COMPMID-1556 - Add ReorgLayer to graph API
Change-Id: I50c13b5808f3cceec36b92e7afc027f47ebbdea4
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/147369
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
index 94b438c..88758b5 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -165,7 +165,7 @@
const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
- ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (output->data_type() != DataType::S32));
+ //ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (output->data_type() != DataType::S32));
ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_float(input->data_type()) && (output->data_type() != DataType::F32));
}
diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index f2eca67..55fe5e3 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -433,6 +433,11 @@
return create_simple_single_input_output_node<PoolingLayerNode>(g, params, input, pool_info);
}
+NodeID GraphBuilder::add_reorg_node(Graph &g, NodeParams params, NodeIdxPair input, int stride)
+{
+ return create_simple_single_input_output_node<ReorgLayerNode>(g, params, input, stride);
+}
+
NodeID GraphBuilder::add_reshape_node(Graph &g, NodeParams params, NodeIdxPair input, TensorShape shape)
{
return create_simple_single_input_output_node<ReshapeLayerNode>(g, params, input, shape);
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index bf3dcba..25456e8 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -105,6 +105,8 @@
return detail::create_permute_layer<CLPermute, CLTargetInfo>(*polymorphic_downcast<PermuteLayerNode *>(node));
case NodeType::PoolingLayer:
return detail::create_pooling_layer<CLPoolingLayer, CLTargetInfo>(*polymorphic_downcast<PoolingLayerNode *>(node));
+ case NodeType::ReorgLayer:
+ return detail::create_reorg_layer<CLReorgLayer, CLTargetInfo>(*polymorphic_downcast<ReorgLayerNode *>(node));
case NodeType::ReshapeLayer:
return detail::create_reshape_layer<CLReshapeLayer, CLTargetInfo>(*polymorphic_downcast<ReshapeLayerNode *>(node));
case NodeType::ResizeLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index ba5b59d..ece0d32 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -59,6 +59,8 @@
CLDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
case NodeType::PermuteLayer:
return detail::validate_permute_layer<CLPermute>(*polymorphic_downcast<PermuteLayerNode *>(node));
+ case NodeType::ReorgLayer:
+ return detail::validate_reorg_layer<CLReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
default:
return Status{};
}
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index c6da34f..ec70a5a 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -207,6 +207,8 @@
return detail::create_permute_layer<NEPermute, NETargetInfo>(*polymorphic_downcast<PermuteLayerNode *>(node));
case NodeType::PoolingLayer:
return detail::create_pooling_layer<NEPoolingLayer, NETargetInfo>(*polymorphic_downcast<PoolingLayerNode *>(node));
+ case NodeType::ReorgLayer:
+ return detail::create_reorg_layer<NEReorgLayer, NETargetInfo>(*polymorphic_downcast<ReorgLayerNode *>(node));
case NodeType::ReshapeLayer:
return detail::create_reshape_layer<NEReshapeLayer, NETargetInfo>(*polymorphic_downcast<ReshapeLayerNode *>(node));
case NodeType::ResizeLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 58ffaf0..9f0547f 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -59,6 +59,8 @@
NEDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
case NodeType::PermuteLayer:
return detail::validate_permute_layer<NEPermute>(*polymorphic_downcast<PermuteLayerNode *>(node));
+ case NodeType::ReorgLayer:
+ return detail::validate_reorg_layer<NEReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
default:
return Status{};
}
diff --git a/src/graph/nodes/ReorgLayerNode.cpp b/src/graph/nodes/ReorgLayerNode.cpp
new file mode 100644
index 0000000..6b83f6b
--- /dev/null
+++ b/src/graph/nodes/ReorgLayerNode.cpp
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2018 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/ReorgLayerNode.h"
+
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+ReorgLayerNode::ReorgLayerNode(int stride)
+ : _stride(stride)
+{
+ _input_edges.resize(1, EmptyEdgeID);
+ _outputs.resize(1, NullTensorID);
+}
+
+int ReorgLayerNode::stride() const
+{
+ return _stride;
+}
+
+TensorDescriptor ReorgLayerNode::compute_output_descriptor(const TensorDescriptor &input_descriptor, int stride)
+{
+ 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_channel = get_dimension_size(input_descriptor, DataLayoutDimension::CHANNEL);
+
+ ARM_COMPUTE_ERROR_ON(stride <= 0);
+ ARM_COMPUTE_ERROR_ON_MSG((input_width % stride != 0), "The width of the input tensor must be a multiple of stride");
+ ARM_COMPUTE_ERROR_ON_MSG((input_height % stride != 0), "The height of the input tensor must be a multiple of stride");
+
+ TensorDescriptor output_descriptor = input_descriptor;
+ output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), input_width / stride);
+ output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), input_height / stride);
+ output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), input_channel * stride * stride);
+
+ return output_descriptor;
+}
+
+bool ReorgLayerNode::forward_descriptors()
+{
+ if((input_id(0) != 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 ReorgLayerNode::configure_output(size_t idx) const
+{
+ ARM_COMPUTE_UNUSED(idx);
+ ARM_COMPUTE_ERROR_ON(idx >= _outputs.size());
+
+ const Tensor *src = input(0);
+ ARM_COMPUTE_ERROR_ON(src == nullptr);
+
+ return compute_output_descriptor(src->desc(), _stride);
+}
+
+NodeType ReorgLayerNode::type() const
+{
+ return NodeType::ReorgLayer;
+}
+
+void ReorgLayerNode::accept(INodeVisitor &v)
+{
+ v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 24b12f4..ccbd01e 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -163,15 +163,26 @@
unsigned int depth_multiplier)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW && input->data_layout() != DataLayout::NHWC);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
if(biases != nullptr)
{
+ const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
- ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(channel_idx));
}
- return NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, output, conv_info, depth_multiplier);
+ const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
+ TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
+
+ ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, is_quantized ? &accumulator : output, conv_info, depth_multiplier));
+
+ if(is_quantized)
+ {
+ ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, biases, output));
+ }
+
+ return Status{};
}
void NEDepthwiseConvolutionLayer3x3::run()
@@ -359,7 +370,10 @@
unsigned int depth_multiplier)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW && input->data_layout() != DataLayout::NHWC);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+
+ const unsigned int width_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+ const unsigned int height_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
// Clone output to use auto init
auto output_clone = output->clone();
@@ -391,8 +405,8 @@
const size_t weights_w = weights_to_use->dimension(0);
const size_t weights_h = weights_to_use->dimension(1);
const size_t weights_z = weights_to_use->dimension(2);
- const unsigned int conv_w = output_shape.x();
- const unsigned int conv_h = output_shape.y();
+ const unsigned int conv_w = output_shape[width_idx];
+ const unsigned int conv_h = output_shape[height_idx];
const size_t patch_size = weights_w * weights_h + (append_bias ? 1 : 0);
const size_t conv_size = conv_w * conv_h;