Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018-2021 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_RESHAPED_KERNEL_H |
| 25 | #define ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_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 | #include "arm_compute/core/KernelDescriptors.h" |
| 32 | |
| 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace opencl |
| 36 | { |
| 37 | namespace kernels |
| 38 | { |
| 39 | /** OpenCL kernel to multiply matrices when both the input matrices LHS (src0) and RHS (src1) have been reshaped |
| 40 | * |
| 41 | * @note The input matrices @p src0 and @p src1 must be reshaped through: |
| 42 | * - @ref ClGemmReshapeLhsMatrixKernel |
| 43 | * - @ref ClGemmReshapeRhsMatrixKernel |
| 44 | */ |
| 45 | class ClGemmMatrixMultiplyReshapedKernel : public IClKernel |
| 46 | { |
| 47 | public: |
Giorgio Arena | 4a95bba | 2021-06-28 11:00:27 +0100 | [diff] [blame] | 48 | ClGemmMatrixMultiplyReshapedKernel(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 49 | ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClGemmMatrixMultiplyReshapedKernel); |
| 50 | /** Initialise the kernel's input and output. |
| 51 | * |
| 52 | * @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag. |
| 53 | * Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the |
| 54 | * multiplications. i.e. float c = (half)a * (half)b |
| 55 | * |
| 56 | * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function. |
| 57 | * Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer, |
| 58 | * the following conditions are required: |
| 59 | * -# rhs_info.n0 can only be 4, 8 and 16 |
| 60 | * -# rhs_info.k0 can only be 4, 8 and 16 |
| 61 | * -# Data type can only be F32 |
| 62 | * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension |
| 63 | * -# The stride Y for the src1 should satisfy the OpenCL pitch alignment requirement |
| 64 | * -# src1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) |
| 65 | * -# src1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT |
| 66 | * |
| 67 | * @param[in] compile_context The compile context to be used. |
| 68 | * @param[in] src0 Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32 (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4 |
| 69 | * @param[in] src1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p src0. The number of dimensions for the RHS matrix must be less or equal than 3 |
| 70 | * @param[in] src2 Input tensor containing the bias matrix. Data type supported: same as @p src0. |
| 71 | * @param[out] dst dst tensor to store the result of matrix multiplication. Data type supported: same as @p src0 |
| 72 | * @param[in] alpha Weight of the matrix product |
| 73 | * @param[in] beta Weight of the matrix bias |
| 74 | * @param[in] lhs_info LHS matrix information used for reshaping the src0 tensor. Only the following values are supported: |
| 75 | * lhs_info.m0: 2,3,4,5,6,7,8 |
| 76 | * lhs_info.k0: 2,3,4,8,16 |
| 77 | * lhs_info.transpose: false |
| 78 | * @param[in] rhs_info RHS matrix information used for reshaping the src1 tensor. Only the following values are supported: |
| 79 | * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) |
| 80 | * rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) |
| 81 | * rhs_info.transpose: true |
| 82 | * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices |
| 83 | * |
| 84 | * @note lhs_info.k0 must be equal to rhs_info.k0 |
| 85 | */ |
| 86 | void configure(const ClCompileContext &compile_context, |
| 87 | ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta, |
| 88 | const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info); |
| 89 | /** Static function to check if given info will lead to a valid configuration |
| 90 | * |
| 91 | * Similar to @ref ClGemmMatrixMultiplyReshapedKernel::configure() |
| 92 | * |
| 93 | * @return a status |
| 94 | */ |
| 95 | static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, |
| 96 | const GEMMRHSMatrixInfo &rhs_info, |
| 97 | const GEMMKernelInfo &gemm_info); |
| 98 | |
| 99 | // Inherited methods overridden: |
| 100 | void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; |
| 101 | |
| 102 | private: |
| 103 | bool _slide_matrix_b{ true }; |
| 104 | bool _reinterpret_output_as_3d{ false }; |
| 105 | bool _use_dummy_work_items{ false }; |
| 106 | bool _add_bias{ false }; |
| 107 | bool _export_to_cl_image{ false }; |
| 108 | unsigned int _k{ 1 }; |
| 109 | }; |
| 110 | } // namespace kernels |
| 111 | } // namespace opencl |
| 112 | } // namespace arm_compute |
| 113 | #endif /* ARM_COMPUTE_CL_GEMM_MATRIXMULTIPLY_RESHAPED_KERNEL_H */ |