diff options
Diffstat (limited to 'arm_compute/graph/backends')
3 files changed, 8 insertions, 360 deletions
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h index 803283e20d..a567427bf1 100644 --- a/arm_compute/graph/backends/FunctionHelpers.h +++ b/arm_compute/graph/backends/FunctionHelpers.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,18 +21,15 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H -#define ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H +#ifndef ACL_ARM_COMPUTE_GRAPH_BACKENDS_FUNCTIONHELPERS_H +#define ACL_ARM_COMPUTE_GRAPH_BACKENDS_FUNCTIONHELPERS_H -#include "arm_compute/core/experimental/IPostOp.h" -#include "arm_compute/core/experimental/PostOps.h" #include "arm_compute/graph/Logger.h" #include "arm_compute/graph/Tensor.h" #include "arm_compute/graph/TypePrinter.h" #include "arm_compute/graph/Types.h" #include "arm_compute/graph/Utils.h" #include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h" -#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h" #include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h" #include "arm_compute/graph/backends/Utils.h" #include "arm_compute/graph/nodes/Nodes.h" @@ -541,183 +538,6 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node, return std::move(func); } -/** Create a backend convolution layer function with post operator - * - * @tparam ConvolutionLayerFunctions Backend convolution functions - * @tparam TargetInfo Target-specific information - * - * @param[in] node Node to create the backend function for - * @param[in] ctx Graph context - * - * @return Backend convolution layer function - */ -template <typename ConvolutionLayerFunctions, typename TargetInfo> -std::unique_ptr<IFunction> create_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node, GraphContext &ctx) -{ - validate_node<TargetInfo>(node, 4 /* expected inputs */, 1 /* expected outputs */); - - // Extract IO and info - typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0)); - typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1)); - typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2)); - typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0)); - - const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); - - if(is_quantized) - { - biases->info()->set_data_type(DataType::S32); - } - - const PadStrideInfo conv_info = node.convolution_info(); - const unsigned int num_groups = node.num_groups(); - const ActivationLayerInfo fused_act = node.fused_activation(); - - experimental::PostOpList<typename TargetInfo::TensorType *> post_ops; - - auto &post_op_info_list = node.post_op_info_list(); - for(const auto &post_op_info : post_op_info_list) - { - switch(post_op_info->type()) - { - case PostOpType::Activation: - { - const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get()); - post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act); - break; - } - case PostOpType::Eltwise_Add: - { - typename TargetInfo::TensorType *add_input = get_backing_tensor<TargetInfo>(node.input(3)); - const auto eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get()); - post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy); - break; - } - default: - { - ARM_COMPUTE_ERROR("Unsupported PostOpType"); - } - } - } - - // Create and configure function (we assume that functions have been validated before creation) - std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType); - std::unique_ptr<IFunction> func; - std::string func_name; - - // Fuse convolution with post ops is only supported for conv1x1, which is only implemented as gemmconv2d - std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::GEMMConvolutionLayer>( - std::string("GEMMConvolutionLayer"), mm, - input, weights, biases, output, conv_info, - WeightsInfo(), Size2D(1U, 1U), fused_act, num_groups, post_ops); - - // Log info - std::ostringstream qss; - if(is_quantized) - { - qss << " Input QuantInfo: " << input->info()->quantization_info() - << " Weights QuantInfo: " << weights->info()->quantization_info() - << " Output QuantInfo: " << output->info()->quantization_info(); - } - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " - << node.name() - << " Type: " << func_name - << " Target: " << TargetInfo::TargetType - << " Data Type: " << input->info()->data_type() - << " Groups: " << num_groups - << " Input shape: " << input->info()->tensor_shape() - << " Weights shape: " << weights->info()->tensor_shape() - << " Output shape: " << output->info()->tensor_shape() - << qss.str() - << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") - << " Post ops" << post_ops - << std::endl); - return std::move(func); -} - -/** Create a backend convolution batch normalization layer function with post operator - * - * @tparam FusedLayerTypes Backend convolution functions - * @tparam TargetInfo Target-specific information - * - * @param[in] node Node to create the backend function for - * @param[in] ctx Graph context - * - * @return Backend fused convolution with batch normalization layer function - */ -template <typename FusedLayerTypes, typename TargetInfo> -std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_with_post_op(FusedConvolutionBatchNormalizationWithPostOpsNode &node, GraphContext &ctx) -{ - validate_node<TargetInfo>(node, 8 /* expected inputs */, 1 /* expected outputs */); - - // Extract IO and info - typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0)); - typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1)); - typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2)); - typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(3)); - typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(4)); - typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(5)); - typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(6)); - - typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0)); - - const PadStrideInfo conv_info = node.convolution_info(); - const unsigned int num_groups = node.num_groups(); - const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled; - const float epsilon = node.epsilon(); - - experimental::PostOpList<typename TargetInfo::TensorType *> post_ops; - - auto &post_op_info_list = node.post_op_info_list(); - for(const auto &post_op_info : post_op_info_list) - { - switch(post_op_info->type()) - { - case PostOpType::Activation: - { - const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get()); - post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act); - break; - } - case PostOpType::Eltwise_Add: - { - typename TargetInfo::TensorType *add_input = get_backing_tensor<TargetInfo>(node.input(3)); - const auto eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get()); - post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy); - break; - } - default: - { - ARM_COMPUTE_ERROR("Unsupported PostOpType"); - } - } - } - - // Create and configure function (we assume that functions have been validated before creation) - std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType); - std::unique_ptr<IFunction> func; - std::string func_name; - - using FType = FusedConvolutionBatchNormalizationWithPostOpsFunction<TargetInfo, FusedLayerTypes>; - - // Create and configure function - std::tie(func, func_name) = create_named_memory_managed_function<FType>( - std::string("FusedConvolutionBatchNormalizationLayerWithPostOpsLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, post_ops); - - // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " - << node.name() - << " Type: " << node.type() - << " Target: " << TargetInfo::TargetType - << " Data Type: " << input->info()->data_type() - << " Input shape: " << input->info()->tensor_shape() - << " Weights shape: " << weights->info()->tensor_shape() - << " Output shape: " << output->info()->tensor_shape() - << " Post Ops:" << post_ops - << std::endl); - return std::move(func); -} - /** Create a backend deconvolution layer function * * @tparam DeconvolutionLayerFunction Backend deconvolution function @@ -2025,4 +1845,4 @@ std::unique_ptr<IFunction> create_strided_slice_layer(StridedSliceLayerNode &nod } // namespace graph } // namespace arm_compute -#endif /* ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H */ +#endif // ACL_ARM_COMPUTE_GRAPH_BACKENDS_FUNCTIONHELPERS_H diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h deleted file mode 100644 index 10f2e5c25e..0000000000 --- a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h +++ /dev/null @@ -1,136 +0,0 @@ -/* - * Copyright (c) 2021 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. - */ - -#ifndef ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H -#define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H - -#include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/IPostOp.h" -#include "arm_compute/runtime/IFunction.h" - -namespace arm_compute -{ -namespace graph -{ -namespace backends -{ -/** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}ConvolutionLayer with the modified weights */ -template <typename TargetInfo, typename FusedLayerTypes> -class FusedConvolutionBatchNormalizationWithPostOpsFunction : public IFunction -{ -public: - using TensorType = typename TargetInfo::TensorType; - using TensorConcreteType = typename TargetInfo::TensorConcreteType; - - FusedConvolutionBatchNormalizationWithPostOpsFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) - : _conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) - { - } - - /** Set the input and output tensors. - * - * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], - * while every optional dimension from 4 and above represent a batch of inputs. - * Data types supported: QASYMM8/F16/F32. - * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. - * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. - * Data type supported: Should match @p input data type. - * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. - * Data types supported: Same as @p input. - * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input - * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input - * @param[in] epsilon Small value to avoid division with zero. Default value is 0.001f. - * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. - * @param[in] num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout - * @param[in] fast_math Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation - * available which may introduce a drop of accuracy as well. Default is false - * @param[in] post_ops A sequence of post operations that are performed after the main operation. - * - */ - void configure(TensorType *input, - TensorType *weights, - TensorType *bias, - TensorType *output, - const TensorType *mean, - const TensorType *var, - const TensorType *beta, - const TensorType *gamma, - float epsilon, const PadStrideInfo &conv_info, unsigned int num_groups, bool fast_math, - const arm_compute::experimental::PostOpList<TensorType *> &post_ops = experimental::PostOpList<TensorType *> {}) - { - // We don't run any validate, as we assume that the layers have been already validated - const bool has_bias = (bias != nullptr); - const TensorType *bias_to_use; - - // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one - // as batch normalization might end up with a bias != 0 - if(has_bias) - { - _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon); - bias_to_use = bias; - } - else - { - _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon); - bias_to_use = &_fused_bias; - } - - ActivationLayerInfo fused_act = ActivationLayerInfo(); // Passing an empty ActivationLayerInfo. - _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups, post_ops); - - if(!has_bias) - { - _fused_bias.allocator()->allocate(); - } - } - - // Inherited methods overridden: - void run() - { - prepare(); - _conv_layer.run(); - } - - void prepare() - { - if(!_is_prepared) - { - _fused_batch_norm_layer.run(); - _is_prepared = true; - } - } - -private: - typename FusedLayerTypes::ConvolutionLayer _conv_layer; - typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer; - TensorConcreteType _fused_bias; - bool _is_prepared; -}; -} // namespace backends -} // namespace graph -} // namespace arm_compute - -#endif /* ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H */ diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h index 89dccd88b7..71a6201554 100644 --- a/arm_compute/graph/backends/ValidateHelpers.h +++ b/arm_compute/graph/backends/ValidateHelpers.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_VALIDATE_HELPERS_H -#define ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_VALIDATE_HELPERS_H +#ifndef ACL_ARM_COMPUTE_GRAPH_BACKENDS_VALIDATEHELPERS_H +#define ACL_ARM_COMPUTE_GRAPH_BACKENDS_VALIDATEHELPERS_H #include "arm_compute/graph/Logger.h" #include "arm_compute/graph/Tensor.h" @@ -183,42 +183,6 @@ Status validate_convolution_layer(ConvolutionLayerNode &node) return status; } -/** Validates a Convolution layer node - * - * @tparam GEMMConvolutionLayer GEMM Convolution layer function type - * - * @param[in] node Node to validate - * - * @return Status - */ -template <typename GEMMConvolutionLayer> -Status validate_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node) -{ - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Validating fused ConvolutionLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); - ARM_COMPUTE_RETURN_ERROR_ON(node.num_inputs() != 4); - ARM_COMPUTE_RETURN_ERROR_ON(node.num_outputs() != 1); - - // Extract IO and info - arm_compute::ITensorInfo *input = get_backing_tensor_info(node.input(0)); - arm_compute::ITensorInfo *weights = get_backing_tensor_info(node.input(1)); - arm_compute::ITensorInfo *biases = get_backing_tensor_info(node.input(2)); - arm_compute::ITensorInfo *output = get_backing_tensor_info(node.output(0)); - - if(is_data_type_quantized_asymmetric(input->data_type())) - { - biases->set_data_type(DataType::S32); - } - - const PadStrideInfo conv_info = node.convolution_info(); - //const ConvolutionMethod conv_algorithm = node.convolution_method(); - //const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled; - const unsigned int num_groups = node.num_groups(); - - // Validate function - return GEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, - WeightsInfo(), Size2D(1, 1), ActivationLayerInfo(), num_groups); -} - /** Validates a Depthwise Convolution layer node * * @tparam DepthwiseConvolutionLayer Default Depthwise Convolution layer type @@ -775,4 +739,4 @@ Status validate_unary_eltwise_layer(UnaryEltwiseLayerNode &node) } // namespace graph } // namespace arm_compute -#endif /* ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_VALIDATE_HELPERS_H */ +#endif // ACL_ARM_COMPUTE_GRAPH_BACKENDS_VALIDATEHELPERS_H |