Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2021 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #ifndef ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_KERNEL_H |
| 25 | #define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_KERNEL_H |
| 26 | |
| 27 | #include "src/core/common/Macros.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 28 | #include "src/gpu/cl/ClCompileContext.h" |
| 29 | #include "src/gpu/cl/IClKernel.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace opencl |
| 34 | { |
| 35 | namespace kernels |
| 36 | { |
| 37 | /** OpenCL kernel to multiply two input matrices "A" and "B" and add a martix "C" if provided. All elements of the output matrix will be multiplied by alpha. In case matrix C is passed, it will be added to the previous result. |
| 38 | * For the matrix C, the broadcast addition is supported if the flag "broadcast_bias" is set in the GEMMReshapeInfo object |
| 39 | * |
| 40 | * @note If the input tensors @p src0 and @p src1 have been reshaped respectively with @ref ClGemmReshapeLhsMatrixKernel" and @ref ClGemmReshapeRhsMatrixKernel, |
| 41 | * the flag @p is_interleaved_transposed must be set to true |
| 42 | * |
| 43 | * @attention @p src1 tensor must have at least 2 dimensions (matrix) |
| 44 | */ |
| 45 | class ClGemmMatrixMultiplyKernel : public IClKernel |
| 46 | { |
| 47 | public: |
Giorgio Arena | 4a95bba | 2021-06-28 11:00:27 +0100 | [diff] [blame] | 48 | ClGemmMatrixMultiplyKernel(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 49 | ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClGemmMatrixMultiplyKernel); |
| 50 | /** Initialise the kernel's input, output and alpha |
| 51 | * |
| 52 | * @param[in] compile_context The compile context to be used. |
| 53 | * @param[in] src0 Input tensor containing the Matrix A. Data types supported: F16/F32 |
| 54 | * @param[in] src1 Input tensor containing the Matrix B. Data type supported: same as @p src0 |
| 55 | * @param[in] src2 Input tensor containing the Matrix C (bias). Can be nullptr. Data type supported: same as @p src0 |
| 56 | * @param[out] dst Output tensor to store the result of matrix multiplication. Data type supported: same as @p src0 |
| 57 | * @param[in] alpha Weight of the matrix product |
| 58 | * @param[in] beta (Optional) Weight of vector C. Default value is 0. Only beta = 1 is currently supported. |
| 59 | * @param[in] is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref ClGemmReshapeLhsMatrixKernel and @ref ClGemmReshapeRhsMatrixKernel |
| 60 | * @param[in] reshape_info (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped |
| 61 | * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy |
| 62 | * @param[in] activation_info (Optional) Activation to apply after the matrix multiplication |
| 63 | * |
| 64 | */ |
| 65 | void configure(const ClCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta = 0.f, |
| 66 | bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(), bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo()); |
| 67 | /** Static function to check if given info will lead to a valid configuration |
| 68 | * |
| 69 | * Similar to @ref ClGemmMatrixMultiplyKernel::configure() |
| 70 | * |
| 71 | * @return a status |
| 72 | */ |
| 73 | static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, |
| 74 | bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision = false, const ActivationLayerInfo &activation_info = ActivationLayerInfo()); |
| 75 | |
| 76 | // Inherited methods overridden: |
| 77 | void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; |
| 78 | |
| 79 | public: |
| 80 | bool _slide_matrix_b{ true }; |
| 81 | bool _reinterpret_input_as_3d{ false }; |
| 82 | bool _reinterpret_output_as_3d{ false }; |
| 83 | bool _add_bias{ false }; |
| 84 | }; |
| 85 | } // namespace kernels |
| 86 | } // namespace opencl |
| 87 | } // namespace arm_compute |
| 88 | #endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_KERNEL_H */ |