IVGCVSW-3162 Support CL workload for TransposeConv2D

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I3b021c0828d30298d99ddb211c9aae17fe3636f0

12 files changed, 267 insertions, 2 deletions

diff --git a/src/backends/aclCommon/ArmComputeTensorUtils.cpp b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
index 4f69c0b7db..e69232c5a1 100644
--- a/src/backends/aclCommon/ArmComputeTensorUtils.cpp
+++ b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
@@ -23,6 +23,8 @@ arm_compute::DataType GetArmComputeDataType(armnn::DataType dataType)
             return arm_compute::DataType::F32;
         case armnn::DataType::QuantisedAsymm8:
             return arm_compute::DataType::QASYMM8;
+        case armnn::DataType::QuantisedSymm16:
+            return arm_compute::DataType::QSYMM16;
         case armnn::DataType::Signed32:
             return arm_compute::DataType::S32;
         case armnn::DataType::Boolean:
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 5582799ff6..12c2efebb8 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -45,6 +45,7 @@
 #include "workloads/ClSplitterWorkload.hpp"
 #include "workloads/ClStridedSliceWorkload.hpp"
 #include "workloads/ClSubtractionWorkload.hpp"
+#include "workloads/ClTransposeConvolution2dWorkload.hpp"
 #endif
 
 using namespace boost;
@@ -685,4 +686,20 @@ bool ClLayerSupport::IsSubtractionSupported(const TensorInfo& input0,
                                    output);
 }
 
+bool ClLayerSupport::IsTransposeConvolution2dSupported(const TensorInfo& input,
+                                                       const TensorInfo& output,
+                                                       const TransposeConvolution2dDescriptor& descriptor,
+                                                       const TensorInfo& weights,
+                                                       const Optional<TensorInfo>& biases,
+                                                       Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClTransposeConvolution2dWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor,
+                                   weights,
+                                   biases);
+}
+
 } // namespace armnn
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index 35c00f9176..f2f25af87d 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -227,6 +227,13 @@ public:
                                 const TensorInfo& input1,
                                 const TensorInfo& output,
                                 Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
+    bool IsTransposeConvolution2dSupported(const TensorInfo& input,
+                                           const TensorInfo& output,
+                                           const TransposeConvolution2dDescriptor& descriptor,
+                                           const TensorInfo& weights,
+                                           const Optional<TensorInfo>& biases,
+                                           Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 };
 
 } // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index 70a4ddad4d..4bce653462 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -399,4 +399,11 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateGather(const armnn::GatherQu
     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 }
 
+std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreateTransposeConvolution2d(
+    const TransposeConvolution2dQueueDescriptor& descriptor,
+    const WorkloadInfo& info) const
+{
+    return MakeWorkload<ClTransposeConvolution2dWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
+}
+
 } // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index c84ce13b8b..8c3e756c0d 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -169,6 +169,9 @@ public:
     std::unique_ptr<IWorkload> CreateGather(const GatherQueueDescriptor& descriptor,
                                             const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateTransposeConvolution2d(const TransposeConvolution2dQueueDescriptor& descriptor,
+                                                            const WorkloadInfo& info) const override;
+
 private:
     template<typename FloatWorkload, typename Uint8Workload, typename QueueDescriptorType, typename... Args>
     static std::unique_ptr<IWorkload> MakeWorkload(const QueueDescriptorType& descriptor,
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 9bd37cc483..137cf7654b 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -54,7 +54,8 @@ BACKEND_SOURCES := \
         workloads/ClSpaceToBatchNdWorkload.cpp \
         workloads/ClSplitterWorkload.cpp \
         workloads/ClStridedSliceWorkload.cpp \
-        workloads/ClSubtractionWorkload.cpp
+        workloads/ClSubtractionWorkload.cpp \
+        workloads/ClTransposeConvolution2dWorkload.cpp
 else
 
 # ARMNN_COMPUTE_CL_ENABLED == 0
diff --git a/src/backends/cl/test/ClEndToEndTests.cpp b/src/backends/cl/test/ClEndToEndTests.cpp
index 174f392eea..ba01c5a980 100644
--- a/src/backends/cl/test/ClEndToEndTests.cpp
+++ b/src/backends/cl/test/ClEndToEndTests.cpp
@@ -9,6 +9,7 @@
 #include <backendsCommon/test/ConcatTestImpl.hpp>
 #include <backendsCommon/test/DequantizeEndToEndTestImpl.hpp>
 #include <backendsCommon/test/SplitterEndToEndTestImpl.hpp>
+#include <backendsCommon/test/TransposeConvolution2dEndToEndTestImpl.hpp>
 
 #include <boost/test/unit_test.hpp>
 
@@ -201,4 +202,29 @@ BOOST_AUTO_TEST_CASE(ClSplitter4dDim3EndToEndUint8Test)
     Splitter4dDim3EndToEnd<armnn::DataType::QuantisedAsymm8>(defaultBackends);
 }
 
-BOOST_AUTO_TEST_SUITE_END()
+// TransposeConvolution2d
+BOOST_AUTO_TEST_CASE(ClTransposeConvolution2dEndToEndFloatNchwTest)
+{
+    TransposeConvolution2dEndToEnd<armnn::DataType::Float32, armnn::DataType::Float32>(
+        defaultBackends, armnn::DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(ClTransposeConvolution2dEndToEndUint8NchwTest)
+{
+    TransposeConvolution2dEndToEnd<armnn::DataType::QuantisedAsymm8, armnn::DataType::Signed32>(
+        defaultBackends, armnn::DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(ClTransposeConvolution2dEndToEndFloatNhwcTest)
+{
+    TransposeConvolution2dEndToEnd<armnn::DataType::Float32, armnn::DataType::Float32>(
+        defaultBackends, armnn::DataLayout::NHWC);
+}
+
+BOOST_AUTO_TEST_CASE(ClTransposeConvolution2dEndToEndUint8NhwcTest)
+{
+    TransposeConvolution2dEndToEnd<armnn::DataType::QuantisedAsymm8, armnn::DataType::Signed32>(
+        defaultBackends, armnn::DataLayout::NHWC);
+}
+
+BOOST_AUTO_TEST_SUITE_END()
\ No newline at end of file
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 37ba39fa7f..3618adc70d 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -460,6 +460,37 @@ ARMNN_AUTO_TEST_CASE(QuantizeClampUint8, QuantizeClampUint8Test)
 ARMNN_AUTO_TEST_CASE(DequantizeSimpleUint8, DequantizeSimpleUint8Test)
 ARMNN_AUTO_TEST_CASE(DequantizeOffsetUint8, DequantizeOffsetUint8Test)
 
+// TransposeConvolution2d
+ARMNN_AUTO_TEST_CASE(SimpleTransposeConvolution2dFloatNchw, SimpleTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(SimpleTransposeConvolution2dFloatNhwc, SimpleTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(SimpleTransposeConvolution2dUint8Nchw, SimpleTransposeConvolution2dUint8NchwTest)
+ARMNN_AUTO_TEST_CASE(SimpleTransposeConvolution2dUint8Nhwc, SimpleTransposeConvolution2dUint8NhwcTest)
+
+ARMNN_AUTO_TEST_CASE(UnbiasedSimpleTransposeConvolution2dFloatNchw, UnbiasedSimpleTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedSimpleTransposeConvolution2dFloatNhwc, UnbiasedSimpleTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedSimpleTransposeConvolution2dUint8Nchw, UnbiasedSimpleTransposeConvolution2dUint8NchwTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedSimpleTransposeConvolution2dUint8Nhwc, UnbiasedSimpleTransposeConvolution2dUint8NhwcTest)
+
+ARMNN_AUTO_TEST_CASE(PaddedTransposeConvolution2dFloatNchw, PaddedTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(PaddedTransposeConvolution2dFloatNhwc, PaddedTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(PaddedTransposeConvolution2dUint8Nchw, PaddedTransposeConvolution2dUint8NchwTest)
+ARMNN_AUTO_TEST_CASE(PaddedTransposeConvolution2dUint8Nhwc, PaddedTransposeConvolution2dUint8NhwcTest)
+
+ARMNN_AUTO_TEST_CASE(UnbiasedPaddedTransposeConvolution2dFloatNchw, UnbiasedPaddedTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedPaddedTransposeConvolution2dFloatNhwc, UnbiasedPaddedTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedPaddedTransposeConvolution2dUint8Nchw, UnbiasedPaddedTransposeConvolution2dUint8NchwTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedPaddedTransposeConvolution2dUint8Nhwc, UnbiasedPaddedTransposeConvolution2dUint8NhwcTest)
+
+ARMNN_AUTO_TEST_CASE(StridedTransposeConvolution2dFloatNhwc, StridedTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(StridedTransposeConvolution2dFloatNchw, StridedTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(StridedTransposeConvolution2dUint8Nhwc, StridedTransposeConvolution2dUint8NhwcTest)
+ARMNN_AUTO_TEST_CASE(StridedTransposeConvolution2dUint8Nchw, StridedTransposeConvolution2dUint8NchwTest)
+
+ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dFloatNhwc, UnbiasedStridedTransposeConvolution2dFloatNhwcTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dFloatNchw, UnbiasedStridedTransposeConvolution2dFloatNchwTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dUint8Nhwc, UnbiasedStridedTransposeConvolution2dUint8NhwcTest)
+ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dUint8Nchw, UnbiasedStridedTransposeConvolution2dUint8NchwTest)
+
 // ============================================================================
 // COMPARE tests
 
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index c055244b1f..1b4c85b947 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -74,6 +74,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClStridedSliceWorkload.hpp
     ClSubtractionWorkload.cpp
     ClSubtractionWorkload.hpp
+    ClTransposeConvolution2dWorkload.cpp
+    ClTransposeConvolution2dWorkload.hpp
     ClWorkloads.hpp
     ClWorkloadUtils.hpp
 )
diff --git a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp
new file mode 100644
index 0000000000..7da6f3bd45
--- /dev/null
+++ b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp
@@ -0,0 +1,122 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClTransposeConvolution2dWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <cl/ClLayerSupport.hpp>
+#include <cl/ClTensorHandle.hpp>
+#include <cl/ClLayerSupport.hpp>
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <backendsCommon/CpuTensorHandle.hpp>
+
+#include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
+                                                             const TensorInfo& output,
+                                                             const TransposeConvolution2dDescriptor& descriptor,
+                                                             const TensorInfo& weights,
+                                                             const Optional<TensorInfo>& biases)
+{
+    arm_compute::TensorInfo aclInputInfo   = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+    arm_compute::TensorInfo aclOutputInfo  = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
+    arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
+
+    arm_compute::TensorInfo aclBiasesInfo;
+    arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
+
+    if (descriptor.m_BiasEnabled)
+    {
+        BOOST_ASSERT(biases.has_value());
+
+        aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
+        optionalAclBiasesInfo = &aclBiasesInfo;
+    }
+
+    arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(descriptor);
+
+    constexpr unsigned int innerBorderRight = 0u;
+    constexpr unsigned int innerBorderTop   = 0u;
+
+    return arm_compute::CLDeconvolutionLayer::validate(&aclInputInfo,
+                                                       &aclWeightsInfo,
+                                                       optionalAclBiasesInfo,
+                                                       &aclOutputInfo,
+                                                       padStrideInfo,
+                                                       innerBorderRight,
+                                                       innerBorderTop);
+}
+
+ClTransposeConvolution2dWorkload::ClTransposeConvolution2dWorkload(
+    const TransposeConvolution2dQueueDescriptor& descriptor,
+    const WorkloadInfo& info,
+    std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) :
+    BaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info),
+    m_Layer(memoryManager)
+{
+    const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
+
+    m_WeightsTensor = std::make_unique<arm_compute::CLTensor>();
+    BuildArmComputeTensor(*m_WeightsTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
+
+    arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
+                                             m_Data.m_Parameters.m_StrideY,
+                                             m_Data.m_Parameters.m_PadLeft,
+                                             m_Data.m_Parameters.m_PadRight,
+                                             m_Data.m_Parameters.m_PadTop,
+                                             m_Data.m_Parameters.m_PadBottom,
+                                             arm_compute::DimensionRoundingType::FLOOR);
+
+    if (m_Data.m_Parameters.m_BiasEnabled)
+    {
+        m_BiasesTensor = std::make_unique<arm_compute::CLTensor>();
+        BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
+    }
+
+    m_Data.ValidateInputsOutputs("ClTransposeConvolution2dWorkload", 1, 1);
+
+    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();
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+
+    input.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
+    m_Layer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, padStrideInfo);
+
+    InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight);
+    if (m_BiasesTensor)
+    {
+        InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias);
+    }
+
+    m_Layer.prepare();
+
+    FreeUnusedTensors();
+}
+
+void ClTransposeConvolution2dWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_CL("ClTransposeConvolution2dWorkload_Execute");
+    RunClFunction(m_Layer, CHECK_LOCATION());
+}
+
+void ClTransposeConvolution2dWorkload::FreeUnusedTensors()
+{
+    FreeTensorIfUnused(m_WeightsTensor);
+    FreeTensorIfUnused(m_BiasesTensor);
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp
new file mode 100644
index 0000000000..b7320bf6ce
--- /dev/null
+++ b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp
@@ -0,0 +1,46 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Tensor.hpp>
+#include <armnn/Descriptors.hpp>
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h>
+#include <arm_compute/runtime/MemoryManagerOnDemand.h>
+
+#include <memory>
+
+namespace armnn
+{
+
+arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input,
+                                                             const TensorInfo& output,
+                                                             const TransposeConvolution2dDescriptor& descriptor,
+                                                             const TensorInfo& weights,
+                                                             const Optional<TensorInfo>& biases);
+
+class ClTransposeConvolution2dWorkload : public BaseWorkload<TransposeConvolution2dQueueDescriptor>
+{
+public:
+    ClTransposeConvolution2dWorkload(const TransposeConvolution2dQueueDescriptor& descriptor,
+                                     const WorkloadInfo& info,
+                                     std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
+
+    void Execute() const override;
+
+private:
+    mutable arm_compute::CLDeconvolutionLayer m_Layer;
+
+    std::unique_ptr<arm_compute::CLTensor> m_WeightsTensor;
+    std::unique_ptr<arm_compute::CLTensor> m_BiasesTensor;
+
+    void FreeUnusedTensors();
+};
+
+} // namespace armnn
+
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index b836781166..e82afc54c3 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -38,3 +38,4 @@
 #include "ClSubtractionWorkload.hpp"
 #include "ClConvertFp16ToFp32Workload.hpp"
 #include "ClConvertFp32ToFp16Workload.hpp"
+#include "ClTransposeConvolution2dWorkload.hpp"