1 /* 2 * Copyright (c) 2018-2020 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_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H 25 #define ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H 26 27 #include "src/core/CL/ICLKernel.h" 28 29 #include "arm_compute/core/KernelDescriptors.h" 30 31 namespace arm_compute 32 { 33 class ICLTensor; 34 35 /** OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped 36 * 37 * @note The input matrices @p input0 and @p input1 must be reshaped through @ref CLGEMMReshapeLHSMatrixKernel and @ref CLGEMMReshapeRHSMatrixKernel 38 */ 39 class CLGEMMMatrixMultiplyReshapedKernel : public ICLKernel 40 { 41 public: 42 /** Default Constructor */ 43 CLGEMMMatrixMultiplyReshapedKernel(); 44 /** Prevent instances of this class from being copied (As this class contains pointers) */ 45 CLGEMMMatrixMultiplyReshapedKernel(const CLGEMMMatrixMultiplyReshapedKernel &) = delete; 46 /** Prevent instances of this class from being copied (As this class contains pointers) */ 47 CLGEMMMatrixMultiplyReshapedKernel &operator=(const CLGEMMMatrixMultiplyReshapedKernel &) = delete; 48 /** Allow instances of this class to be moved */ 49 CLGEMMMatrixMultiplyReshapedKernel(CLGEMMMatrixMultiplyReshapedKernel &&) = default; 50 /** Allow instances of this class to be moved */ 51 CLGEMMMatrixMultiplyReshapedKernel &operator=(CLGEMMMatrixMultiplyReshapedKernel &&) = default; 52 /** Initialise the kernel's input and output. 53 * 54 * @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag. 55 * Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the 56 * multiplications. i.e. float c = (half)a * (half)b 57 * 58 * @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. 59 * Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer, 60 * the following conditions are required: 61 * -# rhs_info.n0 can only be 4, 8 and 16 62 * -# rhs_info.k0 can only be 4, 8 and 16 63 * -# Data type can only be F32 64 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 65 * -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement 66 * -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 67 * -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 68 * 69 * @param[in] input0 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 70 * @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3 71 * @param[in] input2 Input tensor containing the bias matrix. Data type supported: same as @p input0. 72 * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 73 * @param[in] alpha Weight of the matrix product 74 * @param[in] beta Weight of the matrix bias 75 * @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported: 76 * lhs_info.m0: 2,3,4,5,6,7,8 77 * lhs_info.k0: 2,3,4,8,16 78 * lhs_info.transpose: false 79 * @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 80 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 81 * rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 82 * rhs_info.transpose: true 83 * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices 84 * 85 * @note lhs_info.k0 must be equal to rhs_info.k0 86 */ 87 void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, 88 const GEMMRHSMatrixInfo &rhs_info, 89 const GEMMKernelInfo &gemm_info); 90 /** Initialise the kernel's input and output. 91 * 92 * @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag. 93 * Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the 94 * multiplications. i.e. float c = (half)a * (half)b 95 * 96 * @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. 97 * Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer, 98 * the following conditions are required: 99 * -# rhs_info.n0 can only be 4, 8 and 16 100 * -# rhs_info.k0 can only be 4, 8 and 16 101 * -# Data type can only be F32 102 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 103 * -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement 104 * -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 105 * -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 106 * 107 * @param[in] compile_context The compile context to be used. 108 * @param[in] input0 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 4 109 * @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3 110 * @param[in] input2 Input tensor containing the bias matrix. Data type supported: same as @p input0. 111 * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 112 * @param[in] alpha Weight of the matrix product 113 * @param[in] beta Weight of the matrix bias 114 * @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported: 115 * lhs_info.m0: 2,3,4,5,6,7,8 116 * lhs_info.k0: 2,3,4,8,16 117 * lhs_info.transpose: false 118 * @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 119 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 120 * rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 121 * rhs_info.transpose: true 122 * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices 123 * 124 * @note lhs_info.k0 must be equal to rhs_info.k0 125 */ 126 void configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, 127 const GEMMLHSMatrixInfo &lhs_info, 128 const GEMMRHSMatrixInfo &rhs_info, 129 const GEMMKernelInfo &gemm_info); 130 /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyReshapedKernel 131 * 132 * @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag. 133 * Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the 134 * multiplications. i.e. float c = (half)a * (half)b 135 * 136 * @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. 137 * Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer, 138 * the following conditions are required: 139 * -# rhs_info.n0 can only be 4, 8 and 16 140 * -# rhs_info.k0 can only be 4, 8 and 16 141 * -# Data type can only be F32 142 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 143 * -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement 144 * -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 145 * -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 146 * 147 * @param[in] input0 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 4 148 * @param[in] input1 Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3 149 * @param[in] input2 Input tensor info containing the bias matrix. Data type supported: same as @p input0. 150 * @param[in] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 151 * @param[in] alpha Weight of the matrix product 152 * @param[in] beta Weight of the matrix bias 153 * @param[in] lhs_info LHS matrix information used for reshaping the input0 tensor. Only the following values are supported: 154 * lhs_info.m0: 2,3,4,5,6,7,8 155 * lhs_info.k0: 2,3,4,8,16 156 * lhs_info.transpose: false 157 * @param[in] rhs_info RHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 158 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 159 * rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true) 160 * rhs_info.transpose: true 161 * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices 162 * 163 * @note lhs_info.k0 must be equal to rhs_info.k0 164 * 165 * @return a status 166 */ 167 static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, 168 const GEMMRHSMatrixInfo &rhs_info, 169 const GEMMKernelInfo &gemm_info); 170 171 // Inherited methods overridden: 172 void run(const Window &window, cl::CommandQueue &queue) override; 173 174 private: 175 const ICLTensor *_input0; 176 const ICLTensor *_input1; 177 const ICLTensor *_input2; 178 ICLTensor *_output; 179 bool _slide_matrix_b; 180 bool _reinterpret_output_as_3d; 181 bool _use_dummy_work_items; 182 bool _add_bias; 183 bool _broadcast_bias; 184 bool _export_to_cl_image; 185 unsigned int _k; 186 }; 187 } // namespace arm_compute 188 #endif /*ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H*/