Gunes Bayir | ec0113d | 2022-11-09 09:26:27 +0000 | [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 "src/gpu/cl/kernels/ClTransposedConvolutionKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 28 | #include "src/core/CL/CLValidate.h" |
| 29 | #include "src/core/helpers/AutoConfiguration.h" |
| 30 | #include "src/core/helpers/WindowHelpers.h" |
| 31 | #include "support/Cast.h" |
| 32 | |
| 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace opencl |
| 36 | { |
| 37 | namespace kernels |
| 38 | { |
| 39 | namespace |
| 40 | { |
| 41 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, |
| 42 | const PadStrideInfo &deconv_info) |
| 43 | { |
| 44 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); |
| 45 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); |
| 46 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
| 47 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC); |
| 48 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(weights, DataLayout::NHWC); |
| 49 | |
| 50 | constexpr unsigned int channel_idx = 0; |
| 51 | constexpr unsigned int width_idx = 1; |
| 52 | constexpr unsigned int height_idx = 2; |
| 53 | constexpr unsigned int batch_idx = 3; |
| 54 | |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx), "Weights feature map dimension should match the respective src's one"); |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional"); |
| 57 | |
| 58 | if(biases != nullptr) |
| 59 | { |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(channel_idx) != weights->dimension(batch_idx), |
| 62 | "Biases size and number of dst feature maps should match"); |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1, "Biases should be one dimensional"); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC); |
| 65 | } |
| 66 | |
| 67 | // Checks performed when output is configured |
| 68 | if(output->total_size() != 0) |
| 69 | { |
| 70 | const size_t input_width = input->dimension(width_idx); |
| 71 | const size_t input_height = input->dimension(height_idx); |
| 72 | const size_t weights_width = weights->dimension(width_idx); |
| 73 | const size_t weights_height = weights->dimension(height_idx); |
| 74 | |
| 75 | auto out_dims = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info); |
| 76 | TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights); |
| 77 | |
| 78 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); |
| 79 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 80 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(output, DataLayout::NHWC); |
| 81 | } |
| 82 | |
| 83 | return Status{}; |
| 84 | } |
| 85 | } // namespace |
| 86 | |
| 87 | void ClTransposedConvolutionKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights, |
| 88 | const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info) |
| 89 | { |
| 90 | ARM_COMPUTE_UNUSED(biases, deconv_info); |
| 91 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 92 | |
| 93 | // Perform validation |
| 94 | ARM_COMPUTE_ERROR_THROW_ON(validate(input, weights, biases, output, deconv_info)); |
| 95 | |
| 96 | constexpr unsigned int channel_idx = 0; |
| 97 | constexpr unsigned int width_idx = 1; |
| 98 | constexpr unsigned int height_idx = 2; |
| 99 | |
| 100 | const size_t input_channels = input->dimension(channel_idx); // same as weight channels |
| 101 | const size_t input_width = input->dimension(width_idx); |
| 102 | const size_t input_height = input->dimension(height_idx); |
| 103 | const size_t weights_width = weights->dimension(width_idx); |
| 104 | const size_t weights_height = weights->dimension(height_idx); |
| 105 | const size_t output_width = output->dimension(width_idx); |
| 106 | const size_t output_height = output->dimension(height_idx); |
| 107 | const size_t output_channels = output->dimension(channel_idx); |
| 108 | |
| 109 | // Calculate output shape |
| 110 | auto out_dims = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info); |
| 111 | TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights); |
| 112 | auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->quantization_info()); |
| 113 | |
| 114 | // Calculate updated paddings |
| 115 | // p' = k - p - 1 (k: kernel dimensions) |
| 116 | const uint32_t pad_left = weights_width - deconv_info.pad_left() - 1; |
| 117 | const uint32_t pad_top = weights_height - deconv_info.pad_top() - 1; |
| 118 | |
| 119 | // Configure kernel window |
| 120 | Window win; |
| 121 | output_shape.collapse(2U, 1U); // Collapse width and height into single dimension |
| 122 | |
| 123 | // Create window and update padding |
| 124 | win = calculate_max_window(output_shape, Steps(1, 1)); |
| 125 | ICLKernel::configure_internal(win); |
| 126 | |
| 127 | const std::string kernel_name = "transposed_convolution_nhwc"; |
| 128 | CLBuildOptions build_options; |
| 129 | |
| 130 | const DataType input_data_type = input->data_type(); // Fp32 or Fp16 only |
| 131 | const auto strides = deconv_info.stride(); |
| 132 | |
| 133 | const unsigned int n0 = 1; |
| 134 | const unsigned int m0 = 1; |
| 135 | const unsigned int k0 = adjust_vec_size(input_data_type == DataType::F32 ? 4 : 8, input_channels); |
| 136 | const unsigned int partial_store_n0 = output_channels % n0; |
| 137 | |
| 138 | if(biases != nullptr) |
| 139 | { |
| 140 | build_options.add_option(std::string("-DHAS_BIAS")); |
| 141 | build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type()))); |
| 142 | } |
| 143 | |
| 144 | const auto output_data_type = output->data_type(); |
| 145 | |
| 146 | build_options.add_option("-cl-fast-relaxed-math"); |
| 147 | build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER"); |
| 148 | build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type)); |
| 149 | build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(input_channels)); |
| 150 | build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width)); |
| 151 | build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input_height)); |
| 152 | build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(output_channels)); |
| 153 | build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(output_width)); |
| 154 | build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(output_height)); |
| 155 | build_options.add_option("-DDST_TENSOR_TYPE=BUFFER"); |
| 156 | build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(output_data_type)); |
| 157 | build_options.add_option("-DWEI_TENSOR_TYPE=BUFFER"); |
| 158 | build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights_width)); |
| 159 | build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights_height)); |
| 160 | build_options.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->data_type())); |
| 161 | build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(strides.first)); |
| 162 | build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(strides.second)); |
| 163 | build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left)); |
| 164 | build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top)); |
| 165 | build_options.add_option("-DN0=" + support::cpp11::to_string(n0)); |
| 166 | build_options.add_option("-DM0=" + support::cpp11::to_string(m0)); |
| 167 | build_options.add_option("-DK0=" + support::cpp11::to_string(k0)); |
| 168 | build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0)); |
| 169 | build_options.add_option_if((input_channels % k0) != 0, "-DLEFTOVER_LOOP"); |
| 170 | build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type)); |
| 171 | |
| 172 | if(compile_context.get_ddk_version() >= 30) |
| 173 | { |
| 174 | build_options.add_option("-fregister-allocation=64"); |
| 175 | } |
| 176 | |
| 177 | _kernel = create_kernel(compile_context, kernel_name, build_options.options()); |
| 178 | |
| 179 | // Set config_id for enabling LWS tuning |
| 180 | _config_id = kernel_name; |
| 181 | _config_id += "_"; |
| 182 | _config_id += lower_string(string_from_data_type(input_data_type)); |
| 183 | _config_id += "_"; |
| 184 | _config_id += support::cpp11::to_string(weights_width); |
| 185 | _config_id += "_"; |
| 186 | _config_id += support::cpp11::to_string(strides.first); |
| 187 | _config_id += "_"; |
| 188 | _config_id += support::cpp11::to_string(strides.second); |
| 189 | _config_id += "_"; |
| 190 | _config_id += support::cpp11::to_string(output_width); |
| 191 | _config_id += "_"; |
| 192 | _config_id += support::cpp11::to_string(m0); |
| 193 | _config_id += "_"; |
| 194 | _config_id += support::cpp11::to_string(n0); |
| 195 | } |
| 196 | |
| 197 | Status ClTransposedConvolutionKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, |
| 198 | const ITensorInfo *dst, const PadStrideInfo &deconv_info) |
| 199 | { |
| 200 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, deconv_info)); |
| 201 | return Status{}; |
| 202 | } |
| 203 | |
| 204 | void ClTransposedConvolutionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) |
| 205 | { |
| 206 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 207 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 208 | |
| 209 | // Get initial windows |
| 210 | Window slice = window.first_slice_window_3D(); |
| 211 | |
| 212 | const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); |
| 213 | const auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); |
| 214 | const auto biases = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); |
| 215 | auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
| 216 | |
| 217 | unsigned int idx = 0; |
| 218 | add_4d_tensor_nhwc_argument(idx, src); |
| 219 | add_4d_tensor_nhwc_argument(idx, dst); |
| 220 | |
| 221 | add_4d_tensor_nhwc_argument(idx, weights); |
| 222 | if(biases != nullptr) |
| 223 | { |
| 224 | add_1D_tensor_argument(idx, biases, slice); |
| 225 | } |
| 226 | |
| 227 | enqueue(queue, *this, slice, lws_hint()); |
| 228 | } |
| 229 | } // namespace kernels |
| 230 | } // namespace opencl |
| 231 | } // namespace arm_compute |