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_CLGEMMRESHAPERHSMATRIXKERNEL_H 25 #define ARM_COMPUTE_CLGEMMRESHAPERHSMATRIXKERNEL_H 26 27 #include "src/core/CL/ICLKernel.h" 28 29 namespace arm_compute 30 { 31 class ICLTensor; 32 33 /** OpenCL kernel to reshape the RHS matrix when performing the matrix multiplication 34 * In particular, this kernel splits the input matrix in blocks of size K0xN0 and stores each one in 35 * the output matrix unrolling the values */ 36 class CLGEMMReshapeRHSMatrixKernel : public ICLKernel 37 { 38 public: 39 /** Default constructor */ 40 CLGEMMReshapeRHSMatrixKernel(); 41 /** Prevent instances of this class from being copied (As this class contains pointers) */ 42 CLGEMMReshapeRHSMatrixKernel(const CLGEMMReshapeRHSMatrixKernel &) = delete; 43 /** Prevent instances of this class from being copied (As this class contains pointers) */ 44 CLGEMMReshapeRHSMatrixKernel &operator=(const CLGEMMReshapeRHSMatrixKernel &) = delete; 45 /** Allow instances of this class to be moved */ 46 CLGEMMReshapeRHSMatrixKernel(CLGEMMReshapeRHSMatrixKernel &&) = default; 47 /** Allow instances of this class to be moved */ 48 CLGEMMReshapeRHSMatrixKernel &operator=(CLGEMMReshapeRHSMatrixKernel &&) = default; 49 /** Default destructor */ 50 ~CLGEMMReshapeRHSMatrixKernel() = default; 51 /** Initialise the kernel's input and output. 52 * 53 * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor, 54 * required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 55 * Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required: 56 * -# rhs_info.n0 can only be 4, 8 and 16 57 * -# rhs_info.k0 can only be 4, 8 and 16 58 * -# Data type can only be F32, F16 59 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 60 * -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 61 * -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 62 * -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 63 * 64 * @param[in] input Input tensor. Data types supported: All 65 * @param[out] output Output tensor. Data type supported: same as @p input 66 * @param[in] rhs_info RHS matrix information to be used for reshaping. This object contains all the necessary 67 * information to reshape the input tensor. Only the following values are supported: 68 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 69 * rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false), (only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 70 * rhs_info.h0: greater than 0 71 * rhs_info.transpose: true, false 72 * rhs_info.interleave: true, false 73 */ 74 void configure(const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info); 75 /** Initialise the kernel's input and output. 76 * 77 * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor, 78 * required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 79 * Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required: 80 * -# rhs_info.n0 can only be 4, 8 and 16 81 * -# rhs_info.k0 can only be 4, 8 and 16 82 * -# Data type can only be F32, F16 83 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 84 * -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 85 * -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 86 * -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 87 * 88 * @param[in] compile_context The compile context to be used. 89 * @param[in] input Input tensor. Data types supported: All 90 * @param[out] output Output tensor. Data type supported: same as @p input 91 * @param[in] rhs_info RHS matrix information to be used for reshaping. This object contains all the necessary 92 * information to reshape the input tensor. Only the following values are supported: 93 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 94 * rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false), (only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 95 * rhs_info.h0: greater than 0 96 * rhs_info.transpose: true, false 97 * rhs_info.interleave: true, false 98 */ 99 void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info); 100 /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMReshapeRHSMatrixKernel 101 * 102 * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor, 103 * required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 104 * Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required: 105 * -# rhs_info.n0 can only be 4, 8 and 16 106 * -# rhs_info.k0 can only be 4, 8 and 16 107 * -# Data type can only be F32, F16 108 * -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension 109 * -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4) 110 * -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT 111 * -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel 112 * 113 * @param[in] input Input tensor info. Data types supported: All 114 * @param[in] output Output tensor info which stores the interleaved matrix. Data type supported: same as @p input. 115 * @param[in] rhs_info RHS matrix information to be used for reshaping. This object contains all the necessary 116 * information to reshape the input tensor. Only the following values are supported: 117 * rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 118 * rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false),(only 4, 8 and 16 if rhs_info.export_to_cl_image == true) 119 * rhs_info.h0: greater than 0 120 * rhs_info.transpose: true, false 121 * rhs_info.interleave: true, false 122 * 123 * @return a status 124 */ 125 static Status validate(const ITensorInfo *input, const ITensorInfo *output, const GEMMRHSMatrixInfo &rhs_info); 126 127 // Inherited methods overridden 128 void run(const Window &window, cl::CommandQueue &queue) override; 129 130 private: 131 const ICLTensor *_input; 132 ICLTensor *_output; 133 }; 134 } // namespace arm_compute 135 #endif /* ARM_COMPUTE_CLGEMMRESHAPERHSMATRIXKERNEL_H */