From d9eb27597eabe5b7c17520f4f9b3f8a282d72573 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 3 Apr 2018 13:44:29 +0100 Subject: COMPMID-797: Switch to new graph. - Cleaned up build system Change-Id: If2faa27ee5b31fa8b972836960ab3ef671059c8d Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/126435 Tested-by: Jenkins Reviewed-by: Pablo Tello --- src/graph/backends/NEON/NEFunctionFactory.cpp | 563 ++++++++++++++++++++++++++ 1 file changed, 563 insertions(+) create mode 100644 src/graph/backends/NEON/NEFunctionFactory.cpp (limited to 'src/graph/backends/NEON/NEFunctionFactory.cpp') diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp new file mode 100644 index 0000000000..228af9ca6f --- /dev/null +++ b/src/graph/backends/NEON/NEFunctionFactory.cpp @@ -0,0 +1,563 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/graph/backends/NEON/NEFunctionFactory.h" + +#include "arm_compute/core/utils/misc/Cast.h" +#include "arm_compute/graph/Graph.h" +#include "arm_compute/graph/GraphContext.h" +#include "arm_compute/graph/Logger.h" +#include "arm_compute/graph/TypePrinter.h" +#include "arm_compute/graph/backends/Utils.h" +#include "arm_compute/graph/nodes/Nodes.h" +#include "arm_compute/runtime/NEON/NEFunctions.h" +#include "support/ToolchainSupport.h" + +using namespace arm_compute::utils::cast; + +namespace arm_compute +{ +namespace graph +{ +namespace backends +{ +namespace +{ +/** Returns backing tensor of a given tensor + * + * @param[in] tensor Tensor to extract the backing tensor from + * + * @return Backing tensor if present else nullptr + */ +arm_compute::ITensor *get_backing_tensor(arm_compute::graph::Tensor *tensor) +{ + return ((tensor == nullptr) || (tensor->handle() == nullptr)) ? nullptr : &tensor->handle()->tensor(); +} + +/** Create a backend activation layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend activation layer function + */ +std::unique_ptr create_activation_layer(ActivationLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON ActivationLayerNode node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + const ActivationLayerInfo act_info = node.activation_info(); + + // Create function + auto func = support::cpp14::make_unique(); + func->configure(input, output, act_info); + + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEActivationLayer" + << " Data Type: " << input->info()->data_type() + << " Shape: " << input->info()->tensor_shape() + << " Activation function: " << act_info.activation() + << " a: " << act_info.a() + << " b: " << act_info.b() + << " InPlace : " << is_in_place_operation(input, output) + << std::endl); + + return std::move(func); +} + +/** Create a backend batch normalization layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend batch normalization layer function + */ +std::unique_ptr create_batch_normalization_layer(BatchNormalizationLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON BatchNormalization node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + + // TODO (geopin01) : Var and mean are compulsory, switch function to accept nullptr as beta and/or gamma + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 5); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *mean = get_backing_tensor(node.input(1)); + ITensor *var = get_backing_tensor(node.input(2)); + ITensor *beta = get_backing_tensor(node.input(3)); + ITensor *gamma = get_backing_tensor(node.input(4)); + ITensor *output = get_backing_tensor(node.output(0)); + const float epsilon = node.epsilon(); + const ActivationLayerInfo fused_act = node.fused_activation(); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(input, output, mean, var, beta, gamma, epsilon, fused_act); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEBatchNormalizationLayer" + << " Data Type: " << input->info()->data_type() + << " Shape: " << input->info()->tensor_shape() + << " Epsilon: " << epsilon << " " + << (fused_act.enabled() ? to_string(fused_act.activation()) : "") + << " InPlace : " << is_in_place_operation(input, output) + << std::endl); + + return std::move(func); +} + +/** Create a backend convolution layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend convolution layer function + */ +std::unique_ptr create_convolution_layer(ConvolutionLayerNode &node, GraphContext &ctx) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON ConvolutionLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 3); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *weights = get_backing_tensor(node.input(1)); + ITensor *biases = get_backing_tensor(node.input(2)); + ITensor *output = get_backing_tensor(node.output(0)); + const PadStrideInfo conv_info = node.convolution_info(); + const ConvolutionMethod conv_algorithm = node.convolution_method(); + + // Create and configure function (we assume that functions have been validated before creation) + std::shared_ptr mm = get_memory_manager(ctx, Target::NEON); + std::unique_ptr func; + std::string func_name; + if(conv_algorithm == ConvolutionMethod::DIRECT) + { + std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEDirectConvolutionLayer"), mm, + input, weights, biases, output, conv_info); + } + else if(conv_algorithm == ConvolutionMethod::GEMM) + { + std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEGEMMConvolutionLayer"), mm, + input, weights, biases, output, conv_info); + } + else if(conv_algorithm == ConvolutionMethod::WINOGRAD) + { + std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEWinogradLayer"), mm, + input, weights, biases, output, conv_info); + } + else + { + std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEConvolutionLayer"), mm, + input, weights, biases, output, conv_info); + } + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Weights shape: " << weights->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + return func; +} + +/** Create a backend layer depth concatenate function + * + * @param[in] node Node to create the backend function for + * + * @return Backend depth concatenate layer function + */ +std::unique_ptr create_depth_concatenate_layer(DepthConcatenateLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON DepthConcatenate node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Return nullptr if depth concatenate is switched off + if(!node.is_enabled()) + { + return nullptr; + } + + // Extract IO and info + std::vector inputs; + for(unsigned int i = 0; i < node.num_inputs(); ++i) + { + inputs.push_back(get_backing_tensor(node.input(i))); + } + ITensor *output = get_backing_tensor(node.output(0)); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(inputs, output); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEDepthConcatenateLayer" + << " Data Type: " << output->info()->data_type() + << " Shape: " << output->info()->tensor_shape() + << " Num Inputs: " << inputs.size() + << std::endl); + + return std::move(func); +} + +/** Create a backend layer depth-wise convolution function + * + * @param[in] node Node to create the backend function for + * + * @return Backend depth-wise convolution layer function + */ +std::unique_ptr create_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON DepthwiseConvolutionLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 3); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *weights = get_backing_tensor(node.input(1)); + ITensor *biases = get_backing_tensor(node.input(2)); + ITensor *output = get_backing_tensor(node.output(0)); + const PadStrideInfo conv_info = node.convolution_info(); + const DepthwiseConvolutionMethod dwc_algorithm = node.depthwise_convolution_method(); + + // Create and configure function (we assume that functions have been validated before creation) + std::unique_ptr func; + std::string func_name; + if(dwc_algorithm == DepthwiseConvolutionMethod::OPTIMIZED_3x3) + { + std::tie(func, func_name) = create_named_function(std::string("NEDepthwiseConvolutionLayer3x3"), + input, weights, biases, output, conv_info); + } + else + { + std::tie(func, func_name) = create_named_function(std::string("NEDepthwiseConvolutionLayer"), + input, weights, biases, output, conv_info); + } + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Weights shape: " << weights->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + return func; +} + +/** Create a backend element-wise operation layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend element-wise operation layer function + */ +std::unique_ptr create_eltwise_layer(EltwiseLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON EltwiseLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 2); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input1 = get_backing_tensor(node.input(0)); + ITensor *input2 = get_backing_tensor(node.input(1)); + ITensor *output = get_backing_tensor(node.output(0)); + const EltwiseOperation eltwise_op = node.eltwise_operation(); + ARM_COMPUTE_ERROR_ON(input1 == nullptr); + ARM_COMPUTE_ERROR_ON(input2 == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + std::unique_ptr func = nullptr; + std::string func_name; + if(eltwise_op == EltwiseOperation::ADD) + { + std::tie(func, func_name) = create_named_function(std::string("NEArithmeticAddition"), + input1, input2, output, ConvertPolicy::SATURATE); + } + else if(eltwise_op == EltwiseOperation::SUB) + { + std::tie(func, func_name) = create_named_function(std::string("NEArithmeticSubtraction"), + input1, input2, output, ConvertPolicy::SATURATE); + } + else if(eltwise_op == EltwiseOperation::MUL) + { + std::tie(func, func_name) = create_named_function(std::string("NEPixelWiseMultiplication"), + input1, input2, output, 1.f, + ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); + } + else + { + ARM_COMPUTE_ERROR("Unsupported element-wise operation!"); + } + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name + << " Data Type: " << input1->info()->data_type() + << " Shape : " << input1->info()->tensor_shape() + << std::endl); + + return func; +} + +/** Create a backend flatten layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend flatten layer function + */ +std::unique_ptr create_flatten_layer(FlattenLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON FlattenLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(input, output); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEFlattenLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + + return std::move(func); +} + +/** Create a backend fully connected layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend fully connected layer function + */ +std::unique_ptr create_fully_connected_layer(FullyConnectedLayerNode &node, GraphContext &ctx) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON FullyConnectedLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 3); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *weights = get_backing_tensor(node.input(1)); + ITensor *biases = get_backing_tensor(node.input(2)); + ITensor *output = get_backing_tensor(node.output(0)); + + // Create and configure function + auto func = support::cpp14::make_unique(get_memory_manager(ctx, Target::NEON)); + func->configure(input, weights, biases, output); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(weights == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEFullyConnectedLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Weights shape: " << weights->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + + return std::move(func); +} + +/** Create a backend normalization layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend normalization layer function + */ +std::unique_ptr create_normalization_layer(NormalizationLayerNode &node, GraphContext &ctx) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON NormalizationLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + const NormalizationLayerInfo norm_info = node.normalization_info(); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Create and configure function + auto func = support::cpp14::make_unique(get_memory_manager(ctx, Target::NEON)); + func->configure(input, output, norm_info); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NENormalizationLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << " Normalization info: " << norm_info.type() + << std::endl); + + return std::move(func); +} + +/** Create a backend pooling layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend pooling layer function + */ +std::unique_ptr create_pooling_layer(PoolingLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON PoolingLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + const PoolingLayerInfo pool_info = node.pooling_info(); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(input, output, pool_info); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEPoolingLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << " Pooling info: " << pool_info.pool_type() + << std::endl); + + return std::move(func); +} + +/** Create a backend reshape layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend reshape layer function + */ +std::unique_ptr create_reshape_layer(ReshapeLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON ReshapeLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(input, output); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NEReshapeLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + + return std::move(func); +} + +/** Create a backend softmax layer function + * + * @param[in] node Node to create the backend function for + * + * @return Backend softmax layer function + */ +std::unique_ptr create_softmax_layer(SoftmaxLayerNode &node, GraphContext &ctx) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating NEON SoftmaxLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + ITensor *input = get_backing_tensor(node.input(0)); + ITensor *output = get_backing_tensor(node.output(0)); + const float beta = node.beta(); + ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Create and configure function + auto func = support::cpp14::make_unique(get_memory_manager(ctx, Target::NEON)); + func->configure(input, output, beta); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated NESoftmaxLayer" + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << std::endl); + + return std::move(func); +} +} // namespace + +std::unique_ptr NEFunctionFactory::create(INode *node, GraphContext &ctx) +{ + if(node == nullptr) + { + return nullptr; + } + + NodeType type = node->type(); + switch(type) + { + case NodeType::ActivationLayer: + return create_activation_layer(*polymorphic_downcast(node)); + case NodeType::BatchNormalizationLayer: + return create_batch_normalization_layer(*polymorphic_downcast(node)); + case NodeType::ConvolutionLayer: + return create_convolution_layer(*polymorphic_downcast(node), ctx); + case NodeType::DepthConcatenateLayer: + return create_depth_concatenate_layer(*polymorphic_downcast(node)); + case NodeType::DepthwiseConvolutionLayer: + return create_depthwise_convolution_layer(*polymorphic_downcast(node)); + case NodeType::EltwiseLayer: + return create_eltwise_layer(*polymorphic_downcast(node)); + case NodeType::FlattenLayer: + return create_flatten_layer(*polymorphic_downcast(node)); + case NodeType::FullyConnectedLayer: + return create_fully_connected_layer(*polymorphic_downcast(node), ctx); + case NodeType::NormalizationLayer: + return create_normalization_layer(*polymorphic_downcast(node), ctx); + case NodeType::PoolingLayer: + return create_pooling_layer(*polymorphic_downcast(node)); + case NodeType::ReshapeLayer: + return create_reshape_layer(*polymorphic_downcast(node)); + case NodeType::SoftmaxLayer: + return create_softmax_layer(*polymorphic_downcast(node), ctx); + default: + return nullptr; + } +} +} // namespace backends +} // namespace graph +} // namespace arm_compute \ No newline at end of file -- cgit v1.2.1