/* * Copyright (c) 2020 Arm Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "gemm_helpers.h" #include "repeat.h" #if defined(M) && defined(N) && defined(K) && defined(H0) && defined(V0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) /** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) * * @note The number of rows of destination matrix must be passed at compile time using -DM * @note The number of columns of the destination matrix must be passed at compile time using -DN * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2) * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_interleaved_transposed_f32(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_OUTPUT_AS_3D) , uint cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int x = get_global_id(0) / H0; int y = get_global_id(1) / V0; int z = get_global_id(2); // Offset const int offset_row_a = (get_global_id(1) % V0) * 4; const int offset_row_b = (get_global_id(0) % H0) * 4; // src_addr_a = address of matrix A // src_addr_b = address of matrix B int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src1_addr_in_bytes += z * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes); __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes); // Compute end row address for matrix B __global float *src_end_addr_b = src_addr_b + (src1_stride_y / sizeof(float)); src_addr_a += offset_row_a; src_addr_b += offset_row_b; // Reset accumulators float4 c0 = 0.0f; float4 c1 = 0.0f; float4 c2 = 0.0f; float4 c3 = 0.0f; for(; src_addr_b <= (src_end_addr_b - (int)(8 * H0)); src_addr_a += 8 * V0, src_addr_b += 8 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = vload4(0, src_addr_a); float4 b0 = vload4(0, src_addr_b); c0 += (float4)a0.s0 * b0; c1 += (float4)a0.s1 * b0; c2 += (float4)a0.s2 * b0; c3 += (float4)a0.s3 * b0; // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a + 4 * V0); b0 = vload4(0, src_addr_b + 4 * H0); c0 += (float4)a0.s0 * b0; c1 += (float4)a0.s1 * b0; c2 += (float4)a0.s2 * b0; c3 += (float4)a0.s3 * b0; } for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 4 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = vload4(0, src_addr_a); float4 b0 = vload4(0, src_addr_b); c0 += (float4)a0.s0 * b0; c1 += (float4)a0.s1 * b0; c2 += (float4)a0.s2 * b0; c3 += (float4)a0.s3 * b0; } // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); // Compute dst address __global uchar *dst_addr = offset(&dst, 0, 0); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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(4, float, c, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(4, c, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( 2) * src2_stride_z; LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(4, float, bias, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(4, c, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, VEC_SIZE, c, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store 4x4 block const bool cond_y = ((get_global_id(1) + 1) * 4 >= M); const bool cond_x = ((get_global_id(0) + 1) * 4 >= N); STORE_BLOCK_BOUNDARY_AWARE(4, 4, float, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } /** This OpenCL kernel is optimized for Bifrost and tt computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) * * @note The number of rows of destination matrix must be passed at compile time using -DM * @note The number of columns of the destination matrix must be passed at compile time using -DN * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2) * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_OUTPUT_AS_3D) , uint cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int x = get_global_id(0) / H0; int y = get_global_id(1) / V0; int z = get_global_id(2); // Offset const int offset_row_a = (get_global_id(1) % V0) * 4; const int offset_row_b = (get_global_id(0) % H0) * 4; // src_addr_a = address of matrix A // src_addr_b = address of matrix B int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src1_addr_in_bytes += z * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes); __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes); src_addr_a += offset_row_a; src_addr_b += offset_row_b; // Reset accumulators float4 c0 = 0.0f; float4 c1 = 0.0f; float4 c2 = 0.0f; float4 c3 = 0.0f; int i = 0; for(; i <= (int)(K - 4); i += 4) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = vload4(0, src_addr_a); float4 b0 = vload4(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 4 * H0; c0.s0 = fma(a0.s0, b0.s0, c0.s0); c0.s1 = fma(a0.s0, b0.s1, c0.s1); c0.s2 = fma(a0.s0, b0.s2, c0.s2); c0.s3 = fma(a0.s0, b0.s3, c0.s3); c1.s0 = fma(a0.s1, b0.s0, c1.s0); c1.s1 = fma(a0.s1, b0.s1, c1.s1); c1.s2 = fma(a0.s1, b0.s2, c1.s2); c1.s3 = fma(a0.s1, b0.s3, c1.s3); c2.s0 = fma(a0.s2, b0.s0, c2.s0); c2.s1 = fma(a0.s2, b0.s1, c2.s1); c2.s2 = fma(a0.s2, b0.s2, c2.s2); c2.s3 = fma(a0.s2, b0.s3, c2.s3); c3.s0 = fma(a0.s3, b0.s0, c3.s0); c3.s1 = fma(a0.s3, b0.s1, c3.s1); c3.s2 = fma(a0.s3, b0.s2, c3.s2); c3.s3 = fma(a0.s3, b0.s3, c3.s3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload4(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 4 * H0; c0.s0 = fma(a0.s0, b0.s0, c0.s0); c0.s1 = fma(a0.s0, b0.s1, c0.s1); c0.s2 = fma(a0.s0, b0.s2, c0.s2); c0.s3 = fma(a0.s0, b0.s3, c0.s3); c1.s0 = fma(a0.s1, b0.s0, c1.s0); c1.s1 = fma(a0.s1, b0.s1, c1.s1); c1.s2 = fma(a0.s1, b0.s2, c1.s2); c1.s3 = fma(a0.s1, b0.s3, c1.s3); c2.s0 = fma(a0.s2, b0.s0, c2.s0); c2.s1 = fma(a0.s2, b0.s1, c2.s1); c2.s2 = fma(a0.s2, b0.s2, c2.s2); c2.s3 = fma(a0.s2, b0.s3, c2.s3); c3.s0 = fma(a0.s3, b0.s0, c3.s0); c3.s1 = fma(a0.s3, b0.s1, c3.s1); c3.s2 = fma(a0.s3, b0.s2, c3.s2); c3.s3 = fma(a0.s3, b0.s3, c3.s3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload4(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 4 * H0; c0.s0 = fma(a0.s0, b0.s0, c0.s0); c0.s1 = fma(a0.s0, b0.s1, c0.s1); c0.s2 = fma(a0.s0, b0.s2, c0.s2); c0.s3 = fma(a0.s0, b0.s3, c0.s3); c1.s0 = fma(a0.s1, b0.s0, c1.s0); c1.s1 = fma(a0.s1, b0.s1, c1.s1); c1.s2 = fma(a0.s1, b0.s2, c1.s2); c1.s3 = fma(a0.s1, b0.s3, c1.s3); c2.s0 = fma(a0.s2, b0.s0, c2.s0); c2.s1 = fma(a0.s2, b0.s1, c2.s1); c2.s2 = fma(a0.s2, b0.s2, c2.s2); c2.s3 = fma(a0.s2, b0.s3, c2.s3); c3.s0 = fma(a0.s3, b0.s0, c3.s0); c3.s1 = fma(a0.s3, b0.s1, c3.s1); c3.s2 = fma(a0.s3, b0.s2, c3.s2); c3.s3 = fma(a0.s3, b0.s3, c3.s3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload4(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 4 * H0; c0.s0 = fma(a0.s0, b0.s0, c0.s0); c0.s1 = fma(a0.s0, b0.s1, c0.s1); c0.s2 = fma(a0.s0, b0.s2, c0.s2); c0.s3 = fma(a0.s0, b0.s3, c0.s3); c1.s0 = fma(a0.s1, b0.s0, c1.s0); c1.s1 = fma(a0.s1, b0.s1, c1.s1); c1.s2 = fma(a0.s1, b0.s2, c1.s2); c1.s3 = fma(a0.s1, b0.s3, c1.s3); c2.s0 = fma(a0.s2, b0.s0, c2.s0); c2.s1 = fma(a0.s2, b0.s1, c2.s1); c2.s2 = fma(a0.s2, b0.s2, c2.s2); c2.s3 = fma(a0.s2, b0.s3, c2.s3); c3.s0 = fma(a0.s3, b0.s0, c3.s0); c3.s1 = fma(a0.s3, b0.s1, c3.s1); c3.s2 = fma(a0.s3, b0.s2, c3.s2); c3.s3 = fma(a0.s3, b0.s3, c3.s3); } for(; i < (int)K; ++i) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = vload4(0, src_addr_a); float4 b0 = vload4(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 4 * H0; c0.s0 = fma(a0.s0, b0.s0, c0.s0); c0.s1 = fma(a0.s0, b0.s1, c0.s1); c0.s2 = fma(a0.s0, b0.s2, c0.s2); c0.s3 = fma(a0.s0, b0.s3, c0.s3); c1.s0 = fma(a0.s1, b0.s0, c1.s0); c1.s1 = fma(a0.s1, b0.s1, c1.s1); c1.s2 = fma(a0.s1, b0.s2, c1.s2); c1.s3 = fma(a0.s1, b0.s3, c1.s3); c2.s0 = fma(a0.s2, b0.s0, c2.s0); c2.s1 = fma(a0.s2, b0.s1, c2.s1); c2.s2 = fma(a0.s2, b0.s2, c2.s2); c2.s3 = fma(a0.s2, b0.s3, c2.s3); c3.s0 = fma(a0.s3, b0.s0, c3.s0); c3.s1 = fma(a0.s3, b0.s1, c3.s1); c3.s2 = fma(a0.s3, b0.s2, c3.s2); c3.s3 = fma(a0.s3, b0.s3, c3.s3); } // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); // Compute dst address __global uchar *dst_addr = offset(&dst, 0, 0); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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(4, float, c, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(4, c, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( 2) * src2_stride_z; LOAD_BLOCK(4, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(4, float, bias, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(4, c, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(4, ACTIVATION_TYPE, float, VEC_SIZE, c, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store 4x4 block const bool cond_y = ((get_global_id(1) + 1) * 4 >= M); const bool cond_x = ((get_global_id(0) + 1) * 4 >= N); STORE_BLOCK_BOUNDARY_AWARE(4, 4, float, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) /** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) * * @note The number of rows of destination matrix must be passed at compile time using -DM * @note The number of columns of the destination matrix must be passed at compile time using -DN * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2) * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_OUTPUT_AS_3D) , uint cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int x = get_global_id(0) / H0; int y = get_global_id(1) / V0; int z = get_global_id(2); // Offset const int offset_row_a = (get_global_id(1) % V0) * 4; const int offset_row_b = (get_global_id(0) % H0) * 8; // src_addr_a = address of matrix A // src_addr_b = address of matrix B int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src1_addr_in_bytes += z * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); // Compute end row address for matrix B __global half *src_end_addr_b = src_addr_b + (src1_stride_y / sizeof(half)); src_addr_a += offset_row_a; src_addr_b += offset_row_b; // Reset accumulators half8 c0 = 0.0f; half8 c1 = 0.0f; half8 c2 = 0.0f; half8 c3 = 0.0f; for(; src_addr_b <= (src_end_addr_b - (int)(16 * H0)); src_addr_a += 8 * V0, src_addr_b += 16 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) half4 a0 = vload4(0, src_addr_a); half8 b0 = vload8(0, src_addr_b); c0 += (half8)a0.s0 * b0; c1 += (half8)a0.s1 * b0; c2 += (half8)a0.s2 * b0; c3 += (half8)a0.s3 * b0; // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a + 4 * V0); b0 = vload8(0, src_addr_b + 8 * H0); c0 += (half8)a0.s0 * b0; c1 += (half8)a0.s1 * b0; c2 += (half8)a0.s2 * b0; c3 += (half8)a0.s3 * b0; } for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 8 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) half4 a0 = vload4(0, src_addr_a); half8 b0 = vload8(0, src_addr_b); c0 += (half8)a0.s0 * b0; c1 += (half8)a0.s1 * b0; c2 += (half8)a0.s2 * b0; c3 += (half8)a0.s3 * b0; } // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); // Compute dst address __global uchar *dst_addr = offset(&dst, 0, 0); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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(4, half, c, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, half, bias, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(4, c, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( 2) * src2_stride_z; LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(4, half, bias, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(4, c, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store 4x8 block const bool cond_y = ((get_global_id(1) + 1) * 4 >= M); const bool cond_x = ((get_global_id(0) + 1) * 8 >= N); STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } /** This OpenCL kernel computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) while accumulating the result in a 32 floating point variable. * * @note The number of rows of destination matrix must be passed at compile time using -DM * @note The number of columns of the destination matrix must be passed at compile time using -DN * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2) * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_interleaved_transposed_f16_acc32(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_OUTPUT_AS_3D) , uint cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int x = get_global_id(0) / H0; int y = get_global_id(1) / V0; int z = get_global_id(2); // Offset const int offset_row_a = (get_global_id(1) % V0) * 4; const int offset_row_b = (get_global_id(0) % H0) * 8; // src_addr_a = address of matrix A // src_addr_b = address of matrix B int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src1_addr_in_bytes += z * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); // Compute end row address for matrix B __global half *src_end_addr_b = src_addr_b + (src1_stride_y / sizeof(half)); src_addr_a += offset_row_a; src_addr_b += offset_row_b; // Reset accumulators float8 c0 = 0.0f; float8 c1 = 0.0f; float8 c2 = 0.0f; float8 c3 = 0.0f; for(; src_addr_b <= (src_end_addr_b - (int)(16 * H0)); src_addr_a += 8 * V0, src_addr_b += 16 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = convert_float4(vload4(0, src_addr_a)); float8 b0 = convert_float8(vload8(0, src_addr_b)); c0 += (float8)a0.s0 * b0; c1 += (float8)a0.s1 * b0; c2 += (float8)a0.s2 * b0; c3 += (float8)a0.s3 * b0; // Load values from matrix A (interleaved) and matrix B (transposed) a0 = convert_float4(vload4(0, src_addr_a + 4 * V0)); b0 = convert_float8(vload8(0, src_addr_b + 8 * H0)); c0 += (float8)a0.s0 * b0; c1 += (float8)a0.s1 * b0; c2 += (float8)a0.s2 * b0; c3 += (float8)a0.s3 * b0; } for(; src_addr_b < src_end_addr_b; src_addr_a += 4 * V0, src_addr_b += 8 * H0) { // Load values from matrix A (interleaved) and matrix B (transposed) float4 a0 = convert_float4(vload4(0, src_addr_a)); float8 b0 = convert_float8(vload8(0, src_addr_b)); c0 += (float8)a0.s0 * b0; c1 += (float8)a0.s1 * b0; c2 += (float8)a0.s2 * b0; c3 += (float8)a0.s3 * b0; } // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); // Compute dst address __global uchar *dst_addr = offset(&dst, 0, 0); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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(4, float, c, ALPHA); #endif // defined(ALPHA) #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); float8 bias_f0 = convert_float8(bias0); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias_f, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(4, c, bias_f0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( 2) * src2_stride_z; LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); float8 bias_f0 = convert_float8(bias0); float8 bias_f1 = convert_float8(bias1); float8 bias_f2 = convert_float8(bias2); float8 bias_f3 = convert_float8(bias3); #ifndef UNIT_BETA SCALE_BLOCK(4, float, bias_f, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(4, c, bias_f); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) half8 c_h0 = convert_half8(c0); half8 c_h1 = convert_half8(c1); half8 c_h2 = convert_half8(c2); half8 c_h3 = convert_half8(c3); #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c_h, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store 4x8 block const bool cond_y = ((get_global_id(1) + 1) * 4 >= M); const bool cond_x = ((get_global_id(0) + 1) * 8 >= N); STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c_h, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } /** This OpenCL kernel optimized for Bifrost architectures computes the matrix multiplication between matrix A reshaped (src0) and matrix B reshaped (src1) * * @note The number of rows of destination matrix must be passed at compile time using -DM * @note The number of columns of the destination matrix must be passed at compile time using -DN * @note The number of rows of the *un-reshaped* matrix B (K) must be passed at compile time using -DK * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note The multiplication factor for the transposition width (H0) must be passed at compile time using -DH0 (e.g. -DH0=2) * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DV0 (e.g. -DV0=2) * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time: * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_interleaved_transposed_f16_bifrost(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_OUTPUT_AS_3D) , uint cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int x = get_global_id(0) / H0; int y = get_global_id(1) / V0; int z = get_global_id(2); // Offset const int offset_row_a = (get_global_id(1) % V0) * 4; const int offset_row_b = (get_global_id(0) % H0) * 8; // src_addr_a = address of matrix A // src_addr_b = address of matrix B int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes; int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes; #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src1_addr_in_bytes += z * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes); __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes); src_addr_a += offset_row_a; src_addr_b += offset_row_b; // Reset accumulators half8 c0 = 0.0f; half8 c1 = 0.0f; half8 c2 = 0.0f; half8 c3 = 0.0f; int i = 0; for(; i <= (int)(K - 4); i += 4) { #if V0 == 1 // Load values from matrix A (interleaved) and matrix B (transposed) half8 a0 = vload8(0, src_addr_a); half8 b0 = vload8(0, src_addr_b); src_addr_a += 8 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); // Load values from matrix B (transposed) b0 = vload8(0, src_addr_b); src_addr_b += 8 * H0; c0 = fma((half8)a0.s4, b0, c0); c1 = fma((half8)a0.s5, b0, c1); c2 = fma((half8)a0.s6, b0, c2); c3 = fma((half8)a0.s7, b0, c3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload8(0, src_addr_a); b0 = vload8(0, src_addr_b); src_addr_a += 8 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); // Load values from matrix B (transposed) b0 = vload8(0, src_addr_b); src_addr_b += 8 * H0; c0 = fma((half8)a0.s4, b0, c0); c1 = fma((half8)a0.s5, b0, c1); c2 = fma((half8)a0.s6, b0, c2); c3 = fma((half8)a0.s7, b0, c3); #else // V0 == 1 // Load values from matrix A (interleaved) and matrix B (transposed) half4 a0 = vload4(0, src_addr_a); half8 b0 = vload8(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload8(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload8(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); // Load values from matrix A (interleaved) and matrix B (transposed) a0 = vload4(0, src_addr_a); b0 = vload8(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); #endif // V0 == 1 } for(; i < (int)K; ++i) { // Load values from matrix A (interleaved) and matrix B (transposed) half4 a0 = vload4(0, src_addr_a); half8 b0 = vload8(0, src_addr_b); src_addr_a += 4 * V0; src_addr_b += 8 * H0; c0 = fma((half8)a0.s0, b0, c0); c1 = fma((half8)a0.s1, b0, c1); c2 = fma((half8)a0.s2, b0, c2); c3 = fma((half8)a0.s3, b0, c3); } // Compute destination address Image dst = CONVERT_TO_IMAGE_STRUCT(dst); // Compute dst address __global uchar *dst_addr = offset(&dst, 0, 0); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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(4, half, c, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(4, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, half, bias, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(4, c, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (get_global_id(1) * (uint)4 * src2_stride_y) + get_global_id( 2) * src2_stride_z; LOAD_BLOCK(4, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(4, half, bias, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(4, c, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(4, ACTIVATION_TYPE, half, VEC_SIZE, c, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store 4x8 block const bool cond_y = ((get_global_id(1) + 1) * 4 >= M); const bool cond_x = ((get_global_id(0) + 1) * 8 >= N); STORE_BLOCK_BOUNDARY_AWARE(4, 8, half, c, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) #endif // defined(M) && defined(N) && defined(K) && defined(H0) && defined(V0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) #if defined(N) && defined(K) && defined(M0) && defined(N0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) #if defined(DATA_TYPE) #define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, N0) /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped. * * @note This OpenCL kernel works with floating point data types (F16/F32) * @note The floating point data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0 * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16/F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements for the output tensor (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_INPUT_AS_3D) , uint src_cross_plane_pad #endif // REINTERPRET_INPUT_AS_3D #if defined(REINTERPRET_OUTPUT_AS_3D) , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int idx = get_global_id(0) * N0; // Compute starting address for matrix A and Matrix B int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); // Update address for the matrix A src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y; // Update address for the matrix B src_addr.s1 += idx * sizeof(DATA_TYPE); #if defined(REINTERPRET_INPUT_AS_3D) // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zin) is calculated dividing row by HEIGHT_GEMM3D uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zin = min(DEPTH_GEMM3D - 1, zin); // Add offset due to the cross plane paddings zin *= (src_cross_plane_pad * src0_stride_y); // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we // multiply src0_stride_z by DEPTH_GEMM3D src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; #else // defined(REINTERPRET_INPUT_AS_3D) // Add offset for batched GEMM src_addr.s0 += get_global_id(2) * src0_stride_z; #endif // defined(REINTERPRET_INPUT_AS_3D) #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src_addr.s1 += get_global_id(2) * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) int end_row_vec_a = src_addr.s0 + (K * sizeof(DATA_TYPE)); VECTOR_TYPE acc0 = 0.0f; #if M0 > 1 VECTOR_TYPE acc1 = 0.0f; #endif // M0 > 1 #if M0 > 2 VECTOR_TYPE acc2 = 0.0f; #endif // M0 > 2 #if M0 > 3 VECTOR_TYPE acc3 = 0.0f; #endif // M0 > 3 for(; src_addr.s0 <= (end_row_vec_a - 2 * (int)sizeof(DATA_TYPE)); src_addr += (int2)(2 * sizeof(DATA_TYPE), 2 * src1_stride_y)) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A LOAD_BLOCK(M0, 2, DATA_TYPE, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A VEC_DATA_TYPE(DATA_TYPE, 2) a0 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 VEC_DATA_TYPE(DATA_TYPE, 2) a1 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 VEC_DATA_TYPE(DATA_TYPE, 2) a2 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 VEC_DATA_TYPE(DATA_TYPE, 2) a3 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B VECTOR_TYPE b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); VECTOR_TYPE b1 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1 + src1_stride_y)); // Accumulate acc0 += b0 * (VECTOR_TYPE)a0.s0; acc0 += b1 * (VECTOR_TYPE)a0.s1; #if M0 > 1 acc1 += b0 * (VECTOR_TYPE)a1.s0; acc1 += b1 * (VECTOR_TYPE)a1.s1; #endif // M0 > 1 #if M0 > 2 acc2 += b0 * (VECTOR_TYPE)a2.s0; acc2 += b1 * (VECTOR_TYPE)a2.s1; #endif // M0 > 2 #if M0 > 3 acc3 += b0 * (VECTOR_TYPE)a3.s0; acc3 += b1 * (VECTOR_TYPE)a3.s1; #endif // M0 > 3 } for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(DATA_TYPE), src1_stride_y)) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); #if M0 > 1 DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #endif // M0 > 1 #if M0 > 2 DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #endif // M0 > 2 #if M0 > 3 DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #endif // M0 > 3 #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B VECTOR_TYPE b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1)); // Accumulate acc0 += b0 * (VECTOR_TYPE)a0; #if M0 > 1 acc1 += b0 * (VECTOR_TYPE)a1; #endif // M0 > 1 #if M0 > 2 acc2 += b0 * (VECTOR_TYPE)a2; #endif // M0 > 2 #if M0 > 3 acc3 += b0 * (VECTOR_TYPE)a3; #endif // M0 > 3 } int z = get_global_id(2); // Compute dst address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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, acc, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)); LOAD_BLOCK(1, N0, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, DATA_TYPE, bias, BETA); #endif // UNIT_BIAS // c = c + bias[broadcasted] ADD_BLOCK_BROADCAST(M0, acc, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src2_stride_y) + z * src2_stride_z; LOAD_BLOCK(M0, N0, DATA_TYPE, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(M0, DATA_TYPE, bias, BETA); #endif // UNIT_BIAS // c = c + bias ADD_BLOCK(M0, acc, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, acc, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store output block const bool cond_y = get_global_id(1) == 0; const bool cond_x = ((get_global_id(0) + 1) * N0 >= N); STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } #endif // defined(DATA_TYPE) /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped * * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units. * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0. * @note This kernel processed a fixed number of elements along x: -DN0=4. * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_floating_point_f32_bifrost(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_INPUT_AS_3D) , uint src_cross_plane_pad #endif // REINTERPRET_INPUT_AS_3D #if defined(REINTERPRET_OUTPUT_AS_3D) , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int idx = get_global_id(0) * N0; // Compute starting address for matrix A and matrix B int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); // Update address for matrix A src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y; // Update address for matrix B src_addr.s1 += idx * sizeof(float); #if defined(REINTERPRET_INPUT_AS_3D) // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zin) is calculated dividing row by HEIGHT_GEMM3D uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zin = min(DEPTH_GEMM3D - 1, zin); // Add offset due to the cross plane paddings zin *= (src_cross_plane_pad * src0_stride_y); // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we // multiply src0_stride_z by DEPTH_GEMM3D src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; #else // defined(REINTERPRET_INPUT_AS_3D) // Add offset for batched GEMM src_addr.s0 += get_global_id(2) * src0_stride_z; #endif // defined(REINTERPRET_INPUT_AS_3D) #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src_addr.s1 += get_global_id(2) * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) // Initialize accumulators float4 acc0 = 0.0f; #if M0 > 1 float4 acc1 = 0.0f; #endif // M0 > 1 #if M0 > 2 float4 acc2 = 0.0f; #endif // M0 > 2 #if M0 > 3 float4 acc3 = 0.0f; #endif // M0 > 3 // A and B src indices get incremented at the same time. int i = 0; for(; i <= ((int)K - 4); i += 4) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A and matrix B LOAD_BLOCK(M0, 4, float, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A and matrix B float4 a0 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 float4 a1 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 float4 a2 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 float4 a3 = vload4(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0.s0, b0.s0, acc0.s0); acc0.s1 = fma(a0.s0, b0.s1, acc0.s1); acc0.s2 = fma(a0.s0, b0.s2, acc0.s2); acc0.s3 = fma(a0.s0, b0.s3, acc0.s3); #if M0 > 1 acc1.s0 = fma(a1.s0, b0.s0, acc1.s0); acc1.s1 = fma(a1.s0, b0.s1, acc1.s1); acc1.s2 = fma(a1.s0, b0.s2, acc1.s2); acc1.s3 = fma(a1.s0, b0.s3, acc1.s3); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2.s0, b0.s0, acc2.s0); acc2.s1 = fma(a2.s0, b0.s1, acc2.s1); acc2.s2 = fma(a2.s0, b0.s2, acc2.s2); acc2.s3 = fma(a2.s0, b0.s3, acc2.s3); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3.s0, b0.s0, acc3.s0); acc3.s1 = fma(a3.s0, b0.s1, acc3.s1); acc3.s2 = fma(a3.s0, b0.s2, acc3.s2); acc3.s3 = fma(a3.s0, b0.s3, acc3.s3); #endif // M0 > 3 // Load values from matrix A and matrix B b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0.s1, b0.s0, acc0.s0); acc0.s1 = fma(a0.s1, b0.s1, acc0.s1); acc0.s2 = fma(a0.s1, b0.s2, acc0.s2); acc0.s3 = fma(a0.s1, b0.s3, acc0.s3); #if M0 > 1 acc1.s0 = fma(a1.s1, b0.s0, acc1.s0); acc1.s1 = fma(a1.s1, b0.s1, acc1.s1); acc1.s2 = fma(a1.s1, b0.s2, acc1.s2); acc1.s3 = fma(a1.s1, b0.s3, acc1.s3); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2.s1, b0.s0, acc2.s0); acc2.s1 = fma(a2.s1, b0.s1, acc2.s1); acc2.s2 = fma(a2.s1, b0.s2, acc2.s2); acc2.s3 = fma(a2.s1, b0.s3, acc2.s3); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3.s1, b0.s0, acc3.s0); acc3.s1 = fma(a3.s1, b0.s1, acc3.s1); acc3.s2 = fma(a3.s1, b0.s2, acc3.s2); acc3.s3 = fma(a3.s1, b0.s3, acc3.s3); #endif // M0 > 3 // Load values from matrix A and matrix B b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0.s2, b0.s0, acc0.s0); acc0.s1 = fma(a0.s2, b0.s1, acc0.s1); acc0.s2 = fma(a0.s2, b0.s2, acc0.s2); acc0.s3 = fma(a0.s2, b0.s3, acc0.s3); #if M0 > 1 acc1.s0 = fma(a1.s2, b0.s0, acc1.s0); acc1.s1 = fma(a1.s2, b0.s1, acc1.s1); acc1.s2 = fma(a1.s2, b0.s2, acc1.s2); acc1.s3 = fma(a1.s2, b0.s3, acc1.s3); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2.s2, b0.s0, acc2.s0); acc2.s1 = fma(a2.s2, b0.s1, acc2.s1); acc2.s2 = fma(a2.s2, b0.s2, acc2.s2); acc2.s3 = fma(a2.s2, b0.s3, acc2.s3); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3.s2, b0.s0, acc3.s0); acc3.s1 = fma(a3.s2, b0.s1, acc3.s1); acc3.s2 = fma(a3.s2, b0.s2, acc3.s2); acc3.s3 = fma(a3.s2, b0.s3, acc3.s3); #endif // M0 > 3 // Load values from matrix A and matrix B b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0.s3, b0.s0, acc0.s0); acc0.s1 = fma(a0.s3, b0.s1, acc0.s1); acc0.s2 = fma(a0.s3, b0.s2, acc0.s2); acc0.s3 = fma(a0.s3, b0.s3, acc0.s3); #if M0 > 1 acc1.s0 = fma(a1.s3, b0.s0, acc1.s0); acc1.s1 = fma(a1.s3, b0.s1, acc1.s1); acc1.s2 = fma(a1.s3, b0.s2, acc1.s2); acc1.s3 = fma(a1.s3, b0.s3, acc1.s3); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2.s3, b0.s0, acc2.s0); acc2.s1 = fma(a2.s3, b0.s1, acc2.s1); acc2.s2 = fma(a2.s3, b0.s2, acc2.s2); acc2.s3 = fma(a2.s3, b0.s3, acc2.s3); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3.s3, b0.s0, acc3.s0); acc3.s1 = fma(a3.s3, b0.s1, acc3.s1); acc3.s2 = fma(a3.s3, b0.s2, acc3.s2); acc3.s3 = fma(a3.s3, b0.s3, acc3.s3); #endif // M0 > 3 src_addr.s0 += 4 * sizeof(float); } for(; i < (int)K; ++i) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); #if M0 > 1 float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #endif // M0 > 1 #if M0 > 2 float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #endif // M0 > 2 #if M0 > 3 float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #endif // M0 > 3 #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0, b0.s0, acc0.s0); acc0.s1 = fma(a0, b0.s1, acc0.s1); acc0.s2 = fma(a0, b0.s2, acc0.s2); acc0.s3 = fma(a0, b0.s3, acc0.s3); #if M0 > 1 acc1.s0 = fma(a1, b0.s0, acc1.s0); acc1.s1 = fma(a1, b0.s1, acc1.s1); acc1.s2 = fma(a1, b0.s2, acc1.s2); acc1.s3 = fma(a1, b0.s3, acc1.s3); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2, b0.s0, acc2.s0); acc2.s1 = fma(a2, b0.s1, acc2.s1); acc2.s2 = fma(a2, b0.s2, acc2.s2); acc2.s3 = fma(a2, b0.s3, acc2.s3); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3, b0.s0, acc3.s0); acc3.s1 = fma(a3, b0.s1, acc3.s1); acc3.s2 = fma(a3, b0.s2, acc3.s2); acc3.s3 = fma(a3, b0.s3, acc3.s3); #endif // M0 > 3 src_addr.s0 += sizeof(float); } int z = get_global_id(2); // Compute dst address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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, float, acc, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)); LOAD_BLOCK(1, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias[broadcasted] ADD_BLOCK_BROADCAST(M0, acc, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)4 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src2_stride_y) + z * src2_stride_z; LOAD_BLOCK(M0, 4, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(M0, float, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias ADD_BLOCK(M0, acc, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, float, VEC_SIZE, acc, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store the output block const bool cond_y = get_global_id(1) == 0; const bool cond_x = ((get_global_id(0) + 1) * 4 >= N); STORE_BLOCK_BOUNDARY_AWARE(M0, 4, float, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not been reshaped * * @note This OpenCL kernel works with the 32-bit floating point data type (float) and uses the fma units. * This OpenCL kernel is optimized for Bifrost when the number of matrix B columns is less or equal to 1000. * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0. * @note This kernel processed a fixed number of elements along x: -DN0=2. * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_floating_point_f32_bifrost_1000(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_INPUT_AS_3D) , uint src_cross_plane_pad #endif // REINTERPRET_INPUT_AS_3D #if defined(REINTERPRET_OUTPUT_AS_3D) , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { // Requires 2 N0, C vect2, A vect4, B (2 vload2) // to fix for M0 > 1 int idx = get_global_id(0) * N0; // Compute starting address for matrix A and Matrix B int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); // Update address for the matrix A src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y; // Update address for the matrix B src_addr.s1 += idx * sizeof(float); #if defined(REINTERPRET_INPUT_AS_3D) // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zin) is calculated dividing row by HEIGHT_GEMM3D uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zin = min(DEPTH_GEMM3D - 1, zin); // Add offset due to the cross plane paddings zin *= (src_cross_plane_pad * src0_stride_y); // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we // multiply src0_stride_z by DEPTH_GEMM3D src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; #else // defined(REINTERPRET_INPUT_AS_3D) // Add offset for batched GEMM src_addr.s0 += get_global_id(2) * src0_stride_z; #endif // defined(REINTERPRET_INPUT_AS_3D) #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src_addr.s1 += get_global_id(2) * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) // Initialize accumulators float2 acc0 = 0.0f; #if M0 > 1 float2 acc1 = 0.0f; #endif // M0 > 1 #if M0 > 2 float2 acc2 = 0.0f; #endif // M0 > 2 #if M0 > 3 float2 acc3 = 0.0f; #endif // M0 > 3 // A and B src indices get incremented at the same time. int i = 0; for(; i <= ((int)K - 8); i += 8) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + zin.s0)); #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float8 a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0)); #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b1 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b2 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b3 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b4 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b5 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b6 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; float2 b7 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0.s0, b0.s0, acc0.s0); acc0.s0 = fma(a0.s1, b1.s0, acc0.s0); acc0.s0 = fma(a0.s2, b2.s0, acc0.s0); acc0.s0 = fma(a0.s3, b3.s0, acc0.s0); acc0.s0 = fma(a0.s4, b4.s0, acc0.s0); acc0.s0 = fma(a0.s5, b5.s0, acc0.s0); acc0.s0 = fma(a0.s6, b6.s0, acc0.s0); acc0.s0 = fma(a0.s7, b7.s0, acc0.s0); acc0.s1 = fma(a0.s0, b0.s1, acc0.s1); acc0.s1 = fma(a0.s1, b1.s1, acc0.s1); acc0.s1 = fma(a0.s2, b2.s1, acc0.s1); acc0.s1 = fma(a0.s3, b3.s1, acc0.s1); acc0.s1 = fma(a0.s4, b4.s1, acc0.s1); acc0.s1 = fma(a0.s5, b5.s1, acc0.s1); acc0.s1 = fma(a0.s6, b6.s1, acc0.s1); acc0.s1 = fma(a0.s7, b7.s1, acc0.s1); #if M0 > 1 #if defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #else // defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // defined(REINTERPRET_INPUT_AS_3D) acc1.s0 = fma(a0.s0, b0.s0, acc1.s0); acc1.s0 = fma(a0.s1, b1.s0, acc1.s0); acc1.s0 = fma(a0.s2, b2.s0, acc1.s0); acc1.s0 = fma(a0.s3, b3.s0, acc1.s0); acc1.s0 = fma(a0.s4, b4.s0, acc1.s0); acc1.s0 = fma(a0.s5, b5.s0, acc1.s0); acc1.s0 = fma(a0.s6, b6.s0, acc1.s0); acc1.s0 = fma(a0.s7, b7.s0, acc1.s0); acc1.s1 = fma(a0.s0, b0.s1, acc1.s1); acc1.s1 = fma(a0.s1, b1.s1, acc1.s1); acc1.s1 = fma(a0.s2, b2.s1, acc1.s1); acc1.s1 = fma(a0.s3, b3.s1, acc1.s1); acc1.s1 = fma(a0.s4, b4.s1, acc1.s1); acc1.s1 = fma(a0.s5, b5.s1, acc1.s1); acc1.s1 = fma(a0.s6, b6.s1, acc1.s1); acc1.s1 = fma(a0.s7, b7.s1, acc1.s1); #endif // M0 > 1 #if M0 > 2 #if defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #else // defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // defined(REINTERPRET_INPUT_AS_3D) acc2.s0 = fma(a0.s0, b0.s0, acc2.s0); acc2.s0 = fma(a0.s1, b1.s0, acc2.s0); acc2.s0 = fma(a0.s2, b2.s0, acc2.s0); acc2.s0 = fma(a0.s3, b3.s0, acc2.s0); acc2.s0 = fma(a0.s4, b4.s0, acc2.s0); acc2.s0 = fma(a0.s5, b5.s0, acc2.s0); acc2.s0 = fma(a0.s6, b6.s0, acc2.s0); acc2.s0 = fma(a0.s7, b7.s0, acc2.s0); acc2.s1 = fma(a0.s0, b0.s1, acc2.s1); acc2.s1 = fma(a0.s1, b1.s1, acc2.s1); acc2.s1 = fma(a0.s2, b2.s1, acc2.s1); acc2.s1 = fma(a0.s3, b3.s1, acc2.s1); acc2.s1 = fma(a0.s4, b4.s1, acc2.s1); acc2.s1 = fma(a0.s5, b5.s1, acc2.s1); acc2.s1 = fma(a0.s6, b6.s1, acc2.s1); acc2.s1 = fma(a0.s7, b7.s1, acc2.s1); #endif // M0 > 2 #if M0 > 3 #if defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #else // defined(REINTERPRET_INPUT_AS_3D) a0 = vload8(0, (__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // defined(REINTERPRET_INPUT_AS_3D) acc3.s0 = fma(a0.s0, b0.s0, acc3.s0); acc3.s0 = fma(a0.s1, b1.s0, acc3.s0); acc3.s0 = fma(a0.s2, b2.s0, acc3.s0); acc3.s0 = fma(a0.s3, b3.s0, acc3.s0); acc3.s0 = fma(a0.s4, b4.s0, acc3.s0); acc3.s0 = fma(a0.s5, b5.s0, acc3.s0); acc3.s0 = fma(a0.s6, b6.s0, acc3.s0); acc3.s0 = fma(a0.s7, b7.s0, acc3.s0); acc3.s1 = fma(a0.s0, b0.s1, acc3.s1); acc3.s1 = fma(a0.s1, b1.s1, acc3.s1); acc3.s1 = fma(a0.s2, b2.s1, acc3.s1); acc3.s1 = fma(a0.s3, b3.s1, acc3.s1); acc3.s1 = fma(a0.s4, b4.s1, acc3.s1); acc3.s1 = fma(a0.s5, b5.s1, acc3.s1); acc3.s1 = fma(a0.s6, b6.s1, acc3.s1); acc3.s1 = fma(a0.s7, b7.s1, acc3.s1); #endif // M0 > 3 src_addr.s0 += sizeof(float) * 8; } // float size increment for(; i < (int)K; ++i) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); #if M0 > 1 float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #endif // M0 > 1 #if M0 > 2 float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #endif // M0 > 2 #if M0 > 3 float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #endif // M0 > 3 #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A float a0 = *((__global float *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 float a1 = *((__global float *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 float a2 = *((__global float *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 float a3 = *((__global float *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B float2 b0 = vload2(0, (__global float *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Multiply and accumulate acc0.s0 = fma(a0, b0.s0, acc0.s0); acc0.s1 = fma(a0, b0.s1, acc0.s1); #if M0 > 1 acc1.s0 = fma(a1, b0.s0, acc1.s0); acc1.s1 = fma(a1, b0.s1, acc1.s1); #endif // M0 > 1 #if M0 > 2 acc2.s0 = fma(a2, b0.s0, acc2.s0); acc2.s1 = fma(a2, b0.s1, acc2.s1); #endif // M0 > 2 #if M0 > 3 acc3.s0 = fma(a3, b0.s0, acc3.s0); acc3.s1 = fma(a3, b0.s1, acc3.s1); #endif // M0 > 3 src_addr.s0 += sizeof(float); } int z = get_global_id(2); // Compute dst address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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, float, acc, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)); LOAD_BLOCK(1, 2, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias[broadcasted] ADD_BLOCK_BROADCAST(M0, acc, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)2 * sizeof(float)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src2_stride_y) + z * src2_stride_z; LOAD_BLOCK(M0, 2, float, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(M0, float, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias ADD_BLOCK(M0, acc, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, float, VEC_SIZE, acc, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store the output block const bool cond_y = get_global_id(1) == 0; const bool cond_x = ((get_global_id(0) + 1) * 2 >= N); STORE_BLOCK_BOUNDARY_AWARE(M0, 2, float, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped * * @note This OpenCL kernel works with the 16-bit floating point data type (half) and accumulating the result in a 32 floating point variable. * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0. * @note This kernel processed a fixed number of elements along x: -DN0=8. * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_floating_point_f16_bifrost_acc32(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_INPUT_AS_3D) , uint src_cross_plane_pad #endif // REINTERPRET_INPUT_AS_3D #if defined(REINTERPRET_OUTPUT_AS_3D) , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int idx = get_global_id(0) * N0; // Compute starting address for matrix A and Matrix B int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); // Update address for the matrix A src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y; // Update address for the matrix B src_addr.s1 += idx * sizeof(half); #if defined(REINTERPRET_INPUT_AS_3D) // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zin) is calculated dividing row by HEIGHT_GEMM3D uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zin = min(DEPTH_GEMM3D - 1, zin); // Add offset due to the cross plane paddings zin *= (src_cross_plane_pad * src0_stride_y); // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we // multiply src0_stride_z by DEPTH_GEMM3D src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; #else // defined(REINTERPRET_INPUT_AS_3D) // Add offset for batched GEMM src_addr.s0 += get_global_id(2) * src0_stride_z; #endif // defined(REINTERPRET_INPUT_AS_3D) #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src_addr.s1 += get_global_id(2) * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) float8 acc0 = 0.0h; #if M0 > 1 float8 acc1 = 0.0h; #endif // M0 > 1 #if M0 > 2 float8 acc2 = 0.0h; #endif // M0 > 2 #if M0 > 3 float8 acc3 = 0.0h; #endif // M0 > 3 int i = 0; for(; i <= ((int)K - 4); i += 4) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A LOAD_BLOCK(M0, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); src_addr.s1 += src1_stride_y; // Accumulate acc0 = fma(b0, (float8)a0.s0, acc0); #if M0 > 1 acc1 = fma(b0, (float8)a1.s0, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (float8)a2.s0, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (float8)a3.s0, acc3); #endif // M0 > 3 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (float8)a0.s1, acc0); #if M0 > 1 acc1 = fma(b0, (float8)a1.s1, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (float8)a2.s1, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (float8)a3.s1, acc3); #endif // M0 > 3 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (float8)a0.s2, acc0); #if M0 > 1 acc1 = fma(b0, (float8)a1.s2, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (float8)a2.s2, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (float8)a3.s2, acc3); #endif // M0 > 3 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (float8)a0.s3, acc0); #if M0 > 1 acc1 = fma(b0, (float8)a1.s3, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (float8)a2.s3, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (float8)a3.s3, acc3); #endif // M0 > 3 src_addr.s0 += 4 * sizeof(half); } for(; i < (int)K; ++i) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); #if M0 > 1 half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #endif // M0 > 1 #if M0 > 2 half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #endif // M0 > 2 #if M0 > 3 half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #endif // M0 > 3 #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B float8 b0 = convert_float8(vload8(0, (__global half *)(src1_ptr + src_addr.s1))); src_addr += (int2)(sizeof(half), src1_stride_y); // Accumulate acc0 = fma(b0, (float8)a0, acc0); // b0 * (half8)a0; #if M0 > 1 acc1 = fma(b0, (float8)a1, acc1); // b0 * (half8)a1; #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (float8)a2, acc2); // b0 * (half8)a2; #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (float8)a3, acc3); // b0 * (half8)a3; #endif // M0 > 3 } int z = get_global_id(2); // Compute dst address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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, float, acc, ALPHA); #endif // defined(ALPHA) #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); float8 bias_f0 = convert_float8(bias0); #ifndef UNIT_BETA SCALE_BLOCK(1, float, bias_f, BETA); #endif // UNIT_BIAS // acc = acc + bias[broadcasted] ADD_BLOCK_BROADCAST(M0, acc, bias_f0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src2_stride_y) + z * src2_stride_z; LOAD_BLOCK(M0, 8, half, bias, src2_addr, 0, src2_stride_y, zero); float8 bias_f0 = convert_float8(bias0); #if M0 > 1 float8 bias_f1 = convert_float8(bias1); #endif // M0 > 1 #if M0 > 2 float8 bias_f2 = convert_float8(bias2); #endif // M0 > 2 #if M0 > 3 float8 bias_f3 = convert_float8(bias3); #endif // M0 > 3 #ifndef UNIT_BETA SCALE_BLOCK(M0, float, bias_f, BETA); #endif // UNIT_BIAS // acc = acc + bias ADD_BLOCK(M0, acc, bias_f); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) half8 acc_h0 = convert_half8(acc0); #if M0 > 1 half8 acc_h1 = convert_half8(acc1); #endif // M0 > 1 #if M0 > 2 half8 acc_h2 = convert_half8(acc2); #endif // M0 > 2 #if M0 > 3 half8 acc_h3 = convert_half8(acc3); #endif // M0 > 3 #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, half, VEC_SIZE, acc_h, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store the output block const bool cond_y = get_global_id(1) == 0; const bool cond_x = ((get_global_id(0) + 1) * 8 >= N); STORE_BLOCK_BOUNDARY_AWARE(M0, 8, half, acc_h, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } /** This OpenCL kernel computes the matrix by matrix multiplication between the matrix A (src0) and matrix B (src1) in case both matrices have not beed reshaped * * @note This OpenCL kernel works with the 16-bit floating point data type (half) and uses the fma units. * @note The number of elements processed along the x and y directions must be passed at compile time using -DN0 and -DM0. * @note This kernel processed a fixed number of elements along x: -DN0=8. * @note The number of columns of matrix A and the number of columns of the matrix B need to be passed at compile time using -DK and -DN * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1) * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1) * @note The optional alpha's value need to be passed at compile time using -DALPHA * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (e.g. -DMATRIX_B_DEPTH=16) * This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (e.g. a = [K, M, 16, Batches], b = [N, K, 16]) * * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively. * The activation function is performed after the bias addition * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time: * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor. * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns matrix A NOT reshaped * * @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16 * @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src1_ptr Pointer to the source matrix. Supported data types: same as @p src0_ptr * @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes) * @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes) * @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix * @param[in] src2_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr * @param[in] src2_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes) * @param[in] src2_step_x (Optional) src2_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src2_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes) * @param[in] src2_step_y (Optional) src2_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src2_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src0_ptr * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes) * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes) * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix * @param[in] src0_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes) * @param[in] src2_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] src_cross_plane_pad (Optional) Bottom paddings in unit of elements for the input tensor (only if defined REINTERPRET_INPUT_AS_3D) * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D) */ __kernel void gemm_mm_floating_point_f16_bifrost(IMAGE_DECLARATION(src0), IMAGE_DECLARATION(src1), #if defined(BETA) IMAGE_DECLARATION(src2), #endif // defined(BETA) IMAGE_DECLARATION(dst), uint src0_stride_z, uint src1_stride_z, #if defined(BETA) uint src2_stride_z, #endif //defined(BETA) uint dst_stride_z #if defined(REINTERPRET_INPUT_AS_3D) , uint src_cross_plane_pad #endif // REINTERPRET_INPUT_AS_3D #if defined(REINTERPRET_OUTPUT_AS_3D) , uint dst_cross_plane_pad #endif // REINTERPRET_OUTPUT_AS_3D ) { int idx = get_global_id(0) * N0; // Compute starting address for matrix A and Matrix B int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes)); // Update address for the matrix A src_addr.s0 += COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src0_stride_y; // Update address for the matrix B src_addr.s1 += idx * sizeof(half); #if defined(REINTERPRET_INPUT_AS_3D) // Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zin) is calculated dividing row by HEIGHT_GEMM3D uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zin = min(DEPTH_GEMM3D - 1, zin); // Add offset due to the cross plane paddings zin *= (src_cross_plane_pad * src0_stride_y); // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we // multiply src0_stride_z by DEPTH_GEMM3D src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D; #else // defined(REINTERPRET_INPUT_AS_3D) // Add offset for batched GEMM src_addr.s0 += get_global_id(2) * src0_stride_z; #endif // defined(REINTERPRET_INPUT_AS_3D) #if defined(MATRIX_B_DEPTH) // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3 src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z; #else // defined(MATRIX_B_DEPTH) src_addr.s1 += get_global_id(2) * src1_stride_z; #endif // defined(MATRIX_B_DEPTH) half8 acc0 = 0.0h; #if M0 > 1 half8 acc1 = 0.0h; #endif // M0 > 1 #if M0 > 2 half8 acc2 = 0.0h; #endif // M0 > 2 #if M0 > 3 half8 acc3 = 0.0h; #endif // M0 > 3 int i = 0; for(; i <= ((int)K - 4); i += 4) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A LOAD_BLOCK(M0, 4, half, a, src0_ptr, src_addr.s0, src0_stride_y, zin.s); #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 half4 a1 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 half4 a2 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 half4 a3 = vload4(0, (__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; // Accumulate acc0 = fma(b0, (half8)a0.s0, acc0); #if M0 > 1 acc1 = fma(b0, (half8)a1.s0, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (half8)a2.s0, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (half8)a3.s0, acc3); #endif // M0 > 3 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (half8)a0.s1, acc0); #if M0 > 1 acc1 = fma(b0, (half8)a1.s1, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (half8)a2.s1, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (half8)a3.s1, acc3); #endif // M0 > 3 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (half8)a0.s2, acc0); #if M0 > 1 acc1 = fma(b0, (half8)a1.s2, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (half8)a2.s2, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (half8)a3.s2, acc3); #endif // M0 > 3 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); src_addr.s1 += src1_stride_y; acc0 = fma(b0, (half8)a0.s3, acc0); #if M0 > 1 acc1 = fma(b0, (half8)a1.s3, acc1); #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (half8)a2.s3, acc2); #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (half8)a3.s3, acc3); #endif // M0 > 3 src_addr.s0 += 4 * sizeof(half); } for(; i < (int)K; ++i) { #if defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0)); #if M0 > 1 half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1)); #endif // M0 > 1 #if M0 > 2 half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2)); #endif // M0 > 2 #if M0 > 3 half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3)); #endif // M0 > 3 #else // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix A half a0 = *((__global half *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y)); #if M0 > 1 half a1 = *((__global half *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y)); #endif // M0 > 1 #if M0 > 2 half a2 = *((__global half *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y)); #endif // M0 > 2 #if M0 > 3 half a3 = *((__global half *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y)); #endif // M0 > 3 #endif // defined(REINTERPRET_INPUT_AS_3D) // Load values from matrix B half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1)); src_addr += (int2)(sizeof(half), src1_stride_y); // Accumulate acc0 = fma(b0, (half8)a0, acc0); // b0 * (half8)a0; #if M0 > 1 acc1 = fma(b0, (half8)a1, acc1); // b0 * (half8)a1; #endif // M0 > 1 #if M0 > 2 acc2 = fma(b0, (half8)a2, acc2); // b0 * (half8)a2; #endif // M0 > 2 #if M0 > 3 acc3 = fma(b0, (half8)a3, acc3); // b0 * (half8)a3; #endif // M0 > 3 } int z = get_global_id(2); // Compute dst address __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * dst_stride_y); uint4 zout = 0; #if defined(REINTERPRET_OUTPUT_AS_3D) // Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension // in order to take into account the presence of possible cross plane paddings // // | | // | plane0 | // | | // |__________________| // |******************| // | cross_plane_pad | // |******************| // | | // | plane1 | // | | // |__________________| // The plane (zout) is calculated dividing row by HEIGHT_GEMM3D zout = ((uint4)(0, 1, 2, 3) + (uint4)(COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0))) / (uint4)HEIGHT_GEMM3D; zout = min(DEPTH_GEMM3D - 1, zout); // Add offset due to the cross plane paddings zout *= (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, half, acc, ALPHA); #endif // defined(ALPHA) // Add beta*bias #if defined(BETA) REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0); #if defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)); LOAD_BLOCK(1, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(1, half, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias[broadcasted] ADD_BLOCK_BROADCAST(M0, acc, bias0); #else // defined(BROADCAST_BIAS) __global uchar *src2_addr = src2_ptr + src2_offset_first_element_in_bytes + (get_global_id(0) * (uint)8 * sizeof(half)) + (COMPUTE_M0_START_ROW(get_global_id(1), M0, PARTIAL_STORE_M0) * src2_stride_y) + z * src2_stride_z; LOAD_BLOCK(M0, 8, half, bias, src2_addr, 0, src2_stride_y, zero); #ifndef UNIT_BETA SCALE_BLOCK(M0, half, bias, BETA); #endif // UNIT_BIAS // acc = acc + bias ADD_BLOCK(M0, acc, bias); #endif // defined(BROADCAST_BIAS) #endif // defined(BETA) #if defined(ACTIVATION_TYPE) ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, half, VEC_SIZE, acc, A_VAL, B_VAL); #endif // defined(ACTIVATION_TYPE) // Store the output block const bool cond_y = get_global_id(1) == 0; const bool cond_x = ((get_global_id(0) + 1) * 8 >= N); STORE_BLOCK_BOUNDARY_AWARE(M0, 8, half, acc, dst_addr, dst_stride_y, zout.s, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x); } #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) #endif // defined(N) && defined(K) && defined(M0) && defined(N0) && defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0)