diff options
Diffstat (limited to 'src')
10 files changed, 1877 insertions, 36 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 new file mode 100644 index 0000000000..e53ce3d1b2 --- /dev/null +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl @@ -0,0 +1,366 @@ +/* + * 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 "fp_post_ops_act_eltwise_op_act.h" +#include "gemm_helpers.h" +#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(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) + +#define VFMA(a, b, c) \ + ({ \ + c = fma(a, b, c); \ + }) + +#if M0 == 1 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + }) +#elif M0 == 2 // M0 == 2 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + }) +#elif M0 == 3 // M0 == 3 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + }) +#elif M0 == 4 // M0 == 4 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + }) +#elif M0 == 5 // M0 == 5 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + }) +#elif M0 == 6 // M0 == 6 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + }) +#elif M0 == 7 // M0 == 7 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + }) +#elif M0 == 8 // M0 == 8 +#define RHS_VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \ + }) +#else // M0 not supported +#error "M0 not supported" +#endif // M0 not supported + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: + * Post op 1: activation (optional) + * Post op 2: elementwise op + * Post op 3: activation (optional) + * + * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform + * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * + * All parameters are similarly defined in kernel gemm_mm_native, with these additions: + * + * @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32 + * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes) + * @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes) + * @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) + */ +__kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), +#if defined(BETA) + IMAGE_DECLARATION(bias), +#endif // defined(BETA) + IMAGE_DECLARATION(dst), + // Post-Op arguments + IMAGE_DECLARATION(eltwise_operand), + uint lhs_stride_z, + uint rhs_stride_z, +#if defined(BETA) + uint bias_stride_z, +#endif //defined(BETA) + uint dst_stride_z, + uint eltwise_operand_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Block size +#define RHS_BLOCK_SIZE ((K0) * (N0)) + + // RHS offset and step X +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + + // Compute RHS matrix address + uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE); + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); + REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; + + int i = 0; +#if K0 > 1 + for(; i <= (K - K0); i += K0) + { + // Supported cases (M0, K0): + // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 + // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 + // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 + // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 + // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 + // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 + // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 + // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS matrix + LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero); + + RHS_VFMA_M0xN0(0, a, b0, c); + RHS_VFMA_M0xN0(1, a, b1, c); +#if K0 > 2 + RHS_VFMA_M0xN0(2, a, b2, c); +#endif // K0 > 2 +#if K0 > 3 + RHS_VFMA_M0xN0(3, a, b3, c); +#endif // K0 > 3 +#if K0 > 4 + RHS_VFMA_M0xN0(4, a, b4, c); + RHS_VFMA_M0xN0(5, a, b5, c); + RHS_VFMA_M0xN0(6, a, b6, c); + RHS_VFMA_M0xN0(7, a, b7, c); +#endif // K0 > 4 +#if K0 > 8 + RHS_VFMA_M0xN0(8, a, b8, c); + RHS_VFMA_M0xN0(9, a, b9, c); + RHS_VFMA_M0xN0(A, a, bA, c); + RHS_VFMA_M0xN0(B, a, bB, c); + RHS_VFMA_M0xN0(C, a, bC, c); + RHS_VFMA_M0xN0(D, a, bD, c); + RHS_VFMA_M0xN0(E, a, bE, c); + RHS_VFMA_M0xN0(F, a, bF, c); +#endif // K0 > 8 + + lhs_offset += K0 * sizeof(DATA_TYPE); + rhs_offset += K0 * rhs_stride_y; + } +#endif // K0 > 1 + // Left-over accumulations + for(; i < K; ++i) + { + // Load values from LHS matrix + VEC_DATA_TYPE(DATA_TYPE, 2) + a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0)); +#if M0 > 1 + VEC_DATA_TYPE(DATA_TYPE, 2) + a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1)); +#endif // M0 > 1 +#if M0 > 2 + VEC_DATA_TYPE(DATA_TYPE, 2) + a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2)); +#endif // M0 > 2 +#if M0 > 3 + VEC_DATA_TYPE(DATA_TYPE, 2) + a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3)); +#endif // M0 > 3 +#if M0 > 4 + VEC_DATA_TYPE(DATA_TYPE, 2) + a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4)); +#endif // M0 > 4 +#if M0 > 5 + VEC_DATA_TYPE(DATA_TYPE, 2) + a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5)); +#endif // M0 > 5 +#if M0 > 6 + VEC_DATA_TYPE(DATA_TYPE, 2) + a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6)); +#endif // M0 > 6 +#if M0 > 7 + VEC_DATA_TYPE(DATA_TYPE, 2) + a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7)); +#endif // M0 > 7 + + VEC_DATA_TYPE(DATA_TYPE, N0) + b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y)); + RHS_VFMA_M0xN0(0, a, b, c); + + lhs_offset += sizeof(DATA_TYPE); + rhs_offset += rhs_stride_y; + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); + + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary + const bool cond_y = ((y + 1) * M0 >= M); + const bool cond_x = ((x + 1) * N0 >= N); + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Add beta*bias +#if defined(BETA) +#if defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); + + LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(1, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias[broadcasted] + ADD_BLOCK_BROADCAST(M0, c, bias0); + +#else // defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id( + 2) * bias_stride_z; + + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias + ADD_BLOCK(M0, c, bias); + +#endif // defined(BROADCAST_BIAS) +#endif // defined(BETA) + + // c = act(c) + POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + // c = act(c) + POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + + // 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(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) 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 9404c5e6db..758fd327fe 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 @@ -352,6 +352,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_post_act_eltwise_op_act(IMAGE_DECLAR REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary const bool cond_y = ((get_global_id(1) + 1) * M0 >= M); const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); @@ -568,6 +569,7 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act(IMAG REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary const bool cond_y = ((get_global_id(1) + 1) * M0 >= M); const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); @@ -824,6 +826,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLAR const uint y = get_global_id(1); const uint z = get_global_id(2); + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary const bool cond_y = ((get_global_id(1) + 1) * M0 >= M); const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); @@ -1326,6 +1329,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act(IMAG REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary const bool cond_y = ((get_global_id(1) + 1) * M0 >= M); const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); 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 new file mode 100644 index 0000000000..508ee96d2d --- /dev/null +++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl @@ -0,0 +1,1375 @@ +/* + * 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 "fp_post_ops_act_eltwise_op_act.h" +#include "gemm_helpers.h" +#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(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH) + +#define CONCAT(a, b) a##b + +#define ARM_DOT1(a, b, c) \ + ({ \ + c = fma(a, b, c); \ + }) +#define ARM_DOT2(a, b, c) \ + ({ \ + c = fma(a.s0, b.s0, c); \ + c = fma(a.s1, b.s1, c); \ + }) +#define ARM_DOT3(a, b, c) \ + ({ \ + ARM_DOT2(a, b, c); \ + c = fma((a.s2), (b.s2), c); \ + }) +#define ARM_DOT4(a, b, c) \ + ({ \ + ARM_DOT3(a, b, c); \ + c = fma((a.s3), (b.s3), c); \ + }) +#define ARM_DOT8(a, b, c) \ + ({ \ + ARM_DOT4((a.lo), (b.lo), c); \ + ARM_DOT4((a.hi), (b.hi), c); \ + }) +#define ARM_DOT16(a, b, c) \ + ({ \ + ARM_DOT8((a.lo), (b.lo), c); \ + ARM_DOT8((a.hi), (b.hi), c); \ + }) + +#if N0 == 2 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + }) +#elif N0 == 3 // N0 == 3 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + }) +#elif N0 == 4 // N0 == 4 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + }) +#elif N0 == 8 // N0 == 8 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##4), (c.s4)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##5), (c.s5)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##6), (c.s6)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##7), (c.s7)); \ + }) +#elif N0 == 16 // N0 == 16 +#define ARM_DOT_K0XN0(k0, a, b, c) \ + ({ \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##0), (c.s0)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##1), (c.s1)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##2), (c.s2)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##3), (c.s3)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##4), (c.s4)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##5), (c.s5)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##6), (c.s6)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##7), (c.s7)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##8), (c.s8)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##9), (c.s9)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##A), (c.sA)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##B), (c.sB)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##C), (c.sC)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##D), (c.sD)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##E), (c.sE)); \ + CONCAT(ARM_DOT, k0) \ + ((a), (b##F), (c.sF)); \ + }) +#else // N0 not supported +#error "N0 value not supported" +#endif // N0 conditions + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: + * Post op 1: activation (optional) + * Post op 2: elementwise op + * Post op 3: activation (optional) + * + * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform + * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * + * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_t, with these additions: + * + * @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32 + * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes) + * @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes) + * @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) + */ +__kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), +#if defined(BETA) + IMAGE_DECLARATION(bias), +#endif // defined(BETA) + IMAGE_DECLARATION(dst), + // Post-Op arguments + IMAGE_DECLARATION(eltwise_operand), + uint lhs_stride_z, + uint rhs_stride_z, +#if defined(BETA) + uint bias_stride_z, +#endif //defined(BETA) + uint dst_stride_z, + uint eltwise_operand_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Block size +#define RHS_BLOCK_SIZE ((K0) * (N0)) + + // RHS offset and step X +#if defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (K0) +#define RHS_STEP_X ((K0) * (H0)) +#define RHS_STEP_LOOP (1) +#else // defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) +#define RHS_STEP_X (K0) +#define RHS_STEP_LOOP (H0) +#endif // defined(RHS_INTERLEAVE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + + // Compute RHS reshaped matrix address + uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0; + REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0; + + int i = 0; + for(; i <= (K - K0); i += K0) + { + // Supported cases (M0, K0): + // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 + // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 + // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 + // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 + // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 + // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 + // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 + // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS reshaped matrix + LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero); + + // Accumulate + ARM_DOT_K0XN0(K0, a0, b, c0); +#if M0 > 1 + ARM_DOT_K0XN0(K0, a1, b, c1); +#endif // M0 > 1 +#if M0 > 2 + ARM_DOT_K0XN0(K0, a2, b, c2); +#endif // M0 > 2 +#if M0 > 3 + ARM_DOT_K0XN0(K0, a3, b, c3); +#endif // M0 > 3 +#if M0 > 4 + ARM_DOT_K0XN0(K0, a4, b, c4); +#endif // M0 > 4 +#if M0 > 5 + ARM_DOT_K0XN0(K0, a5, b, c5); +#endif // M0 > 5 +#if M0 > 6 + ARM_DOT_K0XN0(K0, a6, b, c6); +#endif // M0 > 6 +#if M0 > 7 + ARM_DOT_K0XN0(K0, a7, b, c7); +#endif // M0 > 7 + + lhs_offset += K0 * sizeof(DATA_TYPE); + rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE); + } + + // Left-over accumulations + for(; i < K; ++i) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS reshaped matrix + LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero); + + // Accumulate + ARM_DOT_K0XN0(1, a0, b, c0); +#if M0 > 1 + ARM_DOT_K0XN0(1, a1, b, c1); +#endif // M0 > 1 +#if M0 > 2 + ARM_DOT_K0XN0(1, a2, b, c2); +#endif // M0 > 2 +#if M0 > 3 + ARM_DOT_K0XN0(1, a3, b, c3); +#endif // M0 > 3 +#if M0 > 4 + ARM_DOT_K0XN0(1, a4, b, c4); +#endif // M0 > 4 +#if M0 > 5 + ARM_DOT_K0XN0(1, a5, b, c5); +#endif // M0 > 5 +#if M0 > 6 + ARM_DOT_K0XN0(1, a6, b, c6); +#endif // M0 > 6 +#if M0 > 7 + ARM_DOT_K0XN0(1, a7, b, c7); +#endif // M0 > 7 + + lhs_offset += sizeof(DATA_TYPE); + rhs_offset += sizeof(DATA_TYPE); + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary + const bool cond_y = ((y + 1) * M0 >= M); + const bool cond_x = ((x + 1) * N0 >= N); + +#if defined(REINTERPRET_OUTPUT_AS_3D) + + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Add beta*bias +#if defined(BETA) +#if defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); + + LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(1, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias[broadcasted] + ADD_BLOCK_BROADCAST(M0, c, bias0); + +#else // defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias + ADD_BLOCK(M0, c, bias); + +#endif // defined(BROADCAST_BIAS) +#endif // defined(BETA) + + // c = act(c) + POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + // c = act(c) + POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + + // 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); + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} + +#if defined(OPENCL_IMAGE_SUPPORT) +/** 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 + * Post op 3: activation (optional) + * + * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform + * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * + * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_t_texture, with these additions: + * + * @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32 + * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes) + * @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes) + * @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) + */ +__kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), + __read_only image2d_t rhs_img, +#if defined(BETA) + IMAGE_DECLARATION(bias), +#endif // defined(BETA) + IMAGE_DECLARATION(dst), + // Post-Op arguments + IMAGE_DECLARATION(eltwise_operand), + uint lhs_stride_z, + uint rhs_stride_z, +#if defined(BETA) + uint bias_stride_z, +#endif //defined(BETA) + uint dst_stride_z, + uint eltwise_operand_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Pixel unit +#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0) + +#define LEFTOVER_K (K % K0) + + // Block size +#define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0)) + + // RHS offset and step X +#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); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH); +#else // defined(MATRIX_B_DEPTH) + const uint z_rhs = get_global_id(2); +#endif // defined(MATRIX_B_DEPTH) + + // Compute RHS matrix coordinates + uint x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X; + const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT; + + REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); + REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); + + int i = 0; + for(; i <= (K - K0); i += K0) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs); + + // Load values from RHS matrix stored in a cl_image + REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0); + LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0); + + // Accumulate + ARM_DOT_K0XN0(K0, a0, b, c0); +#if M0 > 1 + ARM_DOT_K0XN0(K0, a1, b, c1); +#endif // M0 > 1 +#if M0 > 2 + ARM_DOT_K0XN0(K0, a2, b, c2); +#endif // M0 > 2 +#if M0 > 3 + ARM_DOT_K0XN0(K0, a3, b, c3); +#endif // M0 > 3 +#if M0 > 4 + ARM_DOT_K0XN0(K0, a4, b, c4); +#endif // M0 > 4 +#if M0 > 5 + ARM_DOT_K0XN0(K0, a5, b, c5); +#endif // M0 > 5 +#if M0 > 6 + ARM_DOT_K0XN0(K0, a6, b, c6); +#endif // M0 > 6 +#if M0 > 7 + ARM_DOT_K0XN0(K0, a7, b, c7); +#endif // M0 > 7 + + lhs_offset += K0 * sizeof(DATA_TYPE); + 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 + { + 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 }; +#endif // M0 > 1 +#if M0 > 2 + union UNION_VEC_TYPE a2 = {.v = 0 }; +#endif // M0 > 2 +#if M0 > 3 + union UNION_VEC_TYPE a3 = {.v = 0 }; +#endif // M0 > 3 +#if M0 > 4 + union UNION_VEC_TYPE a4 = {.v = 0 }; +#endif // M0 > 4 +#if M0 > 5 + union UNION_VEC_TYPE a5 = {.v = 0 }; +#endif // M0 > 5 +#if M0 > 6 + union UNION_VEC_TYPE a6 = {.v = 0 }; +#endif // M0 > 6 +#if M0 > 7 + union UNION_VEC_TYPE a7 = {.v = 0 }; +#endif // M0 > 7 + + 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 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); +#endif // M0 > 1 +#if M0 > 2 + 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); +#endif // M0 > 3 +#if M0 > 4 + 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); +#endif // M0 > 5 +#if M0 > 6 + 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); +#endif // M0 > 7 + + lhs_offset += sizeof(DATA_TYPE); + } + + // Accumulate + ARM_DOT_K0XN0(K0, a0.v, b, c0); +#if M0 > 1 + ARM_DOT_K0XN0(K0, a1.v, b, c1); +#endif // M0 > 1 +#if M0 > 2 + ARM_DOT_K0XN0(K0, a2.v, b, c2); +#endif // M0 > 2 +#if M0 > 3 + ARM_DOT_K0XN0(K0, a3.v, b, c3); +#endif // M0 > 3 +#if M0 > 4 + ARM_DOT_K0XN0(K0, a4.v, b, c4); +#endif // M0 > 4 +#if M0 > 5 + ARM_DOT_K0XN0(K0, a5.v, b, c5); +#endif // M0 > 5 +#if M0 > 6 + ARM_DOT_K0XN0(K0, a6.v, b, c6); +#endif // M0 > 6 +#if M0 > 7 + 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)) + (y * M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary + const bool cond_y = ((y + 1) * M0 >= M); + const bool cond_x = ((x + 1) * N0 >= N); + +#if defined(REINTERPRET_OUTPUT_AS_3D) + + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Add beta*bias +#if defined(BETA) +#if defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); + + LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(1, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias[broadcasted] + ADD_BLOCK_BROADCAST(M0, c, bias0); + +#else // defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias + ADD_BLOCK(M0, c, bias); + +#endif // defined(BROADCAST_BIAS) +#endif // defined(BETA) + + // c = act(c) + POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + // c = act(c) + POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + + // 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); + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +#undef LEFTOVER_K +#undef PIXEL_UNIT +} +#endif // defined(OPENCL_IMAGE_SUPPORT) + +#define VFMA(a, b, c) \ + ({ \ + c = fma(a, b, c); \ + }) + +#if M0 == 1 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + }) +#elif M0 == 2 // M0 == 2 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + }) +#elif M0 == 3 // M0 == 3 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + }) +#elif M0 == 4 // M0 == 4 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + }) +#elif M0 == 5 // M0 == 5 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + }) +#elif M0 == 6 // M0 == 6 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + }) +#elif M0 == 7 // M0 == 7 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + }) +#elif M0 == 8 // M0 == 8 +#define VFMA_M0xN0(i, a, b, c) \ + ({ \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \ + VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \ + }) +#else // M0 not supported +#error "M0 not supported" +#endif // M0 not supported + +/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops: + * Post op 1: activation (optional) + * Post op 2: elementwise op + * Post op 3: activation (optional) + * + * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform + * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * + * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_nt, with these additions: + * + * @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32 + * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes) + * @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes) + * @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) + */ +__kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), + IMAGE_DECLARATION(rhs), +#if defined(BETA) + IMAGE_DECLARATION(bias), +#endif // defined(BETA) + IMAGE_DECLARATION(dst), + // Post-Op arguments + IMAGE_DECLARATION(eltwise_operand), + uint lhs_stride_z, + uint rhs_stride_z, +#if defined(BETA) + uint bias_stride_z, +#endif //defined(BETA) + uint dst_stride_z, + uint eltwise_operand_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Block size +#define RHS_BLOCK_SIZE ((K0) * (N0)) + + // RHS offset and step X +#if defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (N0) +#define RHS_STEP_X ((N0) * (H0)) +#define RHS_STEP_LOOP (1) +#else // defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) +#define RHS_STEP_X (N0) +#define RHS_STEP_LOOP (H0) +#endif // defined(RHS_INTERLEAVE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + + // Compute RHS reshaped matrix address + uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y; + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z; +#else // defined(MATRIX_B_DEPTH) + rhs_offset += z * rhs_stride_z; +#endif // defined(MATRIX_B_DEPTH) + + REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); //uint zin0=0,zin1=0,zin2=0,... zin7=0; + REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0; + +#if defined(REINTERPRET_INPUT_AS_3D) + + // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(N0-1)=0; + + int i = 0; + for(; i <= (K - K0); i += K0) + { + // Supported cases (M0, K0): + // 1,2 - 1,3 - 1,4 - 1,8 - 1,16 + // 2,2 - 2,3 - 2,4 - 2,8 - 2,16 + // 3,2 - 3,3 - 3,4 - 3,8 - 3,16 + // 4,2 - 4,3 - 4,4 - 4,8 - 4,16 + // 5,2 - 5,3 - 5,4 - 5,8 - 5,16 + // 6,2 - 6,3 - 6,4 - 6,8 - 6,16 + // 7,2 - 7,3 - 7,4 - 7,8 - 7,16 + // 8,2 - 8,3 - 8,4 - 8,8 - 8,16 + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin); + + VEC_DATA_TYPE(DATA_TYPE, N0) + b0; + + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(0, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(1, a, b0, c); +#if K0 > 2 + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(2, a, b0, c); +#endif // K0 > 2 +#if K0 > 3 + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(3, a, b0, c); +#endif // K0 > 3 +#if K0 > 4 + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(4, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(5, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(6, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(7, a, b0, c); +#endif // K0 > 4 +#if K0 > 8 + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(8, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(9, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(A, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(B, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(C, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(D, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(E, a, b0, c); + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(F, a, b0, c); +#endif // K0 > 8 + + lhs_offset += K0 * sizeof(DATA_TYPE); + rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE); + } + + // Left-over accumulations + for(; i < K; ++i) + { + // Load values from LHS matrix + VEC_DATA_TYPE(DATA_TYPE, 2) + a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0)); +#if M0 > 1 + VEC_DATA_TYPE(DATA_TYPE, 2) + a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1)); +#endif // M0 > 1 +#if M0 > 2 + VEC_DATA_TYPE(DATA_TYPE, 2) + a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2)); +#endif // M0 > 2 +#if M0 > 3 + VEC_DATA_TYPE(DATA_TYPE, 2) + a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3)); +#endif // M0 > 3 +#if M0 > 4 + VEC_DATA_TYPE(DATA_TYPE, 2) + a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4)); +#endif // M0 > 4 +#if M0 > 5 + VEC_DATA_TYPE(DATA_TYPE, 2) + a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5)); +#endif // M0 > 5 +#if M0 > 6 + VEC_DATA_TYPE(DATA_TYPE, 2) + a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6)); +#endif // M0 > 6 +#if M0 > 7 + VEC_DATA_TYPE(DATA_TYPE, 2) + a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7)); +#endif // M0 > 7 + + VEC_DATA_TYPE(DATA_TYPE, N0) + b0; + + b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE))); + VFMA_M0xN0(0, a, b0, c); + + lhs_offset += sizeof(DATA_TYPE); + rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE); + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary + const bool cond_y = ((y + 1) * M0 >= M); + const bool cond_x = ((x + 1) * N0 >= N); + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Add beta*bias +#if defined(BETA) +#if defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); + + LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(1, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias[broadcasted] + ADD_BLOCK_BROADCAST(M0, c, bias0); + +#else // defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias + ADD_BLOCK(M0, c, bias); + +#endif // defined(BROADCAST_BIAS) +#endif // defined(BETA) + + // c = act(c) + POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + // c = act(c) + POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + + // 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); + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} + +#if defined(OPENCL_IMAGE_SUPPORT) +/** 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 + * Post op 3: activation (optional) + * + * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform + * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2 + * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3 + * + * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_nt_texture, with these additions: + * + * @param[in] eltwise_operand_ptr Pointer to the eltwise operand matrix. Supported data type: F16/F32 + * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes) + * @param[in] eltwise_operand_step_x eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes) + * @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) + */ +__kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs), + __read_only image2d_t rhs_img, +#if defined(BETA) + IMAGE_DECLARATION(bias), +#endif // defined(BETA) + IMAGE_DECLARATION(dst), + // Post-Op arguments + IMAGE_DECLARATION(eltwise_operand), + uint lhs_stride_z, + uint rhs_stride_z, +#if defined(BETA) + uint bias_stride_z, +#endif //defined(BETA) + uint dst_stride_z, + uint eltwise_operand_stride_z +#if defined(REINTERPRET_INPUT_AS_3D) + , + uint lhs_cross_plane_pad +#endif // REINTERPRET_INPUT_AS_3D +#if defined(REINTERPRET_OUTPUT_AS_3D) + , + uint dst_cross_plane_pad +#endif // REINTERPRET_OUTPUT_AS_3D + ) +{ + // Pixel unit +#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0) + + // Block size +#define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT)) + + // RHS offset and step X +#if defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (PIXEL_UNIT) +#define RHS_STEP_X ((PIXEL_UNIT) * (H0)) +#else // defined(RHS_INTERLEAVE) +#define RHS_OFFSET_X (RHS_BLOCK_SIZE) +#define RHS_STEP_X (PIXEL_UNIT) +#endif // defined(RHS_INTERLEAVE) + + uint x = get_global_id(0); + uint y = get_global_id(1); + uint z = get_global_id(2); + +#if defined(DUMMY_WORK_ITEMS) + if((x * N0 >= N) || (y * M0 >= M)) + { + return; + } +#endif // defined(DUMMY_WORK_ITEMS) + + // Compute LHS matrix address + uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y; + +#if defined(MATRIX_B_DEPTH) + // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 + const uint z_rhs = (z % MATRIX_B_DEPTH); +#else // defined(MATRIX_B_DEPTH) + const uint z_rhs = z; +#endif // defined(MATRIX_B_DEPTH) + + // Compute RHS matrix coordinates + uint x_rhs = (x % H0) * (uint)RHS_OFFSET_X; + const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT; + + REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); + REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); + +#if defined(REINTERPRET_INPUT_AS_3D) + + // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zin, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply lhs_stride_z by DEPTH_GEMM3D + lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_INPUT_AS_3D) + + // Add offset for batched GEMM + lhs_offset += z * lhs_stride_z; + +#endif // defined(REINTERPRET_INPUT_AS_3D) + + // Initialize the accumulators + REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); + + int i = 0; + for(; i <= (K - K0); i += K0) + { + // Load values from LHS matrix + LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin); + + VEC_DATA_TYPE(DATA_TYPE, N0) + b0; + + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(0, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(1, a, b0, c); +#if K0 > 2 + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(2, a, b0, c); +#endif // K0 > 2 +#if K0 > 3 + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(3, a, b0, c); +#endif // K0 > 3 +#if K0 > 4 + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(4, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(5, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(6, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(7, a, b0, c); +#endif // K0 > 4 +#if K0 > 8 + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(8, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(9, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(A, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(B, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(C, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(D, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(E, a, b0, c); + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs)); + VFMA_M0xN0(F, a, b0, c); +#endif // K0 > 8 + + lhs_offset += K0 * sizeof(DATA_TYPE); + x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP; + } + + // Left-over accumulations + for(; i < K; ++i) + { + // Load values from LHS matrix + VEC_DATA_TYPE(DATA_TYPE, 2) + a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0)); +#if M0 > 1 + VEC_DATA_TYPE(DATA_TYPE, 2) + a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1)); +#endif // M0 > 1 +#if M0 > 2 + VEC_DATA_TYPE(DATA_TYPE, 2) + a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2)); +#endif // M0 > 2 +#if M0 > 3 + VEC_DATA_TYPE(DATA_TYPE, 2) + a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3)); +#endif // M0 > 3 +#if M0 > 4 + VEC_DATA_TYPE(DATA_TYPE, 2) + a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4)); +#endif // M0 > 4 +#if M0 > 5 + VEC_DATA_TYPE(DATA_TYPE, 2) + a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5)); +#endif // M0 > 5 +#if M0 > 6 + VEC_DATA_TYPE(DATA_TYPE, 2) + a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6)); +#endif // M0 > 6 +#if M0 > 7 + VEC_DATA_TYPE(DATA_TYPE, 2) + a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7)); +#endif // M0 > 7 + + VEC_DATA_TYPE(DATA_TYPE, N0) + b0; + b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs)); + + VFMA_M0xN0(0, a, b0, c); + + lhs_offset += sizeof(DATA_TYPE); + x_rhs += RHS_STEP_X; + } + + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y); + + REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0; + + // Boundary conditions: detect if current block is at the "bottom" or "right" boundary + const bool cond_y = ((y + 1) * M0 >= M); + const bool cond_x = ((x + 1) * N0 >= N); + +#if defined(REINTERPRET_OUTPUT_AS_3D) + // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D + CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y); + + // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we + // multiply dst_stride_z by DEPTH_GEMM3D + dst_addr += z * dst_stride_z * DEPTH_GEMM3D; + +#else // defined(REINTERPRET_OUTPUT_AS_3D) + + // Add offset for batched GEMM + dst_addr += z * dst_stride_z; + +#endif // defined(REINTERPRET_OUTPUT_AS_3D) + + // Multiply by the weight of matrix-matrix product and store the result +#if defined(ALPHA) + SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA); +#endif // defined(ALPHA) + + // Add beta*bias +#if defined(BETA) +#if defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); + + LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(1, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias[broadcasted] + ADD_BLOCK_BROADCAST(M0, c, bias0); + +#else // defined(BROADCAST_BIAS) + __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * bias_stride_y) + z * bias_stride_z; + + LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + +#ifndef UNIT_BETA + SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); +#endif // UNIT_BIAS + + // c = c + bias + ADD_BLOCK(M0, c, bias); + +#endif // defined(BROADCAST_BIAS) +#endif // defined(BETA) + + // c = act(c) + POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + // c = c + eltwise_operand (mix-precision, broadcast, boundary aware) + POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); + // c = act(c) + POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c); + + // 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); + +#undef RHS_BLOCK_SIZE +#undef RHS_OFFSET_X +#undef RHS_STEP_X +} +#endif // defined(OPENCL_IMAGE_SUPPORT) +#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 diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl index 9732588b31..6502dd496a 100644 --- a/src/core/CL/cl_kernels/common/gemm.cl +++ b/src/core/CL/cl_kernels/common/gemm.cl @@ -1000,6 +1000,7 @@ __kernel void gemm_reshape_rhs_matrix_t(TENSOR3D_DECLARATION(src), /** 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) @@ -1294,6 +1295,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs), /** 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 + * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl * * @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. @@ -1720,6 +1722,7 @@ __kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs), /** 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). @@ -2038,6 +2041,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs), /** 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 -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. @@ -4025,6 +4029,7 @@ __kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs), /** 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) diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp index cbc4caf5f6..c47cf8ef11 100644 --- a/src/gpu/cl/ClKernelLibrary.cpp +++ b/src/gpu/cl/ClKernelLibrary.cpp @@ -272,6 +272,7 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map = { "gemm_mv", "common/gemv.cl" }, { "gemm_mv_quantized", "common/gemv.cl" }, { "gemm_mm_native", "common/gemm.cl" }, + { "gemm_mm_native_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl" }, { "gemm_mm_reshaped_lhs_nt_rhs_t", "common/gemm.cl" }, { "gemm_mm_reshaped_lhs_nt_rhs_t_texture", "common/gemm.cl" }, { "gemm_mm_reshaped_lhs_t_rhs_nt", "common/gemm.cl" }, @@ -284,6 +285,10 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map = { "gemm_mm_reshaped_only_rhs_nt_texture", "common/gemm.cl" }, { "gemm_mm_reshaped_only_rhs_t", "common/gemm.cl" }, { "gemm_mm_reshaped_only_rhs_t_texture", "common/gemm.cl" }, + { "gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" }, + { "gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" }, + { "gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" }, + { "gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" }, { "gemm_lc_vm_f32", "common/gemm.cl" }, { "gemm_reshape_lhs_matrix_nt", "common/gemm.cl" }, { "gemm_reshape_lhs_matrix_t", "common/gemm.cl" }, @@ -585,10 +590,18 @@ const std::map<std::string, std::string> ClKernelLibrary::_program_source_map = #include "./cl_kernels/common/gemm.clembed" }, { + "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl", +#include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.clembed" + }, + { "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl", #include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.clembed" }, { + "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl", +#include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.clembed" + }, + { "common/gemmlowp.cl", #include "./cl_kernels/common/gemmlowp.clembed" }, diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp index e389ce5b0c..7ad3d55fe0 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp @@ -33,6 +33,8 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/AccessWindowStatic.h" +#include "src/core/CL/CLUtils.h" +#include "src/core/experimental/PostOp.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/helpers/float_ops.h" @@ -49,6 +51,17 @@ namespace { using ElementsProcessed = Steps; +const auto post_op_utils = experimental::PostOpCLKernelUtils( +{ + // PostOp sequence -> {Kernel Postfix, PostOp Slots} + { {}, { "", {} } }, + { { experimental::PostOpType::Activation }, { "", { 1 } } }, + { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } }, + { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } }, + { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, + { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } } +}); + Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) @@ -68,6 +81,7 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -110,6 +124,7 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant"); } return Status{}; @@ -170,16 +185,17 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile { ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); - // dst tensor auto initialization if not yet initialized auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); + auto padding_info = get_padding_info({ src0, src1, src2, dst }); _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); _add_bias = src2 != nullptr; + _num_post_op_args = gemm_info.post_ops.total_num_arguments(); // In case both input and dst have to be reinterpreted as 3D tensors, // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. @@ -237,11 +253,20 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); + // If post_ops are used, then we disable the use of gemm_info.activation_info + if(gemm_info.post_ops.size() > 0) + { + post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops); + } + else + { + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); + } std::string kernel_name("gemm_mm_native"); + post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops); // Create kernel _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); @@ -323,11 +348,11 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window unsigned int idx0; if(_add_bias) { - idx0 = 4 * num_arguments_per_2D_tensor() + 4; + idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + (4 + _num_post_op_args); } else { - idx0 = 3 * num_arguments_per_2D_tensor() + 3; + idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + (3 + _num_post_op_args); } const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom; _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); @@ -339,11 +364,11 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window unsigned int idx0; if(_add_bias) { - idx0 = 4 * num_arguments_per_2D_tensor() + 4 + (_reinterpret_input_as_3d ? 1 : 0); + idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + 4 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args; } else { - idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0); + idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args; } const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom; _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); @@ -367,6 +392,12 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window add_2D_tensor_argument(idx, src2, slice); } add_2D_tensor_argument(idx, dst, slice); + // post op argument buffers + for(size_t i = 0; i < _num_post_op_args; ++i) + { + const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); + add_2D_tensor_argument(idx, post_op_arg, slice); + } _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2])); if(_add_bias) @@ -374,6 +405,12 @@ void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2])); } _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2])); + // post op argument stride_z + for(size_t i = 0; i < _num_post_op_args; ++i) + { + const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); + _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2])); + } enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); } while(window.slide_window_slice_3D(slice)); diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h index 89837cc515..415eb7bf3b 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h @@ -76,11 +76,12 @@ public: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; private: - bool _slide_matrix_b{ true }; - bool _reinterpret_input_as_3d{ false }; - bool _reinterpret_output_as_3d{ false }; - bool _use_dummy_work_items{ false }; - bool _add_bias{ false }; + bool _slide_matrix_b{ true }; + bool _reinterpret_input_as_3d{ false }; + bool _reinterpret_output_as_3d{ false }; + bool _use_dummy_work_items{ false }; + bool _add_bias{ false }; + unsigned int _num_post_op_args{ 0 }; // (EXPERIMENTAL_POST_OPS) total number of post op arguments }; } // namespace kernels } // namespace opencl diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp index 04c1cd66c9..260ed134e4 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp @@ -27,6 +27,7 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "src/core/CL/CLUtils.h" #include "src/core/CL/CLValidate.h" +#include "src/core/experimental/PostOp.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "src/core/utils/helpers/float_ops.h" @@ -44,6 +45,17 @@ namespace { using ElementsProcessed = Steps; +const auto post_op_utils = experimental::PostOpCLKernelUtils( +{ + // PostOp sequence -> {Kernel Postfix, PostOp Slots} + { {}, { "", {} } }, + { { experimental::PostOpType::Activation }, { "", { 1 } } }, + { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } }, + { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } }, + { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } }, + { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } } +}); + Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) { @@ -64,6 +76,7 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info)); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -109,6 +122,7 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant"); } return Status{}; @@ -168,6 +182,9 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext { ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); + // dst tensor auto initialization if not yet initialized + auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; @@ -176,9 +193,7 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext _add_bias = src2 != nullptr; _export_to_cl_image = rhs_info.export_to_cl_image; _has_pad_y = gemm_info.has_pad_y; - - // dst tensor auto initialization if not yet initialized - auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); + _num_post_op_args = gemm_info.post_ops.total_num_arguments(); auto padding_info = get_padding_info({ src0, src1, src2, dst }); @@ -239,9 +254,6 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); - build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); if(_has_pad_y) { build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); @@ -249,10 +261,22 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d)); build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d)); } + // If post_ops are used, then we disable the use of gemm_info.activation_info + if(gemm_info.post_ops.size() > 0) + { + post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops); + } + else + { + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); + build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); + } std::string kernel_name("gemm_mm_reshaped_only_rhs_"); kernel_name += rhs_info.transpose ? "t" : "nt"; kernel_name += rhs_info.export_to_cl_image ? "_texture" : ""; + post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops); // Create kernel _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); @@ -375,6 +399,13 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con // dst buffer add_2D_tensor_argument(idx, dst, slice); + // post op argument buffers + for(size_t i = 0; i < _num_post_op_args; ++i) + { + const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); + add_2D_tensor_argument(idx, post_op_arg, slice); + } + // LHS stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[lhs_idx_batch_size])); @@ -389,6 +420,12 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con // dst stride_z _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size])); + // post op argument stride_z + for(size_t i = 0; i < _num_post_op_args; ++i) + { + const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i))); + _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2])); + } // Cross-plan padding (if _reinterpret_input_as_3d = true) if(_reinterpret_input_as_3d && _has_pad_y) diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h index cb82b4af5e..a8f0c4c3a0 100644 --- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h +++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h @@ -90,13 +90,14 @@ public: void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; private: - bool _slide_matrix_b{ true }; - bool _reinterpret_input_as_3d{ false }; - bool _reinterpret_output_as_3d{ false }; - bool _use_dummy_work_items{ false }; - bool _add_bias{ false }; - bool _export_to_cl_image{ false }; - bool _has_pad_y{ false }; + bool _slide_matrix_b{ true }; + bool _reinterpret_input_as_3d{ false }; + bool _reinterpret_output_as_3d{ false }; + bool _use_dummy_work_items{ false }; + bool _add_bias{ false }; + bool _export_to_cl_image{ false }; + bool _has_pad_y{ false }; + unsigned int _num_post_op_args{ 0 }; // (EXPERIMENTAL_POST_OPS) total number of post op arguments }; } // namespace kernels } // namespace opencl diff --git a/src/gpu/cl/operators/ClGemm.cpp b/src/gpu/cl/operators/ClGemm.cpp index e05256ee2f..50ecb214e3 100644 --- a/src/gpu/cl/operators/ClGemm.cpp +++ b/src/gpu/cl/operators/ClGemm.cpp @@ -204,7 +204,6 @@ ClGemm::ClGemm() void ClGemm::configure_native(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) { - ARM_COMPUTE_ERROR_ON_MSG(gemm_info.post_ops().size() > 0, "PostOps are not supported in this kernel"); DataType data_type = a->data_type(); bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); @@ -223,6 +222,7 @@ void ClGemm::configure_native(const CLCompileContext &compile_context, ITensorIn kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; kernel_info.broadcast_bias = broadcast_bias; kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); // Set the target for the kernels _mm_native_kernel->set_target(gpu_target); @@ -281,7 +281,6 @@ void ClGemm::configure_reshaped(const CLCompileContext &compile_context, ITensor void ClGemm::configure_reshaped_only_rhs(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) { - ARM_COMPUTE_ERROR_ON_MSG(gemm_info.post_ops().size() > 0, "PostOps are not supported in this kernel"); DataType data_type = a->data_type(); bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); @@ -300,6 +299,7 @@ void ClGemm::configure_reshaped_only_rhs(const CLCompileContext &compile_context kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; kernel_info.broadcast_bias = broadcast_bias; kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); // Set the target for the kernels _mm_reshaped_only_rhs_kernel->set_target(gpu_target); @@ -334,7 +334,6 @@ Status ClGemm::validate_native(const ITensorInfo *a, const ITensorInfo *b, const { ARM_COMPUTE_UNUSED(alpha); ARM_COMPUTE_UNUSED(output); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.post_ops().size() > 0, "PostOps are not supported in this kernel"); // Get the GPU target const GPUTarget gpu_target = CLScheduler::get().target(); @@ -355,6 +354,7 @@ Status ClGemm::validate_native(const ITensorInfo *a, const ITensorInfo *b, const kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; kernel_info.broadcast_bias = broadcast_bias; kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); @@ -418,7 +418,6 @@ Status ClGemm::validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInf { ARM_COMPUTE_UNUSED(alpha); ARM_COMPUTE_UNUSED(output); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.post_ops().size() > 0, "PostOps are not supported in this kernel"); TensorInfo tmp_b_info{}; @@ -441,6 +440,7 @@ Status ClGemm::validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInf kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; kernel_info.broadcast_bias = broadcast_bias; kernel_info.activation_info = gemm_info.activation_info(); + kernel_info.post_ops = gemm_info.post_ops(); GEMMLHSMatrixInfo lhs_info; GEMMRHSMatrixInfo rhs_info; @@ -562,10 +562,9 @@ Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso void ClGemm::run(ITensorPack &tensors) { - const ITensor *lhs = tensors.get_const_tensor(ACL_SRC_0); - const ITensor *rhs = tensors.get_const_tensor(ACL_SRC_1); - const ITensor *src2 = tensors.get_const_tensor(ACL_SRC_2); - ITensor *dst = tensors.get_tensor(ACL_DST); + const ITensor *lhs = tensors.get_const_tensor(ACL_SRC_0); + const ITensor *rhs = tensors.get_const_tensor(ACL_SRC_1); + ITensor *dst = tensors.get_tensor(ACL_DST); ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, dst); @@ -620,7 +619,10 @@ void ClGemm::run(ITensorPack &tensors) const unsigned int cross_plane_pad_dst = dst->info()->padding().top + dst->info()->padding().bottom; bool has_pad_y = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0); - ITensorPack gemm_reshaped_onlyrhs_pack{ { ACL_SRC_0, lhs }, { ACL_SRC_1, rhs_reshaped.get() }, { ACL_SRC_2, src2 }, { ACL_DST, dst } }; + // Copy original tensor pack and overwrite rhs with reshaped counterpart + ITensorPack gemm_reshaped_onlyrhs_pack(tensors); + gemm_reshaped_onlyrhs_pack.add_const_tensor(ACL_SRC_1, rhs_reshaped.get()); + if(has_pad_y) { CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_fallback_kernel, gemm_reshaped_onlyrhs_pack, true); |