IVGCVSW-3885 Add reference workload for DepthToSpace
Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: Id937dc4425884ad1985dcdfaae8bf3fb64f0c766
diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index 83444ed..c2eb025 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -24,6 +24,8 @@
Debug.cpp
Debug.hpp
Decoders.hpp
+ DepthToSpace.cpp
+ DepthToSpace.hpp
DetectionPostProcess.cpp
DetectionPostProcess.hpp
ElementwiseFunction.cpp
@@ -71,6 +73,8 @@
RefElementwiseWorkload.hpp
RefDebugWorkload.cpp
RefDebugWorkload.hpp
+ RefDepthToSpaceWorkload.cpp
+ RefDepthToSpaceWorkload.hpp
RefDepthwiseConvolution2dWorkload.cpp
RefDepthwiseConvolution2dWorkload.hpp
RefDequantizeWorkload.cpp
diff --git a/src/backends/reference/workloads/DepthToSpace.cpp b/src/backends/reference/workloads/DepthToSpace.cpp
new file mode 100644
index 0000000..046bd47
--- /dev/null
+++ b/src/backends/reference/workloads/DepthToSpace.cpp
@@ -0,0 +1,79 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "DepthToSpace.hpp"
+
+#include <DataLayoutIndexed.hpp>
+#include <Permute.hpp>
+
+#include <boost/assert.hpp>
+
+using namespace armnnUtils;
+
+namespace armnn
+{
+
+void DepthToSpace(const TensorInfo& inputInfo,
+ const DepthToSpaceDescriptor& descriptor,
+ const void* inputData,
+ void* outputData,
+ unsigned int dataTypeSize)
+{
+ const unsigned int blockSize = descriptor.m_BlockSize;
+ BOOST_ASSERT(blockSize != 0u);
+
+ const TensorShape& inputShape = inputInfo.GetShape();
+ const unsigned int batches = inputShape[0];
+
+ armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout);
+ const unsigned int inDepth = inputShape[dataLayoutIndexed.GetChannelsIndex()];
+ const unsigned int inHeight = inputShape[dataLayoutIndexed.GetHeightIndex()];
+ const unsigned int inWidth = inputShape[dataLayoutIndexed.GetWidthIndex()];
+
+ const unsigned int outDepth = inDepth / (blockSize * blockSize);
+
+ // The 4D input data can be interpreted as 6D (implicitly reshaped) as follows:
+ //
+ // [batch, block size, block size, inDepth, inHeight, inWidth] for NCHW and
+ // [batch, inHeight, inWidth, blockSize, blockSize, outDepth] for NHWC.
+ //
+ // DepthToSpace can then be implemented as a permutation in 6D resulting in
+ // the following shapes:
+ //
+ // [batch, outDepth, inHeight, blockSize, inWidth, blockSize] for NCHW and
+ // [batch, inHeight, blockSize, inWidth, blockSize, outDepth] for NHWC.
+ //
+ // NOTE:
+ // Since 6D tensors are not currently supported, in practice we need to handle each
+ // batch separately and execute 5D permutations
+
+ TensorShape permDestShape;
+ std::initializer_list<unsigned int> permVector;
+ if (descriptor.m_DataLayout == DataLayout::NCHW)
+ {
+ permDestShape = TensorShape({ outDepth, inHeight, blockSize, inWidth, blockSize });
+ permVector = { 2, 4, 0, 1, 3 };
+ }
+ else
+ {
+ permDestShape = TensorShape({ inHeight, blockSize, inWidth, blockSize, outDepth });
+ permVector = { 0, 2, 1, 3, 4 };
+ }
+
+ const unsigned int numElementsPerBatch = inputShape.GetNumElements() / batches;
+
+ for (unsigned int batchIndex = 0u; batchIndex < batches; ++batchIndex)
+ {
+ const uintptr_t batchDataOffset = batchIndex * (numElementsPerBatch * dataTypeSize);
+
+ armnnUtils::Permute(permDestShape,
+ PermutationVector(permVector),
+ static_cast<const void*>(reinterpret_cast<const uint8_t*>(inputData) + batchDataOffset),
+ static_cast<void*>(reinterpret_cast<uint8_t*>(outputData) + batchDataOffset),
+ dataTypeSize);
+ }
+}
+
+} // namespace armnn
diff --git a/src/backends/reference/workloads/DepthToSpace.hpp b/src/backends/reference/workloads/DepthToSpace.hpp
new file mode 100644
index 0000000..a1805c0
--- /dev/null
+++ b/src/backends/reference/workloads/DepthToSpace.hpp
@@ -0,0 +1,20 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+#include <armnn/Tensor.hpp>
+
+namespace armnn
+{
+
+void DepthToSpace(const TensorInfo& inputInfo,
+ const DepthToSpaceDescriptor& descriptor,
+ const void* inputData,
+ void* outputData,
+ unsigned int dataTypeSize);
+
+} // namespace armnn
diff --git a/src/backends/reference/workloads/RefDepthToSpaceWorkload.cpp b/src/backends/reference/workloads/RefDepthToSpaceWorkload.cpp
new file mode 100644
index 0000000..93c1120
--- /dev/null
+++ b/src/backends/reference/workloads/RefDepthToSpaceWorkload.cpp
@@ -0,0 +1,27 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefDepthToSpaceWorkload.hpp"
+
+#include "DepthToSpace.hpp"
+#include "RefWorkloadUtils.hpp"
+
+namespace armnn
+{
+
+void RefDepthToSpaceWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefDepthToSpaceWorkload_Execute");
+
+ const TensorInfo inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
+
+ DepthToSpace(inputInfo,
+ m_Data.m_Parameters,
+ m_Data.m_Inputs[0]->Map(),
+ m_Data.m_Outputs[0]->Map(),
+ GetDataTypeSize(inputInfo.GetDataType()));
+}
+
+} // namespace armnn
diff --git a/src/backends/reference/workloads/RefDepthToSpaceWorkload.hpp b/src/backends/reference/workloads/RefDepthToSpaceWorkload.hpp
new file mode 100644
index 0000000..327cd9d
--- /dev/null
+++ b/src/backends/reference/workloads/RefDepthToSpaceWorkload.hpp
@@ -0,0 +1,20 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "backendsCommon/Workload.hpp"
+
+namespace armnn
+{
+
+class RefDepthToSpaceWorkload : public BaseWorkload<DepthToSpaceQueueDescriptor>
+{
+public:
+ using BaseWorkload<DepthToSpaceQueueDescriptor>::BaseWorkload;
+ virtual void Execute() const override;
+};
+
+} // namespace armnn
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index b4721b1..94592cb 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -25,9 +25,11 @@
#include "RefConcatWorkload.hpp"
#include "RefConvertFp16ToFp32Workload.hpp"
#include "RefConvertFp32ToFp16Workload.hpp"
-#include "RefDepthwiseConvolution2dWorkload.hpp"
-#include "RefDetectionPostProcessWorkload.hpp"
#include "RefDebugWorkload.hpp"
+#include "RefDepthToSpaceWorkload.hpp"
+#include "RefDepthwiseConvolution2dWorkload.hpp"
+#include "RefDequantizeWorkload.hpp"
+#include "RefDetectionPostProcessWorkload.hpp"
#include "RefDequantizeWorkload.hpp"
#include "RefElementwiseWorkload.hpp"
#include "RefFullyConnectedWorkload.hpp"