IVGCVSW-4375 Add support for Transpose
* Added TransposeLayer
* Added CL, Neon and Ref Workloads
* Added Transpose utilities
* Added Serializer and Deserializer support
* Added Quantizer support
Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Change-Id: I04c755ba7cb5b1edf72b3c9f3c0314878032e3c7
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index e8548e4..d3ac986 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -60,6 +60,7 @@
#include "workloads/ClStridedSliceWorkload.hpp"
#include "workloads/ClSubtractionWorkload.hpp"
#include "workloads/ClTransposeConvolution2dWorkload.hpp"
+#include "workloads/ClTransposeWorkload.hpp"
#endif
using namespace boost;
@@ -819,4 +820,12 @@
biases);
}
+bool ClLayerSupport::IsTransposeSupported(const TensorInfo& input,
+ const TensorInfo& output,
+ const TransposeDescriptor& descriptor,
+ Optional<std::string&> reasonIfUnsupported) const
+{
+ FORWARD_WORKLOAD_VALIDATE_FUNC(ClTransposeWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
+}
+
} // namespace armnn
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index 819d086..60899d0 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -286,6 +286,12 @@
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
+ bool IsTransposeSupported(const TensorInfo& input,
+ const TensorInfo& output,
+ const TransposeDescriptor& descriptor,
+ Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
};
} // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index 4bb2e2a..21c2629 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -534,6 +534,12 @@
return MakeWorkload<ClSubtractionWorkload>(descriptor, info);
}
+std::unique_ptr<IWorkload> ClWorkloadFactory::CreateTranspose(const TransposeQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const
+{
+ return MakeWorkload<ClTransposeWorkload>(descriptor, info);
+}
+
std::unique_ptr<IWorkload> ClWorkloadFactory::CreateTransposeConvolution2d(
const TransposeConvolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info) const
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index 980be91..a716801 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -210,6 +210,9 @@
std::unique_ptr<IWorkload> CreateSubtraction(const SubtractionQueueDescriptor& descriptor,
const WorkloadInfo& info) const override;
+ std::unique_ptr<IWorkload> CreateTranspose(const TransposeQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const override;
+
std::unique_ptr<IWorkload> CreateTransposeConvolution2d(const TransposeConvolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info) const override;
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 4182b94..e326add 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -67,7 +67,8 @@
workloads/ClStackWorkload.cpp \
workloads/ClStridedSliceWorkload.cpp \
workloads/ClSubtractionWorkload.cpp \
- workloads/ClTransposeConvolution2dWorkload.cpp
+ workloads/ClTransposeConvolution2dWorkload.cpp \
+ workloads/ClTransposeWorkload.cpp
else
# ARMNN_COMPUTE_CL_ENABLED == 0
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index cfec81a..d8b0fd1 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -770,6 +770,20 @@
ARMNN_AUTO_TEST_CASE(DequantizeSimpleUint8ToFp16, DequantizeSimpleUint8ToFp16Test)
ARMNN_AUTO_TEST_CASE(DequantizeSimpleInt16ToFp16, DequantizeSimpleInt16ToFp16Test)
+// Transpose
+ARMNN_AUTO_TEST_CASE(SimpleTransposeFloat32, SimpleTransposeTest<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(TransposeFloat32ValueSet1Test, TransposeValueSet1Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(TransposeFloat32ValueSet2Test, TransposeValueSet2Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(TransposeFloat32ValueSet3Test, TransposeValueSet3Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE(SimpleTransposeQASymm8, SimpleTransposeTest<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE(TransposeQASymm8ValueSet1Test, TransposeValueSet1Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE(TransposeQASymm8ValueSet2Test, TransposeValueSet2Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE(TransposeQASymm8ValueSet3Test, TransposeValueSet3Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE(SimpleTransposeQSymm16, SimpleTransposeTest<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE(TransposeQSymm16ValueSet1Test, TransposeValueSet1Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE(TransposeQSymm16ValueSet2Test, TransposeValueSet2Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE(TransposeQSymm16ValueSet3Test, TransposeValueSet3Test<DataType::QSymmS16>)
+
// TransposeConvolution2d
ARMNN_AUTO_TEST_CASE(SimpleTransposeConvolution2dFloatNchw,
SimpleTransposeConvolution2dTest<DataType::Float32, DataType::Float32>,
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index de62ca9..17d69b1 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -96,6 +96,8 @@
ClSubtractionWorkload.hpp
ClTransposeConvolution2dWorkload.cpp
ClTransposeConvolution2dWorkload.hpp
+ ClTransposeWorkload.cpp
+ ClTransposeWorkload.hpp
ClWorkloads.hpp
ClWorkloadUtils.hpp
)
diff --git a/src/backends/cl/workloads/ClTransposeWorkload.cpp b/src/backends/cl/workloads/ClTransposeWorkload.cpp
new file mode 100644
index 0000000..b276b22
--- /dev/null
+++ b/src/backends/cl/workloads/ClTransposeWorkload.cpp
@@ -0,0 +1,49 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClTransposeWorkload.hpp"
+#include <cl/ClTensorHandle.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/core/Error.h>
+
+#include "ClWorkloadUtils.hpp"
+
+namespace armnn
+{
+
+arm_compute::Status ClTransposeWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& output,
+ const TransposeDescriptor& descriptor)
+{
+ const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+ const armnn::PermutationVector& mappings = descriptor.m_DimMappings;
+
+ return arm_compute::CLPermute::validate(&aclInputInfo, &aclOutputInfo,
+ armcomputetensorutils::BuildArmComputeTransposeVector(mappings));
+}
+
+ClTransposeWorkload::ClTransposeWorkload(const TransposeQueueDescriptor& descriptor,
+ const WorkloadInfo& info)
+ : BaseWorkload<TransposeQueueDescriptor>(descriptor, info)
+{
+ m_Data.ValidateInputsOutputs(GetName(), 1, 1);
+
+ const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+ const armnn::PermutationVector& mappings = m_Data.m_Parameters.m_DimMappings;
+ // Run the layer.
+ m_PermuteFunction.configure(&input, &output,
+ armcomputetensorutils::BuildArmComputeTransposeVector(mappings));
+}
+
+void ClTransposeWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_CL(GetName() + "_Execute");
+ RunClFunction(m_PermuteFunction, CHECK_LOCATION());
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClTransposeWorkload.hpp b/src/backends/cl/workloads/ClTransposeWorkload.hpp
new file mode 100644
index 0000000..c1bed93
--- /dev/null
+++ b/src/backends/cl/workloads/ClTransposeWorkload.hpp
@@ -0,0 +1,40 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
+
+#include <armnn/TypesUtils.hpp>
+#include <arm_compute/runtime/CL/functions/CLPermute.h>
+
+#include <string>
+
+namespace armnn
+{
+
+arm_compute::Status ClTransposeWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& output,
+ const TransposeDescriptor& descriptor);
+
+class ClTransposeWorkload : public BaseWorkload<TransposeQueueDescriptor>
+{
+public:
+ static const std::string& GetName()
+ {
+ static const std::string name = std::string("ClTransposeWorkload");
+ return name;
+ }
+
+ ClTransposeWorkload(const TransposeQueueDescriptor& descriptor, const WorkloadInfo& info);
+ void Execute() const override;
+
+private:
+ using BaseWorkload<TransposeQueueDescriptor>::m_Data;
+ mutable arm_compute::CLPermute m_PermuteFunction;
+};
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index 014dc3f..ec193d5 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -49,3 +49,4 @@
#include "ClConvertFp16ToFp32Workload.hpp"
#include "ClConvertFp32ToFp16Workload.hpp"
#include "ClTransposeConvolution2dWorkload.hpp"
+#include "ClTransposeWorkload.hpp"