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>

6 files changed, 243 insertions, 12 deletions

diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index c14100ab42..b9f22f6199 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -74,8 +74,9 @@ struct CLEltwiseFunctions
 /** Function and tensor types to be used inside a CL fused convolution/batch normalization layer */
 struct CLFusedLayerTypes
 {
-    using ConvolutionLayer       = CLConvolutionLayer;
-    using FuseBatchNormalization = CLFuseBatchNormalization;
+    using ConvolutionLayer          = CLConvolutionLayer;
+    using DepthwiseConvolutionLayer = CLDepthwiseConvolutionLayer;
+    using FuseBatchNormalization    = CLFuseBatchNormalization;
 };
 
 // TODO (isagot01): Remove once we support heterogeneous scheduling at function level
@@ -203,6 +204,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx);
         case NodeType::FusedConvolutionBatchNormalizationLayer:
             return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node));
+        case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer:
+            return detail::create_fused_depthwise_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx);
         case NodeType::NormalizationLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index d4892f53a6..ad96240a4b 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -80,8 +80,9 @@ struct NEEltwiseFunctions
 /** Function and tensor types to be used inside a NEON fused convolution/batch normalization layer */
 struct NEFusedLayerTypes
 {
-    using ConvolutionLayer       = NEConvolutionLayer;
-    using FuseBatchNormalization = NEFuseBatchNormalization;
+    using ConvolutionLayer          = NEConvolutionLayer;
+    using DepthwiseConvolutionLayer = NEDepthwiseConvolutionLayer;
+    using FuseBatchNormalization    = NEFuseBatchNormalization;
 };
 
 namespace detail
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 427d7b5095..83177a8431 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -64,7 +64,6 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
         // Extract conv inputs
         const auto   conv_input_id   = conv_node->input_edge(0)->producer_id();
         const auto   conv_weights_id = conv_node->input_edge(1)->producer_id();
-        const auto   out_quant_info  = conv_node->output(0)->desc().quant_info;
         const auto   conv_info       = conv_node->convolution_info();
         const auto   conv_method     = conv_node->convolution_method();
         const auto   num_groups      = conv_node->num_groups();
@@ -79,7 +78,7 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
         const auto epsilon     = bn_node->epsilon();
 
         // Create the fused node
-        const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint, out_quant_info, act_info);
+        const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint, act_info);
 
         if(conv_node->input_edge(2) != nullptr)
         {
@@ -125,6 +124,83 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
     }
 }
 
+void fuse_depthwise_convolution_with_batch_normalization(Graph &g, const Edge *output_edge)
+{
+    ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
+
+    auto *depth_conv_node = arm_compute::utils::cast::polymorphic_downcast<DepthwiseConvolutionLayerNode *>(output_edge->producer());
+    auto *bn_node         = arm_compute::utils::cast::polymorphic_downcast<BatchNormalizationLayerNode *>(output_edge->consumer());
+
+    ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing depthwise convolution node with ID : " << output_edge->producer_id()
+                                  << " with BatchNormalization Layer node with ID : " << output_edge->consumer_id() << std::endl);
+
+    // Prevent fusion if fused node has an output accessor
+    if(depth_conv_node->output(0)->accessor() == nullptr)
+    {
+        const Target assigned_target = depth_conv_node->assigned_target();
+
+        // Extract conv inputs
+        const auto depth_conv_input_id = depth_conv_node->input_edge(0)->producer_id();
+        const auto conv_weights_id     = depth_conv_node->input_edge(1)->producer_id();
+        const auto conv_info           = depth_conv_node->convolution_info();
+        const auto depth_conv_method   = depth_conv_node->depthwise_convolution_method();
+        const auto depth_multiplier    = depth_conv_node->depth_multiplier();
+        const auto act_info            = bn_node->fused_activation();
+
+        // Extract bn inputs
+        const auto bn_mean_id  = bn_node->input_edge(1)->producer_id();
+        const auto bn_var_id   = bn_node->input_edge(2)->producer_id();
+        const auto bn_beta_id  = bn_node->input_edge(3)->producer_id();
+        const auto bn_gamma_id = bn_node->input_edge(4)->producer_id();
+        const auto epsilon     = bn_node->epsilon();
+
+        // Create the fused node
+        const NodeID fused_id = g.add_node<FusedDepthwiseConvolutionBatchNormalizationNode>(epsilon, conv_info, depth_multiplier, depth_conv_method, act_info);
+
+        if(depth_conv_node->input_edge(2) != nullptr)
+        {
+            const auto conv_bias_id = depth_conv_node->input_edge(2)->producer_id();
+            g.add_connection(conv_bias_id, 0, fused_id, 2);
+        }
+
+        // Add connections from the conv/batch_norm inputs to the fused node
+        g.add_connection(depth_conv_input_id, 0, fused_id, 0);
+        g.add_connection(conv_weights_id, 0, fused_id, 1);
+        g.add_connection(bn_mean_id, 0, fused_id, 3);
+        g.add_connection(bn_var_id, 0, fused_id, 4);
+        g.add_connection(bn_beta_id, 0, fused_id, 5);
+        g.add_connection(bn_gamma_id, 0, fused_id, 6);
+
+        auto                     fused_node       = g.node(fused_id);
+        std::vector<NodeIdxPair> bn_driving_nodes = get_driving_nodes(*bn_node);
+
+        // Extract batch normalization node accessor if any
+        auto bn_node_accessor = bn_node->output(0)->extract_accessor();
+        auto bn_node_name     = bn_node->name();
+
+        // Remove batch normalization node
+        g.remove_node(bn_node->id());
+
+        // Get driving nodes of batch normalization node
+        for(auto &driving_node : bn_driving_nodes)
+        {
+            g.add_connection(fused_id, 0, driving_node.node_id, driving_node.index);
+            configure_tensor(fused_node->output(0));
+        }
+        // Update fused node outputs
+        fused_node->output(0)->set_accessor(std::move(bn_node_accessor));
+        fused_node->set_assigned_target(assigned_target);
+        fused_node->set_common_node_parameters(NodeParams{ depth_conv_node->name() + "+" + bn_node_name, assigned_target });
+
+        // Remove convolution node
+        g.remove_node(depth_conv_node->id());
+    }
+    else
+    {
+        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of depthwise convolution with batch normalization due to the presence of an output accessor\n");
+    }
+}
+
 template <typename N>
 void fuse_node_with_activation(Graph &g, const Edge *output_edge, const std::set<Activation> &supported_fused_activations)
 {
@@ -224,6 +300,8 @@ void NodeFusionMutator::mutate(Graph &g)
         return (output_qasymm8 && same_qinfo) || !output_qasymm8;
     };
 
+    Target target = g.nodes()[0].get()->output(0)->desc().target;
+
     // Fusion mutations
     detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations);
     detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations);
@@ -231,6 +309,11 @@ void NodeFusionMutator::mutate(Graph &g)
 
     // TODO (COMPMID-2055): re-enable once we fuse bias and activations to convolution
     // detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
+    if(target == Target::CL)
+    {
+        //Depthwise Convolution and Batch Normalization Fusion active only for CL
+        detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
+    }
 }
 } // namespace graph
 } // namespace arm_compute
diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
index 6496a71251..0a0c0c50e8 100644
--- a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
+++ b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
@@ -36,8 +36,8 @@ FusedConvolutionBatchNormalizationNode::FusedConvolutionBatchNormalizationNode(f
                                                                                unsigned int      num_groups,
                                                                                ConvolutionMethod method,
                                                                                FastMathHint      fast_math_hint,
-                                                                               QuantizationInfo out_quant_info, ActivationLayerInfo fused_activation)
-    : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(std::move(out_quant_info)), _fused_activation(fused_activation)
+                                                                               ActivationLayerInfo fused_activation)
+    : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _fused_activation(fused_activation)
 {
     _input_edges.resize(7, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -132,10 +132,6 @@ TensorDescriptor FusedConvolutionBatchNormalizationNode::configure_output(size_t
     ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
 
     TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info);
-    if(!_out_quant_info.empty())
-    {
-        output_info.quant_info = _out_quant_info;
-    }
 
     return output_info;
 }
diff --git a/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp
new file mode 100644
index 0000000000..a04d75407a
--- /dev/null
+++ b/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp
@@ -0,0 +1,140 @@
+/*
+ * 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.
+ */
+#include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
+
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+FusedDepthwiseConvolutionBatchNormalizationNode::FusedDepthwiseConvolutionBatchNormalizationNode(float                      epsilon,
+                                                                                                 PadStrideInfo              info,
+                                                                                                 unsigned int               depth_multiplier,
+                                                                                                 DepthwiseConvolutionMethod method,
+                                                                                                 ActivationLayerInfo        fused_activation)
+    : _epsilon(epsilon), _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _fused_activation(fused_activation)
+{
+    _input_edges.resize(7, EmptyEdgeID);
+    _outputs.resize(1, NullTensorID);
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::set_depthwise_convolution_method(DepthwiseConvolutionMethod method)
+{
+    _method = method;
+}
+
+DepthwiseConvolutionMethod FusedDepthwiseConvolutionBatchNormalizationNode::depthwise_convolution_method() const
+{
+    return _method;
+}
+
+float FusedDepthwiseConvolutionBatchNormalizationNode::epsilon() const
+{
+    return _epsilon;
+}
+
+PadStrideInfo FusedDepthwiseConvolutionBatchNormalizationNode::convolution_info() const
+{
+    return _info;
+}
+
+unsigned int FusedDepthwiseConvolutionBatchNormalizationNode::depth_multiplier() const
+{
+    return _depth_multiplier;
+}
+
+ActivationLayerInfo FusedDepthwiseConvolutionBatchNormalizationNode::fused_activation() const
+{
+    return _fused_activation;
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::set_fused_activation(ActivationLayerInfo fused_activation)
+{
+    _fused_activation = fused_activation;
+}
+
+TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
+                                                                                            const TensorDescriptor &weights_descriptor,
+                                                                                            const PadStrideInfo    &info,
+                                                                                            int                     depth_multiplier)
+{
+    unsigned int output_width  = 0;
+    unsigned int output_height = 0;
+
+    const unsigned int input_width    = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH);
+    const unsigned int input_height   = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT);
+    const unsigned int input_channels = get_dimension_size(input_descriptor, DataLayoutDimension::CHANNEL);
+    const unsigned int kernel_width   = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH);
+    const unsigned int kernel_height  = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT);
+
+    std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
+
+    TensorDescriptor output_descriptor = input_descriptor;
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::WIDTH), output_width);
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::HEIGHT), output_height);
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::CHANNEL), input_channels * depth_multiplier);
+
+    return output_descriptor;
+}
+
+bool FusedDepthwiseConvolutionBatchNormalizationNode::forward_descriptors()
+{
+    if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (output_id(0) != NullTensorID))
+    {
+        Tensor *dst = output(0);
+        ARM_COMPUTE_ERROR_ON(dst == nullptr);
+        dst->desc() = configure_output(0);
+        return true;
+    }
+    return false;
+}
+
+TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_UNUSED(idx);
+    const Tensor *src     = input(0);
+    const Tensor *weights = input(1);
+
+    ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
+
+    TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+
+    return output_info;
+}
+
+NodeType FusedDepthwiseConvolutionBatchNormalizationNode::type() const
+{
+    return FusedDepthwiseConvolutionBatchNormalizationNode::node_type;
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
diff --git a/src/graph/printers/DotGraphPrinter.cpp b/src/graph/printers/DotGraphPrinter.cpp
index c939de1b64..46f6ee828e 100644
--- a/src/graph/printers/DotGraphPrinter.cpp
+++ b/src/graph/printers/DotGraphPrinter.cpp
@@ -85,6 +85,14 @@ void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n)
     _info = ss.str();
 }
 
+void DotGraphVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n)
+{
+    ARM_COMPUTE_UNUSED(n);
+    std::stringstream ss;
+    ss << "FusedDepthwiseConvolutionBatchNormalizationNode";
+    _info = ss.str();
+}
+
 void DotGraphVisitor::visit(NormalizationLayerNode &n)
 {
     std::stringstream ss;