS_3D) - - // Add offset for batched GEMM - output_ptr += z * (uint)dst_stride_z; - - // ---------------------------Load input values -------------------------------- - REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, K0), a, 0); - - LOAD_TENSOR_BOUNDARY_AWARE_M0XK0(M0, K0, DATA_TYPE, a, input_ptr, src_stride_y, zin); - - // ---------------------------Transpose and store block ----------------------- - - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 0); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 1); -#if K0 > 2 - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 2); -#endif // K0 > 2 -#if K0 > 3 - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 3); -#endif // K0 > 3 -#if K0 > 4 - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 4); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 5); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 6); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 7); -#endif // K0 > 4 -#if K0 > 8 - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 8); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, 9); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, A); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, B); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, C); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, D); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, E); - TRANSPOSE_COLUMN_AND_STORE(output_ptr, OUTPUT_STEP_X, F); -#endif // K0 > 8 - -#undef BLOCK_SIZE -#undef OUTPUT_OFFSET_X -#undef OUTPUT_STEP_X -} -#endif // defined(M0) && defined(K0) && defined(V0) && defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PARTIAL_LOAD_M0) && defined(PARTIAL_LOAD_K0) - -#if defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT) -/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (not transposed) in - * the output matrix unrolling the values. - * - * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) - * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) - * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). - * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2) - * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. - * @note Only the following values for K0, N0 and H0 are supported: - * N0: 2,3,4,8,16 - * K0: 1,2,3,4,8,16 - * H0: greater than 0 - * - * @param[in] src_ptr Pointer to the source RHS tensor. Supported data types: All - * @param[in] src_stride_x Stride of the source RHS tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source RHS tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source RHS tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - */ -__kernel void gemm_reshape_rhs_matrix_nt(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Block size -#define BLOCK_SIZE ((K0) * (N0)) - - // Output offset X -#if defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (N0) -#else // defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (BLOCK_SIZE) -#endif // defined(INTERLEAVE) - - // Output step X -#if defined(INTERLEAVE) -#define OUTPUT_STEP_X (N0) * (H0) -#else // Do not interleave -#define OUTPUT_STEP_X (N0) -#endif // defined(INTERLEAVE) - - // Compute source and destination addresses - uint x = get_global_id(0); - uint y = get_global_id(1); - uint z = get_global_id(2); - - // ------------------ Compute input/output addresses --------------------------- - - // Compute the input address - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z; - - // Compute the output address - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % (uint)H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + (( - x / (uint)H0) - * (uint)dst_stride_y) - + z * (uint)dst_stride_z; - - // ---------------------------Load input values -------------------------------- - - REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); ////uint a0=0, a1=0, a2=0...a(M0-1)=0; - - // Load values from the RHS matrix - a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y)); -#if K0 > 1 - if(y * (uint)K0 + 1 < SRC_HEIGHT) - { - a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y)); - } -#endif // K0 > 1 -#if K0 > 2 - if(y * (uint)K0 + 2 < SRC_HEIGHT) - { - a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y)); - } -#endif // K0 > 2 -#if K0 > 3 - if(y * (uint)K0 + 3 < SRC_HEIGHT) - { - a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y)); - } -#endif // K0 > 3 -#if K0 > 4 - if(y * (uint)K0 + 4 < SRC_HEIGHT) - { - a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y)); - } - if(y * (uint)K0 + 5 < SRC_HEIGHT) - { - a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y)); - } - if(y * (uint)K0 + 6 < SRC_HEIGHT) - { - a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y)); - } - if(y * (uint)K0 + 7 < SRC_HEIGHT) - { - a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y)); - } -#endif // K0 > 4 -#if K0 > 8 - if(y * (uint)K0 + 8 < SRC_HEIGHT) - { - a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y)); - } - if(y * (uint)K0 + 9 < SRC_HEIGHT) - { - a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y)); - } - if(y * (uint)K0 + 10 < SRC_HEIGHT) - { - aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y)); - } - if(y * (uint)K0 + 11 < SRC_HEIGHT) - { - aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y)); - } - if(y * (uint)K0 + 12 < SRC_HEIGHT) - { - aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y)); - } - if(y * (uint)K0 + 13 < SRC_HEIGHT) - { - aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y)); - } - if(y * (uint)K0 + 14 < SRC_HEIGHT) - { - aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y)); - } - if(y * (uint)K0 + 15 < SRC_HEIGHT) - { - aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y)); - } -#endif // K0 > 8 - - // ---------------------------Store output values ------------------------------ - REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); - STORE_BLOCK(K0, N0, DATA_TYPE, a, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); - -#undef BLOCK_SIZE -#undef OUTPUT_OFFSET_X -#undef OUTPUT_STEP_X -} - -#if defined(TRANSPOSE) -/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (transposed) in - * the output matrix unrolling the values. - * - * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) - * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) - * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). - * @note The number of K0xN0 vertical blocks to store on the same output row must be passed at compile time using -DH0 (e.g. -DH0=2) - * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. - * @note The option -DTRANSPOSE must passed at compile time. - * @note Only the following values for K0, N0 and H0 are supported: - * N0: 2,3,4,8,16 - * K0: 2,3,4,8,16 - * H0: greater than 0 - * - * @param[in] src_ptr Pointer to the source RHS tensor. Supported data types: All - * @param[in] src_stride_x Stride of the source RHS tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source RHS tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source RHS tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source RHS tensor - * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - */ -__kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst)) -{ - // Block size -#define BLOCK_SIZE ((K0) * (N0)) - - // Output offset X -#if defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (K0) -#else // defined(INTERLEAVE) -#define OUTPUT_OFFSET_X (BLOCK_SIZE) -#endif // defined(INTERLEAVE) - - // Output step X -#if defined(INTERLEAVE) -#define OUTPUT_STEP_X (K0) * (H0) -#else // Do not interleave -#define OUTPUT_STEP_X (K0) -#endif // defined(INTERLEAVE) - - // Compute source and destination addresses - uint x = get_global_id(0); - uint y = get_global_id(1); - uint z = get_global_id(2); - - // ------------------ Compute input/output addresses --------------------------- - - // Compute the input address - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + x * (uint)N0 * sizeof(DATA_TYPE) + y * (uint)K0 * src_stride_y + z * (uint)src_stride_z; - - // Compute the output address - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y * (uint)BLOCK_SIZE * (uint)H0 * sizeof(DATA_TYPE)) + ((x % H0) * (uint)OUTPUT_OFFSET_X * sizeof(DATA_TYPE)) + ((x / - (uint)H0) * (uint)dst_stride_y) + z * (uint)dst_stride_z; - - // ---------------------------Load input values -------------------------------- - REPEAT_VAR_INIT_TO_CONST(K0, VEC_DATA_TYPE(DATA_TYPE, N0), a, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) a0=0, a1=0, ... a(K0-1)=0; - - // Load values from the RHS matrix - a0 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 0 * src_stride_y)); - if(y * (uint)K0 + 1 < SRC_HEIGHT) - { - a1 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y)); - } -#if K0 > 2 - if(y * (uint)K0 + 2 < SRC_HEIGHT) - { - a2 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y)); - } -#endif // K0 > 2 -#if K0 > 3 - if(y * (uint)K0 + 3 < SRC_HEIGHT) - { - a3 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y)); - } -#endif // K0 > 3 -#if K0 > 4 - if(y * (uint)K0 + 4 < SRC_HEIGHT) - { - a4 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 4 * src_stride_y)); - } - if(y * (uint)K0 + 5 < SRC_HEIGHT) - { - a5 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 5 * src_stride_y)); - } - if(y * (uint)K0 + 6 < SRC_HEIGHT) - { - a6 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 6 * src_stride_y)); - } - if(y * (uint)K0 + 7 < SRC_HEIGHT) - { - a7 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 7 * src_stride_y)); - } -#endif // K0 > 4 -#if K0 > 8 - if(y * (uint)K0 + 8 < SRC_HEIGHT) - { - a8 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 8 * src_stride_y)); - } - if(y * (uint)K0 + 9 < SRC_HEIGHT) - { - a9 = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 9 * src_stride_y)); - } - if(y * (uint)K0 + 10 < SRC_HEIGHT) - { - aA = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 10 * src_stride_y)); - } - if(y * (uint)K0 + 11 < SRC_HEIGHT) - { - aB = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 11 * src_stride_y)); - } - if(y * (uint)K0 + 12 < SRC_HEIGHT) - { - aC = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 12 * src_stride_y)); - } - if(y * (uint)K0 + 13 < SRC_HEIGHT) - { - aD = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 13 * src_stride_y)); - } - if(y * (uint)K0 + 14 < SRC_HEIGHT) - { - aE = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 14 * src_stride_y)); - } - if(y * (uint)K0 + 15 < SRC_HEIGHT) - { - aF = VLOAD(N0)(0, (__global DATA_TYPE *)(input_ptr + 15 * src_stride_y)); - } -#endif // K0 > 8 - - // ---------------------------Transpose the block ------------------------------ - REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), res, 0); //VEC_DATA_TYPE(DATA_TYPE, K0) res0=0, res1=0, res2=0,... res(N0-1)=0; - -#if K0 == 2 - // This part computes the following transpositions: - // 2x2 -> 2x2 - // 2x4 -> 4x2 - // 2x8 -> 8x2 - // 2x16 -> 16x2 - res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0); - res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1); -#if N0 > 2 - res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2); -#endif // N0 > 2 -#if N0 > 3 - res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3); -#endif // N0 > 3 -#if N0 > 4 - res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4); - res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5); - res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6); - res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7); -#endif // N0 > 4 -#if N0 > 8 - res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8); - res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9); - resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA); - resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB); - resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC); - resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD); - resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE); - resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF); -#endif // N0 > 8 - -#elif K0 == 3 // K0 == 2 - // This part computes the following transpositions: - // 3x2 -> 2x3 - // 3x4 -> 4x3 - // 3x8 -> 8x3 - // 3x16 -> 16x3 - res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0); - res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1); -#if N0 > 2 - res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2); -#endif // N0 > 2 -#if N0 > 3 - res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3); -#endif // N0 > 3 -#if N0 > 4 - res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4); - res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5); - res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6); - res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7); -#endif // N0 > 4 -#if N0 > 8 - res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8); - res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9); - resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA); - resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB); - resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC); - resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD); - resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE); - resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF); -#endif // N0 > 8 - -#elif K0 == 4 // K0 == 4 - // This part computes the following transpositions: - // 4x2 -> 2x4 - // 4x4 -> 4x4 - // 4x8 -> 8x4 - // 4x16 -> 16x4 - res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0); - res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1); -#if N0 > 2 - res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2); -#endif // N0 > 2 -#if N0 > 3 - res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3); -#endif // N0 > 3 -#if N0 > 4 - res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4); - res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5); - res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6); - res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7); -#endif // N0 > 4 -#if N0 > 8 - res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8); - res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9); - resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA); - resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB); - resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC); - resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD); - resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE); - resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF); -#endif // N0 > 8 - -#elif K0 == 8 // K0 == 8 - // This part computes the following transpositions: - // 8x2 -> 2x8 - // 8x4 -> 4x8 - // 8x8 -> 8x8 - // 8x16 -> 16x8 - res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0); - res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1); -#if N0 > 2 - res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2); -#endif // N0 > 2 -#if N0 > 3 - res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3); -#endif // N0 > 3 -#if N0 > 4 - res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4); - res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5); - res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6); - res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7); -#endif // N0 > 4 -#if N0 > 8 - res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8); - res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9); - resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA); - resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB); - resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC); - resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD); - resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE); - resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF); -#endif // N0 > 8 - -#elif K0 == 16 // K0 == 16 - - // This part computes the following transpositions: - // 16x2 -> 2x16 - // 16x4 -> 4x16 - // 16x8 -> 8x16 - // 16x16 -> 16x16 - res0 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s0, a1.s0, a2.s0, a3.s0, a4.s0, a5.s0, a6.s0, a7.s0, - a8.s0, a9.s0, aA.s0, aB.s0, aC.s0, aD.s0, aE.s0, aF.s0); - res1 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s1, a1.s1, a2.s1, a3.s1, a4.s1, a5.s1, a6.s1, a7.s1, - a8.s1, a9.s1, aA.s1, aB.s1, aC.s1, aD.s1, aE.s1, aF.s1); -#if N0 > 2 - res2 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s2, a1.s2, a2.s2, a3.s2, a4.s2, a5.s2, a6.s2, a7.s2, - a8.s2, a9.s2, aA.s2, aB.s2, aC.s2, aD.s2, aE.s2, aF.s2); -#endif // N0 > 2 -#if N0 > 3 - res3 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s3, a1.s3, a2.s3, a3.s3, a4.s3, a5.s3, a6.s3, a7.s3, - a8.s3, a9.s3, aA.s3, aB.s3, aC.s3, aD.s3, aE.s3, aF.s3); -#endif // N0 > 3 -#if N0 > 4 - res4 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s4, a1.s4, a2.s4, a3.s4, a4.s4, a5.s4, a6.s4, a7.s4, - a8.s4, a9.s4, aA.s4, aB.s4, aC.s4, aD.s4, aE.s4, aF.s4); - res5 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s5, a1.s5, a2.s5, a3.s5, a4.s5, a5.s5, a6.s5, a7.s5, - a8.s5, a9.s5, aA.s5, aB.s5, aC.s5, aD.s5, aE.s5, aF.s5); - res6 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s6, a1.s6, a2.s6, a3.s6, a4.s6, a5.s6, a6.s6, a7.s6, - a8.s6, a9.s6, aA.s6, aB.s6, aC.s6, aD.s6, aE.s6, aF.s6); - res7 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s7, a1.s7, a2.s7, a3.s7, a4.s7, a5.s7, a6.s7, a7.s7, - a8.s7, a9.s7, aA.s7, aB.s7, aC.s7, aD.s7, aE.s7, aF.s7); -#endif // N0 > 4 -#if N0 > 8 - res8 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s8, a1.s8, a2.s8, a3.s8, a4.s8, a5.s8, a6.s8, a7.s8, - a8.s8, a9.s8, aA.s8, aB.s8, aC.s8, aD.s8, aE.s8, aF.s8); - res9 = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.s9, a1.s9, a2.s9, a3.s9, a4.s9, a5.s9, a6.s9, a7.s9, - a8.s9, a9.s9, aA.s9, aB.s9, aC.s9, aD.s9, aE.s9, aF.s9); - resA = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sA, a1.sA, a2.sA, a3.sA, a4.sA, a5.sA, a6.sA, a7.sA, - a8.sA, a9.sA, aA.sA, aB.sA, aC.sA, aD.sA, aE.sA, aF.sA); - resB = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sB, a1.sB, a2.sB, a3.sB, a4.sB, a5.sB, a6.sB, a7.sB, - a8.sB, a9.sB, aA.sB, aB.sB, aC.sB, aD.sB, aE.sB, aF.sB); - resC = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sC, a1.sC, a2.sC, a3.sC, a4.sC, a5.sC, a6.sC, a7.sC, - a8.sC, a9.sC, aA.sC, aB.sC, aC.sC, aD.sC, aE.sC, aF.sC); - resD = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sD, a1.sD, a2.sD, a3.sD, a4.sD, a5.sD, a6.sD, a7.sD, - a8.sD, a9.sD, aA.sD, aB.sD, aC.sD, aD.sD, aE.sD, aF.sD); - resE = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sE, a1.sE, a2.sE, a3.sE, a4.sE, a5.sE, a6.sE, a7.sE, - a8.sE, a9.sE, aA.sE, aB.sE, aC.sE, aD.sE, aE.sE, aF.sE); - resF = (VEC_DATA_TYPE(DATA_TYPE, K0))(a0.sF, a1.sF, a2.sF, a3.sF, a4.sF, a5.sF, a6.sF, a7.sF, - a8.sF, a9.sF, aA.sF, aB.sF, aC.sF, aD.sF, aE.sF, aF.sF); -#endif // N0 > 8 - -#else // N0 == 16 -#error "Not supported N0 value" -#endif // N0 > 2 - - // ---------------------------Store the output values ------------------------------ - REPEAT_VAR_INIT_TO_CONST(16, uint, zout, 0); - STORE_BLOCK(N0, K0, DATA_TYPE, res, output_ptr, OUTPUT_STEP_X * sizeof(DATA_TYPE), zout); - -#undef BLOCK_SIZE -#undef OUTPUT_OFFSET_X -#undef OUTPUT_STEP_X -} -#endif // defined(TRANSPOSE) -#endif // defined(K0) && defined(N0) && defined(H0) && defined(DATA_TYPE) && defined(SRC_HEIGHT) - -#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K) +#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) #define CONCAT(a, b) a##b @@ -997,14 +148,14 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src), #error "N0 value not supported" #endif // N0 conditions +#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix is NOT reshaped * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl * * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) - * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64) + * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters. * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) @@ -1056,6 +207,9 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src), * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -1077,7 +231,10 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define RHS_BLOCK_SIZE ((K0) * (N0)) @@ -1288,9 +445,11 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X +#undef RHS_STEP_LOOP } +#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_T) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image * The LHS matrix is NOT reshaped * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed @@ -1298,7 +457,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), * * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) + * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters. * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT * could be different from the value returned by get_image_height(rhs_img). @@ -1348,6 +507,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -1369,12 +531,15 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) -#define LEFTOVER_K (K % K0) + const uint LEFTOVER_K = K % K0; // Block size #define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0)) @@ -1477,99 +642,100 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP; } -#if LEFTOVER_K != 0 - // Note: We cannot read out-of-bound elements from the RHS matrix because - // the RHS width is always multiple of K0. This is not be true for the LHS matrix - - union UNION_VEC_TYPE + if(LEFTOVER_K != 0) { - DATA_TYPE s[K0]; - VEC_DATA_TYPE(DATA_TYPE, K0) - v; - }; - - union UNION_VEC_TYPE a0 = {.v = 0 }; + // Note: We cannot read out-of-bound elements from the RHS matrix because + // the RHS width is always multiple of K0. This is not be true for the LHS matrix + // Left-over accumulations for LHS matrix + + union UNION_VEC_TYPE + { + DATA_TYPE s[K0]; + VEC_DATA_TYPE(DATA_TYPE, K0) + v; + }; + + union UNION_VEC_TYPE a0 = {.v = 0 }; #if M0 > 1 - union UNION_VEC_TYPE a1 = {.v = 0 }; + union UNION_VEC_TYPE a1 = {.v = 0 }; #endif // M0 > 1 #if M0 > 2 - union UNION_VEC_TYPE a2 = {.v = 0 }; + union UNION_VEC_TYPE a2 = {.v = 0 }; #endif // M0 > 2 #if M0 > 3 - union UNION_VEC_TYPE a3 = {.v = 0 }; + union UNION_VEC_TYPE a3 = {.v = 0 }; #endif // M0 > 3 #if M0 > 4 - union UNION_VEC_TYPE a4 = {.v = 0 }; + union UNION_VEC_TYPE a4 = {.v = 0 }; #endif // M0 > 4 #if M0 > 5 - union UNION_VEC_TYPE a5 = {.v = 0 }; + union UNION_VEC_TYPE a5 = {.v = 0 }; #endif // M0 > 5 #if M0 > 6 - union UNION_VEC_TYPE a6 = {.v = 0 }; + union UNION_VEC_TYPE a6 = {.v = 0 }; #endif // M0 > 6 #if M0 > 7 - union UNION_VEC_TYPE a7 = {.v = 0 }; + union UNION_VEC_TYPE a7 = {.v = 0 }; #endif // M0 > 7 - REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); + REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); - // Load from RHS matrix - LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); + // Load from RHS matrix + LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); - // Load from LHS matrix - for(int k = 0; k < LEFTOVER_K; ++k) - { - a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0); + // Load from LHS matrix + for(int k = 0; k < LEFTOVER_K; ++k) + { + a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0); #if M0 > 1 - a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1); + a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1); #endif // M0 > 1 #if M0 > 2 - a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2); + a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2); #endif // M0 > 2 #if M0 > 3 - a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3); + a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3); #endif // M0 > 3 #if M0 > 4 - a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4); + a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4); #endif // M0 > 4 #if M0 > 5 - a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5); + a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5); #endif // M0 > 5 #if M0 > 6 - a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6); + a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6); #endif // M0 > 6 #if M0 > 7 - a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7); + a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7); #endif // M0 > 7 - lhs_offset += sizeof(DATA_TYPE); - } + lhs_offset += sizeof(DATA_TYPE); + } - // Accumulate - ARM_DOT_K0XN0(K0, a0.v, b, c0); + // Accumulate + ARM_DOT_K0XN0(K0, a0.v, b, c0); #if M0 > 1 - ARM_DOT_K0XN0(K0, a1.v, b, c1); + ARM_DOT_K0XN0(K0, a1.v, b, c1); #endif // M0 > 1 #if M0 > 2 - ARM_DOT_K0XN0(K0, a2.v, b, c2); + ARM_DOT_K0XN0(K0, a2.v, b, c2); #endif // M0 > 2 #if M0 > 3 - ARM_DOT_K0XN0(K0, a3.v, b, c3); + ARM_DOT_K0XN0(K0, a3.v, b, c3); #endif // M0 > 3 #if M0 > 4 - ARM_DOT_K0XN0(K0, a4.v, b, c4); + ARM_DOT_K0XN0(K0, a4.v, b, c4); #endif // M0 > 4 #if M0 > 5 - ARM_DOT_K0XN0(K0, a5.v, b, c5); + ARM_DOT_K0XN0(K0, a5.v, b, c5); #endif // M0 > 5 #if M0 > 6 - ARM_DOT_K0XN0(K0, a6.v, b, c6); + ARM_DOT_K0XN0(K0, a6.v, b, c6); #endif // M0 > 6 #if M0 > 7 - ARM_DOT_K0XN0(K0, a7.v, b, c7); + ARM_DOT_K0XN0(K0, a7.v, b, c7); #endif // M0 > 7 - -#endif // LEFTOVER_K != 0 + } __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y); @@ -1635,10 +801,10 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X -#undef LEFTOVER_K +#undef RHS_STEP_LOOP #undef PIXEL_UNIT } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE) #define VFMA(a, b, c) \ ({ \ @@ -1717,13 +883,14 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), #error "M0 not supported" #endif // M0 not supported +#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix is NOT reshaped * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl * * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90). + * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters. * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4). * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2) @@ -1775,6 +942,9 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -1796,7 +966,10 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define RHS_BLOCK_SIZE ((K0) * (N0)) @@ -2032,9 +1205,11 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X +#undef RHS_STEP_LOOP } +#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_NT) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix is NOT reshaped * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed @@ -2042,7 +1217,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), * * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90). + * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters. * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT * could be different from the value returned by get_image_height(rhs_img). @@ -2092,6 +1267,9 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -2113,7 +1291,10 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) @@ -2125,9 +1306,11 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), #if defined(RHS_INTERLEAVE) #define RHS_OFFSET_X (PIXEL_UNIT) #define RHS_STEP_X ((PIXEL_UNIT) * (H0)) +#define RHS_STEP_LOOP 1 #else // defined(RHS_INTERLEAVE) #define RHS_OFFSET_X (RHS_BLOCK_SIZE) #define RHS_STEP_X (PIXEL_UNIT) +#define RHS_STEP_LOOP (H0) #endif // defined(RHS_INTERLEAVE) uint x = get_global_id(0); @@ -2342,11 +1525,12 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), #undef RHS_BLOCK_SIZE #undef RHS_OFFSET_X #undef RHS_STEP_X +#undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) -#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(M) && defined(N) && defined(K) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE) +#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) -#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N) +#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) #if defined(MIXED_PRECISION) #if K0 == 2 @@ -2525,6 +1709,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), #error "N0 value not supported" #endif // N0 conditions +#if defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed @@ -2581,12 +1766,14 @@ __kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -2594,7 +1781,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(bias), #endif // defined(BETA) IMAGE_DECLARATION(dst), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -2605,7 +1791,10 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define LHS_BLOCK_SIZE ((K0) * (M0)) @@ -2661,7 +1850,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); - for(int i = 0; i < k; i += K0) + for(int i = 0; i < K; i += K0) { // Supported cases (M0, K0): // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 @@ -2798,8 +1987,9 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } +#endif // defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object. * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed @@ -2855,12 +2045,14 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -2868,7 +2060,6 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(bias), #endif // defined(BETA) IMAGE_DECLARATION(dst), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -2879,7 +2070,10 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) @@ -3070,7 +2264,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE) #if defined(LHS_TRANSPOSE) @@ -3182,6 +2376,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), CONCAT(ARM_MM_T_NT_M0xN0x, K0) \ (M0, N0, TYPE, A, B, C) +#if defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed @@ -3236,12 +2431,14 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(rhs), @@ -3249,7 +2446,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(bias), #endif // defined(BETA) IMAGE_DECLARATION(dst), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -3260,7 +2456,10 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Block size #define LHS_BLOCK_SIZE ((K0) * (M0)) @@ -3322,7 +2521,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr); __global DATA_TYPE *rhs = (__global DATA_TYPE *)(rhs_addr); - for(int i = 0; i < k; i += K0) + for(int i = 0; i < K; i += K0) { VEC_DATA_TYPE(DATA_TYPE, M0) a0; @@ -3562,8 +2761,9 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), #undef RHS_OFFSET_X #undef RHS_STEP_X } +#endif // defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT) -#if defined(OPENCL_IMAGE_SUPPORT) +#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object. * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed @@ -3572,7 +2772,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE). * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24). + * @note The GEMM's dimensions M, N and K must be passed at runtime. * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32) * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT * could be different from the value returned by get_image_height(rhs_img). @@ -3617,12 +2817,14 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs), * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix - * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped. * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes) * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes) * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. */ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), __read_only image2d_t rhs_img, @@ -3630,7 +2832,6 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), IMAGE_DECLARATION(bias), #endif // defined(BETA) IMAGE_DECLARATION(dst), - uint k, uint lhs_stride_z, uint rhs_stride_z, #if defined(BETA) @@ -3641,7 +2842,10 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D - ) + , + const int M, + const int N, + const int K) { // Pixel unit #define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) @@ -3933,13 +3137,13 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), #undef LHS_STEP_LOOP #undef RHS_STEP_LOOP } -#endif // defined(OPENCL_IMAGE_SUPPORT) +#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE) #endif // defined(LHS_TRANSPOSE) -#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) && defined(M) && defined(N) +#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR) -#if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) +#if defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE) #define VFMA(a, b, c) \ ({ \ @@ -4018,14 +3222,14 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), #error "M0 not supported" #endif // M0 not supported +#if defined(GEMM_MM_NATIVE) /** This OpenCL kernel computes the matrix multiplication between 2 matrices. * The LHS matrix is NOT reshaped * The RHS matrix is NOT reshaped * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl * * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time. - * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (e.g. -DM=52, -DN=30 and -DK=90) - * @note The number of columns of LHS matrix must be passed at compile time using -DK (e.g. -DK=64) + * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters. * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2) * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (e.g., -DK0=2) * @note The number of N0 columns to process must be passed at compile time using -DN0 (e.g. -DN0=2) @@ -4073,6 +3277,9 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), * @param[in] rhs_stride_z Stride of the RHS matrix in Z dimension (in bytes) * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] M Number of rows in LHS matrix not reshaped. + * @param[in] N Number of columns in RHS matrix not reshaped. + * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped. * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ @@ -4087,7 +3294,10 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs), #if defined(BETA) uint bias_stride_z, #endif //defined(BETA) - uint dst_stride_z + uint dst_stride_z, + const int M, + const int N, + const int K #if defined(REINTERPRET_INPUT_AS_3D) , uint lhs_cross_plane_pad @@ -4303,7 +3513,8 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs), // Store output block STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } -#endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE) +#endif // defined(GEMM_MM_NATIVE) +#endif // defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE) #if defined(BETA) /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta: @@ -4389,4 +3600,4 @@ __kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src), vstore8(out, 0, (__global half *)dst.ptr); } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) -#endif // defined(BETA)
\ No newline at end of file +#endif // defined(BETA) diff --git a/src/core/CL/cl_kernels/common/gemm_utils.cl b/src/core/CL/cl_kernels/common/gemm_utils.cl new file mode 100644 index 000000000..be57d94ce --- /dev/null +++ b/src/core/CL/cl_kernels/common/gemm_utils.cl @@ -0,0 +1,458 @@ +/* + * Copyright (c) 2017-2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "gemm_helpers.h" +#include "helpers.h" +#include "repeat.h" +#include "tile_helpers.h" + +#if defined(RESHAPE_LHS_NT) +/** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (not transposed) in + * the output matrix unrolling the values. + * + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) + * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) + * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). + * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_M0 (e.g. -DPARTIAL_M0=1) + * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_K0 (e.g. -DPARTIAL_K0=1) + * @note Only the following values for M0, K0 and V0 are supported: + * M0: 2,3,4,5,6,7,8 + * K0: 2,3,4,8,16 + * V0: greater than 0 + * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: All + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the depth dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: All + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the depth dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] M The size of height dimension of the source tensor, affected by reinterpret_input_as_3d + * @param[in] V0 The number of blocks to place on the same row. It must be greater than 0. + */ +__kernel void gemm_reshape_lhs_matrix_nt(TENSOR3D_T(src, BUFFER), + TENSOR3D_T(dst, BUFFER), + const int M, + const int V0) +{ + // Block size +#define BLOCK_SIZE ((M0) * (K0)) + + // Output offset X +#if defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (K0) +#else // defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (BLOCK_SIZE) +#endif // defined(INTERLEAVE) + + // Output step X +#if defined(INTERLEAVE) +#define OUTPUT_STEP_X (K0) * (V0) +#else // Do not interleave +#define OUTPUT_STEP_X (K0) +#endif // defined(INTERLEAVE) + + const int x = GET_SPATIAL_IDX(0, 1, 0); // K + const int y = GET_SPATIAL_IDX(1, 1, 0); // M + const int z = GET_SPATIAL_IDX(2, 1, 0); // Batch size + + const int xi = x * K0; + const int yi = y * M0; + + const int xo = x * BLOCK_SIZE * V0 + (y % V0) * OUTPUT_OFFSET_X; + const int yo = (y / V0); + + // src_stride_z is expressed as M * src_stride_y, to handle case where reinterpret_input_as_3d=true + src_offset_first_element_in_bytes += yi * src_stride_y + z * M * src_stride_y; + dst_offset_first_element_in_bytes += yo * dst_stride_y + z * dst_stride_z; + + TILE(DATA_TYPE, M0, K0, in); + + // Initialize the input tile to zero + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + in[_i].v = 0; + }); + + bool x_cond = (xi + K0 >= src_w) && (PARTIAL_K0 != 0); + bool y_cond = (yi + M0 >= M) && (PARTIAL_M0 != 0); + // Load input tile + TILE(uint, M0, 1, in_indirect_y); + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + in_indirect_y[_i].v = _i; + + }); +#if PARTIAL_M0 != 0 + if(y_cond) + { + T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, PARTIAL_M0, K0, PARTIAL_K0, BUFFER, src, xi, src_stride_y, x_cond, in, in_indirect_y); + } + else +#endif // PARTIAL_M0 != 0 + { + T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, K0, PARTIAL_K0, BUFFER, src, xi, src_stride_y, x_cond, in, in_indirect_y); + } + + // Store output tile + TILE(uint, M0, 1, dst_indirect_y); + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + dst_indirect_y[_i].v = _i; + }); + + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, K0, 0, BUFFER, dst, xo, (OUTPUT_STEP_X * sizeof(DATA_TYPE)), false, in, dst_indirect_y); +#undef BLOCK_SIZE +#undef OUTPUT_OFFSET_X +#undef OUTPUT_STEP_X +} +#endif // defined(RESHAPE_LHS_NT) + +#if defined(RESHAPE_LHS_T) +/** This OpenCL kernel reshapes the lhs input matrix. The kernel splits the input matrix in blocks of size M0xK0 and stores each one (transposed) in + * the output matrix unrolling the values. + * + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The width of the input tensor must be passed at compile time using -DSRC_WIDTH (e.g. -DSRC_WIDTH=16) + * @note The height of the input tensor must be passed at compile time using -DSRC_HEIGHT (e.g. -DSRC_HEIGHT=16) + * @note The block's dimensions (M0 and K0) must be passed at compile time using -DM0 and -DK0 (e.g. -DM0=2, -DK0=2). + * @note The size of the partial load block in y must be passed at compile time using -DPARTIAL_M0 (e.g. -DPARTIAL_M0=1) + * @note The size of the partial load block in x must be passed at compile time using -DPARTIAL_K0 (e.g. -DPARTIAL_K0=1) + * @note Only the following values for M0, K0 and V0 are supported: + * M0: 2,3,4,8,16 + * K0: 2,3,4,8,16 + * V0: greater than 0 + * @note If the M0xK0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: All + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the depth dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: All + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the depth dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] M The size of height dimension of the source tensor, affected by reinterpret_input_as_3d + * @param[in] V0 The number of blocks to place on the same row. It must be greater than 0 + */ +__kernel void gemm_reshape_lhs_matrix_t(TENSOR3D_T(src, BUFFER), + TENSOR3D_T(dst, BUFFER), + const int M, + const int V0) +{ + // Block size +#define BLOCK_SIZE ((M0) * (K0)) + + // Output offset X +#if defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (M0) +#else // defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (BLOCK_SIZE) +#endif // defined(INTERLEAVE) + + // Output step X +#if defined(INTERLEAVE) +#define OUTPUT_STEP_X (M0) * (V0) +#else // Do not interleave +#define OUTPUT_STEP_X (M0) +#endif // defined(INTERLEAVE) + + const int x = GET_SPATIAL_IDX(0, 1, 0); // K + const int y = GET_SPATIAL_IDX(1, 1, 0); // M + const int z = GET_SPATIAL_IDX(2, 1, 0); // Batch size + + const int xi = x * K0; + const int yi = y * M0; + + const int xo = x * BLOCK_SIZE * V0 + ((y % V0) * OUTPUT_OFFSET_X); + const int yo = (y / V0); + + // src_stride_z is expressed as M * src_stride_y, to handle case where reinterpret_input_as_3d=true + src_offset_first_element_in_bytes += yi * src_stride_y + z * M * src_stride_y; + dst_offset_first_element_in_bytes += yo * dst_stride_y + z * dst_stride_z; + + TILE(DATA_TYPE, M0, K0, in); + TILE(DATA_TYPE, K0, M0, in_tr); + + // Initialize the tile to zero + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + in[_i].v = 0; + }); + + // Load input tile + bool x_cond = (xi + K0 >= src_w) && (PARTIAL_K0 != 0); + bool y_cond = (yi + M0 >= M) && (PARTIAL_M0 != 0); + + TILE(uint, M0, 1, in_indirect_y); + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + in_indirect_y[_i].v = _i; + + }); +#if PARTIAL_M0 != 0 + if(y_cond) + { + T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, PARTIAL_M0, K0, PARTIAL_K0, BUFFER, src, xi, src_stride_y, x_cond, in, in_indirect_y); + } + else +#endif // PARTIAL_M0 != 0 + { + T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, K0, PARTIAL_K0, BUFFER, src, xi, src_stride_y, x_cond, in, in_indirect_y); + } + // Transpose input tile + LOOP_UNROLLING(int, m0, 0, 1, M0, + { + LOOP_UNROLLING(int, k0, 0, 1, K0, + { + in_tr[k0].s[m0] = in[m0].s[k0]; + }) + }); + + TILE(uint, K0, 1, dst_indirect_y); + LOOP_UNROLLING(int, _i, 0, 1, K0, + { + dst_indirect_y[_i].v = _i; + }); + + // Store output tile + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, K0, M0, 0, BUFFER, dst, xo, (OUTPUT_STEP_X * sizeof(DATA_TYPE)), false, in_tr, dst_indirect_y); + +#undef BLOCK_SIZE +#undef OUTPUT_OFFSET_X +#undef OUTPUT_STEP_X +} +#endif // defined(RESHAPE_LHS_T) + +#if defined(RESHAPE_RHS_NT) +/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (not transposed) in + * the output matrix unrolling the values. + * + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). + * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. + * @note Only the following values for K0, N0 and H0 are supported: + * N0: 2,3,4,8,16 + * K0: 1,2,3,4,8,16 + * H0: greater than 0 + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: All + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the depth dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: All + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the depth dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] H0 The number of blocks to place on the same row. It must be greater than 0 + */ +__kernel void gemm_reshape_rhs_matrix_nt(TENSOR3D_T(src, BUFFER), + TENSOR3D_T(dst, BUFFER), + const int H0) +{ + // Block size +#define BLOCK_SIZE ((K0) * (N0)) + + // Output offset X +#if defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (N0) +#else // defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (BLOCK_SIZE) +#endif // defined(INTERLEAVE) + + // Output step X +#if defined(INTERLEAVE) +#define OUTPUT_STEP_X (N0) * (H0) +#else // Do not interleave +#define OUTPUT_STEP_X (N0) +#endif // defined(INTERLEAVE) + + const int x = GET_SPATIAL_IDX(0, 1, 0); + const int y = GET_SPATIAL_IDX(1, 1, 0); + const int z = GET_SPATIAL_IDX(2, 1, 0); + + const int xi = x * N0; + const int yi = y * K0; + + const int xo = y * BLOCK_SIZE * H0 + (x % H0) * OUTPUT_OFFSET_X; + const int yo = (x / H0); + + src_offset_first_element_in_bytes += yi * src_stride_y + z * src_stride_z; + dst_offset_first_element_in_bytes += yo * dst_stride_y + z * dst_stride_z; + + TILE(DATA_TYPE, K0, N0, in); + + // Initialize the tile to zero + for(int i = 0; i < K0; ++i) + { + in[i].v = 0; + } + + // Load input tile + for(int i = 0; i < K0; ++i) + { + if(yi + i < src_h) + { + in[i].v = V_LOAD(DATA_TYPE, N0, BUFFER, src, xi, i, src_stride_y); + } + } + + TILE(uint, K0, 1, dst_indirect_y); + for(int i = 0; i < K0; ++i) + { + dst_indirect_y[i].v = i; + } + + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, K0, N0, 0, BUFFER, dst, xo, (OUTPUT_STEP_X * sizeof(DATA_TYPE)), false, in, dst_indirect_y); + +#undef BLOCK_SIZE +#undef OUTPUT_OFFSET_X +#undef OUTPUT_STEP_X +} +#endif // defined(RESHAPE_RHS_NT) + +#if defined(RESHAPE_RHS_T) +/** This OpenCL kernel reshapes the rhs input matrix. The kernel splits the input matrix in blocks of size K0xN0 and stores each one (transposed) in + * the output matrix unrolling the values. + * + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The block's dimensions (K0 and N0) must be passed at compile time using -DK0 and -DN0 (e.g. -DK0=2, -DN0=2). + * @note If the K0xN0 blocks have to be interleaved, the option -DINTERLEAVE must passed at compile time. + * @note The option -DTRANSPOSE must passed at compile time. + * @note Only the following values for K0, N0 and H0 are supported: + * N0: 2,3,4,8,16 + * K0: 2,3,4,8,16 + * H0: greater than 0 + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: All + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the depth dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: All + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the depth dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] H0 The number of blocks to place on the same row. It must be greater than 0. + */ +__kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_T(src, BUFFER), + TENSOR3D_T(dst, BUFFER), + const int H0) +{ + // Block size +#define BLOCK_SIZE ((K0) * (N0)) + + // Output offset X +#if defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (K0) +#else // defined(INTERLEAVE) +#define OUTPUT_OFFSET_X (BLOCK_SIZE) +#endif // defined(INTERLEAVE) + + // Output step X +#if defined(INTERLEAVE) +#define OUTPUT_STEP_X (K0) * (H0) +#else // Do not interleave +#define OUTPUT_STEP_X (K0) +#endif // defined(INTERLEAVE) + + const int x = GET_SPATIAL_IDX(0, 1, 0); + const int y = GET_SPATIAL_IDX(1, 1, 0); + const int z = GET_SPATIAL_IDX(2, 1, 0); + + const int xi = x * N0; + const int yi = y * K0; + + const int xo = y * BLOCK_SIZE * H0 + (x % H0) * OUTPUT_OFFSET_X; + const int yo = (x / H0); + + src_offset_first_element_in_bytes += yi * src_stride_y + z * src_stride_z; + dst_offset_first_element_in_bytes += yo * dst_stride_y + z * dst_stride_z; + + TILE(DATA_TYPE, K0, N0, in); + TILE(DATA_TYPE, N0, K0, in_tr); + + // Initialize the tile to zero + for(int i = 0; i < K0; ++i) + { + in[i].v = 0; + } + + // Load input tile + for(int i = 0; i < K0; ++i) + { + if(yi + i < src_h) + { + in[i].v = V_LOAD(DATA_TYPE, N0, BUFFER, src, xi, i, src_stride_y); + } + } + + // Transpose input tile + for(int k0 = 0; k0 < K0; ++k0) + { + for(int n0 = 0; n0 < N0; ++n0) + { + in_tr[n0].s[k0] = in[k0].s[n0]; + } + } + + TILE(uint, N0, 1, dst_indirect_y); + for(int i = 0; i < N0; ++i) + { + dst_indirect_y[i].v = i; + } + + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, N0, K0, 0, BUFFER, dst, xo, (OUTPUT_STEP_X * sizeof(DATA_TYPE)), false, in_tr, dst_indirect_y); + +#undef BLOCK_SIZE +#undef OUTPUT_OFFSET_X +#undef OUTPUT_STEP_X +} + +#endif // defined(RESHAPE_RHS_T)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h index 88a7665ee..bfb693e37 100644 --- a/src/core/CL/cl_kernels/helpers.h +++ b/src/core/CL/cl_kernels/helpers.h @@ -392,18 +392,18 @@ #define vload_partial_12(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ vload_partial_4(DATA.s89AB, OFFSET, PTR + 8); - +// For vload_partial_{13,14,15}, an 8-vector size has been passed, because vectors size of size 5,6,7 are not supported #define vload_partial_13(DATA, OFFSET, PTR) \ vload_partial_8(DATA.s01234567, OFFSET, PTR); \ - vload_partial_5(DATA.s89ABC, OFFSET, PTR + 8); + 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.s89ABCD, OFFSET, PTR + 8); + 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.s89ABCDE, OFFSET, PTR + 8); + vload_partial_7(DATA.s89ABCDEF, OFFSET, PTR + 8); #define vload_partial_16(DATA, OFFSET, PTR) \ DATA = vload16(OFFSET, PTR); diff --git a/src/core/CL/cl_kernels/nchw/direct_convolution.cl b/src/core/CL/cl_kernels/nchw/direct_convolution.cl new file mode 100644 index 000000000..866f62da9 --- /dev/null +++ b/src/core/CL/cl_kernels/nchw/direct_convolution.cl @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" +#include "helpers_asymm.h" + +/** This kernel performs a direct convolution to convolve the low three dimensions. + * + * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 + * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 + * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH + * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. + * @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234 + * @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4 + * @note The input offset quantization parameter must be passed at compile time using -DINPUT_OFFSET e.g. -DINPUT_OFFSET=3 + * @note The weights offset quantization parameter must be passed at compile time using -DWEIGHTS_OFFSET e.g. -DWEIGHTS_OFFSET=3 + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr + * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) + * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) + * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) + * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) + * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor + * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr + * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) + * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor + * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension + */ +__kernel void direct_convolution_nchw( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), + TENSOR3D_DECLARATION(weights), +#ifdef HAS_BIAS + VECTOR_DECLARATION(biases), +#endif /* defined(HAS_BIAS) */ + unsigned int weights_stride_w) +{ + const int id0 = get_global_id(0); + const int id1 = get_global_id(1); + const int id2 = get_global_id(2); + + const int x_coords = (id0 * STRIDE_X) - PAD_LEFT; + const int y_coords = (id1 * STRIDE_Y) - PAD_TOP; + + const int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) * sizeof(DATA_TYPE); + + __global uchar *src_addr = (__global uchar *)(src_ptr + src_offset_first_element_in_bytes); + __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + id2 * weights_stride_w); + __global uchar *dst_addr = (__global uchar *)dst_ptr + dst_offset_first_element_in_bytes + x_offs + id1 * dst_stride_y + id2 * dst_stride_z; + +#ifdef IS_QUANTIZED + int acc_value = 0; +#else /* IS_QUANTIZED */ + DATA_TYPE acc_value = 0; +#endif /* IS_QUANTIZED */ + for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) + { + for(int y = 0; y < WEI_HEIGHT; ++y) + { + for(int x = 0; x < WEI_WIDTH; ++x) + { + const int idx_x = (x_coords + x); + const int idx_y = (y_coords + y); + if((idx_x >= 0 && idx_x < SRC_WIDTH) && (idx_y >= 0 && idx_y < SRC_HEIGHT)) + { + const int weight_offset = x + (WEI_HEIGHT * y); + const int input_offset = idx_x + SRC_WIDTH * idx_y; +#ifdef IS_QUANTIZED + int weight = convert_int(*((__global DATA_TYPE *)weights_addr + weight_offset)); + int input = convert_int(*((__global DATA_TYPE *)src_addr + input_offset)); + acc_value += (input + INPUT_OFFSET) * (weight + WEIGHTS_OFFSET); +#else /* IS_QUANTIZED */ + DATA_TYPE weight = *((__global DATA_TYPE *)weights_addr + weight_offset); + DATA_TYPE input = *((__global DATA_TYPE *)src_addr + input_offset); + acc_value += input * weight; +#endif /* IS_QUANTIZED */ + } + } + } + src_addr += src_stride_z; + weights_addr += weights_stride_z; + } + +#ifdef HAS_BIAS + + Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); +#ifdef IS_QUANTIZED + int bias = *((__global int *)(vector_offset(&biases, id2))); +#else /* IS_QUANTIZED */ + DATA_TYPE bias = *((__global DATA_TYPE *)(vector_offset(&biases, id2))); +#endif /* IS_QUANTIZED */ + acc_value += bias; + +#endif /* defined(HAS_BIAS) */ + +#ifdef IS_QUANTIZED + +#if OUTPUT_SHIFT < 0 + acc_value = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc_value, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 1); +#else // OUTPUT_SHIFT < 0 + acc_value = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(acc_value, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 1); +#endif // OUTPUT_SHIFT < 0 + acc_value = acc_value + OUTPUT_OFFSET; +#endif /* IS_QUANTIZED */ + + *(__global DATA_TYPE *)dst_addr = CONVERT_SAT(acc_value, DATA_TYPE); +}
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nchw/direct_convolution1x1.cl b/src/core/CL/cl_kernels/nchw/direct_convolution1x1.cl deleted file mode 100644 index 8ab2d1d4e..000000000 --- a/src/core/CL/cl_kernels/nchw/direct_convolution1x1.cl +++ /dev/null @@ -1,316 +0,0 @@ -/* - * Copyright (c) 2016-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" - -#undef CONVERT_SAT - -#define ADD_OP(a, b) ((a) + (b)) -#define MUL_OP(a, b) ((a) * (b)) -#define CONVERT_SAT(a, b) ((a)) - -#if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if STRIDE_X == 3 -#define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size -#define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size) -#elif STRIDE_X == 2 -#define INPUT_PIXEL(data_size) extract_input_stride2 -#elif STRIDE_X == 1 -#define INPUT_PIXEL(data_size) extract_input_stride1 -#else /* STRIDE_X not equals 1, 2 or 3 */ -#error "Only support strides 1, 2 and 3" -#endif /* STRIDE_X == 3 */ - -/** Extracts a 1D horizontal vector from the input tensor with stride as 1. - * - * @param[in] input_pixel Pointer to the first pixel. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_pixel) -{ - return vload8(0, input_pixel); -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 2. - * - * @param[in] input_pixel Pointer to the first pixel. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_pixel) -{ - VEC_DATA_TYPE(DATA_TYPE, 16) - temp = vload16(0, input_pixel); - return temp.s02468ace; -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 32-bit data size. - * - * @param[in] input_pixel Pointer to the first pixel. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_TYPE *input_pixel) -{ - VEC_DATA_TYPE(DATA_TYPE, 4) - temp1 = vload4(0, input_pixel); - VEC_DATA_TYPE(DATA_TYPE, 4) - temp2 = vload4(0, input_pixel + 6); - VEC_DATA_TYPE(DATA_TYPE, 4) - temp3 = vload4(0, input_pixel + 12); - VEC_DATA_TYPE(DATA_TYPE, 4) - temp4 = vload4(0, input_pixel + 18); - return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s03, temp2.s03, temp3.s03, temp4.s03); -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 16-bit data size. - * - * @param[in] input_pixel Pointer to the first pixel. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_TYPE *input_pixel) -{ - VEC_DATA_TYPE(DATA_TYPE, 8) - temp1 = vload8(0, input_pixel); - VEC_DATA_TYPE(DATA_TYPE, 8) - temp2 = vload8(0, input_pixel + 8); - VEC_DATA_TYPE(DATA_TYPE, 8) - temp3 = vload8(0, input_pixel + 16); - return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s036, temp2.s147, temp3.s25); -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. - * - * @param[in] input_pixel Pointer to the first pixel. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_TYPE *input_pixel) -{ - VEC_DATA_TYPE(DATA_TYPE, 16) - temp1 = vload16(0, input_pixel); - VEC_DATA_TYPE(DATA_TYPE, 16) - temp2 = vload16(0, input_pixel + 12); - return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369); -} - -/** This kernel performs a direct convolution to convolve the low three dimensions. - * - * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float - * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 - * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution1x1( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); -#endif /* defined(HAS_BIAS) */ - - VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8) - values = 0; - - const uint z_index = get_global_id(2); - - weights.ptr += z_index * weights_stride_w; - for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) - { - DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr; - VEC_DATA_TYPE(DATA_TYPE, 8) - input_pixel = INPUT_PIXEL(DATA_SIZE)((__global DATA_TYPE *)src.ptr); - values = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, input_pixel)); - src.ptr += src_stride_z; - weights.ptr += weights_stride_z; - } - -#ifdef HAS_BIAS - values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index)))); -#endif /* defined(HAS_BIAS) */ - - vstore8(CONVERT_SAT(values, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr); -} -#endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if defined(WEIGHTS_DEPTH) - -#define CONVOLUTION1x1_BIFROST(acc, src, weight_value) \ - ({ \ - acc.s0 = mad(src.s0, weight_value, acc.s0); \ - acc.s1 = mad(src.s1, weight_value, acc.s1); \ - acc.s2 = mad(src.s2, weight_value, acc.s2); \ - acc.s3 = mad(src.s3, weight_value, acc.s3); \ - }) - -/** An optimized direct convolution 1x1 OpenCL kernel for Bifrost architectures when the data type is F32 - * - * @note This OpenCL kernel works only with stride_x and stride_y equal to 1 - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note In case biases, -DHAS_BIAS must to be passed at compile - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution1x1_f32_bifrost( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - // Get the kernel index - const int kernel_index = get_global_id(2); - - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - float4 acc0 = 0.0f; - float4 acc1 = 0.0f; - float4 acc2 = 0.0f; - float4 acc3 = 0.0f; - - __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w); - __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); - - for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d) - { - // Load the weights - float weight = *((__global float *)weights_addr); - - // Load values from row0 of input tensor - float4 src0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); - float4 src1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); - float4 src2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); - float4 src3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); - - CONVOLUTION1x1_BIFROST(acc0, src0, weight); - CONVOLUTION1x1_BIFROST(acc1, src1, weight); - CONVOLUTION1x1_BIFROST(acc2, src2, weight); - CONVOLUTION1x1_BIFROST(acc3, src3, weight); - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - - float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index))); - - acc0.s0 += bias; - acc0.s1 += bias; - acc0.s2 += bias; - acc0.s3 += bias; - acc1.s0 += bias; - acc1.s1 += bias; - acc1.s2 += bias; - acc1.s3 += bias; - acc2.s0 += bias; - acc2.s1 += bias; - acc2.s2 += bias; - acc2.s3 += bias; - acc3.s0 += bias; - acc3.s1 += bias; - acc3.s2 += bias; - acc3.s3 += bias; -#endif /* defined(HAS_BIAS) */ - - vstore4(acc0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); - vstore4(acc1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); - vstore4(acc2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y)); - vstore4(acc3, 0, (__global float *)(dst.ptr + 3 * dst_stride_y)); -} -#endif // defined(WEIGHTS_DEPTH) diff --git a/src/core/CL/cl_kernels/nchw/direct_convolution3x3.cl b/src/core/CL/cl_kernels/nchw/direct_convolution3x3.cl deleted file mode 100644 index 811df053c..000000000 --- a/src/core/CL/cl_kernels/nchw/direct_convolution3x3.cl +++ /dev/null @@ -1,291 +0,0 @@ -/* - * Copyright (c) 2016-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" - -#undef CONVERT_SAT - -#define ADD_OP(a, b) ((a) + (b)) -#define MUL_OP(a, b) ((a) * (b)) -#define CONVERT_SAT(a, b) ((a)) - -#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if STRIDE_X == 1 -#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) -#elif STRIDE_X == 2 /* STRIDE_X == 1 */ -#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X == 2 */ - -#define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 3) \ - weights_values0 = vload3(0, weights_row_ptr); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - src0 = vload8(0, src_row_ptr); \ - VEC_DATA_TYPE(DATA_TYPE, 2) \ - src1 = vload2(0, src_row_ptr + 8); \ - \ - acc = ADD_OP(acc, MUL_OP(src0, (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2)); \ - }) - -#define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 3) \ - weights_values0 = vload3(0, weights_row_ptr); \ - VEC_DATA_TYPE(DATA_TYPE, 16) \ - src0 = vload16(0, src_row_ptr); \ - DATA_TYPE src1 = *(src_row_ptr + 16); \ - \ - acc = ADD_OP(acc, MUL_OP(src0.even, (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1)); \ - acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2)); \ - }) - -/** This kernel performs a direct convolution to convolve the low three dimensions. - * - * @note This OpenCL kernel works with stride_x = 1 and 2 - * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note If biases are used then -DHAS_BIAS has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution3x3( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8) - values0 = 0; - - __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0); - __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); - - const int kernel_index = get_global_id(2); - weights_addr += kernel_index * weights_stride_w; - - for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) - { - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y)); - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - - values0 = ADD_OP(values0, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index)))); -#endif /* defined(HAS_BIAS) */ - - vstore8(CONVERT_SAT(values0, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr); -} -#endif //defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if defined(WEIGHTS_DEPTH) - -#define CONVOLUTION1x3_BIFROST(acc, src0, src1, weights_row0) \ - ({ \ - acc.s0 = mad(src0.s0, weights_row0.s0, acc.s0); \ - acc.s1 = mad(src0.s1, weights_row0.s0, acc.s1); \ - acc.s2 = mad(src0.s2, weights_row0.s0, acc.s2); \ - acc.s3 = mad(src0.s3, weights_row0.s0, acc.s3); \ - acc.s0 = mad(src0.s1, weights_row0.s1, acc.s0); \ - acc.s1 = mad(src0.s2, weights_row0.s1, acc.s1); \ - acc.s2 = mad(src0.s3, weights_row0.s1, acc.s2); \ - acc.s3 = mad(src1.s0, weights_row0.s1, acc.s3); \ - acc.s0 = mad(src0.s2, weights_row0.s2, acc.s0); \ - acc.s1 = mad(src0.s3, weights_row0.s2, acc.s1); \ - acc.s2 = mad(src1.s0, weights_row0.s2, acc.s2); \ - acc.s3 = mad(src1.s1, weights_row0.s2, acc.s3); \ - }) - -/** An optimized direct convolution 3x3 OpenCL kernel for Bifrost architectures when the data type is F32 - * - * @note This OpenCL kernel works only with stride_x and stride_y equal to 1 - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note In case biases, -DHAS_BIAS must to be passed at compile - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution3x3_f32_bifrost( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - // Get the kernel index - const int kernel_index = get_global_id(2); - - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - float4 values0 = 0; - float4 values1 = 0; - float4 values2 = 0; - - __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w); - __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); - - // Note: Since each work-item computes 4x3 elements, we need to load 5 rows from the input tensor - - for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d) - { - // Load the weights - float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y)); - float3 weights_row1 = vload3(0, (__global float *)(weights_addr + 1 * weights_stride_y)); - float3 weights_row2 = vload3(0, (__global float *)(weights_addr + 2 * weights_stride_y)); - float4 src0; - float2 src1; - - // Load values from row0 of input tensor - src0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); - src1 = vload2(0, (__global float *)(src_addr + 0 * src_stride_y) + 4); - - CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row0); - - // Load values from row1 of input tensor - src0 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); - src1 = vload2(0, (__global float *)(src_addr + 1 * src_stride_y) + 4); - - // Accumulate - CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row1); - CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row0); - - // Load values from row2 of input tensor - src0 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); - src1 = vload2(0, (__global float *)(src_addr + 2 * src_stride_y) + 4); - - // Accumulate - CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row2); - CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row1); - CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row0); - - // Load values from row3 of input tensor - src0 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); - src1 = vload2(0, (__global float *)(src_addr + 3 * src_stride_y) + 4); - - // Accumulate - CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row2); - CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row1); - - // Row4 - src0 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); - src1 = vload2(0, (__global float *)(src_addr + 4 * src_stride_y) + 4); - - // Accumulate - CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row2); - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - - float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index))); - - values0 += (float4)bias; - values1 += (float4)bias; - values2 += (float4)bias; -#endif /* defined(HAS_BIAS) */ - - vstore4(values0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); - vstore4(values1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); - vstore4(values2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y)); -} -#endif // defined(WEIGHTS_DEPTH) diff --git a/src/core/CL/cl_kernels/nchw/direct_convolution5x5.cl b/src/core/CL/cl_kernels/nchw/direct_convolution5x5.cl deleted file mode 100644 index 59d668f0b..000000000 --- a/src/core/CL/cl_kernels/nchw/direct_convolution5x5.cl +++ /dev/null @@ -1,313 +0,0 @@ -/* - * Copyright (c) 2016-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" - -#undef CONVERT_SAT - -#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if STRIDE_X == 1 -#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) -#elif STRIDE_X == 2 /* STRIDE_X == 1 */ -#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X == 2 */ - -#define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - weights_values0 = vload4(0, weights_row_ptr); \ - DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \ - VEC_DATA_TYPE(DATA_TYPE, 8) \ - src0 = vload8(0, src_row_ptr); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - src1 = vload4(0, src_row_ptr + 8); \ - \ - acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s45, src0.s67, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ - }) - -#define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - weights_values0 = vload4(0, weights_row_ptr); \ - DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \ - VEC_DATA_TYPE(DATA_TYPE, 16) \ - src0 = vload16(0, src_row_ptr); \ - VEC_DATA_TYPE(DATA_TYPE, 4) \ - src1 = vload4(0, src_row_ptr + 16); \ - acc += src0.even * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ - \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ - acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ - }) - -/** This kernel performs a direct convolution to convolve the low three dimensions. - * - * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note If biases are used then -DHAS_BIAS has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution5x5( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - VEC_DATA_TYPE(DATA_TYPE, 8) - values0 = 0; - - __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0); - __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); - - const int kernel_index = get_global_id(2); - weights_addr += kernel_index * weights_stride_w; - - for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) - { - CONVOLUTION1x5(values0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y)); - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - - values0 += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index))); -#endif /* defined(HAS_BIAS) */ - - vstore8(values0, 0, (__global DATA_TYPE *)dst.ptr); -} -#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) - -#if defined(WEIGHTS_DEPTH) - -#define CONVOLUTION1x5_BIFROST(acc, src0, weights_row00, weights_row01) \ - ({ \ - acc.s0 = mad(src0.s0, weights_row00.s0, acc.s0); \ - acc.s1 = mad(src0.s1, weights_row00.s0, acc.s1); \ - acc.s2 = mad(src0.s2, weights_row00.s0, acc.s2); \ - acc.s3 = mad(src0.s3, weights_row00.s0, acc.s3); \ - acc.s0 = mad(src0.s1, weights_row00.s1, acc.s0); \ - acc.s1 = mad(src0.s2, weights_row00.s1, acc.s1); \ - acc.s2 = mad(src0.s3, weights_row00.s1, acc.s2); \ - acc.s3 = mad(src0.s4, weights_row00.s1, acc.s3); \ - acc.s0 = mad(src0.s2, weights_row00.s2, acc.s0); \ - acc.s1 = mad(src0.s3, weights_row00.s2, acc.s1); \ - acc.s2 = mad(src0.s4, weights_row00.s2, acc.s2); \ - acc.s3 = mad(src0.s5, weights_row00.s2, acc.s3); \ - acc.s0 = mad(src0.s3, weights_row00.s3, acc.s0); \ - acc.s1 = mad(src0.s4, weights_row00.s3, acc.s1); \ - acc.s2 = mad(src0.s5, weights_row00.s3, acc.s2); \ - acc.s3 = mad(src0.s6, weights_row00.s3, acc.s3); \ - acc.s0 = mad(src0.s4, weights_row01, acc.s0); \ - acc.s1 = mad(src0.s5, weights_row01, acc.s1); \ - acc.s2 = mad(src0.s6, weights_row01, acc.s2); \ - acc.s3 = mad(src0.s7, weights_row01, acc.s3); \ - }) - -/** An optimized direct convolution 5x5 OpenCL kernel for Bifrost architectures when the data type is F32 - * - * @note This OpenCL kernel works only with stride_x and stride_y equal to 1 - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note If biases are used then -DHAS_BIAS has to be passed at compile time - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution5x5_f32_bifrost( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - // Get the kernel index - const int kernel_index = get_global_id(2); - - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - float4 values0 = 0.0f; - float4 values1 = 0.0f; - - __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w); - __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0); - - // Note: Since each work-item computes 4x2 elements, we need to load 6 rows from the input tensor - - for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d) - { - // Load the weights from row0 and row1 - float4 weights_row00 = vload4(0, (__global float *)(weights_addr + 0 * weights_stride_y)); - float weights_row01 = *((__global float *)(weights_addr + 0 * weights_stride_y) + 4); - float4 weights_row10 = vload4(0, (__global float *)(weights_addr + 1 * weights_stride_y)); - float weights_row11 = *((__global float *)(weights_addr + 1 * weights_stride_y) + 4); - float8 src0; - - // Load values from row0 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 0 * src_stride_y)); - - // Accumulate - CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01); - - // Load values from row1 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 1 * src_stride_y)); - - // Accumulate - CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11); - CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01); - - // Load values from row2 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 2 * src_stride_y)); - - // Load weights from row2 - weights_row00 = vload4(0, (__global float *)(weights_addr + 2 * weights_stride_y)); - weights_row01 = *((__global float *)(weights_addr + 2 * weights_stride_y) + 4); - - // Accumulate - CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01); - CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11); - - // Load values from row3 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 3 * src_stride_y)); - - // Load weights from row3 - weights_row10 = vload4(0, (__global float *)(weights_addr + 3 * weights_stride_y)); - weights_row11 = *((__global float *)(weights_addr + 3 * weights_stride_y) + 4); - - // Accumulate - CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11); - CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01); - - // Load values from row4 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 4 * src_stride_y)); - - // Load weights from row4 - weights_row00 = vload4(0, (__global float *)(weights_addr + 4 * weights_stride_y)); - weights_row01 = *((__global float *)(weights_addr + 4 * weights_stride_y) + 4); - - CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01); - CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11); - - // Load values from row5 of input tensor - src0 = vload8(0, (__global float *)(src_addr + 5 * src_stride_y)); - - // Accumulate - CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01); - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - - float4 bias = (float4) * ((__global float *)(vector_offset(&biases, kernel_index))); - - values0 += bias; - values1 += bias; -#endif /* defined(HAS_BIAS) */ - - vstore4(values0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y)); - vstore4(values1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y)); -} -#endif // defined(WEIGHTS_DEPTH) diff --git a/src/core/CL/cl_kernels/nchw/direct_convolution_quantized.cl b/src/core/CL/cl_kernels/nchw/direct_convolution_quantized.cl deleted file mode 100644 index b80d4f587..000000000 --- a/src/core/CL/cl_kernels/nchw/direct_convolution_quantized.cl +++ /dev/null @@ -1,308 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers_asymm.h" - -#undef CONVERT_SAT_STR -#undef CONVERT_SAT - -#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) - -#define CONVERT_SAT_STR(x, type) (convert_##type##8_sat((x))) -#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type) - -#if KERNEL_SIZE == 9 - -#if STRIDE_X == 1 -#define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr) -#elif STRIDE_X == 2 -#define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X */ - -#define CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \ - int weights_value1 = convert_int(*(weights_row_ptr + 8)); \ - int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ - acc += (src0.lo + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1234, src0.s5678) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s2345, src0.s6789) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s3456, src0.s789A) + INPUT_OFFSET) * ((int8)weights_values0.s3 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s4567, src0.s89AB) + INPUT_OFFSET) * ((int8)weights_values0.s4 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s5678, src0.s9ABC) + INPUT_OFFSET) * ((int8)weights_values0.s5 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s6789, src0.sABCD) + INPUT_OFFSET) * ((int8)weights_values0.s6 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s789A, src0.sBCDE) + INPUT_OFFSET) * ((int8)weights_values0.s7 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s89AB, src0.sCDEF) + INPUT_OFFSET) * ((int8)weights_value1 + WEIGHTS_OFFSET); \ - }) - -#define CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \ - int weights_value1 = convert_int(*(weights_row_ptr + 8)); \ - int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ - int8 src1 = convert_int8(vload8(0, src_row_ptr + 16)); \ - acc += (src0.even + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1357, src0.s9BDF) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + INPUT_OFFSET) * ((int8)weights_values0.s3 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s468A, src0.sCE, src1.s02) + INPUT_OFFSET) * ((int8)weights_values0.s4 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s579B, src0.sDF, src1.s13) + INPUT_OFFSET) * ((int8)weights_values0.s5 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s68AC, src0.sE, src1.s024) + INPUT_OFFSET) * ((int8)weights_values0.s6 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s79BD, src0.sF, src1.s135) + INPUT_OFFSET) * ((int8)weights_values0.s7 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s8ACE, src1.s0246) + INPUT_OFFSET) * ((int8)weights_value1 + WEIGHTS_OFFSET); \ - }) - -#elif KERNEL_SIZE == 5 - -#if STRIDE_X == 1 -#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) -#elif STRIDE_X == 2 -#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X */ - -#define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \ - int weights_value1 = convert_int(*(weights_row_ptr + 4)); \ - int8 src0 = convert_int8(vload8(0, src_row_ptr)); \ - int4 src1 = convert_int4(vload4(0, src_row_ptr + 8)); \ - acc += (src0 + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1234, src0.s567, src1.s0) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s234, src0.s567, src1.s01) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s345, src0.s67, src1.s012) + INPUT_OFFSET) * ((int8)weights_values0.s3 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s45, src0.s67, src1.s0123) + INPUT_OFFSET) * ((int8)weights_value1 + WEIGHTS_OFFSET); \ - }) - -#define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \ - int weights_value1 = convert_int(*(weights_row_ptr + 4)); \ - int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ - int4 src1 = convert_int4(vload4(0, src_row_ptr + 16)); \ - acc += (src0.even + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1357, src0.s9BDF) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + INPUT_OFFSET) * ((int8)weights_values0.s3 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + INPUT_OFFSET) * ((int8)weights_value1 + WEIGHTS_OFFSET); \ - }) - -#elif KERNEL_SIZE == 3 - -#if STRIDE_X == 1 -#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) -#elif STRIDE_X == 2 -#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) -#else /* STRIDE_X not equals 1 or 2 */ -#error "STRIDE_X larger than 2 is not supported" -#endif /* STRIDE_X */ - -#define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \ - int8 src0 = convert_int8(vload8(0, src_row_ptr)); \ - int2 src1 = convert_int2(vload2(0, src_row_ptr + 8)); \ - acc += (src0 + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1234, src0.s567, src1.s0) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s234, src0.s567, src1.s01) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - }) - -#define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ - ({ \ - int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \ - int16 src0 = convert_int16(vload16(0, src_row_ptr)); \ - int src1 = convert_int(*(src_row_ptr + 16)); \ - acc += (src0.even + INPUT_OFFSET) * ((int8)weights_values0.s0 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s1357, src0.s9BDF) + INPUT_OFFSET) * ((int8)weights_values0.s1 + WEIGHTS_OFFSET); \ - acc += ((int8)(src0.s2468, src0.sACE, src1) + INPUT_OFFSET) * ((int8)weights_values0.s2 + WEIGHTS_OFFSET); \ - }) - -#elif KERNEL_SIZE == 1 - -#if STRIDE_X == 3 -#define INPUT_VALUE extract_input_stride3 -#elif STRIDE_X == 2 -#define INPUT_VALUE extract_input_stride2 -#elif STRIDE_X == 1 -#define INPUT_VALUE extract_input_stride1 - -#else /* STRIDE_X not equals 1, 2 or 3 */ -#error "Only support strides 1, 2 and 3" -#endif /* STRIDE_X */ - -/** Extracts a 1D horizontal vector from the input tensor with stride as 1. - * - * @param[in] input_value Pointer to the first value. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_value) -{ - return vload8(0, input_value); -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 2. - * - * @param[in] input_value Pointer to the first value. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_value) -{ - VEC_DATA_TYPE(DATA_TYPE, 16) - temp = vload16(0, input_value); - return temp.s02468ace; -} - -/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. - * - * @param[in] input_value Pointer to the first value. - * - * @return extracted input values. - */ -inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3(__global const DATA_TYPE *input_value) -{ - VEC_DATA_TYPE(DATA_TYPE, 16) - temp1 = vload16(0, input_value); - VEC_DATA_TYPE(DATA_TYPE, 16) - temp2 = vload16(0, input_value + 12); - return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369); -} - -#else /* KERNEL_SIZE not equals 1, 3 , 5, 9 */ -#error "Only kernel sizes 1, 3, 5 and 9 are supported" -#endif /* KERNEL_SIZE */ - -/** This kernel performs a direct convolution to convolve the low three dimensions. - * - * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 - * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH - * @note If biases are used then -DHAS_BIAS has to be passed at compile time - * @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234 - * @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4 - * @note The input offset quantization parameter must be passed at compile time using -DINPUT_OFFSET e.g. -DINPUT_OFFSET=3 - * @note The weights offset quantization parameter must be passed at compile time using -DWEIGHTS_OFFSET e.g. -DWEIGHTS_OFFSET=3 - * @note The destination offset quantization parameter must be passed at compile time using -DOUTPUT_OFFSET e.g. -DOUTPUT_OFFSET=3 - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr - * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) - * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) - * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) - * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) - * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor - * @param[in] biases_ptr Pointer to the biases tensor. Supported data types: S32 - * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) - * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor - * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension - */ -__kernel void direct_convolution_quantized( - TENSOR3D_DECLARATION(src), - TENSOR3D_DECLARATION(dst), - TENSOR3D_DECLARATION(weights), -#ifdef HAS_BIAS - VECTOR_DECLARATION(biases), -#endif /* defined(HAS_BIAS) */ - unsigned int weights_stride_w) -{ - Image src = CONVERT_TO_IMAGE_STRUCT(src); - Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); - - int8 values0 = 0; - - __global DATA_TYPE *weights_addr = (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0); - __global DATA_TYPE *src_addr = (__global DATA_TYPE *)offset(&src, 0, 0); - - const int kernel_index = get_global_id(2); - weights_addr += kernel_index * weights_stride_w; - - for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d) - { -#if KERNEL_SIZE == 9 - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 5 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 6 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 6 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 7 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 7 * weights_stride_y)); - CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 8 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 8 * weights_stride_y)); -#elif KERNEL_SIZE == 5 - CONVOLUTION1x5(values0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y)); - CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y)); -#elif KERNEL_SIZE == 3 - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y)); - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y)); - CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y)); -#elif KERNEL_SIZE == 1 - int weight = convert_int(*(__global DATA_TYPE *)weights_addr); - int8 input_value = convert_int8(INPUT_VALUE((__global DATA_TYPE *)src_addr)); - values0 += (input_value + INPUT_OFFSET) * ((int8)weight + WEIGHTS_OFFSET); -#endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */ - - src_addr += src_stride_z; - weights_addr += weights_stride_z; - } - -#ifdef HAS_BIAS - Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); - __global int *bias_addr = ((__global int *)(vector_offset(&biases, kernel_index))); - values0 += (int8)(*bias_addr); -#endif /* defined(HAS_BIAS) */ - -#if OUTPUT_SHIFT < 0 - values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); -#else // OUTPUT_SHIFT < 0 - values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); -#endif // OUTPUT_SHIFT < 0 - values0 = values0 + OUTPUT_OFFSET; - - vstore8(CONVERT_SAT(values0, DATA_TYPE), 0, (__global DATA_TYPE *)dst.ptr); -} -#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) diff --git a/src/core/CL/cl_kernels/nchw/remap.cl b/src/core/CL/cl_kernels/nchw/remap.cl deleted file mode 100644 index fab88a168..000000000 --- a/src/core/CL/cl_kernels/nchw/remap.cl +++ /dev/null @@ -1,133 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" -#include "warp_helpers.h" - -#ifndef DEPTH_OUT -/** Performs a remapping of an input image to an output given two remapping image using nearest neighbor as interpolation. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - */ -__kernel void remap_nearest_neighbour_nchw( - IMAGE_DECLARATION(in), - IMAGE_DECLARATION(out), - IMAGE_DECLARATION(mapx), - IMAGE_DECLARATION(mapy), - const float width, - const float height) -{ - Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); - Image out = CONVERT_TO_IMAGE_STRUCT(out); - Image mapx = CONVERT_TO_IMAGE_STRUCT(mapx); - Image mapy = CONVERT_TO_IMAGE_STRUCT(mapy); - - float4 mapx_coords = vload4(0, (__global float *)mapx.ptr); - float4 mapy_coords = vload4(0, (__global float *)mapy.ptr); - float8 map_coords = (float8)(mapx_coords.s0, mapy_coords.s0, mapx_coords.s1, mapy_coords.s1, - mapx_coords.s2, mapy_coords.s2, mapx_coords.s3, mapy_coords.s3); - - vstore4(read_texels4(&in, convert_int8(clamp_to_border(map_coords, width, height))), 0, out.ptr); -} - -/** Performs a remapping of an input image to an output given two remapping image using bilinear as interpolation. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - */ -__kernel void remap_bilinear_nchw( - IMAGE_DECLARATION(in), - IMAGE_DECLARATION(out), - IMAGE_DECLARATION(mapx), - IMAGE_DECLARATION(mapy), - const float width, - const float height) -{ - Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); - Image out = CONVERT_TO_IMAGE_STRUCT(out); - Image mapx = CONVERT_TO_IMAGE_STRUCT(mapx); - Image mapy = CONVERT_TO_IMAGE_STRUCT(mapy); - - float4 mapx_coords = vload4(0, (__global float *)mapx.ptr); - float4 mapy_coords = vload4(0, (__global float *)mapy.ptr); - float8 map_coords = (float8)(mapx_coords.s0, mapy_coords.s0, mapx_coords.s1, mapy_coords.s1, - mapx_coords.s2, mapy_coords.s2, mapx_coords.s3, mapy_coords.s3); - - vstore4(bilinear_interpolate(&in, clamp_to_border(map_coords, width, height), width, height), 0, out.ptr); -} -#endif // DEPTH_OUT
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/direct_convolution.cl b/src/core/CL/cl_kernels/nhwc/direct_convolution.cl index 75a7a0f00..f1b422a68 100644 --- a/src/core/CL/cl_kernels/nhwc/direct_convolution.cl +++ b/src/core/CL/cl_kernels/nhwc/direct_convolution.cl @@ -103,9 +103,9 @@ */ //! @endcond __kernel void direct_convolution_nhwc( - TENSOR4D(src, SRC_TENSOR_TYPE), - TENSOR4D(dst, DST_TENSOR_TYPE), - TENSOR4D(wei, WEI_TENSOR_TYPE) + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), + TENSOR4D_T(wei, WEI_TENSOR_TYPE) #if defined(HAS_BIAS) , VECTOR_DECLARATION(bia) @@ -116,12 +116,12 @@ __kernel void direct_convolution_nhwc( // In case of dynamic tensor support, the following dimensions should be passed as function argument. #define _IWEI_WIDTH WEI_WIDTH #define _IWEI_HEIGHT WEI_HEIGHT -#define _ISRC_WIDTH SRC_WIDTH -#define _ISRC_HEIGHT SRC_HEIGHT -#define _ISRC_CHANNELS SRC_CHANNELS -#define _IDST_WIDTH DST_WIDTH -#define _IDST_HEIGHT DST_HEIGHT -#define _IDST_CHANNELS DST_CHANNELS +#define _ISRC_WIDTH src_w +#define _ISRC_HEIGHT src_h +#define _ISRC_CHANNELS src_c +#define _IDST_WIDTH dst_w +#define _IDST_HEIGHT dst_h +#define _IDST_CHANNELS dst_c #define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT) // If quantized, the output tile has to be quantized first before being stored to global memory @@ -192,7 +192,7 @@ __kernel void direct_convolution_nhwc( // We voluntarily use SRC_CHANNELS rather than _DSRC_CHANNELS // This #if directive should be removed in case of dynamic tensor support -#if((SRC_CHANNELS % K0) != 0) +#if defined(LEFTOVER_LOOP) // Left-over accumulations for(; k < _ISRC_CHANNELS; ++k) { @@ -220,7 +220,7 @@ __kernel void direct_convolution_nhwc( ++ck; } -#endif // ((SRC_CHANNELS % K0) != 0) +#endif // defined(LEFTOVER_LOOP) } // Offset correction required for the quantized asymmetric computation diff --git a/src/core/CL/cl_kernels/nhwc/dwc_native_fp_nhwc.cl b/src/core/CL/cl_kernels/nhwc/dwc_native_fp_nhwc.cl index 58f01fa3e..4f57a81e7 100644 --- a/src/core/CL/cl_kernels/nhwc/dwc_native_fp_nhwc.cl +++ b/src/core/CL/cl_kernels/nhwc/dwc_native_fp_nhwc.cl @@ -26,7 +26,7 @@ #include "helpers.h" #include "tile_helpers.h" -#if defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_WIDTH) && defined(DST_HEIGHT) && defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) +#if defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) //! @cond Doxygen_Suppress /** OpenCL kernel to compute the depthwise convolution for floating-point data types (F32/F16) * @@ -37,10 +37,6 @@ * @note The convolution strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y (e.g. -DSTRIDE_X=2, -DSTRIDE_Y=2) * @note The convolution dilations must be passed at compile time using -DDILATION_X and -DDILATION_Y (e.g. -DDILATION_X=2, -DDILATION_Y=2) * @note The spatial dimensions of the weights must be passed at compile time using -DWEI_WIDTH and -DWEI_HEIGHT (e.g. -DWEI_WIDTH=9, -DWEI_HEIGHT=9) - * @note The spatial dimensions of the source tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT (e.g. -DSRC_WIDTH=96, -DSRC_HEIGHT=64) - * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64) - * @note The channels of the source tensor must be passed at compile time using -DSRC_CHANNELS (e.g. -DSRC_CHANNELS=64) - * @note The channels of the destination tensor must be passed at compile time using -DDST_CHANNELS (e.g. -DDDST_CHANNELS=64) * @note The tensor type ("BUFFER" or "IMAGE") of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" or "IMAGE") of the weights tensor must be passed at compile time using -DWEI_TENSOR_TYPE (e.g. -DWEI_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" or "IMAGE") of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER) @@ -57,24 +53,22 @@ * @note The number of rows to read from the src tensor must be passed at compile time using -DM0_A (e.g., -DM0_A=3). M0_A must be equal to WEI_WIDTH + (M0 - 1) * * @param[in] src_ptr Pointer to the source tensor. Supported data type: F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_c The size of the channels dimension of the source tensor + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the batches dimension of the source tensor * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data type: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] dst_c The size of the channels dimension of the destination tensor + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the batches dimension of the destination tensor * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[in] wei_ptr Pointer to the weights tensor. Supported data type: same as @p src_ptr * @param[in] wei_stride_x Stride of the weights tensor in X dimension (in bytes) @@ -93,8 +87,8 @@ */ //! @endcond __kernel void dwc_native_fp_nhwc( - TENSOR4D(src, SRC_TENSOR_TYPE), - TENSOR4D(dst, DST_TENSOR_TYPE), + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), TENSOR4D(wei, WEI_TENSOR_TYPE) #if defined(HAS_BIAS) , @@ -102,15 +96,10 @@ __kernel void dwc_native_fp_nhwc( #endif // defined(HAS_BIAS) ) { - // All the tensor dimensions are passed at compile time. + // Only the weight tensor dimensions are passed at compile time. // In case of dynamic tensor support, the following dimensions should be passed as function argument. #define _IWEI_WIDTH WEI_WIDTH #define _IWEI_HEIGHT WEI_HEIGHT -#define _ISRC_WIDTH SRC_WIDTH -#define _ISRC_HEIGHT SRC_HEIGHT -#define _IDST_WIDTH DST_WIDTH -#define _IDST_HEIGHT DST_HEIGHT -#define _IDST_CHANNELS DST_CHANNELS #define _IM0_A M0_A // _IWEI_WIDTH + (M0 - 1) Rows tile A (If M0 != 1, the tiles overlap of 1 element on the X dimension) #define _IN0_A N0 // Cols tile A #define _IM0_B _IWEI_WIDTH // Rows tile B @@ -120,12 +109,12 @@ __kernel void dwc_native_fp_nhwc( const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM const int xo = GET_SPATIAL_IDX(1, M0, 0); // WIDTH #if defined(BATCHED_EXECUTION) - const int yo = GET_SPATIAL_IDX(2, 1, 0) % _IDST_HEIGHT; // HEIGHT - const int bout = GET_SPATIAL_IDX(2, 1, 0) / _IDST_HEIGHT; // BATCH SIZE IDX -#else // defined(BATCHED_EXECUTION) + const int yo = GET_SPATIAL_IDX(2, 1, 0) % dst_h; // HEIGHT + const int bout = GET_SPATIAL_IDX(2, 1, 0) / dst_h; // BATCH SIZE IDX +#else // defined(BATCHED_EXECUTION) const int yo = GET_SPATIAL_IDX(2, 1, 0); // HEIGHT const int bout = 0; // BATCH SIZE IDX -#endif // defined(BATCHED_EXECUTION) +#endif // defined(BATCHED_EXECUTION) int xi = xo * STRIDE_X; int yi = yo * STRIDE_Y; @@ -145,7 +134,7 @@ __kernel void dwc_native_fp_nhwc( c[i].v = 0; }) -#if _IWEI_HEIGHT <= 5 +#if _IWEI_HEIGHT < 5 LOOP_UNROLLING(int, yk, 0, 1, _IWEI_HEIGHT, #else // _IWEI_HEIGHT <= 5 for(int yk = 0; yk < _IWEI_HEIGHT; yk++) @@ -159,7 +148,7 @@ __kernel void dwc_native_fp_nhwc( }) // Load tile from the src tensor (TILE A) - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, _IM0_A, _IN0_A, SRC_TENSOR_TYPE, src, bout, yi + yk * DILATION_Y, xi, cout, _ISRC_WIDTH, _ISRC_HEIGHT, DILATION_X, 1, _IBOUNDARY_CHECK, a); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, _IM0_A, _IN0_A, SRC_TENSOR_TYPE, src, bout, yi + yk * DILATION_Y, xi, cout, src_w, src_h, DILATION_X, 1, _IBOUNDARY_CHECK, a); TILE(WEI_DATA_TYPE, _IM0_B, _IN0_B, b); @@ -176,7 +165,7 @@ __kernel void dwc_native_fp_nhwc( }) }) } -#if _IWEI_HEIGHT <= 5 +#if _IWEI_HEIGHT < 5 ) #endif // _IWEI_HEIGHT <= 5 @@ -199,7 +188,7 @@ __kernel void dwc_native_fp_nhwc( { LOOP_UNROLLING(int, m0, 0, 1, M0, { - int xi_out = min(xo + M0 - 1 - m0, (int)(_IDST_WIDTH) - 1); + int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1); VSTORE_PARTIAL(N0, PARTIAL_N0) (c[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)((cout * DEPTH_MULTIPLIER) + d) * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w)); }) @@ -208,11 +197,11 @@ __kernel void dwc_native_fp_nhwc( { LOOP_UNROLLING(int, m0, 0, 1, M0, { - int xi_out = min(xo + M0 - 1 - m0, (int)(_IDST_WIDTH) - 1); + int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1); VSTORE(N0) (c[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)((cout * DEPTH_MULTIPLIER) + d) * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w)); }) } } } -#endif // defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_WIDTH) && defined(DST_HEIGHT) && defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
\ No newline at end of file +#endif // defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/dwc_native_quantized_nhwc.cl b/src/core/CL/cl_kernels/nhwc/dwc_native_quantized_nhwc.cl index 1bc58b6e2..ec2593af7 100644 --- a/src/core/CL/cl_kernels/nhwc/dwc_native_quantized_nhwc.cl +++ b/src/core/CL/cl_kernels/nhwc/dwc_native_quantized_nhwc.cl @@ -44,7 +44,7 @@ #define T_LOAD_MULTIPLIERS_SHIFT(QUANTIZATION_TYPE) T_LOAD_MULTIPLIERS_SHIFT_STR(QUANTIZATION_TYPE) #define T_LOAD_MULTIPLIERS_SHIFT_STR(QUANTIZATION_TYPE) T_LOAD_MULTIPLIERS_SHIFT_##QUANTIZATION_TYPE() -#if defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_WIDTH) && defined(DST_HEIGHT) && defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) +#if defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) //! @cond Doxygen_Suppress /** OpenCL kernel to compute the depthwise convolution for quantized data types * @@ -54,10 +54,6 @@ * @note The convolution strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y (e.g. -DSTRIDE_X=2, -DSTRIDE_Y=2) * @note The convolution dilations must be passed at compile time using -DDILATION_X and -DDILATION_Y (e.g. -DDILATION_X=2, -DDILATION_Y=2) * @note The spatial dimensions of the weights must be passed at compile time using -DWEI_WIDTH and -DWEI_HEIGHT (e.g. -DWEI_WIDTH=9, -DWEI_HEIGHT=9) - * @note The spatial dimensions of the source tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT (e.g. -DSRC_WIDTH=96, -DSRC_HEIGHT=64) - * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64) - * @note The channels of the source tensor must be passed at compile time using -DSRC_CHANNELS (e.g. -DSRC_CHANNELS=64) - * @note The channels of the destination tensor must be passed at compile time using -DDST_CHANNELS (e.g. -DDDST_CHANNELS=64) * @note The tensor type ("BUFFER" or "IMAGE") of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" or "IMAGE") of the weights tensor must be passed at compile time using -DWEI_TENSOR_TYPE (e.g. -DWEI_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" or "IMAGE") of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER) @@ -79,24 +75,22 @@ * @note The number of rows to read from the src tensor must be passed at compile time using -DM0_A (e.g., -DM0_A=3). M0_A must be equal to WEI_WIDTH + (M0 - 1) * * @param[in] src_ptr Pointer to the source tensor. Supported data type: QSYMM8/QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] src_c The size of the channels dimension of the source tensor + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the batches dimension of the source tensor * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data type: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] dst_c The size of the channels dimension of the destination tensor + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the batches dimension of the destination tensor * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[in] wei_ptr Pointer to the weights tensor. Supported data type: same as @p src_ptr * @param[in] wei_stride_x Stride of the weights tensor in X dimension (in bytes) @@ -123,8 +117,8 @@ */ //! @endcond __kernel void dwc_native_quantized_nhwc( - TENSOR4D(src, SRC_TENSOR_TYPE), - TENSOR4D(dst, DST_TENSOR_TYPE), + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), TENSOR4D(wei, WEI_TENSOR_TYPE), VECTOR_DECLARATION(dst_multipliers), VECTOR_DECLARATION(dst_shifts) @@ -134,15 +128,10 @@ __kernel void dwc_native_quantized_nhwc( #endif // defined(HAS_BIAS) ) { - // All the tensor dimensions are passed at compile time. + // Only the weight tensor dimensions are passed at compile time. // In case of dynamic tensor support, the following dimensions should be passed as function argument. #define _IWEI_WIDTH WEI_WIDTH #define _IWEI_HEIGHT WEI_HEIGHT -#define _ISRC_WIDTH SRC_WIDTH -#define _ISRC_HEIGHT SRC_HEIGHT -#define _IDST_WIDTH DST_WIDTH -#define _IDST_HEIGHT DST_HEIGHT -#define _IDST_CHANNELS DST_CHANNELS #define _IM0_A M0_A // _IWEI_WIDTH + (M0 - 1) Rows tile A (If M0 != 1, the tiles overlap of 1 element on the X dimension) #define _IN0_A N0 // Cols tile A #define _IM0_B _IWEI_WIDTH // Rows tile B @@ -152,12 +141,12 @@ __kernel void dwc_native_quantized_nhwc( const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM const int xo = GET_SPATIAL_IDX(1, M0, 0); // WIDTH #if defined(BATCHED_EXECUTION) - const int yo = GET_SPATIAL_IDX(2, 1, 0) % _IDST_HEIGHT; // HEIGHT - const int bout = GET_SPATIAL_IDX(2, 1, 0) / _IDST_HEIGHT; // BATCH SIZE IDX -#else // defined(BATCHED_EXECUTION) + const int yo = GET_SPATIAL_IDX(2, 1, 0) % dst_h; // HEIGHT + const int bout = GET_SPATIAL_IDX(2, 1, 0) / dst_h; // BATCH SIZE IDX +#else // defined(BATCHED_EXECUTION) const int yo = GET_SPATIAL_IDX(2, 1, 0); // HEIGHT const int bout = 0; // BATCH SIZE IDX -#endif // defined(BATCHED_EXECUTION) +#endif // defined(BATCHED_EXECUTION) int xi = xo * STRIDE_X; int yi = yo * STRIDE_Y; @@ -191,7 +180,7 @@ __kernel void dwc_native_quantized_nhwc( }) // Load tile from the src tensor (TILE A) - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, _IM0_A, _IN0_A, SRC_TENSOR_TYPE, src, bout, yi + yk * DILATION_Y, xi, cout, _ISRC_WIDTH, _ISRC_HEIGHT, DILATION_X, 1, _IBOUNDARY_CHECK, a); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, _IM0_A, _IN0_A, SRC_TENSOR_TYPE, src, bout, yi + yk * DILATION_Y, xi, cout, src_w, src_h, DILATION_X, 1, _IBOUNDARY_CHECK, a); TILE(WEI_DATA_TYPE, _IM0_B, _IN0_B, b); @@ -265,7 +254,7 @@ __kernel void dwc_native_quantized_nhwc( { LOOP_UNROLLING(int, m0, 0, 1, M0, { - int xi_out = min(xo + M0 - 1 - m0, (int)(_IDST_WIDTH) - 1); + int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1); VSTORE_PARTIAL(N0, PARTIAL_N0) (cq[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)((cout * DEPTH_MULTIPLIER) + d) * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w)); }) @@ -274,11 +263,11 @@ __kernel void dwc_native_quantized_nhwc( { LOOP_UNROLLING(int, m0, 0, 1, M0, { - int xi_out = min(xo + M0 - 1 - m0, (int)(_IDST_WIDTH) - 1); + int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1); VSTORE(N0) (cq[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)((cout * DEPTH_MULTIPLIER) + d) * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w)); }) } } } -#endif // defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_WIDTH) && defined(DST_HEIGHT) && defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
\ No newline at end of file +#endif // defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/im2col.cl b/src/core/CL/cl_kernels/nhwc/im2col.cl index ac00c1128..a23e943fa 100644 --- a/src/core/CL/cl_kernels/nhwc/im2col.cl +++ b/src/core/CL/cl_kernels/nhwc/im2col.cl @@ -22,23 +22,11 @@ * SOFTWARE. */ #include "helpers.h" -#if defined(DATA_TYPE) && defined(ELEMENT_SIZE) - -#if ELEMENT_SIZE == 1 -#define COND_DATA_TYPE char -#elif ELEMENT_SIZE == 2 -#define COND_DATA_TYPE short -#elif ELEMENT_SIZE == 4 -#define COND_DATA_TYPE int -#else // ELEMENT_SIZE -#error "Element size not support" -#endif // ELEMENT_SIZE - -#if defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) && defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) #define VECTOR_N VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE) -#define COND_N VEC_DATA_TYPE(COND_DATA_TYPE, VECTOR_SIZE) +#define COND_N SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE) +#if defined(IM2COL_3X3) || defined(IM2COL_9X9) /** Store a 1x9 row or a 3x3 block in a boundary-aware manner to avoid paddings in the channel dimension * @name IM2COL1X9_NHWC_STORE * @@ -109,7 +97,9 @@ VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \ (DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH); /** @}*/ +#endif // defined(IM2COL_3X3) || defined(IM2COL_9X9) +#if defined(IM2COL_3X3) /** This kernel performs im2col when the kernel size is 3x3 and the data layout is NHWC * * @note This kernel computes VECTOR_SIZE elements @@ -269,7 +259,9 @@ __kernel void im2col3x3_nhwc( } #endif // HAS_BIAS } +#endif // defined(IM2COL_3X3) +#if defined(IM2COL_9X9) #if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0 #define IM2COL1x9(i) \ ({ \ @@ -416,7 +408,9 @@ __kernel void im2col9x9_nhwc( } #endif // HAS_BIAS } +#endif // defined(IM2COL_9X9) +#if defined(IM2COL_GENERIC) /** This opencl kernel performs a generic im2col implementation when the data layout is NHWC * * @note This kernel computes VECTOR_SIZE elements @@ -463,19 +457,20 @@ __kernel void im2col_generic_nhwc( const int batch = get_global_id(2); // batch size // Calculate input indices - const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X; - const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y; + const int xi = (yo % CONVOLVED_WIDTH) * STRIDE_X; + const int yi = (yo / (int)CONVOLVED_WIDTH) * STRIDE_Y; // Get input and output address - __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w; - __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w; + const int stride_x = ch * sizeof(DATA_TYPE); + __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + stride_x + batch * (int)src_stride_w; + __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + stride_x + yo * (int)dst_stride_y + batch * (int)dst_stride_w; int i = 0; for(int yk = 0; yk < KERNEL_HEIGHT; ++yk) { // Clamp yi_coord int yi_coord = yi + yk * DILATION_Y - (int)PAD_TOP; - yi_coord = CLAMP(yi_coord, (int)0, (int)(SRC_HEIGHT - 1)); + yi_coord = clamp(yi_coord, (int)0, (int)(SRC_HEIGHT - 1)); // Out-of-bound condition for Y int y_border_condition = ((yi + yk * DILATION_Y - (int)PAD_TOP) < (int)0) || ((yi + yk * DILATION_Y - (int)PAD_TOP) >= (int)SRC_HEIGHT); @@ -484,7 +479,7 @@ __kernel void im2col_generic_nhwc( { // Clamp xi_coord int xi_coord = (xi + xk * DILATION_X - (int)PAD_LEFT); - xi_coord = CLAMP(xi_coord, (int)0, (int)(SRC_WIDTH - 1)); + xi_coord = clamp(xi_coord, (int)0, (int)(SRC_WIDTH - 1)); // Out-of-bound condition for X int x_border_condition = ((xi + xk * DILATION_X - (int)PAD_LEFT) < (int)0) || ((xi + xk * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH); @@ -528,5 +523,4 @@ __kernel void im2col_generic_nhwc( } #endif // HAS_BIAS } -#endif // defined(CONVOLVED_WIDTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_DEPTH) && defined(PAD_LEFT) && defined(PAD_RIGHT) && defined(PAD_TOP) && defined(PAD_BOTTOM) && defined(PAD_VALUE) && defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) -#endif // defined(DATA_TYPE) && defined(ELEMENT_SIZE)
\ No newline at end of file +#endif // defined(IM2COL_GENERIC)
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/remap.cl b/src/core/CL/cl_kernels/nhwc/remap.cl deleted file mode 100644 index 0b629fe6c..000000000 --- a/src/core/CL/cl_kernels/nhwc/remap.cl +++ /dev/null @@ -1,180 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "helpers.h" -#include "warp_helpers.h" - -#ifdef DEPTH_OUT -/** Performs a remapping of an input image to an output given two remapping image using nearest neighbor as interpolation. - * Also applies constant border value, "border_val", if "CONSTANT_BORDER" is set. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8,F16. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8,F16. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - * @param[in] border_val Value to use for border around input tensor when in CONSTANT border is selected - */ -__kernel void remap_nearest_neighbour_nhwc( - TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - TENSOR4D_DECLARATION(mapx), - TENSOR4D_DECLARATION(mapy), - const float width, - const float height -#ifdef CONSTANT_BORDER - , - const DATA_TYPE border_val -#endif // CONSTANT_BORDER -) -{ - Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); - Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); - Tensor4D mapx = CONVERT_TO_TENSOR4D_STRUCT(mapx, DEPTH_OUT); - Tensor4D mapy = CONVERT_TO_TENSOR4D_STRUCT(mapy, DEPTH_OUT); - - float mapx_coord = (float) * (__global float *)mapx.ptr; - float mapy_coord = (float) * (__global float *)mapy.ptr; - -#ifdef CONSTANT_BORDER - if(mapx_coord < 0 || mapx_coord > width - 1 || mapy_coord < 0 || mapy_coord > height - 1) - { - *((__global DATA_TYPE *)out.ptr) = border_val; - return; - } -#else // CONSTANT_BORDER - mapx_coord = clamp(mapx_coord, 0.0f, width - 1); - mapy_coord = clamp(mapy_coord, 0.0f, height - 1); -#endif // CONSTANT_BORDER - *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(mapx_coord), convert_int(mapy_coord), (get_global_id(2) / DEPTH_OUT))); -} - -/** Performs a remapping of an input image to an output given two remapping image using bilinear as interpolation. - * Also applies constant border value, "border_val", if "CONSTANT_BORDER" is set. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8,F16. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8,F16. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - * @param[in] border_val Value to use for border around input tensor when in CONSTANT border is selected - */ -__kernel void remap_bilinear_nhwc( - TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - TENSOR4D_DECLARATION(mapx), - TENSOR4D_DECLARATION(mapy), - const float width, - const float height -#ifdef CONSTANT_BORDER - , - const DATA_TYPE border_val -#endif // CONSTANT_BORDER -) -{ - Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); - Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); - Tensor4D mapx = CONVERT_TO_TENSOR4D_STRUCT(mapx, DEPTH_OUT); - Tensor4D mapy = CONVERT_TO_TENSOR4D_STRUCT(mapy, DEPTH_OUT); - - float mapx_coord = (float) * (__global float *)mapx.ptr; - float mapy_coord = (float) * (__global float *)mapy.ptr; - -#ifdef CONSTANT_BORDER - if(mapx_coord < 0 || mapx_coord > width - 1 || mapy_coord < 0 || mapy_coord > height - 1) - { - *((__global DATA_TYPE *)out.ptr) = border_val; - return; - } -#endif // CONSTANT_BORDER - - const float new_xf = floor(mapx_coord); - const float new_yf = floor(mapy_coord); - const float clamped_x = clamp(new_xf, 0.0f, width - 1); - const float clamped_x1 = clamp(new_xf + 1, 0.0f, width - 1); - const float clamped_y = clamp(new_yf, 0.0f, height - 1); - const float clamped_y1 = clamp(new_yf + 1, 0.0f, height - 1); - - float4 ins = (float4)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); - - const float a = mapx_coord - new_xf; - const float b = 1.f - a; - const float a1 = mapy_coord - new_yf; - const float b1 = 1.f - a1; - const float fr = ((ins.s0 * b * b1) + (ins.s1 * a * b1) + (ins.s2 * b * a1) + (ins.s3 * a * a1)); - - *((__global DATA_TYPE *)out.ptr) = CONVERT(fr, DATA_TYPE); -} - -#endif // DEPTH_OUT
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/scale.cl b/src/core/CL/cl_kernels/nhwc/scale.cl index 21579aed9..bccfd6543 100644 --- a/src/core/CL/cl_kernels/nhwc/scale.cl +++ b/src/core/CL/cl_kernels/nhwc/scale.cl @@ -24,12 +24,11 @@ #include "helpers.h" #include "tile_helpers.h" +#if defined(SCALE_NEAREST_NEIGHBOUR) //! @cond Doxygen_Suppress /** Performs scale on a tensor by interpolating with the NEAREAST NEIGHBOUR method. (NHWC) * * @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT - * @note The spatial dimensions of the source tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT (e.g. -DSRC_WIDTH=96, -DSRC_HEIGHT=64) - * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64) * @note The tensor type ("BUFFER" only is supported) of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" only is supported) of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER) * @note The data type of the source tensor must be passed at compile time using -DSRC_DATA_TYPE (e.g. -DSRC_DATA_TYPE=float) @@ -37,61 +36,52 @@ * @note The number of N0 output channels to process must be passed at compile time using -DN0 (e.g. -DN0=2) * @note The border value value must be passed at compile time using -DCONSTANT_VALUE (e.g. -DCONSTANT_VALUE=0) * @note In case of F32/F16, -DIS_FLOATING_POINT must be passed at compile time - * @note The scale value to apply on the source width must be passed at compile using -DSCALE_X (e.g., -DSCALE_X=0.5) - * @note The scale value to apply on the source height must be passed at compile using -DSCALE_Y (e.g., -DSCALE_Y=0.5) * @note If the source tensor has more than 3 dimensions, -DBATCHED_EXECUTION must be passed at compile time * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32. - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32. + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_c The size of the channels dimension of the source tensor + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the batches dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: U8/S16/F16/F32. + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) + * @param[in] dst_c The size of the channels dimension of the destination tensor + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the batches dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] scale_x The scale value to apply on the source width + * @param[in] scale_y The scale value to apply on the source height */ - //! @endcond +//! @endcond __kernel void scale_nearest_neighbour_nhwc( - TENSOR4D(src, SRC_TENSOR_TYPE), - TENSOR4D(dst, DST_TENSOR_TYPE)) + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), + const float scale_x, + const float scale_y) { - // All the tensor dimensions are passed at compile time. - // In case of dynamic tensor support, the following dimensions should be passed as function argument. -#define _ISRC_WIDTH SRC_WIDTH -#define _ISRC_HEIGHT SRC_HEIGHT -#define _IDST_WIDTH DST_WIDTH -#define _IDST_HEIGHT DST_HEIGHT -#define _ISCALE_X SCALE_X -#define _ISCALE_Y SCALE_Y - const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM const int xo = GET_SPATIAL_IDX(1, 1, 0); // WIDTH #if defined(BATCHED_EXECUTION) - const int yo = GET_SPATIAL_IDX(2, 1, 0) % _IDST_HEIGHT; // HEIGHT - const int bout = GET_SPATIAL_IDX(2, 1, 0) / _IDST_HEIGHT; // BATCH SIZE IDX -#else // defined(BATCHED_EXECUTION) + const int yo = GET_SPATIAL_IDX(2, 1, 0) % dst_h; // HEIGHT + const int bout = GET_SPATIAL_IDX(2, 1, 0) / dst_h; // BATCH SIZE IDX +#else // defined(BATCHED_EXECUTION) const int yo = GET_SPATIAL_IDX(2, 1, 0); // HEIGHT - const int bout = 0; // BATCH SIZE IDX -#endif // defined(BATCHED_EXECUTION) + const int bout = 0; // BATCH SIZE IDX +#endif // defined(BATCHED_EXECUTION) #ifdef SAMPLING_POLICY_TOP_LEFT - float xi_f = (xo * (float)SCALE_X); - float yi_f = (yo * (float)SCALE_Y); + float xi_f = (xo * scale_x); + float yi_f = (yo * scale_y); #elif SAMPLING_POLICY_CENTER - float xi_f = ((xo + 0.5f) * (float)SCALE_X); - float yi_f = ((yo + 0.5f) * (float)SCALE_Y); + float xi_f = ((xo + 0.5f) * scale_x); + float yi_f = ((yo + 0.5f) * scale_y); #else // SAMPLING_POLICY #error("Unsupported sampling policy"); #endif // SAMPLING_POLICY @@ -101,30 +91,30 @@ __kernel void scale_nearest_neighbour_nhwc( yi_f = round(yi_f); #endif // ALIGN_CORNERS - const int xi0 = clamp((int)xi_f, 0, _ISRC_WIDTH - 1); - const int yi0 = clamp((int)yi_f, 0, _ISRC_HEIGHT - 1); + const int xi0 = clamp((int)xi_f, 0, (int)src_w - 1); + const int yi0 = clamp((int)yi_f, 0, (int)src_h - 1); TILE(SRC_DATA_TYPE, 1, N0, in00); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi0, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, false, in00); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi0, cout, src_w, src_h, 1, 1, false, in00); TILE(uint, 1, 1, dst_indirect_y); // Calculate the destination indirect Y - dst_indirect_y[0].v = xo + (yo * (int)(_IDST_WIDTH)) + bout * (int)(_IDST_WIDTH * _IDST_HEIGHT); + dst_indirect_y[0].v = xo + (yo * (int)(dst_w)) + bout * (int)(dst_w * dst_h); bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0; T_STORE_INDIRECT_WIDTH_SELECT(DST_DATA_TYPE, 1, N0, PARTIAL_N0, DST_TENSOR_TYPE, dst, cout, dst_stride_y, x_cond, in00, dst_indirect_y); } +#endif /* SCALE_NEAREST_NEIGHBOUR */ +#if defined(SCALE_BILINEAR) //! @cond Doxygen_Suppress /** Performs scale on a tensor by interpolating with the BILINEAR method. (NHWC) * * @note If border mode replicate is used, is should be passed as -DBORDER_MODE_REPLICATE * @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT - * @note The spatial dimensions of the source tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT (e.g. -DSRC_WIDTH=96, -DSRC_HEIGHT=64) - * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64) * @note The tensor type ("BUFFER" only is supported) of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER) * @note The tensor type ("BUFFER" only is supported) of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER) * @note The data type of the source tensor must be passed at compile time using -DSRC_DATA_TYPE (e.g. -DSRC_DATA_TYPE=float) @@ -132,65 +122,56 @@ __kernel void scale_nearest_neighbour_nhwc( * @note The number of N0 output channels to process must be passed at compile time using -DN0 (e.g. -DN0=2) * @note The border value value must be passed at compile time using -DCONSTANT_VALUE (e.g. -DCONSTANT_VALUE=0) * @note In case of F32/F16, -DIS_FLOATING_POINT must be passed at compile time - * @note The scale value to apply on the source width must be passed at compile using -DSCALE_X (e.g., -DSCALE_X=0.5) - * @note The scale value to apply on the source height must be passed at compile using -DSCALE_Y (e.g., -DSCALE_Y=0.5) * @note If the source tensor has more than 3 dimensions, -DBATCHED_EXECUTION must be passed at compile time * * @note In case of QASYMM8, the following extra information must be passed at compile time: * - The source offset e.g. -DOFFSET=4 * - The source scale e.g. -DSCALE=4 * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32. - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) - * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) - * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] src_ptr Pointer to the source tensor. Supported data types: U8/S16/F16/F32. + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] src_c The size of the channels dimension of the source tensor + * @param[in] src_w The size of the width dimension of the source tensor + * @param[in] src_h The size of the height dimension of the source tensor + * @param[in] src_n The size of the batches dimension of the source tensor + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: U8/S16/F16/F32. + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes) + * @param[in] dst_c The size of the channels dimension of the destination tensor + * @param[in] dst_w The size of the width dimension of the destination tensor + * @param[in] dst_h The size of the height dimension of the destination tensor + * @param[in] dst_n The size of the batches dimension of the destination tensor + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] scale_x The scale value to apply on the source width + * @param[in] scale_y The scale value to apply on the source height */ - //! @endcond +//! @endcond __kernel void scale_bilinear_nhwc( - TENSOR4D(src, SRC_TENSOR_TYPE), - TENSOR4D(dst, DST_TENSOR_TYPE)) + TENSOR4D_T(src, SRC_TENSOR_TYPE), + TENSOR4D_T(dst, DST_TENSOR_TYPE), + const float scale_x, + const float scale_y) { - // All the tensor dimensions are passed at compile time. - // In case of dynamic tensor support, the following dimensions should be passed as function argument. -#define _ISRC_WIDTH SRC_WIDTH -#define _ISRC_HEIGHT SRC_HEIGHT -#define _IDST_WIDTH DST_WIDTH -#define _IDST_HEIGHT DST_HEIGHT -#define _ISCALE_X SCALE_X -#define _ISCALE_Y SCALE_Y - const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM const int xo = GET_SPATIAL_IDX(1, 1, 0); // WIDTH #if defined(BATCHED_EXECUTION) - const int yo = GET_SPATIAL_IDX(2, 1, 0) % _IDST_HEIGHT; // HEIGHT - const int bout = GET_SPATIAL_IDX(2, 1, 0) / _IDST_HEIGHT; // BATCH SIZE IDX -#else // defined(BATCHED_EXECUTION) + const int yo = GET_SPATIAL_IDX(2, 1, 0) % dst_h; // HEIGHT + const int bout = GET_SPATIAL_IDX(2, 1, 0) / dst_h; // BATCH SIZE IDX +#else // defined(BATCHED_EXECUTION) const int yo = GET_SPATIAL_IDX(2, 1, 0); // HEIGHT const int bout = 0; // BATCH SIZE IDX -#endif // defined(BATCHED_EXECUTION) +#endif // defined(BATCHED_EXECUTION) #ifdef SAMPLING_POLICY_TOP_LEFT - float xi_f = (xo * (float)SCALE_X); - float yi_f = (yo * (float)SCALE_Y); + float xi_f = (xo * scale_x); + float yi_f = (yo * scale_y); #elif SAMPLING_POLICY_CENTER - float xi_f = ((xo + 0.5f) * (float)SCALE_X - 0.5f); - float yi_f = ((yo + 0.5f) * (float)SCALE_Y - 0.5f); + float xi_f = ((xo + 0.5f) * scale_x - 0.5f); + float yi_f = ((yo + 0.5f) * scale_y - 0.5f); #else // SAMPLING_POLICY #error("Unsupported sampling policy"); #endif // SAMPLING_POLICY @@ -210,20 +191,20 @@ __kernel void scale_bilinear_nhwc( in11[0].v = CONSTANT_VALUE; #ifndef BORDER_MODE_REPLICATE - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi, xi, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, true, in00); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi, xi + 1, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, true, in01); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi + 1, xi, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, true, in10); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi + 1, xi + 1, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, true, in11); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi, xi, cout, src_w, src_h, 1, 1, true, in00); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi, xi + 1, cout, src_w, src_h, 1, 1, true, in01); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi + 1, xi, cout, src_w, src_h, 1, 1, true, in10); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi + 1, xi + 1, cout, src_w, src_h, 1, 1, true, in11); #else // BORDER_MODE_REPLICATE - const int xi0 = clamp(xi, 0, _ISRC_WIDTH - 1); - const int yi0 = clamp(yi, 0, _ISRC_HEIGHT - 1); - const int xi1 = clamp(xi + 1, 0, _ISRC_WIDTH - 1); - const int yi1 = clamp(yi + 1, 0, _ISRC_HEIGHT - 1); + const int xi0 = clamp(xi, 0, (int)src_w - 1); + const int yi0 = clamp(yi, 0, (int)src_h - 1); + const int xi1 = clamp(xi + 1, 0, (int)src_w - 1); + const int yi1 = clamp(yi + 1, 0, (int)src_h - 1); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi0, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, false, in00); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi1, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, false, in01); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi1, xi0, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, false, in10); - T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi1, xi1, cout, _ISRC_WIDTH, _ISRC_HEIGHT, 1, 1, false, in11); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi0, cout, src_w, src_h, 1, 1, false, in00); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi0, xi1, cout, src_w, src_h, 1, 1, false, in01); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi1, xi0, cout, src_w, src_h, 1, 1, false, in10); + T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, 1, N0, SRC_TENSOR_TYPE, src, bout, yi1, xi1, cout, src_w, src_h, 1, 1, false, in11); #endif // BORDER_MODE_REPLICATE TILE(DST_DATA_TYPE, 1, N0, out); @@ -270,9 +251,10 @@ __kernel void scale_bilinear_nhwc( TILE(uint, 1, 1, dst_indirect_y); // Calculate the destination indirect Y - dst_indirect_y[0].v = xo + (yo * (int)(_IDST_WIDTH)) + bout * (int)(_IDST_WIDTH * _IDST_HEIGHT); + dst_indirect_y[0].v = xo + (yo * (int)(dst_w)) + bout * (int)(dst_w * dst_h); bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0; T_STORE_INDIRECT_WIDTH_SELECT(DST_DATA_TYPE, 1, N0, PARTIAL_N0, DST_TENSOR_TYPE, dst, cout, dst_stride_y, x_cond, out, dst_indirect_y); -}
\ No newline at end of file +} +#endif /* SCALE_BILINEAR */
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h index f36f273e1..eba231624 100644 --- a/src/core/CL/cl_kernels/tile_helpers.h +++ b/src/core/CL/cl_kernels/tile_helpers.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -104,6 +104,54 @@ #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 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) @@ -448,6 +496,42 @@ }) \ }) +/** Load a tile from global memory (tensor) using an indirect Y index tile and conditionally use a different length for the load + * + * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length for the store + * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones + * + * @param[in] DATA_TYPE Data type + * @param[in] HEIGHT Number of dst rows + * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false + * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true + * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). + * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) + * @param[in] TENSOR Tensor basename + * @param[in] X Starting X position + * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row. + * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store + * @param[out] dst Output tile + * @param[out] indirect_y Indirect Y index tile + */ +#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); \ + }) \ + } \ + }) /** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout * * @param[in] DATA_TYPE Data type |