From c1f6b09cc22b6e2de3d9eb88aec1778d0308a2b3 Mon Sep 17 00:00:00 2001 From: Teresa Charlin Date: Mon, 11 May 2020 16:10:38 +0100 Subject: IVGCVSW-4753 Refactor CL Softmax workload generalizing for different datatype * Change ComputeSoftmaxAclAxis to work with int and uint axis Signed-off-by: Teresa Charlin Change-Id: Ibbfa9ec7e2f0416e6885673212a767419c871cca --- src/backends/aclCommon/ArmComputeUtils.hpp | 5 ++- src/backends/cl/ClLayerSupport.cpp | 2 +- src/backends/cl/ClWorkloadFactory.cpp | 3 +- src/backends/cl/backend.mk | 4 +- src/backends/cl/test/ClCreateWorkloadTests.cpp | 28 +++++++++++-- src/backends/cl/workloads/CMakeLists.txt | 8 +--- .../cl/workloads/ClSoftmaxBaseWorkload.cpp | 27 ------------ .../cl/workloads/ClSoftmaxBaseWorkload.hpp | 19 --------- .../cl/workloads/ClSoftmaxFloatWorkload.cpp | 36 ---------------- .../cl/workloads/ClSoftmaxFloatWorkload.hpp | 30 -------------- .../cl/workloads/ClSoftmaxUint8Workload.cpp | 46 --------------------- .../cl/workloads/ClSoftmaxUint8Workload.hpp | 31 -------------- src/backends/cl/workloads/ClSoftmaxWorkload.cpp | 48 ++++++++++++++++++++++ src/backends/cl/workloads/ClSoftmaxWorkload.hpp | 34 +++++++++++++++ src/backends/cl/workloads/ClWorkloads.hpp | 3 +- .../neon/workloads/NeonSoftmaxWorkload.cpp | 4 +- 16 files changed, 118 insertions(+), 210 deletions(-) delete mode 100644 src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp delete mode 100644 src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp delete mode 100644 src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp delete mode 100644 src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp delete mode 100644 src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp delete mode 100644 src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp create mode 100644 src/backends/cl/workloads/ClSoftmaxWorkload.cpp create mode 100644 src/backends/cl/workloads/ClSoftmaxWorkload.hpp diff --git a/src/backends/aclCommon/ArmComputeUtils.hpp b/src/backends/aclCommon/ArmComputeUtils.hpp index 80bb7623e8..d8818ce209 100644 --- a/src/backends/aclCommon/ArmComputeUtils.hpp +++ b/src/backends/aclCommon/ArmComputeUtils.hpp @@ -150,7 +150,8 @@ inline arm_compute::InterpolationPolicy ConvertResizeMethodToAclInterpolationPol } } -inline unsigned int ComputeSoftmaxAclAxis(const SoftmaxDescriptor& softmaxDesc, const armnn::TensorInfo& tensor) +template +inline T ComputeSoftmaxAclAxis(const SoftmaxDescriptor& softmaxDesc, const armnn::TensorInfo& tensor) { // Detect the Android default value of -1 and return the ACL default value of 1. if (softmaxDesc.m_Axis == -1) @@ -163,7 +164,7 @@ inline unsigned int ComputeSoftmaxAclAxis(const SoftmaxDescriptor& softmaxDesc, ARMNN_ASSERT(dim != 0); // Currently ArmNN support axis 1. - return dim - 1; + return static_cast(dim) - 1; } inline std::set ComputeSplitAxis(const armnn::SplitterDescriptor& desc, const TensorShape& input) diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp index 546cbc1e59..eb68a80765 100644 --- a/src/backends/cl/ClLayerSupport.cpp +++ b/src/backends/cl/ClLayerSupport.cpp @@ -53,7 +53,7 @@ #include "workloads/ClQuantizedLstmWorkload.hpp" #include "workloads/ClQuantizeWorkload.hpp" #include "workloads/ClSliceWorkload.hpp" -#include "workloads/ClSoftmaxBaseWorkload.hpp" +#include "workloads/ClSoftmaxWorkload.hpp" #include "workloads/ClSpaceToBatchNdWorkload.hpp" #include "workloads/ClSpaceToDepthWorkload.hpp" #include "workloads/ClSplitterWorkload.hpp" diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp index b0d2fdf835..f584272e10 100644 --- a/src/backends/cl/ClWorkloadFactory.cpp +++ b/src/backends/cl/ClWorkloadFactory.cpp @@ -490,8 +490,7 @@ std::unique_ptr ClWorkloadFactory::CreateSlice(const SliceQueueDescri std::unique_ptr ClWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info) const { - return MakeWorkload(descriptor, info, - m_MemoryManager->GetIntraLayerManager()); + return std::make_unique(descriptor, info, m_MemoryManager->GetIntraLayerManager()); } std::unique_ptr ClWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor, diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk index 3f2e80824d..8b464bb1d6 100644 --- a/src/backends/cl/backend.mk +++ b/src/backends/cl/backend.mk @@ -59,9 +59,7 @@ BACKEND_SOURCES := \ workloads/ClResizeWorkload.cpp \ workloads/ClRsqrtWorkload.cpp \ workloads/ClSliceWorkload.cpp \ - workloads/ClSoftmaxBaseWorkload.cpp \ - workloads/ClSoftmaxFloatWorkload.cpp \ - workloads/ClSoftmaxUint8Workload.cpp \ + workloads/ClSoftmaxWorkload.cpp \ workloads/ClSpaceToBatchNdWorkload.cpp \ workloads/ClSpaceToDepthWorkload.cpp \ workloads/ClSplitterWorkload.cpp \ diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp index b09b26f9b3..b7522547d4 100644 --- a/src/backends/cl/test/ClCreateWorkloadTests.cpp +++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp @@ -569,19 +569,41 @@ static void ClSoftmaxWorkloadTest() auto inputHandle = PolymorphicDowncast(queueDescriptor.m_Inputs[0]); auto outputHandle = PolymorphicDowncast(queueDescriptor.m_Outputs[0]); + 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(CompareIClTensorHandleShape(inputHandle, {4, 1})); BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {4, 1})); } -BOOST_AUTO_TEST_CASE(CreateSoftmaxFloatWorkloadTest) +BOOST_AUTO_TEST_CASE(CreateSoftmaxFloat32WorkloadTest) { - ClSoftmaxWorkloadTest(); + ClSoftmaxWorkloadTest(); } BOOST_AUTO_TEST_CASE(CreateSoftmaxFloat16WorkloadTest) { - ClSoftmaxWorkloadTest(); + ClSoftmaxWorkloadTest(); +} + +BOOST_AUTO_TEST_CASE(CreateSoftmaxQAsymmU8Workload) +{ + ClSoftmaxWorkloadTest(); +} + +BOOST_AUTO_TEST_CASE(CreateSoftmaxQAsymmS8Workload) +{ + ClSoftmaxWorkloadTest(); } template diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt index 161ad96361..e595028cbb 100644 --- a/src/backends/cl/workloads/CMakeLists.txt +++ b/src/backends/cl/workloads/CMakeLists.txt @@ -78,12 +78,8 @@ list(APPEND armnnClBackendWorkloads_sources ClRsqrtWorkload.hpp ClSliceWorkload.cpp ClSliceWorkload.hpp - ClSoftmaxBaseWorkload.cpp - ClSoftmaxBaseWorkload.hpp - ClSoftmaxFloatWorkload.cpp - ClSoftmaxFloatWorkload.hpp - ClSoftmaxUint8Workload.cpp - ClSoftmaxUint8Workload.hpp + ClSoftmaxWorkload.cpp + ClSoftmaxWorkload.hpp ClSpaceToBatchNdWorkload.hpp ClSpaceToBatchNdWorkload.cpp ClSpaceToDepthWorkload.cpp diff --git a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp deleted file mode 100644 index a355ba0c2d..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp +++ /dev/null @@ -1,27 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxBaseWorkload.hpp" - -#include -#include - -#include - -namespace armnn -{ - -arm_compute::Status ClSoftmaxWorkloadValidate(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::CLSoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis); -} - -} diff --git a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp deleted file mode 100644 index 8d73060162..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp +++ /dev/null @@ -1,19 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include -#include -#include - -namespace armnn -{ - -arm_compute::Status ClSoftmaxWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SoftmaxDescriptor& descriptor); - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp deleted file mode 100644 index adb4872b80..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp +++ /dev/null @@ -1,36 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxFloatWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include -#include -#include - -namespace armnn -{ - -ClSoftmaxFloatWorkload::ClSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr& memoryManager) - : FloatWorkload(descriptor, info) - , m_SoftmaxLayer(memoryManager) -{ - m_Data.ValidateInputsOutputs("ClSoftmaxFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); - - unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); - m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis); -} - -void ClSoftmaxFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSoftmaxFloatWorkload_Execute"); - RunClFunction(m_SoftmaxLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp deleted file mode 100644 index 7efdae858a..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp +++ /dev/null @@ -1,30 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include - -#include -#include - -#include - -namespace armnn -{ - -class ClSoftmaxFloatWorkload : public FloatWorkload -{ -public: - ClSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr& memoryManager); - void Execute() const override; - -private: - mutable arm_compute::CLSoftmaxLayer m_SoftmaxLayer; -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp deleted file mode 100644 index f14ea11c82..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp +++ /dev/null @@ -1,46 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxUint8Workload.hpp" -#include "ClWorkloadUtils.hpp" - -#include -#include -#include - -namespace armnn -{ - -ClSoftmaxUint8Workload::ClSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr& memoryManager) - : Uint8Workload(descriptor, info) - , m_SoftmaxLayer(memoryManager) -{ - m_Data.ValidateInputsOutputs("ClSoftmaxUint8Workload", 1, 1); - - arm_compute::ICLTensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast(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"); - } - - unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); - m_SoftmaxLayer.configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis); -} - -void ClSoftmaxUint8Workload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSoftmaxUint8Workload_Execute"); - RunClFunction(m_SoftmaxLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp deleted file mode 100644 index f378b89a5d..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp +++ /dev/null @@ -1,31 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include - -#include -#include - -#include - -namespace armnn -{ -// Softmax -class ClSoftmaxUint8Workload : public Uint8Workload -{ -public: - ClSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr& memoryManager); - - void Execute() const override; -private: - - mutable arm_compute::CLSoftmaxLayer m_SoftmaxLayer; -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClSoftmaxWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxWorkload.cpp new file mode 100644 index 0000000000..cbca7668ed --- /dev/null +++ b/src/backends/cl/workloads/ClSoftmaxWorkload.cpp @@ -0,0 +1,48 @@ +// +// Copyright © 2020 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "ClSoftmaxWorkload.hpp" +#include "ClWorkloadUtils.hpp" + +#include +#include + +#include + +namespace armnn +{ + +arm_compute::Status ClSoftmaxWorkloadValidate(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::CLSoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis); +} + +ClSoftmaxWorkload::ClSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr& memoryManager) + : BaseWorkload(descriptor, info) + , m_SoftmaxLayer(memoryManager) +{ + m_Data.ValidateInputsOutputs("ClSoftmaxWorkload", 1, 1); + + arm_compute::ICLTensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ICLTensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); + + unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); + m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis); +} + +void ClSoftmaxWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_CL("ClSoftmaxWorkload_Execute"); + RunClFunction(m_SoftmaxLayer, CHECK_LOCATION()); +} + +} // namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxWorkload.hpp b/src/backends/cl/workloads/ClSoftmaxWorkload.hpp new file mode 100644 index 0000000000..158bf46c32 --- /dev/null +++ b/src/backends/cl/workloads/ClSoftmaxWorkload.hpp @@ -0,0 +1,34 @@ +// +// Copyright © 2020 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +#include +#include +#include + +#include + +namespace armnn +{ + +arm_compute::Status ClSoftmaxWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const SoftmaxDescriptor& descriptor); + +class ClSoftmaxWorkload : public BaseWorkload +{ +public: + ClSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr& memoryManager); + void Execute() const override; + +private: + mutable arm_compute::CLSoftmaxLayer m_SoftmaxLayer; +}; + +} // namespace armnn diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp index ffe66a0716..62b73daef3 100644 --- a/src/backends/cl/workloads/ClWorkloads.hpp +++ b/src/backends/cl/workloads/ClWorkloads.hpp @@ -39,8 +39,7 @@ #include "ClResizeWorkload.hpp" #include "ClRsqrtWorkload.hpp" #include "ClSliceWorkload.hpp" -#include "ClSoftmaxFloatWorkload.hpp" -#include "ClSoftmaxUint8Workload.hpp" +#include "ClSoftmaxWorkload.hpp" #include "ClSpaceToBatchNdWorkload.hpp" #include "ClSpaceToDepthWorkload.hpp" #include "ClSplitterWorkload.hpp" diff --git a/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp b/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp index 149804bdd6..35e2e0e3fc 100644 --- a/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp +++ b/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp @@ -23,7 +23,7 @@ arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input, const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); - unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input); + unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input); return arm_compute::NESoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis); } @@ -38,7 +38,7 @@ NeonSoftmaxWorkload::NeonSoftmaxWorkload(const SoftmaxQueueDescriptor& descripto arm_compute::ITensor& output = PolymorphicDowncast(m_Data.m_Outputs[0])->GetTensor(); auto layer = std::make_unique(memoryManager); - unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); + 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()); } -- cgit v1.2.1