Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +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 "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h" |
| 25 | |
| 26 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 27 | |
| 28 | #include "src/common/utils/Log.h" |
| 29 | #include "src/core/helpers/AutoConfiguration.h" |
| 30 | #include "src/dynamic_fusion/sketch/ArgumentPack.h" |
| 31 | #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" |
| 32 | #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.h" |
| 33 | #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h" |
Gian Marco Iodice | 9d3bd41 | 2022-12-30 09:45:00 +0000 | [diff] [blame] | 34 | #include "src/runtime/heuristics/dwc_native/ClDWCNativeKernelConfig.h" |
| 35 | #include "src/runtime/heuristics/dwc_native/IClDWCNativeKernelConfig.h" |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace experimental |
| 40 | { |
| 41 | namespace dynamic_fusion |
| 42 | { |
| 43 | namespace |
| 44 | { |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 45 | void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *src, const ITensorInfo *wei, const DepthwiseConv2dAttributes &attributes) |
| 46 | { |
| 47 | if(dst->total_size() == 0U) |
| 48 | { |
| 49 | const PadStrideInfo pad_stride_info(attributes.stride().x(), |
| 50 | attributes.stride().y(), |
| 51 | attributes.pad().left, |
| 52 | attributes.pad().right, |
| 53 | attributes.pad().top, |
| 54 | attributes.pad().bottom, |
| 55 | attributes.dimension_rounding_type()); |
| 56 | |
| 57 | const ConvolutionInfo conv_info{ pad_stride_info, attributes.depth_multiplier(), ActivationLayerInfo(), attributes.dilation() }; |
| 58 | const TensorShape shape = misc::shape_calculator::compute_depthwise_convolution_shape(*src, *wei, conv_info); |
| 59 | |
| 60 | auto_init_if_empty(*dst, src->clone()->set_tensor_shape(shape)); |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | constexpr GpuOperatorType operator_type = GpuOperatorType::Complex; |
| 65 | } // namespace |
| 66 | |
| 67 | Status GpuDepthwiseConv2d::is_supported_op(const GpuWorkloadContext &context, |
| 68 | const ITensorInfo *src, |
| 69 | const ITensorInfo *wei, |
| 70 | const ITensorInfo *bia, |
| 71 | const ITensorInfo *dst, |
| 72 | const DepthwiseConv2dAttributes &attributes) |
| 73 | { |
| 74 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei, dst); |
| 75 | |
| 76 | // Auto initialize dst tensor info |
| 77 | TensorInfo dst_info_to_validate = *dst; |
| 78 | calculate_and_init_dst_if_empty(&dst_info_to_validate, src, wei, attributes); |
| 79 | |
| 80 | // Check support level |
| 81 | // Data type |
| 82 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 83 | // Data layout |
| 84 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); |
| 85 | |
| 86 | const GpuTarget gpu_target = context.gpu_target(); |
| 87 | |
| 88 | if(context.gpu_language() == GpuLanguage::OpenCL) |
| 89 | { |
| 90 | const CLCompileContext *cl_compile_ctx = context.cl_compile_context(); |
| 91 | ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr); |
| 92 | |
| 93 | // Validate Depthwise Conv2d Component |
| 94 | { |
| 95 | const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 96 | auto settings = ClComponentDepthwiseConv2d::Settings(); |
| 97 | |
Gian Marco Iodice | 9d3bd41 | 2022-12-30 09:45:00 +0000 | [diff] [blame] | 98 | const PadStrideInfo legacy_conv_info(attributes.stride().x(), attributes.stride().y(), attributes.pad().left, |
| 99 | attributes.pad().right, |
| 100 | attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR); |
| 101 | |
| 102 | // Get the depthwise convolution compute parameters |
| 103 | auto t = arm_compute::cl_dwc::ClDWCNativeKernelConfigurationFactory::create(gpu_target); |
| 104 | const DWCComputeKernelInfo dwc_info = t->configure(src, wei, legacy_conv_info, attributes.dilation(), attributes.depth_multiplier()); |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 105 | |
| 106 | settings.fast_relaxed_math( |
| 107 | (gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST) |
| 108 | && (dst_info_to_validate.data_type() == DataType::F32 || dst_info_to_validate.data_type() == DataType::F16)); |
| 109 | |
| 110 | settings.is_fma_available(get_arch_from_target(gpu_target) == GPUTarget::MIDGARD) |
| 111 | .m0(dwc_info.m0) |
| 112 | .n0(dwc_info.n0) |
| 113 | .export_input_to_cl_image(dwc_info.export_input_to_cl_image) |
| 114 | .export_weights_to_cl_image(dwc_info.export_weights_to_cl_image); |
| 115 | |
| 116 | ArgumentPack<ITensorInfo> arguments; |
| 117 | arguments.add_const_tensor(ACL_SRC_0, src); |
| 118 | arguments.add_const_tensor(ACL_SRC_1, wei); |
| 119 | arguments.add_const_tensor(ACL_SRC_2, bia); |
| 120 | arguments.add_const_tensor(ACL_DST_0, &dst_info_to_validate); |
| 121 | ARM_COMPUTE_RETURN_ON_ERROR(ClComponentDepthwiseConv2d::validate(properties, arguments, attributes, settings)); |
| 122 | } |
| 123 | } |
| 124 | else |
| 125 | { |
| 126 | ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); |
| 127 | } |
| 128 | |
| 129 | return Status{}; |
| 130 | } |
| 131 | |
| 132 | Status GpuDepthwiseConv2d::validate_op(const GpuWorkloadSketch &sketch, |
| 133 | const ITensorInfo *src, |
| 134 | const ITensorInfo *wei, |
| 135 | const ITensorInfo *bia, |
| 136 | const ITensorInfo *dst, |
| 137 | const DepthwiseConv2dAttributes &attributes) |
| 138 | { |
| 139 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei, dst); |
| 140 | ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !wei->has_valid_id() || !dst->has_valid_id()); |
| 141 | |
| 142 | if(bia != nullptr) |
| 143 | { |
| 144 | ARM_COMPUTE_RETURN_ERROR_ON(!bia->has_valid_id()); |
| 145 | } |
| 146 | |
| 147 | // Auto initialize dst tensor info |
| 148 | TensorInfo dst_info_to_validate = *dst; |
| 149 | calculate_and_init_dst_if_empty(&dst_info_to_validate, src, wei, attributes); |
| 150 | |
| 151 | // Perform fusion test |
| 152 | // Pack tensor infos |
| 153 | ArgumentPack<ITensorInfo> tensors; |
| 154 | tensors.add_const_tensor(ACL_SRC_0, src); |
| 155 | tensors.add_const_tensor(ACL_SRC_1, wei); |
| 156 | tensors.add_const_tensor(ACL_SRC_2, bia); |
| 157 | tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate); |
| 158 | const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 159 | |
| 160 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op), |
| 161 | "Operator fusion test failed. This operator cannot be fused into the workload"); |
| 162 | |
| 163 | // Check if configuration is supported |
| 164 | return is_supported_op(*sketch.gpu_context(), src, wei, bia, &dst_info_to_validate, attributes); |
| 165 | } |
| 166 | |
| 167 | void GpuDepthwiseConv2d::create_op(GpuWorkloadSketch &sketch, |
| 168 | ITensorInfo *src, |
| 169 | ITensorInfo *wei, |
| 170 | ITensorInfo *bia, |
| 171 | ITensorInfo *dst, |
| 172 | const DepthwiseConv2dAttributes &attributes) |
| 173 | { |
| 174 | // Assert validation |
| 175 | ARM_COMPUTE_ERROR_THROW_ON(GpuDepthwiseConv2d::validate_op(sketch, src, wei, bia, dst, attributes)); |
| 176 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, wei, dst); |
| 177 | ARM_COMPUTE_LOG_PARAMS(src, wei, bia, dst, attributes); |
| 178 | |
| 179 | calculate_and_init_dst_if_empty(dst, src, wei, attributes); |
| 180 | |
| 181 | // Translate into components and add to component graph |
| 182 | GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph(); |
| 183 | const auto *sketch_ctx = sketch.implementation().context(); |
| 184 | const GpuTarget gpu_target = sketch_ctx->gpu_target(); |
| 185 | |
| 186 | if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) |
| 187 | { |
Viet-Hoa Do | 81f796b | 2022-12-23 14:42:55 +0000 | [diff] [blame] | 188 | ARM_COMPUTE_ERROR_ON_NULLPTR(sketch_ctx->cl_compile_context()); |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 189 | |
| 190 | // Add Depthwise Conv2d Component |
| 191 | { |
| 192 | const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 193 | auto settings = ClComponentDepthwiseConv2d::Settings(); |
| 194 | |
Gian Marco Iodice | 9d3bd41 | 2022-12-30 09:45:00 +0000 | [diff] [blame] | 195 | const PadStrideInfo legacy_conv_info(attributes.stride().x(), attributes.stride().y(), attributes.pad().left, |
| 196 | attributes.pad().right, |
| 197 | attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR); |
| 198 | |
| 199 | // Get the depthwise convolution compute parameters |
| 200 | auto t = arm_compute::cl_dwc::ClDWCNativeKernelConfigurationFactory::create(gpu_target); |
| 201 | const DWCComputeKernelInfo dwc_info = t->configure(src, wei, legacy_conv_info, attributes.dilation(), attributes.depth_multiplier()); |
Gunes Bayir | 7dc0234 | 2022-11-21 21:46:50 +0000 | [diff] [blame] | 202 | |
| 203 | settings.is_fma_available(get_arch_from_target(gpu_target) != GPUTarget::MIDGARD) |
| 204 | .m0(dwc_info.m0) |
| 205 | .n0(dwc_info.n0) |
| 206 | .export_input_to_cl_image(dwc_info.export_input_to_cl_image) |
| 207 | .export_weights_to_cl_image(dwc_info.export_weights_to_cl_image); |
| 208 | |
| 209 | if(settings.export_input_to_cl_image()) |
| 210 | { |
| 211 | arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(src); |
| 212 | } |
| 213 | |
| 214 | if(settings.export_weights_to_cl_image()) |
| 215 | { |
| 216 | arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(wei); |
| 217 | } |
| 218 | |
| 219 | ArgumentPack<ITensorInfo> arguments; |
| 220 | arguments.add_const_tensor(ACL_SRC_0, src); |
| 221 | arguments.add_const_tensor(ACL_SRC_1, wei); |
| 222 | arguments.add_const_tensor(ACL_SRC_2, bia); |
| 223 | arguments.add_const_tensor(ACL_DST_0, dst); |
| 224 | comp_graph.add_new_component<ClComponentDepthwiseConv2d>(properties, arguments, attributes, settings); |
| 225 | } |
| 226 | } |
| 227 | else |
| 228 | { |
| 229 | ARM_COMPUTE_ERROR("Unimplemented Gpu language"); |
| 230 | } |
| 231 | |
| 232 | // Set up fusion test by adding to the Operator Group |
| 233 | // Note this has to be performed after all the components have been successfully added to the component graph |
| 234 | |
| 235 | // Pack tensor infos |
| 236 | ArgumentPack<ITensorInfo> tensors; |
| 237 | tensors.add_const_tensor(ACL_SRC_0, src); |
| 238 | tensors.add_const_tensor(ACL_SRC_1, wei); |
| 239 | tensors.add_const_tensor(ACL_SRC_2, bia); |
| 240 | tensors.add_const_tensor(ACL_DST_0, dst); |
| 241 | |
| 242 | const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 243 | sketch.implementation().operator_group().add_operator(op); |
| 244 | } |
| 245 | |
| 246 | } // namespace dynamic_fusion |
| 247 | } // namespace experimental |
| 248 | } // namespace arm_compute |