// Connect up
armnn::TensorInfo tensorInfo({4, 1}, DataType);
+ if (DataType == armnn::DataType::QAsymmU8)
+ {
+ tensorInfo.SetQuantizationOffset(0);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
+ else if (DataType == armnn::DataType::QAsymmS8)
+ {
+ tensorInfo.SetQuantizationOffset(-128);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
+
Connect(input, layer, tensorInfo);
Connect(layer, output, tensorInfo);
CreateTensorHandles(graph, factory);
#include "workloads/NeonResizeWorkload.hpp"
#include "workloads/NeonRsqrtWorkload.hpp"
#include "workloads/NeonSliceWorkload.hpp"
-#include "workloads/NeonSoftmaxBaseWorkload.hpp"
+#include "workloads/NeonSoftmaxWorkload.hpp"
#include "workloads/NeonSpaceToBatchNdWorkload.hpp"
#include "workloads/NeonSpaceToDepthWorkload.hpp"
#include "workloads/NeonSplitterWorkload.hpp"
std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
const WorkloadInfo& info) const
{
- return MakeWorkloadHelper<NeonSoftmaxFloatWorkload, NeonSoftmaxUint8Workload>(
- descriptor, info, m_MemoryManager->GetIntraLayerManager());
+ return std::make_unique<NeonSoftmaxWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
}
std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor,
workloads/NeonResizeWorkload.cpp \
workloads/NeonRsqrtWorkload.cpp \
workloads/NeonSliceWorkload.cpp \
- workloads/NeonSoftmaxBaseWorkload.cpp \
- workloads/NeonSoftmaxFloatWorkload.cpp \
- workloads/NeonSoftmaxUint8Workload.cpp \
+ workloads/NeonSoftmaxWorkload.cpp \
workloads/NeonSpaceToBatchNdWorkload.cpp \
workloads/NeonSpaceToDepthWorkload.cpp \
workloads/NeonSplitterWorkload.cpp \
SoftmaxQueueDescriptor queueDescriptor = workload->GetData();
auto inputHandle = PolymorphicDowncast<IAclTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto outputHandle = PolymorphicDowncast<IAclTensorHandle*>(queueDescriptor.m_Outputs[0]);
- BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({4, 1}, DataType)));
- BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({4, 1}, DataType)));
+ armnn::TensorInfo tensorInfo({4, 1}, DataType);
+ if (DataType == armnn::DataType::QAsymmU8)
+ {
+ tensorInfo.SetQuantizationOffset(0);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
+ else if (DataType == armnn::DataType::QAsymmS8)
+ {
+ tensorInfo.SetQuantizationOffset(-128);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
+ BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, tensorInfo));
+ BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, tensorInfo));
}
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
BOOST_AUTO_TEST_CASE(CreateSoftmaxFloat16Workload)
{
- NeonCreateSoftmaxWorkloadTest<NeonSoftmaxFloatWorkload, DataType::Float16>();
+ NeonCreateSoftmaxWorkloadTest<NeonSoftmaxWorkload, DataType::Float16>();
}
#endif
BOOST_AUTO_TEST_CASE(CreateSoftmaxFloatWorkload)
{
- NeonCreateSoftmaxWorkloadTest<NeonSoftmaxFloatWorkload, DataType::Float32>();
+ NeonCreateSoftmaxWorkloadTest<NeonSoftmaxWorkload, DataType::Float32>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateSoftmaxQAsymmU8Workload)
+{
+ NeonCreateSoftmaxWorkloadTest<NeonSoftmaxWorkload, DataType::QAsymmU8>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateSoftmaxQAsymmS8Workload)
+{
+ NeonCreateSoftmaxWorkloadTest<NeonSoftmaxWorkload, DataType::QAsymmS8>();
}
template <typename SpaceToDepthWorkloadType, typename armnn::DataType DataType>
NeonRsqrtWorkload.hpp
NeonSliceWorkload.cpp
NeonSliceWorkload.hpp
- NeonSoftmaxBaseWorkload.cpp
- NeonSoftmaxBaseWorkload.hpp
- NeonSoftmaxFloatWorkload.cpp
- NeonSoftmaxFloatWorkload.hpp
- NeonSoftmaxUint8Workload.cpp
- NeonSoftmaxUint8Workload.hpp
+ NeonSoftmaxWorkload.cpp
+ NeonSoftmaxWorkload.hpp
NeonSpaceToBatchNdWorkload.cpp
NeonSpaceToBatchNdWorkload.hpp
NeonSpaceToDepthWorkload.cpp
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "NeonSoftmaxBaseWorkload.hpp"
-
-#include <aclCommon/ArmComputeTensorUtils.hpp>
-#include <aclCommon/ArmComputeUtils.hpp>
-
-#include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
-
-namespace armnn
-{
-
-arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
- const TensorInfo& output,
- const SoftmaxDescriptor& descriptor)
-{
- const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
- const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
-
- unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input);
- return arm_compute::NESoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis);
-}
-
-} //namespace armnn
-
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <armnn/Descriptors.hpp>
-#include <arm_compute/core/Error.h>
-
-namespace armnn
-{
-
-arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
- const TensorInfo& output,
- const SoftmaxDescriptor& descriptor);
-
-} //namespace armnn
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "NeonSoftmaxFloatWorkload.hpp"
-
-#include "NeonWorkloadUtils.hpp"
-
-#include <aclCommon/ArmComputeUtils.hpp>
-#include <armnn/utility/PolymorphicDowncast.hpp>
-
-#include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
-
-namespace armnn
-{
-
-NeonSoftmaxFloatWorkload::NeonSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor,
- const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
- : FloatWorkload<SoftmaxQueueDescriptor>(descriptor, info)
-{
- m_Data.ValidateInputsOutputs("NeonSoftmaxFloatWorkload", 1, 1);
-
- // The ArmCompute softmax layer uses 2D input/output tensors, so flatten the first three dimensions.
- arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
- arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
-
- auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
- layer->configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
- m_SoftmaxLayer.reset(layer.release());
-}
-
-void NeonSoftmaxFloatWorkload::Execute() const
-{
- ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSoftmaxFloatWorkload_Execute");
- m_SoftmaxLayer->run();
-}
-
-} //namespace armnn
-
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <backendsCommon/Workload.hpp>
-
-#include <arm_compute/runtime/IFunction.h>
-#include <arm_compute/runtime/MemoryManagerOnDemand.h>
-
-#include <memory>
-
-namespace armnn
-{
-
-class NeonSoftmaxFloatWorkload : public FloatWorkload<SoftmaxQueueDescriptor>
-{
-public:
- NeonSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info,
- std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
- virtual void Execute() const override;
-
-private:
- std::unique_ptr<arm_compute::IFunction> m_SoftmaxLayer;
-};
-
-} //namespace armnn
-
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "NeonSoftmaxUint8Workload.hpp"
-#include "NeonWorkloadUtils.hpp"
-
-#include <aclCommon/ArmComputeUtils.hpp>
-#include <armnn/utility/PolymorphicDowncast.hpp>
-
-#include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
-
-namespace armnn
-{
-
-NeonSoftmaxUint8Workload::NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor,
- const WorkloadInfo& info,
- std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
- : Uint8Workload<SoftmaxQueueDescriptor>(descriptor, info)
-{
- m_Data.ValidateInputsOutputs("NeonSoftmaxUint8Workload", 1, 1);
-
- arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
- arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
-
- const auto outputQuantization = output.info()->quantization_info();
-
- if ((!outputQuantization.scale().empty() && outputQuantization.scale()[0] != (1.0f / 256.0f)) ||
- (!outputQuantization.offset().empty() && outputQuantization.offset()[0] != 0) ||
- outputQuantization.scale().empty() || outputQuantization.offset().empty())
- {
- throw InvalidArgumentException(
- "Invalid quantization for output. Only scale = 1.0f / 256.0f and offset = 0 supported");
- }
-
- auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
- layer->configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis);
- m_SoftmaxLayer.reset(layer.release());
-}
-
-void NeonSoftmaxUint8Workload::Execute() const
-{
- ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSoftmaxUint8Workload_Execute");
-
- m_SoftmaxLayer->run();
-}
-
-} //namespace armnn
-
+++ /dev/null
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <backendsCommon/Workload.hpp>
-
-#include <arm_compute/runtime/IFunction.h>
-#include <arm_compute/runtime/MemoryManagerOnDemand.h>
-
-#include <memory>
-
-namespace armnn
-{
-
-class NeonSoftmaxUint8Workload : public Uint8Workload<SoftmaxQueueDescriptor>
-{
-public:
- NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info,
- std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
- virtual void Execute() const override;
-
-private:
- std::unique_ptr<arm_compute::IFunction> m_SoftmaxLayer;
-};
-
-} //namespace armnn
-
--- /dev/null
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonSoftmaxWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <armnn/utility/PolymorphicDowncast.hpp>
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& output,
+ const SoftmaxDescriptor& descriptor)
+{
+ const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input);
+ return arm_compute::NESoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis);
+}
+
+NeonSoftmaxWorkload::NeonSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor,
+ const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
+ : BaseWorkload<SoftmaxQueueDescriptor>(descriptor, info)
+{
+ m_Data.ValidateInputsOutputs("NeonSoftmaxWorkload", 1, 1);
+
+ // The ArmCompute softmax layer uses 2D input/output tensors, so flatten the first three dimensions.
+ arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+ auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
+ layer->configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
+ m_SoftmaxLayer.reset(layer.release());
+}
+
+void NeonSoftmaxWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSoftmaxWorkload_Execute");
+ m_SoftmaxLayer->run();
+}
+
+} //namespace armnn
+
--- /dev/null
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/IFunction.h>
+#include <arm_compute/runtime/MemoryManagerOnDemand.h>
+
+#include <memory>
+
+namespace armnn
+{
+
+arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& output,
+ const SoftmaxDescriptor& descriptor);
+
+class NeonSoftmaxWorkload : public BaseWorkload<SoftmaxQueueDescriptor>
+{
+public:
+ NeonSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info,
+ std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
+ virtual void Execute() const override;
+
+private:
+ std::unique_ptr<arm_compute::IFunction> m_SoftmaxLayer;
+};
+
+} //namespace armnn
+
#include "NeonResizeWorkload.hpp"
#include "NeonRsqrtWorkload.hpp"
#include "NeonSliceWorkload.hpp"
-#include "NeonSoftmaxFloatWorkload.hpp"
-#include "NeonSoftmaxUint8Workload.hpp"
+#include "NeonSoftmaxWorkload.hpp"
#include "NeonSpaceToBatchNdWorkload.hpp"
#include "NeonSpaceToDepthWorkload.hpp"
#include "NeonSplitterWorkload.hpp"
auto workload = CreateSoftmaxWorkloadTest<SoftmaxWorkloadType, DataType>(factory, graph);
// Checks that outputs and inputs are as we expect them (see definition of CreateSoftmaxWorkloadTest).
+
+ armnn::TensorInfo tensorInfo({4, 1}, DataType);
+ if (DataType == armnn::DataType::QAsymmU8)
+ {
+ tensorInfo.SetQuantizationOffset(0);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
+ else if (DataType == armnn::DataType::QAsymmS8)
+ {
+ tensorInfo.SetQuantizationOffset(-128);
+ tensorInfo.SetQuantizationScale(1.f / 256);
+ }
CheckInputOutput(
std::move(workload),
- TensorInfo({4, 1}, DataType),
- TensorInfo({4, 1}, DataType));
+ tensorInfo,
+ tensorInfo);
}
BOOST_AUTO_TEST_CASE(CreateSoftmaxFloat32Workload)