SiCong Li | f44bbc5 | 2022-08-29 18:25:51 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 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 | #include "ClKernelRuntime.h" |
| 25 | #include "arm_compute/core/CL/ICLTensor.h" |
| 26 | #include "src/core/CL/CLUtils.h" |
| 27 | #include "src/dynamic_fusion/sketch/gpu/GpuKernelSourceCode.h" |
| 28 | #include "src/gpu/cl/ClKernelLibrary.h" |
| 29 | |
| 30 | #include "support/Cast.h" |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace experimental |
| 34 | { |
| 35 | namespace dynamic_fusion |
| 36 | { |
| 37 | using namespace arm_compute::opencl; |
| 38 | |
| 39 | void ClKernelRuntime::configure(const ClCompileContext &compile_ctx, const GpuKernelSourceCode &code) |
| 40 | { |
| 41 | // Create kernel from kernel source string |
| 42 | opencl::ClKernelLibrary &klib = opencl::ClKernelLibrary::get(); |
| 43 | _kernel = static_cast<cl::Kernel>(compile_ctx.create_kernel(code.name(), |
| 44 | "" /* Program name: Used to as part of a unique string for built kernel cache. Not needed */, |
| 45 | code.code(), |
| 46 | klib.kernel_path() /* Kernel path: Used in cases of embedded kernels */, |
| 47 | code.build_options().options(), |
| 48 | false /* Is source binary */)); |
| 49 | |
| 50 | // Configure execution window |
| 51 | IClKernel::configure_internal(code.window()); |
| 52 | |
| 53 | // Set config id for lws tuning |
| 54 | _config_id = code.config_id(); |
| 55 | |
| 56 | // Set kernel arguments |
| 57 | _arguments = code.arguments(); |
| 58 | } |
| 59 | |
| 60 | inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images) |
| 61 | { |
| 62 | switch(arg.type) |
| 63 | { |
| 64 | case GpuKernelArgumentInfo::Type::Scalar: |
| 65 | { |
| 66 | ARM_COMPUTE_ERROR("Unsupported yet"); |
| 67 | break; |
| 68 | } |
| 69 | |
| 70 | case GpuKernelArgumentInfo::Type::Vector: |
| 71 | { |
| 72 | add_1D_tensor_argument(idx, tensor, arg_slice); |
| 73 | break; |
| 74 | } |
| 75 | |
| 76 | case GpuKernelArgumentInfo::Type::Image: |
| 77 | { |
| 78 | add_2D_tensor_argument(idx, tensor, arg_slice); |
| 79 | break; |
| 80 | } |
| 81 | case GpuKernelArgumentInfo::Type::Image_Reinterpret_As_3D: |
| 82 | { |
| 83 | add_2D_tensor_argument(idx, tensor, arg_slice); |
| 84 | const unsigned int total_cross_plane_pad = tensor->info()->padding().top + tensor->info()->padding().bottom; |
| 85 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad)); |
| 86 | break; |
| 87 | } |
| 88 | case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D: |
| 89 | { |
| 90 | const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3)); |
| 91 | const size_t image_row_pitch = tensor->info()->strides_in_bytes()[1]; |
| 92 | cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch); |
| 93 | cl_images.push_back(tensor_image2d); |
| 94 | _kernel.setArg(idx++, tensor_image2d); |
| 95 | break; |
| 96 | } |
| 97 | |
| 98 | case GpuKernelArgumentInfo::Type::Image_3D: |
| 99 | { |
| 100 | add_2D_tensor_argument(idx, tensor, arg_slice); |
| 101 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(tensor->info()->strides_in_bytes()[2])); |
| 102 | break; |
| 103 | } |
| 104 | case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D: |
| 105 | { |
| 106 | const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3)); |
| 107 | const size_t image_row_pitch = tensor->info()->strides_in_bytes()[1]; |
| 108 | cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch); |
| 109 | cl_images.push_back(tensor_image2d); |
| 110 | _kernel.setArg(idx++, tensor_image2d); |
| 111 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(tensor->info()->strides_in_bytes()[2])); |
| 112 | break; |
| 113 | } |
| 114 | |
| 115 | case GpuKernelArgumentInfo::Type::Tensor_3D: |
| 116 | { |
| 117 | add_3D_tensor_argument(idx, tensor, arg_slice); |
| 118 | break; |
| 119 | } |
| 120 | |
| 121 | case GpuKernelArgumentInfo::Type::Tensor_4D: |
| 122 | { |
| 123 | add_4D_tensor_argument(idx, tensor, arg_slice); |
| 124 | break; |
| 125 | } |
| 126 | case GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer: |
| 127 | { |
| 128 | add_4d_tensor_nhwc_argument(idx, tensor); |
| 129 | break; |
| 130 | } |
| 131 | case GpuKernelArgumentInfo::Type::Tensor_4D_t_Image: |
| 132 | { |
| 133 | const size_t image_w = tensor->info()->dimension(0) / 4; |
| 134 | const size_t image_h = tensor->info()->tensor_shape().total_size_upper(1); |
| 135 | const size_t image_stride_y = tensor->info()->strides_in_bytes()[1]; |
| 136 | |
| 137 | cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), |
| 138 | TensorShape(image_w, image_h), tensor->info()->data_type(), image_stride_y); |
| 139 | cl_images.push_back(tensor_image2d); |
| 140 | |
| 141 | _kernel.setArg(idx++, tensor_image2d); |
| 142 | add_4d_tensor_nhwc_argument(idx, tensor); |
| 143 | break; |
| 144 | } |
| 145 | default: |
| 146 | { |
| 147 | ARM_COMPUTE_ERROR("Unsupported"); |
| 148 | } |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) |
| 153 | { |
| 154 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 155 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 156 | |
| 157 | Window slice = window.first_slice_window_3D(); |
| 158 | // Don't slice matrix along the z dimension if matrix has just 2 dimensions and matrix A more than 2 |
| 159 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 160 | Window slice_fixed_z = slice; |
| 161 | slice_fixed_z.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 162 | slice_fixed_z.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 163 | |
| 164 | /// NOTE: Parameters extracted from old kernels. So far they seem to be constant |
| 165 | /// but we may need to make them into another configuration passed from GpuWorkloadSourceCode if needed in the future |
| 166 | constexpr bool slide_along_dimz = true; |
| 167 | constexpr bool skip_sliding_window = false; |
| 168 | constexpr bool use_dummy_work_items = false; |
| 169 | |
| 170 | unsigned int idx = 0; |
| 171 | do |
| 172 | { |
| 173 | // Set kernel arguments |
| 174 | Window arg_slice = slice; |
| 175 | // CLImages created from tensor arguments. Need to be retained until enqueue |
| 176 | std::vector<cl::Image2D> cl_images; |
| 177 | for(auto id_arg : _arguments) |
| 178 | { |
| 179 | const auto arg = id_arg.second; |
| 180 | auto tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(id_arg.first)); |
| 181 | ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); |
| 182 | ARM_COMPUTE_ERROR_ON_NULLPTR(tensor->info()); |
| 183 | if(!slide_along_dimz) |
| 184 | { |
| 185 | // The stride_z for matrix must be zero if we do not slice |
| 186 | ARM_COMPUTE_ERROR_ON(tensor->info()->strides_in_bytes()[3] != 0); |
| 187 | arg_slice = slice_fixed_z; |
| 188 | } |
| 189 | add_tensor_argument(idx, *arg.kernel_argument_info(), tensor, arg_slice, cl_images); |
| 190 | } |
| 191 | |
| 192 | // Dispatch kernel |
| 193 | enqueue(queue, *this, slice, lws_hint(), use_dummy_work_items); |
| 194 | } |
| 195 | while(skip_sliding_window && window.slide_window_slice_3D(slice)); |
| 196 | } |
| 197 | |
| 198 | } // namespace dynamic_fusion |
| 199 | } // namespace experimental |
| 200 | } // namespace arm_compute |