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/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 3c161d5..7e58dab 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -59,6 +59,7 @@
#include "workloads/NeonStridedSliceWorkload.hpp"
#include "workloads/NeonSubtractionWorkload.hpp"
#include "workloads/NeonTransposeConvolution2dWorkload.hpp"
+#include "workloads/NeonTransposeWorkload.hpp"
#endif
using namespace boost;
@@ -803,4 +804,12 @@
biases);
}
+bool NeonLayerSupport::IsTransposeSupported(const TensorInfo& input,
+ const TensorInfo& output,
+ const TransposeDescriptor& descriptor,
+ Optional<std::string&> reasonIfUnsupported) const
+{
+ FORWARD_WORKLOAD_VALIDATE_FUNC(NeonTransposeWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
+}
+
} // namespace armnn
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index 9cb64ea..f45db35 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -288,6 +288,11 @@
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;
+
}; // class NeonLayerSupport
} // namespace armnn
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index c3e0dc8..dc3ee84 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -503,6 +503,12 @@
return std::make_unique<NeonSubtractionWorkload>(descriptor, info);
}
+std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateTranspose(const TransposeQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const
+{
+ return std::make_unique<NeonTransposeWorkload>(descriptor, info);
+}
+
std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateTransposeConvolution2d(
const TransposeConvolution2dQueueDescriptor &descriptor,
const WorkloadInfo &info) const
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index 95271e2..bc4107d 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -213,6 +213,9 @@
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/neon/backend.mk b/src/backends/neon/backend.mk
index 1c572e6..d9a5405 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -67,7 +67,8 @@
workloads/NeonStackWorkload.cpp \
workloads/NeonStridedSliceWorkload.cpp \
workloads/NeonSubtractionWorkload.cpp \
- workloads/NeonTransposeConvolution2dWorkload.cpp
+ workloads/NeonTransposeConvolution2dWorkload.cpp \
+ workloads/NeonTransposeWorkload.cpp
else
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 18658a3..482bc25 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -891,6 +891,20 @@
ARMNN_AUTO_TEST_CASE(StackOutput3DInputs3, StackOutput3DInputs3Float32Test)
ARMNN_AUTO_TEST_CASE(StackOutput5D, StackOutput5DFloat32Test)
+// 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/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 02ffedc..a932f8b 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -98,6 +98,8 @@
NeonSubtractionWorkload.hpp
NeonTransposeConvolution2dWorkload.cpp
NeonTransposeConvolution2dWorkload.hpp
+ NeonTransposeWorkload.cpp
+ NeonTransposeWorkload.hpp
NeonWorkloads.hpp
NeonWorkloadUtils.hpp
)
diff --git a/src/backends/neon/workloads/NeonTransposeWorkload.cpp b/src/backends/neon/workloads/NeonTransposeWorkload.cpp
new file mode 100644
index 0000000..c11f2df
--- /dev/null
+++ b/src/backends/neon/workloads/NeonTransposeWorkload.cpp
@@ -0,0 +1,48 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonTransposeWorkload.hpp"
+#include <neon/NeonTensorHandle.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/core/Error.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonTransposeWorkloadValidate(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::NEPermute::validate(&aclInputInfo, &aclOutputInfo,
+ armcomputetensorutils::BuildArmComputeTransposeVector(mappings));
+}
+
+NeonTransposeWorkload::NeonTransposeWorkload(const TransposeQueueDescriptor& descriptor,
+ const WorkloadInfo& info)
+ : BaseWorkload<TransposeQueueDescriptor>(descriptor, info)
+{
+ m_Data.ValidateInputsOutputs(GetName(), 1, 1);
+
+ const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(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 NeonTransposeWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_NEON(GetName() + "_Execute");
+ m_PermuteFunction.run();
+}
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonTransposeWorkload.hpp b/src/backends/neon/workloads/NeonTransposeWorkload.hpp
new file mode 100644
index 0000000..aab7b70
--- /dev/null
+++ b/src/backends/neon/workloads/NeonTransposeWorkload.hpp
@@ -0,0 +1,39 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
+#include <neon/workloads/NeonWorkloadUtils.hpp>
+
+#include <armnn/TypesUtils.hpp>
+#include <arm_compute/runtime/NEON/functions/NEPermute.h>
+
+#include <string>
+
+namespace armnn
+{
+arm_compute::Status NeonTransposeWorkloadValidate(const TensorInfo& input, const TensorInfo& output,
+ const TransposeDescriptor& descriptor);
+
+class NeonTransposeWorkload : public BaseWorkload<TransposeQueueDescriptor>
+{
+public:
+ static const std::string& GetName()
+ {
+ static const std::string name = std::string("NeonTransposeWorkload");
+ return name;
+ }
+
+ NeonTransposeWorkload(const TransposeQueueDescriptor& descriptor, const WorkloadInfo& info);
+ void Execute() const override;
+
+private:
+ using BaseWorkload<TransposeQueueDescriptor>::m_Data;
+ mutable arm_compute::NEPermute m_PermuteFunction;
+};
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index b08483c..52cd76f 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -50,3 +50,4 @@
#include "NeonStridedSliceWorkload.hpp"
#include "NeonSubtractionWorkload.hpp"
#include "NeonTransposeConvolution2dWorkload.hpp"
+#include "NeonTransposeWorkload.hpp"