diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2019-11-08 12:13:48 +0000 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2019-11-12 11:52:32 +0000 |
commit | 5dea19e58a5521b05e95375c8618a37072697bc0 (patch) | |
tree | eac3414eb9a5f2290b3bb597e4c6d8e398678355 /arm_compute/graph/backends | |
parent | eaa01ab593428bc7267ebbe107b2d813a11b64b5 (diff) | |
download | ComputeLibrary-5dea19e58a5521b05e95375c8618a37072697bc0.tar.gz |
COMPMID-2579: Fuse batch normalization with convolution and depthwise convolution at graph level on NEON
Change-Id: Ib263a680bbd2dc1a4947102ee8d6da76b95f02bf
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2252
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'arm_compute/graph/backends')
3 files changed, 28 insertions, 12 deletions
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h index ee257e3abf..02bfe9dc22 100644 --- a/arm_compute/graph/backends/FunctionHelpers.h +++ b/arm_compute/graph/backends/FunctionHelpers.h @@ -174,11 +174,12 @@ std::unique_ptr<IFunction> create_batch_normalization_layer(BatchNormalizationLa * @tparam TargetInfo Target-specific information * * @param[in] node Node to create the backend function for + * @param[in] ctx Graph context * * @return Backend batch normalization layer function */ template <typename FusedLayerTypes, typename TargetInfo> -std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node) +std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node, GraphContext &ctx) { validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */); @@ -199,9 +200,16 @@ std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(Fu const ActivationLayerInfo fused_act = node.fused_activation(); const float epsilon = node.epsilon(); + // 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 = FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>; + // Create and configure function - auto func = support::cpp14::make_unique<FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>(); - func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act); + std::tie(func, func_name) = create_named_memory_managed_function<FType>( + std::string("FusedConvolutionBatchNormalizationLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act); // Log info ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " @@ -214,7 +222,7 @@ std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(Fu << " Output shape: " << output->info()->tensor_shape() << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") << std::endl); - return std::move(func); + return func; } /** Create a backend fused depthwise convolution batch normalization layer function @@ -223,11 +231,12 @@ std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(Fu * @tparam TargetInfo Target-specific information * * @param[in] node Node to create the backend function for + * @param[in] ctx Graph context * * @return Backend fused depthwise convolution batch normalization layer function */ template <typename FusedLayerTypes, typename TargetInfo> -std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node) +std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node, GraphContext &ctx) { validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */); @@ -247,9 +256,16 @@ std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalizatio const ActivationLayerInfo fused_act = node.fused_activation(); const float epsilon = node.epsilon(); + // 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 = FusedDepthwiseConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>; + // Create and configure function - auto func = support::cpp14::make_unique<FusedDepthwiseConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>(); - func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, depth_multiplier, fused_act); + std::tie(func, func_name) = create_named_memory_managed_function<FType>( + std::string("FusedDepthwiseConvolutionBatchNormalizationLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, depth_multiplier, fused_act); // Log info ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " @@ -262,7 +278,7 @@ std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalizatio << " Output shape: " << output->info()->tensor_shape() << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") << std::endl); - return std::move(func); + return func; } /** Create a backend bounding box transform layer function diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h index a6da76bb06..0af3abc547 100644 --- a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h +++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h @@ -42,8 +42,8 @@ public: using TensorType = typename TargetInfo::TensorType; using TensorConcreteType = typename TargetInfo::TensorConcreteType; - FusedConvolutionBatchNormalizationFunction() - : _conv_layer(), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) + FusedConvolutionBatchNormalizationFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) + : _conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) { } diff --git a/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h index 6f70d3c3a0..14474f4ee5 100644 --- a/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h +++ b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h @@ -42,8 +42,8 @@ public: using TensorType = typename TargetInfo::TensorType; using TensorConcreteType = typename TargetInfo::TensorConcreteType; - FusedDepthwiseConvolutionBatchNormalizationFunction() - : _depth_conv_layer(), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) + FusedDepthwiseConvolutionBatchNormalizationFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr) + : _depth_conv_layer(memory_manager), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) { } |