COMPMID-2273: Fuse Batch Normalization with Depthwise Convolution layer at graph level (only for CL)
Change-Id: I1d941c6e66722f39583bf68148c980bb28ff89a1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1423
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
index 9b0f5b7..c124c982 100644
--- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
@@ -41,14 +41,13 @@
* @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
- * @param[in] out_quant_info (Optional) Output quantization info
* @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
*/
FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info,
unsigned int num_groups = 1,
ConvolutionMethod method = ConvolutionMethod::Default,
FastMathHint fast_math_hint = FastMathHint::Disabled,
- QuantizationInfo out_quant_info = QuantizationInfo(), ActivationLayerInfo fused_activation = ActivationLayerInfo());
+ ActivationLayerInfo fused_activation = ActivationLayerInfo());
/** Epsilon parameter accessor
*
@@ -135,7 +134,6 @@
unsigned int _num_groups;
ConvolutionMethod _method;
FastMathHint _fast_math_hint;
- QuantizationInfo _out_quant_info;
ActivationLayerInfo _fused_activation;
};
diff --git a/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
new file mode 100644
index 0000000..a2241ef
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2019 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_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+#define __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+
+#include "arm_compute/graph/INode.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Fused Depthwise Convolution Batch Normalization node */
+class FusedDepthwiseConvolutionBatchNormalizationNode final : public INode
+{
+public:
+ /** Constructor
+ *
+ * @param[in] epsilon Epsilon parameter.
+ * @param[in] info Convolution layer attributes.
+ * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] method (Optional) Convolution method to use
+ * @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
+ */
+ FusedDepthwiseConvolutionBatchNormalizationNode(float epsilon,
+ PadStrideInfo info,
+ unsigned int depth_multiplier,
+ DepthwiseConvolutionMethod method,
+ ActivationLayerInfo fused_activation = ActivationLayerInfo());
+
+ /** Sets the depthwise convolution layer method to use
+ *
+ * @param[in] method Method to use for depthwise convolution
+ */
+ void set_depthwise_convolution_method(DepthwiseConvolutionMethod method);
+
+ /** Depthwise convolution layer method accessor
+ *
+ * @note This is an indication on which depthwise convolution layer implementation to use,
+ * if it fails to be created the library's heuristic approach will be used
+ *
+ * @return Depthwise convolution layer method to be used by the node
+ */
+ DepthwiseConvolutionMethod depthwise_convolution_method() const;
+
+ /** Epsilon parameter accessor
+ *
+ * @return Epsilon parameter
+ */
+ float epsilon() const;
+
+ /** Returns fused activation
+ *
+ * @return Fused activation
+ */
+ ActivationLayerInfo fused_activation() const;
+
+ /** Sets fused activation
+ *
+ * @param[in] fused_activation Fused activation to set
+ */
+ void set_fused_activation(ActivationLayerInfo fused_activation);
+
+ /** Computes convolution output descriptor
+ *
+ * @param[in] input_descriptor Input descriptor
+ * @param[in] weights_descriptor Weights descriptor
+ * @param[in] info Convolution operation attributes
+ * @param[in] depth_multiplier Depth multiplier
+ *
+ * @return Output descriptor
+ */
+ static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+ const TensorDescriptor &weights_descriptor,
+ const PadStrideInfo &info,
+ int depth_multiplier);
+
+ /** Sets the convolution layer method to use
+ *
+ * @param[in] method Method to use for convolution
+ */
+ void set_convolution_method(ConvolutionMethod method);
+
+ /** Depth multiplier accessor
+ *
+ * @return Depth multiplier
+ */
+ unsigned int depth_multiplier() 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::FusedDepthwiseConvolutionBatchNormalizationLayer;
+
+private:
+ float _epsilon;
+
+ PadStrideInfo _info;
+ unsigned int _depth_multiplier;
+ DepthwiseConvolutionMethod _method;
+ ActivationLayerInfo _fused_activation;
+};
+
+} // namespace graph
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__ */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index c891bc2..52e2f88 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -39,6 +39,7 @@
#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/FusedDepthwiseConvolutionBatchNormalizationNode.h"
#include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
#include "arm_compute/graph/nodes/InputNode.h"
#include "arm_compute/graph/nodes/NormalizationLayerNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 0f3450b..2c89679 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -45,6 +45,7 @@
class FlattenLayerNode;
class FullyConnectedLayerNode;
class FusedConvolutionBatchNormalizationNode;
+class FusedDepthwiseConvolutionBatchNormalizationNode;
class GenerateProposalsLayerNode;
class InputNode;
class NormalizationLayerNode;