diff options
author | ramelg01 <ramy.elgammal@arm.com> | 2021-11-26 19:12:40 +0000 |
---|---|---|
committer | Ramy Elgammal <ramy.elgammal@arm.com> | 2021-12-09 13:55:06 +0000 |
commit | b75d62430e9871fcc6f19cf82879f65d2e7fb201 (patch) | |
tree | 5914cb360f90f1f34ca1eb27ef6946b4b55e257a /arm_compute/graph/nodes | |
parent | 78baa48308cba4101b4bcb4680f2f4ca90aeefd7 (diff) | |
download | ComputeLibrary-b75d62430e9871fcc6f19cf82879f65d2e7fb201.tar.gz |
Graph Fusion With Post Ops Fix
- Fusing ConvolutionBatchNormalization Nodes with post ops (activation
or element wise ops)
Resolves: COMPMID-4982
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I5b2d32cad00f710fd744cb5aa2d59fd7e5c97e0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6766
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Diffstat (limited to 'arm_compute/graph/nodes')
4 files changed, 131 insertions, 2 deletions
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h index b3661c300f..b0051b1385 100644 --- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h +++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2019, 2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -100,7 +100,7 @@ public: */ ConvolutionMethod convolution_method() const; - /** Sets the fast math fast hint + /** Sets the fast math hint * * @param[in] hint Hint to use for convolution */ diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h new file mode 100644 index 0000000000..a42e06d889 --- /dev/null +++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h @@ -0,0 +1,127 @@ +/* + * 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_FUSED_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H +#define ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H + +#include "arm_compute/graph/INode.h" + +namespace arm_compute +{ +namespace graph +{ +/** Batch Normalization node */ +class FusedConvolutionBatchNormalizationWithPostOpsNode final : public INode +{ +public: + /** Constructor + * + * @param[in] epsilon Epsilon parameter. + * @param[in] info Convolution layer attributes. + * @param[in] num_groups (Optional) Number of groups (Defaults to 1) + * @param[in] method (Optional) Convolution method to use + * @param[in] fast_math_hint (Optional) Fast math hint + */ + FusedConvolutionBatchNormalizationWithPostOpsNode(float epsilon, PadStrideInfo info, + unsigned int num_groups = 1, + ConvolutionMethod method = ConvolutionMethod::Default, + FastMathHint fast_math_hint = FastMathHint::Disabled); + + /** Epsilon parameter accessor + * + * @return Epsilon parameter + */ + float epsilon() const; + + /** Computes convolution output descriptor + * + * @param[in] input_descriptor Input descriptor + * @param[in] weights_descriptor Weights descriptor + * @param[in] info Convolution operation attributes + * + * @return Output descriptor + */ + static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor, + const TensorDescriptor &weights_descriptor, + const PadStrideInfo &info); + + /** Sets the convolution layer method to use + * + * @param[in] method Method to use for convolution + */ + void set_convolution_method(ConvolutionMethod method); + + /** Number of groups in convolution accessor + * + * @return Number of groups in convolution + */ + unsigned int num_groups() const; + + /** Convolution layer method accessor + * + * @note This is an indication on which convolution layer implementation to use, + * if it fails to be created the library's heuristic approach will be used + * + * @return Convolution layer method to be used by the node + */ + ConvolutionMethod convolution_method() const; + + /** Sets the fast math hint + * + * @param[in] hint Hint to use for convolution + */ + void set_fast_math_hint(FastMathHint hint); + + /** Fast math hint accessor + * + * @return Fast math hint to be used by the node + */ + FastMathHint fast_math_hint() const; + + /** Convolution metadata accessor + * + * @return Convolution information + */ + PadStrideInfo convolution_info() const; + + // Inherited overridden methods: + NodeType type() const override; + bool forward_descriptors() override; + TensorDescriptor configure_output(size_t idx) const override; + void accept(INodeVisitor &v) override; + +public: + static constexpr NodeType node_type = NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer; + +private: + float _epsilon; + + PadStrideInfo _info; + unsigned int _num_groups; + ConvolutionMethod _method; + FastMathHint _fast_math_hint; +}; + +} // namespace graph +} // namespace arm_compute +#endif /* ARM_COMPUTE_GRAPH_BATCH_NORMALIZATION_LAYER_NODE_H */ diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h index fb0eb155fd..3887eaeac6 100644 --- a/arm_compute/graph/nodes/Nodes.h +++ b/arm_compute/graph/nodes/Nodes.h @@ -43,6 +43,7 @@ #include "arm_compute/graph/nodes/FlattenLayerNode.h" #include "arm_compute/graph/nodes/FullyConnectedLayerNode.h" #include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h" +#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h" #include "arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h" #include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h" #include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h" diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h index 6393b1d5d4..f1576d6336 100644 --- a/arm_compute/graph/nodes/NodesFwd.h +++ b/arm_compute/graph/nodes/NodesFwd.h @@ -51,6 +51,7 @@ class FullyConnectedLayerNode; class FusedConvolutionBatchNormalizationNode; class FusedConvolutionWithPostOpNode; class FusedDepthwiseConvolutionBatchNormalizationNode; +class FusedConvolutionBatchNormalizationWithPostOpsNode; class GenerateProposalsLayerNode; class InputNode; class L2NormalizeLayerNode; |