diff options
Diffstat (limited to 'src/graph/backends/CL')
| -rw-r--r-- | src/graph/backends/CL/CLDeviceBackend.cpp | 65 | ||||
| -rw-r--r-- | src/graph/backends/CL/CLFunctionsFactory.cpp | 183 | ||||
| -rw-r--r-- | src/graph/backends/CL/CLNodeValidator.cpp | 85 | ||||
| -rw-r--r-- | src/graph/backends/CL/CLSubTensorHandle.cpp | 11 | ||||
| -rw-r--r-- | src/graph/backends/CL/CLTensorHandle.cpp | 11 |
5 files changed, 228 insertions, 127 deletions
diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp index 0159592af6..e27a4109d1 100644 --- a/src/graph/backends/CL/CLDeviceBackend.cpp +++ b/src/graph/backends/CL/CLDeviceBackend.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 ARM Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,19 +23,17 @@ */ #include "arm_compute/graph/backends/CL/CLDeviceBackend.h" -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/GraphContext.h" -#include "arm_compute/graph/INode.h" -#include "arm_compute/graph/Logger.h" -#include "arm_compute/graph/Tensor.h" +#include "arm_compute/core/TensorInfo.h" #include "arm_compute/graph/backends/BackendRegistrar.h" #include "arm_compute/graph/backends/CL/CLFunctionFactory.h" #include "arm_compute/graph/backends/CL/CLNodeValidator.h" #include "arm_compute/graph/backends/CL/CLSubTensorHandle.h" #include "arm_compute/graph/backends/CL/CLTensorHandle.h" - -#include "arm_compute/core/CL/CLCoreRuntimeContext.h" -#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/graph/Graph.h" +#include "arm_compute/graph/GraphContext.h" +#include "arm_compute/graph/INode.h" +#include "arm_compute/graph/Logger.h" +#include "arm_compute/graph/Tensor.h" #include "arm_compute/runtime/BlobLifetimeManager.h" #include "arm_compute/runtime/CL/CLBufferAllocator.h" #include "arm_compute/runtime/CL/CLScheduler.h" @@ -65,17 +63,18 @@ bool file_exists(const std::string &filename) static detail::BackendRegistrar<CLDeviceBackend> CLDeviceBackend_registrar(Target::CL); CLDeviceBackend::CLDeviceBackend() - : _context_count(0), _tuner(), _allocator(nullptr), _tuner_file() + : _context_count(0), + _tuner(), + _gemm_heuristics(), + _allocator(nullptr), + _tuner_file(), + _backend_type(CLBackendType::Native) { } CLDeviceBackend::~CLDeviceBackend() { - // TODO (geopin01) : Shouldn't call non exception safe stuff here - if(_tuner.tune_new_kernels() && !_tuner.lws_table().empty() && !_tuner_file.empty()) - { - _tuner.save_to_file(_tuner_file); - } + _tuner.save_to_file(_tuner_file); } void CLDeviceBackend::set_kernel_tuning(bool enable_tuning) @@ -91,16 +90,16 @@ void CLDeviceBackend::set_kernel_tuning_mode(CLTunerMode tuning_mode) void CLDeviceBackend::initialize_backend() { // Setup Scheduler - CLScheduler::get().default_init(&_tuner); + CLScheduler::get().default_init(&_tuner, &_gemm_heuristics, _backend_type); // Create allocator with new context - _allocator = support::cpp14::make_unique<CLBufferAllocator>(nullptr /* legacy path for CLCoreRuntimeContext */); + _allocator = std::make_unique<CLBufferAllocator>(); } void CLDeviceBackend::release_backend_context(GraphContext &ctx) { ARM_COMPUTE_UNUSED(ctx); _context_count--; - if(_context_count == 0) // No more context using the backend: free resources + if (_context_count == 0) // No more context using the backend: free resources { _allocator = nullptr; } @@ -110,15 +109,17 @@ void CLDeviceBackend::setup_backend_context(GraphContext &ctx) { // Force backend initialization _context_count++; - if(_context_count == 1) + if (_context_count == 1) { + _backend_type = ctx.config().backend_type; initialize_backend(); } // Setup tuner _tuner_file = ctx.config().tuner_file; + // Load tuner data if available - if(file_exists(_tuner_file)) + if (file_exists(_tuner_file)) { _tuner.load_from_file(_tuner_file); } @@ -126,8 +127,12 @@ void CLDeviceBackend::setup_backend_context(GraphContext &ctx) set_kernel_tuning(ctx.config().use_tuner); set_kernel_tuning_mode(ctx.config().tuner_mode); + // Attempt to load mlgo heuristics + ARM_COMPUTE_ERROR_ON(CLScheduler::get().gemm_heuristics() == nullptr); + CLScheduler::get().gemm_heuristics()->reload_from_file(ctx.config().mlgo_file); + // Setup a management backend - if(ctx.memory_management_ctx(Target::CL) == nullptr) + if (ctx.memory_management_ctx(Target::CL) == nullptr) { MemoryManagerContext mm_ctx; mm_ctx.target = Target::CL; @@ -140,7 +145,7 @@ void CLDeviceBackend::setup_backend_context(GraphContext &ctx) } // Create function level weights manager - if(ctx.weights_management_ctx(Target::CL) == nullptr) + if (ctx.weights_management_ctx(Target::CL) == nullptr) { WeightsManagerContext wm_ctx; wm_ctx.target = Target::CL; @@ -170,17 +175,18 @@ std::unique_ptr<ITensorHandle> CLDeviceBackend::create_tensor(const Tensor &tens TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type, tensor_desc.quant_info); info.set_data_layout(tensor_desc.layout); - return support::cpp14::make_unique<CLTensorHandle>(info); + return std::make_unique<CLTensorHandle>(info); } -std::unique_ptr<ITensorHandle> CLDeviceBackend::create_subtensor(ITensorHandle *parent, TensorShape shape, Coordinates coords, bool extend_parent) +std::unique_ptr<ITensorHandle> +CLDeviceBackend::create_subtensor(ITensorHandle *parent, TensorShape shape, Coordinates coords, bool extend_parent) { - if(parent == nullptr) + if (parent == nullptr) { return nullptr; } - return support::cpp14::make_unique<CLSubTensorHandle>(parent, shape, coords, extend_parent); + return std::make_unique<CLSubTensorHandle>(parent, shape, coords, extend_parent); } std::unique_ptr<arm_compute::IFunction> CLDeviceBackend::configure_node(INode &node, GraphContext &ctx) @@ -202,7 +208,7 @@ arm_compute::Status CLDeviceBackend::validate_node(INode &node) std::shared_ptr<arm_compute::IMemoryManager> CLDeviceBackend::create_memory_manager(MemoryManagerAffinity affinity) { - if(affinity == MemoryManagerAffinity::Offset) + if (affinity == MemoryManagerAffinity::Offset) { ARM_COMPUTE_LOG_GRAPH_WARNING("CL Backend does not support offset affinity memory management!"); return nullptr; @@ -220,6 +226,11 @@ std::shared_ptr<arm_compute::IWeightsManager> CLDeviceBackend::create_weights_ma auto weights_mgr = std::make_shared<IWeightsManager>(); return weights_mgr; } + +void CLDeviceBackend::sync() +{ + CLScheduler::get().sync(); +} } // namespace backends } // namespace graph } // namespace arm_compute diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp index 312e09a49a..d4e1aa880f 100644 --- a/src/graph/backends/CL/CLFunctionsFactory.cpp +++ b/src/graph/backends/CL/CLFunctionsFactory.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 ARM Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,14 +22,15 @@ * SOFTWARE. */ #include "arm_compute/graph/backends/CL/CLFunctionFactory.h" - -#include "arm_compute/core/utils/misc/Cast.h" +#include "arm_compute/graph/backends/FunctionHelpers.h" #include "arm_compute/graph/Graph.h" #include "arm_compute/graph/GraphContext.h" -#include "arm_compute/graph/backends/FunctionHelpers.h" #include "arm_compute/runtime/CL/CLFunctions.h" #include "arm_compute/runtime/CPP/CPPFunctions.h" +#include "src/core/CL/CLKernels.h" +#include "support/Cast.h" + using namespace arm_compute::utils::cast; namespace arm_compute @@ -42,6 +43,7 @@ namespace backends struct CLTargetInfo { using TensorType = arm_compute::ICLTensor; + using SrcTensorType = const arm_compute::ICLTensor; using TensorConcreteType = CLTensor; static Target TargetType; }; @@ -63,6 +65,14 @@ struct CLEltwiseFunctions using Addition = CLArithmeticAddition; using Subtraction = CLArithmeticSubtraction; using Multiplication = CLPixelWiseMultiplication; + using Maximum = CLElementwiseMax; + using Division = CLArithmeticDivision; +}; + +/** Collection of CL unary element-wise functions */ +struct CLUnaryEltwiseFunctions +{ + using Exp = CLExpLayer; }; /** Function and tensor types to be used inside a CL fused convolution/batch normalization layer */ @@ -71,28 +81,27 @@ struct CLFusedLayerTypes using ConvolutionLayer = CLConvolutionLayer; using DepthwiseConvolutionLayer = CLDepthwiseConvolutionLayer; using FuseBatchNormalization = CLFuseBatchNormalization; + using GEMMConvolutionLayer = CLGEMMConvolutionLayer; }; -// TODO (isagot01): Remove once we support heterogeneous scheduling at function level /** Wrapper for the CPP Function in the OpenCL backend **/ class CPPWrapperFunction : public IFunction { public: /* Default constructor */ - CPPWrapperFunction() - : _tensors(), _func(nullptr) + CPPWrapperFunction() : _tensors(), _func(nullptr) { } void run() override { - for(auto &tensor : _tensors) + for (auto &tensor : _tensors) { tensor->map(CLScheduler::get().queue()); } _func->run(); - for(auto &tensor : _tensors) + for (auto &tensor : _tensors) { tensor->unmap(CLScheduler::get().queue()); } @@ -117,7 +126,8 @@ namespace detail { // Specialized functions template <> -std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(DetectionOutputLayerNode &node) +std::unique_ptr<IFunction> +create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(DetectionOutputLayerNode &node) { validate_node<CLTargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */); @@ -134,23 +144,19 @@ std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer ARM_COMPUTE_ERROR_ON(output == nullptr); // Create and configure function - auto func = support::cpp14::make_unique<CPPDetectionOutputLayer>(); + auto func = std::make_unique<CPPDetectionOutputLayer>(); func->configure(input0, input1, input2, output, detect_info); // Log info ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " - << node.name() - << " Type: " << node.type() - << " Target: " << CLTargetInfo::TargetType - << " Data Type: " << input0->info()->data_type() - << " Input0 shape: " << input0->info()->tensor_shape() - << " Input1 shape: " << input1->info()->tensor_shape() + << node.name() << " Type: " << node.type() << " Target: " << CLTargetInfo::TargetType + << " Data Type: " << input0->info()->data_type() << " Input0 shape: " + << input0->info()->tensor_shape() << " Input1 shape: " << input1->info()->tensor_shape() << " Input2 shape: " << input2->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() - << " DetectionOutputLayer info: " << detect_info - << std::endl); + << " DetectionOutputLayer info: " << detect_info << std::endl); - auto wrap_function = support::cpp14::make_unique<CPPWrapperFunction>(); + auto wrap_function = std::make_unique<CPPWrapperFunction>(); wrap_function->register_function(std::move(func)); wrap_function->register_tensor(input0); @@ -158,10 +164,11 @@ std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer wrap_function->register_tensor(input2); wrap_function->register_tensor(output); - return RETURN_UNIQUE_PTR(wrap_function); + return std::move(wrap_function); } template <> -std::unique_ptr<IFunction> create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(DetectionPostProcessLayerNode &node) +std::unique_ptr<IFunction> +create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(DetectionPostProcessLayerNode &node) { validate_node<CLTargetInfo>(node, 3 /* expected inputs */, 4 /* expected outputs */); @@ -184,26 +191,22 @@ std::unique_ptr<IFunction> create_detection_post_process_layer<CPPDetectionPostP ARM_COMPUTE_ERROR_ON(output3 == nullptr); // Create and configure function - auto func = support::cpp14::make_unique<CPPDetectionPostProcessLayer>(); + auto func = std::make_unique<CPPDetectionPostProcessLayer>(); func->configure(input0, input1, input2, output0, output1, output2, output3, detect_info); // Log info ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " - << node.name() - << " Type: " << node.type() - << " Target: " << CLTargetInfo::TargetType - << " Data Type: " << input0->info()->data_type() - << " Input0 shape: " << input0->info()->tensor_shape() - << " Input1 shape: " << input1->info()->tensor_shape() + << node.name() << " Type: " << node.type() << " Target: " << CLTargetInfo::TargetType + << " Data Type: " << input0->info()->data_type() << " Input0 shape: " + << input0->info()->tensor_shape() << " Input1 shape: " << input1->info()->tensor_shape() << " Input2 shape: " << input2->info()->tensor_shape() << " Output0 shape: " << output0->info()->tensor_shape() << " Output1 shape: " << output1->info()->tensor_shape() << " Output2 shape: " << output2->info()->tensor_shape() << " Output3 shape: " << output3->info()->tensor_shape() - << " DetectionPostProcessLayer info: " << detect_info - << std::endl); + << " DetectionPostProcessLayer info: " << detect_info << std::endl); - auto wrap_function = support::cpp14::make_unique<CPPWrapperFunction>(); + auto wrap_function = std::make_unique<CPPWrapperFunction>(); wrap_function->register_function(std::move(func)); wrap_function->register_tensor(input0); @@ -214,90 +217,134 @@ std::unique_ptr<IFunction> create_detection_post_process_layer<CPPDetectionPostP wrap_function->register_tensor(output2); wrap_function->register_tensor(output3); - return RETURN_UNIQUE_PTR(wrap_function); + return std::move(wrap_function); } } // namespace detail std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &ctx) { - if(node == nullptr) + if (node == nullptr) { return nullptr; } NodeType type = node->type(); - switch(type) + switch (type) { case NodeType::ActivationLayer: - return detail::create_activation_layer<CLActivationLayer, CLTargetInfo>(*polymorphic_downcast<ActivationLayerNode *>(node)); + return detail::create_activation_layer<CLActivationLayer, CLTargetInfo>( + *polymorphic_downcast<ActivationLayerNode *>(node)); + case NodeType::ArgMinMaxLayer: + return detail::create_arg_min_max_layer<CLArgMinMaxLayer, CLTargetInfo>( + *polymorphic_downcast<ArgMinMaxLayerNode *>(node)); case NodeType::BatchNormalizationLayer: - return detail::create_batch_normalization_layer<CLBatchNormalizationLayer, CLTargetInfo>(*polymorphic_downcast<BatchNormalizationLayerNode *>(node)); + return detail::create_batch_normalization_layer<CLBatchNormalizationLayer, CLTargetInfo>( + *polymorphic_downcast<BatchNormalizationLayerNode *>(node)); case NodeType::BoundingBoxTransformLayer: - return detail::create_bounding_box_transform_layer<CLBoundingBoxTransform, CLTargetInfo>(*polymorphic_downcast<BoundingBoxTransformLayerNode *>(node)); + return detail::create_bounding_box_transform_layer<CLBoundingBoxTransform, CLTargetInfo>( + *polymorphic_downcast<BoundingBoxTransformLayerNode *>(node)); case NodeType::ChannelShuffleLayer: - return detail::create_channel_shuffle_layer<CLChannelShuffleLayer, CLTargetInfo>(*polymorphic_downcast<ChannelShuffleLayerNode *>(node)); + return detail::create_channel_shuffle_layer<CLChannelShuffleLayer, CLTargetInfo>( + *polymorphic_downcast<ChannelShuffleLayerNode *>(node)); case NodeType::ConvolutionLayer: - return detail::create_convolution_layer<CLConvolutionLayerFunctions, CLTargetInfo>(*polymorphic_downcast<ConvolutionLayerNode *>(node), ctx); + return detail::create_convolution_layer<CLConvolutionLayerFunctions, CLTargetInfo>( + *polymorphic_downcast<ConvolutionLayerNode *>(node), ctx); case NodeType::DeconvolutionLayer: - return detail::create_deconvolution_layer<CLDeconvolutionLayer, CLTargetInfo>(*polymorphic_downcast<DeconvolutionLayerNode *>(node), ctx); + return detail::create_deconvolution_layer<CLDeconvolutionLayer, CLTargetInfo>( + *polymorphic_downcast<DeconvolutionLayerNode *>(node), ctx); case NodeType::ConcatenateLayer: - return detail::create_concatenate_layer<CLConcatenateLayer, CLTargetInfo>(*polymorphic_downcast<ConcatenateLayerNode *>(node)); + return detail::create_concatenate_layer<CLConcatenateLayer, CLTargetInfo>( + *polymorphic_downcast<ConcatenateLayerNode *>(node)); + case NodeType::DepthToSpaceLayer: + return detail::create_depth_to_space_layer<CLDepthToSpaceLayer, CLTargetInfo>( + *polymorphic_downcast<DepthToSpaceLayerNode *>(node)); case NodeType::DepthwiseConvolutionLayer: - return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayer, CLTargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node)); + return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayer, CLTargetInfo>( + *polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node)); case NodeType::DequantizationLayer: - return detail::create_dequantization_layer<CLDequantizationLayer, CLTargetInfo>(*polymorphic_downcast<DequantizationLayerNode *>(node)); + return detail::create_dequantization_layer<CLDequantizationLayer, CLTargetInfo>( + *polymorphic_downcast<DequantizationLayerNode *>(node)); case NodeType::DetectionOutputLayer: - return detail::create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node)); + return detail::create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>( + *polymorphic_downcast<DetectionOutputLayerNode *>(node)); case NodeType::DetectionPostProcessLayer: - return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node)); + return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>( + *polymorphic_downcast<DetectionPostProcessLayerNode *>(node)); case NodeType::EltwiseLayer: - return detail::create_eltwise_layer<CLEltwiseFunctions, CLTargetInfo>(*polymorphic_downcast<EltwiseLayerNode *>(node)); + return detail::create_eltwise_layer<CLEltwiseFunctions, CLTargetInfo>( + *polymorphic_downcast<EltwiseLayerNode *>(node)); + case NodeType::UnaryEltwiseLayer: + return detail::create_unary_eltwise_layer<CLUnaryEltwiseFunctions, CLTargetInfo>( + *polymorphic_downcast<UnaryEltwiseLayerNode *>(node)); case NodeType::FlattenLayer: - return detail::create_flatten_layer<CLFlattenLayer, CLTargetInfo>(*polymorphic_downcast<FlattenLayerNode *>(node)); + return detail::create_flatten_layer<CLFlattenLayer, CLTargetInfo>( + *polymorphic_downcast<FlattenLayerNode *>(node)); case NodeType::FullyConnectedLayer: - return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx); + return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>( + *polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx); case NodeType::FusedConvolutionBatchNormalizationLayer: - return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node), ctx); + return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>( + *polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node), ctx); case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer: - return detail::create_fused_depthwise_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node), ctx); + return detail::create_fused_depthwise_convolution_batch_normalization_layer<CLFusedLayerTypes, + CLTargetInfo>( + *polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node), ctx); case NodeType::GenerateProposalsLayer: - return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx); + return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>( + *polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx); + case NodeType::L2NormalizeLayer: + return detail::create_l2_normalize_layer<CLL2NormalizeLayer, CLTargetInfo>( + *polymorphic_downcast<L2NormalizeLayerNode *>(node), ctx); case NodeType::NormalizationLayer: - return detail::create_normalization_layer<CLNormalizationLayer, CLTargetInfo>(*polymorphic_downcast<NormalizationLayerNode *>(node), ctx); + return detail::create_normalization_layer<CLNormalizationLayer, CLTargetInfo>( + *polymorphic_downcast<NormalizationLayerNode *>(node), ctx); case NodeType::NormalizePlanarYUVLayer: - return detail::create_normalize_planar_yuv_layer<CLNormalizePlanarYUVLayer, CLTargetInfo>(*polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node)); + return detail::create_normalize_planar_yuv_layer<CLNormalizePlanarYUVLayer, CLTargetInfo>( + *polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node)); case NodeType::PadLayer: return detail::create_pad_layer<CLPadLayer, CLTargetInfo>(*polymorphic_downcast<PadLayerNode *>(node)); case NodeType::PermuteLayer: - return detail::create_permute_layer<CLPermute, CLTargetInfo>(*polymorphic_downcast<PermuteLayerNode *>(node)); + return detail::create_permute_layer<CLPermute, CLTargetInfo>( + *polymorphic_downcast<PermuteLayerNode *>(node)); case NodeType::PoolingLayer: - return detail::create_pooling_layer<CLPoolingLayer, CLTargetInfo>(*polymorphic_downcast<PoolingLayerNode *>(node)); + return detail::create_pooling_layer<CLPoolingLayer, CLTargetInfo>( + *polymorphic_downcast<PoolingLayerNode *>(node)); case NodeType::PReluLayer: - return detail::create_prelu_layer<CLPReluLayer, CLTargetInfo>(*polymorphic_downcast<PReluLayerNode *>(node)); + return detail::create_prelu_layer<CLPReluLayer, CLTargetInfo>( + *polymorphic_downcast<PReluLayerNode *>(node)); case NodeType::PrintLayer: return detail::create_print_layer<CLTargetInfo>(*polymorphic_downcast<PrintLayerNode *>(node)); case NodeType::PriorBoxLayer: - return detail::create_priorbox_layer<CLPriorBoxLayer, CLTargetInfo>(*polymorphic_downcast<PriorBoxLayerNode *>(node)); + return detail::create_priorbox_layer<CLPriorBoxLayer, CLTargetInfo>( + *polymorphic_downcast<PriorBoxLayerNode *>(node)); case NodeType::QuantizationLayer: - return detail::create_quantization_layer<CLQuantizationLayer, CLTargetInfo>(*polymorphic_downcast<QuantizationLayerNode *>(node)); + return detail::create_quantization_layer<CLQuantizationLayer, CLTargetInfo>( + *polymorphic_downcast<QuantizationLayerNode *>(node)); + case NodeType::ReductionOperationLayer: + return detail::create_reduction_operation_layer<CLReductionOperation, CLTargetInfo>( + *polymorphic_downcast<ReductionLayerNode *>(node), ctx); case NodeType::ReorgLayer: - return detail::create_reorg_layer<CLReorgLayer, CLTargetInfo>(*polymorphic_downcast<ReorgLayerNode *>(node)); + return detail::create_reorg_layer<CLReorgLayer, CLTargetInfo>( + *polymorphic_downcast<ReorgLayerNode *>(node)); case NodeType::ReshapeLayer: - return detail::create_reshape_layer<CLReshapeLayer, CLTargetInfo>(*polymorphic_downcast<ReshapeLayerNode *>(node)); + return detail::create_reshape_layer<CLReshapeLayer, CLTargetInfo>( + *polymorphic_downcast<ReshapeLayerNode *>(node)); case NodeType::ResizeLayer: return detail::create_resize_layer<CLScale, CLTargetInfo>(*polymorphic_downcast<ResizeLayerNode *>(node)); case NodeType::ROIAlignLayer: - return detail::create_roi_align_layer<CLROIAlignLayer, CLTargetInfo>(*polymorphic_downcast<ROIAlignLayerNode *>(node)); + return detail::create_roi_align_layer<CLROIAlignLayer, CLTargetInfo>( + *polymorphic_downcast<ROIAlignLayerNode *>(node)); case NodeType::SliceLayer: return detail::create_slice_layer<CLSlice, CLTargetInfo>(*polymorphic_downcast<SliceLayerNode *>(node)); case NodeType::SoftmaxLayer: - return detail::create_softmax_layer<CLSoftmaxLayer, CLTargetInfo>(*polymorphic_downcast<SoftmaxLayerNode *>(node), ctx); + return detail::create_softmax_layer<CLSoftmaxLayer, CLTargetInfo>( + *polymorphic_downcast<SoftmaxLayerNode *>(node), ctx); case NodeType::StackLayer: - return detail::create_stack_layer<CLStackLayer, CLTargetInfo>(*polymorphic_downcast<StackLayerNode *>(node)); - case NodeType::UpsampleLayer: - return detail::create_upsample_layer<CLUpsampleLayer, CLTargetInfo>(*polymorphic_downcast<UpsampleLayerNode *>(node), ctx); - case NodeType::YOLOLayer: - return detail::create_yolo_layer<CLYOLOLayer, CLTargetInfo>(*polymorphic_downcast<YOLOLayerNode *>(node), ctx); + return detail::create_stack_layer<CLStackLayer, CLTargetInfo>( + *polymorphic_downcast<StackLayerNode *>(node)); + case NodeType::StridedSliceLayer: + return detail::create_strided_slice_layer<CLStridedSlice, CLTargetInfo>( + *polymorphic_downcast<StridedSliceLayerNode *>(node)); default: return nullptr; } diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp index ddb8e3d1ac..510eda7935 100644 --- a/src/graph/backends/CL/CLNodeValidator.cpp +++ b/src/graph/backends/CL/CLNodeValidator.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 ARM Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,11 +25,11 @@ #include "arm_compute/graph/backends/ValidateHelpers.h" #include "arm_compute/graph/nodes/Nodes.h" - -#include "arm_compute/core/utils/misc/Cast.h" #include "arm_compute/runtime/CL/CLFunctions.h" #include "arm_compute/runtime/CPP/CPPFunctions.h" +#include "support/Cast.h" + using namespace arm_compute::utils::cast; namespace arm_compute @@ -38,37 +38,69 @@ namespace graph { namespace backends { +/** Collection of CL element-wise functions */ +struct CLEltwiseLayerFunctions +{ + using ArithmeticAddition = CLArithmeticAddition; + using ArithmeticSubtraction = CLArithmeticSubtraction; + using PixelWiseMultiplication = CLPixelWiseMultiplication; + using ElementwiseMax = CLElementwiseMax; + using ArithmeticDivision = CLArithmeticDivision; +}; + +/** Collection of CL unary element-wise functions */ +struct CLUnaryEltwiseLayerFunctions +{ + using ExpLayer = CLExpLayer; +}; + Status CLNodeValidator::validate(INode *node) { - if(node == nullptr) + if (node == nullptr) { return Status{}; } NodeType type = node->type(); - switch(type) + switch (type) { + case NodeType::ArgMinMaxLayer: + return detail::validate_arg_min_max_layer<CLArgMinMaxLayer>( + *polymorphic_downcast<ArgMinMaxLayerNode *>(node)); case NodeType::BoundingBoxTransformLayer: - return detail::validate_bounding_box_transform_layer<CLBoundingBoxTransform>(*polymorphic_downcast<BoundingBoxTransformLayerNode *>(node)); + return detail::validate_bounding_box_transform_layer<CLBoundingBoxTransform>( + *polymorphic_downcast<BoundingBoxTransformLayerNode *>(node)); case NodeType::ChannelShuffleLayer: - return detail::validate_channel_shuffle_layer<CLChannelShuffleLayer>(*polymorphic_downcast<ChannelShuffleLayerNode *>(node)); + return detail::validate_channel_shuffle_layer<CLChannelShuffleLayer>( + *polymorphic_downcast<ChannelShuffleLayerNode *>(node)); case NodeType::ConvolutionLayer: - return detail::validate_convolution_layer<CLConvolutionLayer, - CLDirectConvolutionLayer, - CLGEMMConvolutionLayer, - CLWinogradConvolutionLayer>(*polymorphic_downcast<ConvolutionLayerNode *>(node)); + return detail::validate_convolution_layer<CLConvolutionLayer, CLDirectConvolutionLayer, + CLGEMMConvolutionLayer, CLWinogradConvolutionLayer>( + *polymorphic_downcast<ConvolutionLayerNode *>(node)); + case NodeType::DepthToSpaceLayer: + return detail::validate_depth_to_space_layer<CLDepthToSpaceLayer>( + *polymorphic_downcast<DepthToSpaceLayerNode *>(node)); case NodeType::DepthwiseConvolutionLayer: - return detail::validate_depthwise_convolution_layer<CLDepthwiseConvolutionLayer>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node)); + return detail::validate_depthwise_convolution_layer<CLDepthwiseConvolutionLayer>( + *polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node)); case NodeType::DequantizationLayer: - return detail::validate_dequantization_layer<CLDequantizationLayer>(*polymorphic_downcast<DequantizationLayerNode *>(node)); + return detail::validate_dequantization_layer<CLDequantizationLayer>( + *polymorphic_downcast<DequantizationLayerNode *>(node)); case NodeType::DetectionOutputLayer: - return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node)); + return detail::validate_detection_output_layer<CPPDetectionOutputLayer>( + *polymorphic_downcast<DetectionOutputLayerNode *>(node)); case NodeType::DetectionPostProcessLayer: - return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node)); + return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>( + *polymorphic_downcast<DetectionPostProcessLayerNode *>(node)); case NodeType::GenerateProposalsLayer: - return detail::validate_generate_proposals_layer<CLGenerateProposalsLayer>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node)); + return detail::validate_generate_proposals_layer<CLGenerateProposalsLayer>( + *polymorphic_downcast<GenerateProposalsLayerNode *>(node)); + case NodeType::L2NormalizeLayer: + return detail::validate_l2_normalize_layer<CLL2NormalizeLayer>( + *polymorphic_downcast<L2NormalizeLayerNode *>(node)); case NodeType::NormalizePlanarYUVLayer: - return detail::validate_normalize_planar_yuv_layer<CLNormalizePlanarYUVLayer>(*polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node)); + return detail::validate_normalize_planar_yuv_layer<CLNormalizePlanarYUVLayer>( + *polymorphic_downcast<NormalizePlanarYUVLayerNode *>(node)); case NodeType::PadLayer: return detail::validate_pad_layer<CLPadLayer>(*polymorphic_downcast<PadLayerNode *>(node)); case NodeType::PermuteLayer: @@ -78,7 +110,11 @@ Status CLNodeValidator::validate(INode *node) case NodeType::PriorBoxLayer: return detail::validate_priorbox_layer<CLPriorBoxLayer>(*polymorphic_downcast<PriorBoxLayerNode *>(node)); case NodeType::QuantizationLayer: - return detail::validate_quantization_layer<CLQuantizationLayer>(*polymorphic_downcast<QuantizationLayerNode *>(node)); + return detail::validate_quantization_layer<CLQuantizationLayer>( + *polymorphic_downcast<QuantizationLayerNode *>(node)); + case NodeType::ReductionOperationLayer: + return detail::validate_reduction_operation_layer<CLReductionOperation>( + *polymorphic_downcast<ReductionLayerNode *>(node)); case NodeType::ReorgLayer: return detail::validate_reorg_layer<CLReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node)); case NodeType::ReshapeLayer: @@ -87,10 +123,15 @@ Status CLNodeValidator::validate(INode *node) return detail::validate_roi_align_layer<CLROIAlignLayer>(*polymorphic_downcast<ROIAlignLayerNode *>(node)); case NodeType::SliceLayer: return detail::validate_slice_layer<CLSlice>(*polymorphic_downcast<SliceLayerNode *>(node)); - case NodeType::UpsampleLayer: - return detail::validate_upsample_layer<CLUpsampleLayer>(*polymorphic_downcast<UpsampleLayerNode *>(node)); - case NodeType::YOLOLayer: - return detail::validate_yolo_layer<CLYOLOLayer>(*polymorphic_downcast<YOLOLayerNode *>(node)); + case NodeType::StridedSliceLayer: + return detail::validate_strided_slice_layer<CLStridedSlice>( + *polymorphic_downcast<StridedSliceLayerNode *>(node)); + case NodeType::EltwiseLayer: + return detail::validate_eltwise_Layer<CLEltwiseLayerFunctions>( + *polymorphic_downcast<EltwiseLayerNode *>(node)); + case NodeType::UnaryEltwiseLayer: + return detail::validate_unary_eltwise_layer<CLUnaryEltwiseLayerFunctions>( + *polymorphic_downcast<UnaryEltwiseLayerNode *>(node)); default: return Status{}; } diff --git a/src/graph/backends/CL/CLSubTensorHandle.cpp b/src/graph/backends/CL/CLSubTensorHandle.cpp index 016dca753b..ccdc877a18 100644 --- a/src/graph/backends/CL/CLSubTensorHandle.cpp +++ b/src/graph/backends/CL/CLSubTensorHandle.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,7 +23,7 @@ */ #include "arm_compute/graph/backends/CL/CLSubTensorHandle.h" -#include "arm_compute/core/utils/misc/Cast.h" +#include "support/Cast.h" namespace arm_compute { @@ -31,7 +31,10 @@ namespace graph { namespace backends { -CLSubTensorHandle::CLSubTensorHandle(ITensorHandle *parent_handle, const TensorShape &shape, const Coordinates &coords, bool extend_parent) +CLSubTensorHandle::CLSubTensorHandle(ITensorHandle *parent_handle, + const TensorShape &shape, + const Coordinates &coords, + bool extend_parent) : _sub_tensor(), _parent_handle(nullptr) { ARM_COMPUTE_ERROR_ON(!parent_handle); @@ -98,4 +101,4 @@ Target CLSubTensorHandle::target() const } } // namespace backends } // namespace graph -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/src/graph/backends/CL/CLTensorHandle.cpp b/src/graph/backends/CL/CLTensorHandle.cpp index 891c784cbe..1b69f9dede 100644 --- a/src/graph/backends/CL/CLTensorHandle.cpp +++ b/src/graph/backends/CL/CLTensorHandle.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -31,8 +31,7 @@ namespace graph { namespace backends { -CLTensorHandle::CLTensorHandle(const ITensorInfo &info) - : _tensor() +CLTensorHandle::CLTensorHandle(const ITensorInfo &info) : _tensor() { _tensor.allocator()->init(info); } @@ -49,7 +48,7 @@ void CLTensorHandle::free() void CLTensorHandle::manage(IMemoryGroup *mg) { - if(mg != nullptr) + if (mg != nullptr) { mg->manage(&_tensor); } @@ -68,7 +67,7 @@ void CLTensorHandle::unmap() void CLTensorHandle::release_if_unused() { // TODO (geopin01): Release tensor only if all sub-tensors are marked as not used - if(!_tensor.is_used()) + if (!_tensor.is_used()) { _tensor.allocator()->free(); } @@ -100,4 +99,4 @@ Target CLTensorHandle::target() const } } // namespace backends } // namespace graph -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute |