COMPMID-554 Add Nodes

- DepthwiseConvolutionLayer

Change-Id: Icaef85d7474f7532bf7d93d11b5c787712e57bdd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110524
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 184 insertions, 0 deletions

diff --git a/src/graph/nodes/DepthwiseConvolutionLayer.cpp b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
new file mode 100644
index 0000000000..1c006d61db
--- /dev/null
+++ b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
@@ -0,0 +1,84 @@
+/*
+ * Copyright (c) 2017 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/DepthwiseConvolutionLayer.h"
+
+#include "arm_compute/graph/Error.h"
+#include "arm_compute/graph/NodeContext.h"
+#include "arm_compute/graph/OperationRegistry.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute::graph;
+
+std::unique_ptr<arm_compute::IFunction> DepthwiseConvolutionLayer::instantiate_node(GraphContext &ctx, ITensorObject *input, ITensorObject *output)
+{
+    ARM_COMPUTE_ERROR_ON_UNALLOCATED_TENSOR_OBJECT(input, output);
+
+    arm_compute::ITensor *in  = input->tensor();
+    arm_compute::ITensor *out = output->tensor();
+    _target_hint              = ctx.hints().target_hint();
+
+    if(_weights.tensor() == nullptr)
+    {
+        TensorShape shape = in->info()->tensor_shape();
+        shape.set(Window::DimX, _conv_width);
+        shape.set(Window::DimY, _conv_height);
+        _weights.set_info(TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+    }
+    if(_biases.has_accessor() && _biases.tensor() == nullptr)
+    {
+        _biases.set_info(TensorInfo(TensorShape(in->info()->dimension(2)), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+    }
+
+    bool weights_is_loaded = _weights.tensor() != nullptr;
+    bool biases_is_loaded  = _biases.has_accessor() ? _biases.tensor() != nullptr : false;
+
+    _weights.set_target(_target_hint);
+    _biases.set_target(_target_hint);
+
+    // Create node context
+    NodeContext node_ctx(OperationType::DepthwiseConvolutionLayer);
+    node_ctx.set_target(_target_hint);
+    node_ctx.add_input(in);
+    node_ctx.add_input(_weights.tensor());
+    if(_biases.has_accessor())
+    {
+        node_ctx.add_input(_biases.tensor());
+    }
+    node_ctx.add_output(out);
+    node_ctx.add_parameter<PadStrideInfo>("ConvolutionInfo", _conv_info);
+    node_ctx.add_parameter<bool>("Optimized3x3", _opt3x3);
+
+    // Fill tensors
+    if(!weights_is_loaded)
+    {
+        _weights.allocate_and_fill_if_needed();
+    }
+    if(!biases_is_loaded)
+    {
+        _biases.allocate_and_fill_if_needed();
+    }
+
+    // Get function
+    return OperationRegistry::get().find_operation(OperationType::DepthwiseConvolutionLayer, _target_hint)->configure(node_ctx);
+}
diff --git a/src/graph/operations/CLSimpleOperations.cpp b/src/graph/operations/CLSimpleOperations.cpp
index 4ec3a22f37..881f4910ad 100644
--- a/src/graph/operations/CLSimpleOperations.cpp
+++ b/src/graph/operations/CLSimpleOperations.cpp
@@ -135,6 +135,56 @@ REGISTER_SIMPLE_OPERATION(CLDepthConvertLayerOperation, OPENCL, OperationType::D
     return std::move(depthconvert);
 }
 
+/* DepthwiseConvolutionLayer Layer */
+REGISTER_SIMPLE_OPERATION(CLDepthwiseConvolutionOperation, OPENCL, OperationType::DepthwiseConvolutionLayer)
+{
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 || ctx.num_inputs() != 3);
+    ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0)) == nullptr);
+
+    // Extract IO and info
+    auto      *in        = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0));
+    auto      *weights   = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(1));
+    auto      *biases    = ctx.num_inputs() == 3 ? dynamic_cast<arm_compute::ICLTensor *>(ctx.input(2)) : nullptr;
+    auto      *out       = dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0));
+    const auto conv_info = ctx.parameter<PadStrideInfo>("ConvolutionInfo");
+    const auto opt3x3    = ctx.parameter<bool>("Optimized3x3");
+
+    // Create and configure function
+    std::unique_ptr<arm_compute::IFunction> func;
+    bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
+    if(run_3x3_opt)
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolution>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
+    else
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolution3x3>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLDepthwiseConvolutionLayer"
+                               << " Data Type: " << in->info()->data_type()
+                               << " Input shape: " << in->info()->tensor_shape()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << out->info()->tensor_shape());
+    if(biases == nullptr)
+    {
+        ARM_COMPUTE_LOG_GRAPH_INFO(" Biases shape: No biases provided" << std::endl);
+    }
+    else
+    {
+        ARM_COMPUTE_LOG_GRAPH_INFO(" Biases shape: " << biases->info()->tensor_shape() << std::endl);
+    }
+
+    return func;
+}
+
 /* DeQuantizationLayer Layer */
 REGISTER_SIMPLE_OPERATION(CLDequantizationLayerOperation, OPENCL, OperationType::DequantizationLayer)
 {
diff --git a/src/graph/operations/NESimpleOperations.cpp b/src/graph/operations/NESimpleOperations.cpp
index 12f8c6c76b..c77aeeca11 100644
--- a/src/graph/operations/NESimpleOperations.cpp
+++ b/src/graph/operations/NESimpleOperations.cpp
@@ -135,6 +135,56 @@ REGISTER_SIMPLE_OPERATION(NEDepthConvertLayerOperation, NEON, OperationType::Dep
     return std::move(depthconvert);
 }
 
+/* DepthwiseConvolutionLayer Layer */
+REGISTER_SIMPLE_OPERATION(NEDepthwiseConvolutionOperation, NEON, OperationType::DepthwiseConvolutionLayer)
+{
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 || ctx.num_inputs() != 3);
+    ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.input(0)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.output(0)) == nullptr);
+
+    // Extract IO and info
+    auto      *in        = dynamic_cast<arm_compute::ITensor *>(ctx.input(0));
+    auto      *weights   = dynamic_cast<arm_compute::ITensor *>(ctx.input(1));
+    auto      *biases    = ctx.num_inputs() == 3 ? dynamic_cast<arm_compute::ITensor *>(ctx.input(2)) : nullptr;
+    auto      *out       = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
+    const auto conv_info = ctx.parameter<PadStrideInfo>("ConvolutionInfo");
+    const auto opt3x3    = ctx.parameter<bool>("Optimized3x3");
+
+    // Create and configure function
+    std::unique_ptr<arm_compute::IFunction> func;
+    bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
+    if(run_3x3_opt)
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolution>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
+    else
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolution3x3>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDepthwiseConvolutionLayer"
+                               << " Data Type: " << in->info()->data_type()
+                               << " Input shape: " << in->info()->tensor_shape()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << out->info()->tensor_shape());
+    if(biases == nullptr)
+    {
+        ARM_COMPUTE_LOG_GRAPH_INFO(" Biases shape: No biases provided" << std::endl);
+    }
+    else
+    {
+        ARM_COMPUTE_LOG_GRAPH_INFO(" Biases shape: " << biases->info()->tensor_shape() << std::endl);
+    }
+
+    return func;
+}
+
 /* DeQuantizationLayer Layer */
 REGISTER_SIMPLE_OPERATION(NEDequantizationLayerOperation, NEON, OperationType::DequantizationLayer)
 {