From a2747487fbe7eb6d9f5357c6d16c32355ed6e01c Mon Sep 17 00:00:00 2001 From: Sadik Armagan Date: Tue, 9 Feb 2021 10:28:54 +0000 Subject: MLCE-347 'REDUCE_MIN, REDUCE_MAX, REDUCE_SUM Support' * Added TfLiteParser support for REDUCE_MIN and REDUCE_MAX operators * Added ACL workloads support for REDUCE_MIN, REDUCE_MAX, and REDUCE_SUM operators * Added TfLite Delegate support for REDUCE_MIN, REDUCE_MAX, and REDUCE_SUM operators Signed-off-by: Sadik Armagan Change-Id: I8085d59946bfd4ab78a59a61f899031ae53371a8 --- src/backends/neon/NeonLayerSupport.cpp | 13 +++++ src/backends/neon/NeonLayerSupport.hpp | 5 ++ src/backends/neon/NeonWorkloadFactory.cpp | 6 ++ src/backends/neon/NeonWorkloadFactory.hpp | 3 + src/backends/neon/backend.mk | 1 + src/backends/neon/test/NeonLayerTests.cpp | 15 +++++ src/backends/neon/workloads/CMakeLists.txt | 2 + src/backends/neon/workloads/NeonReduceWorkload.cpp | 66 ++++++++++++++++++++++ src/backends/neon/workloads/NeonReduceWorkload.hpp | 30 ++++++++++ src/backends/neon/workloads/NeonWorkloads.hpp | 1 + 10 files changed, 142 insertions(+) create mode 100644 src/backends/neon/workloads/NeonReduceWorkload.cpp create mode 100644 src/backends/neon/workloads/NeonReduceWorkload.hpp (limited to 'src/backends/neon') diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp index 2d22576e57..66999c1a30 100644 --- a/src/backends/neon/NeonLayerSupport.cpp +++ b/src/backends/neon/NeonLayerSupport.cpp @@ -58,6 +58,7 @@ #include "workloads/NeonQLstmWorkload.hpp" #include "workloads/NeonQuantizeWorkload.hpp" #include "workloads/NeonQuantizedLstmWorkload.hpp" +#include "workloads/NeonReduceWorkload.hpp" #include "workloads/NeonReshapeWorkload.hpp" #include "workloads/NeonResizeWorkload.hpp" #include "workloads/NeonRsqrtWorkload.hpp" @@ -784,6 +785,18 @@ bool NeonLayerSupport::IsQuantizedLstmSupported(const TensorInfo& input, paramsInfo); } +bool NeonLayerSupport::IsReduceSupported(const TensorInfo& input, + const TensorInfo& output, + const ReduceDescriptor& descriptor, + Optional reasonIfUnsupported) const +{ + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonReduceWorkloadValidate, + reasonIfUnsupported, + input, + output, + descriptor); +} + bool NeonLayerSupport::IsReshapeSupported(const TensorInfo& input, const TensorInfo& output, const ReshapeDescriptor& descriptor, diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp index dc13cc2e4e..2ae1b0d489 100644 --- a/src/backends/neon/NeonLayerSupport.hpp +++ b/src/backends/neon/NeonLayerSupport.hpp @@ -263,6 +263,11 @@ public: const QuantizedLstmInputParamsInfo& paramsInfo, Optional reasonIfUnsupported = EmptyOptional()) const override; + bool IsReduceSupported(const TensorInfo& input, + const TensorInfo& output, + const ReduceDescriptor& descriptor, + Optional reasonIfUnsupported = EmptyOptional()) const override; + bool IsReshapeSupported(const TensorInfo& input, const TensorInfo& output, const ReshapeDescriptor& descriptor, diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp index 0d36110da5..7d0942874e 100644 --- a/src/backends/neon/NeonWorkloadFactory.cpp +++ b/src/backends/neon/NeonWorkloadFactory.cpp @@ -497,6 +497,12 @@ std::unique_ptr NeonWorkloadFactory::CreateRank(const RankQueueDescri return std::make_unique(descriptor, info); } +std::unique_ptr NeonWorkloadFactory::CreateReduce(const ReduceQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + return std::make_unique(descriptor, info); +} + std::unique_ptr NeonWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor, const WorkloadInfo& info) const { diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp index 745dece659..4817a06540 100644 --- a/src/backends/neon/NeonWorkloadFactory.hpp +++ b/src/backends/neon/NeonWorkloadFactory.hpp @@ -207,6 +207,9 @@ public: std::unique_ptr CreateRank(const RankQueueDescriptor& descriptor, const WorkloadInfo& info) const override; + std::unique_ptr CreateReduce(const ReduceQueueDescriptor& descriptor, + const WorkloadInfo& info) const override; + std::unique_ptr CreateReshape(const ReshapeQueueDescriptor& descriptor, const WorkloadInfo& info) const override; diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk index 54560cb0fa..6feeeb5f2c 100644 --- a/src/backends/neon/backend.mk +++ b/src/backends/neon/backend.mk @@ -66,6 +66,7 @@ BACKEND_SOURCES := \ workloads/NeonQLstmWorkload.cpp \ workloads/NeonQuantizedLstmWorkload.cpp \ workloads/NeonQuantizeWorkload.cpp \ + workloads/NeonReduceWorkload.cpp \ workloads/NeonReshapeWorkload.cpp \ workloads/NeonResizeWorkload.cpp \ workloads/NeonRsqrtWorkload.cpp \ diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp index d351870645..8434a67082 100644 --- a/src/backends/neon/test/NeonLayerTests.cpp +++ b/src/backends/neon/test/NeonLayerTests.cpp @@ -1372,6 +1372,21 @@ ARMNN_AUTO_TEST_CASE_WITH_THF(LogicalOrBroadcast1, LogicalOrBroadcast1Test) ARMNN_AUTO_TEST_CASE_WITH_THF(LogicalOrBroadcast2, LogicalOrBroadcast2Test) ARMNN_AUTO_TEST_CASE_WITH_THF(LogicalOrBroadcast3, LogicalOrBroadcast3Test) +// ReduceSum +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceSumFloat32, ReduceSumSimpleTest) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceSumSingleAxisFloat32_1, ReduceSumSingleAxisTest1) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceSumSingleAxisFloat32_2, ReduceSumSingleAxisTest2) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceSumSingleAxisFloat32_3, ReduceSumSingleAxisTest3) + +// ReduceMax +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceMaxFloat32, ReduceMaxSimpleTest) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceMaxNegativeAxisFloat32, ReduceMaxNegativeAxisTest) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceMax2Float32, ReduceMaxSimpleTest2) + +// ReduceMin +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceMinFloat32, ReduceMinSimpleTest) +ARMNN_AUTO_TEST_CASE_WITH_THF(ReduceMinNegativeAxisFloat32, ReduceMinNegativeAxisTest) + #if defined(ARMNNREF_ENABLED) // The ARMNN_COMPARE_REF_AUTO_TEST_CASE and the ARMNN_COMPARE_REF_FIXTURE_TEST_CASE test units are not available diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt index f1a723b324..7c2b185ec3 100644 --- a/src/backends/neon/workloads/CMakeLists.txt +++ b/src/backends/neon/workloads/CMakeLists.txt @@ -93,6 +93,8 @@ list(APPEND armnnNeonBackendWorkloads_sources NeonQuantizeWorkload.cpp NeonQuantizeWorkload.hpp NeonRankWorkload.hpp + NeonReduceWorkload.cpp + NeonReduceWorkload.hpp NeonReshapeWorkload.cpp NeonReshapeWorkload.hpp NeonResizeWorkload.cpp diff --git a/src/backends/neon/workloads/NeonReduceWorkload.cpp b/src/backends/neon/workloads/NeonReduceWorkload.cpp new file mode 100644 index 0000000000..0e1b46a3a1 --- /dev/null +++ b/src/backends/neon/workloads/NeonReduceWorkload.cpp @@ -0,0 +1,66 @@ +// +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "NeonReduceWorkload.hpp" + +#include +#include + +#include + +#include "NeonWorkloadUtils.hpp" + +namespace armnn +{ +using namespace armcomputetensorutils; + +arm_compute::Status NeonReduceWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const ReduceDescriptor& desc) +{ + const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); + const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); + if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1) + { + return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, + "NeonReduceWorkload: Reduction is supported only on 1 axis."); + } + + arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(), + input.GetNumDimensions(), + desc.m_vAxis); + + return arm_compute::NEReductionOperation::validate(&aclInputInfo, + &aclOutputInfo, + static_cast(coords[0]), + ConvertReductionOperationToAcl(desc), + desc.m_KeepDims); +} + +NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info) + : BaseWorkload(descriptor, info) +{ + m_Data.ValidateInputsOutputs("NeonReduceWorkload", 1, 1); + + arm_compute::ITensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ITensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); + + arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(), + info.m_InputTensorInfos[0].GetNumDimensions(), + m_Data.m_Parameters.m_vAxis); + m_Layer.configure(&input, + &output, + static_cast(coords[0]), + ConvertReductionOperationToAcl(m_Data.m_Parameters), + m_Data.m_Parameters.m_KeepDims); +} + +void NeonReduceWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonReduceWorkload_Execute"); + m_Layer.run(); +} + +} //namespace armnn diff --git a/src/backends/neon/workloads/NeonReduceWorkload.hpp b/src/backends/neon/workloads/NeonReduceWorkload.hpp new file mode 100644 index 0000000000..0472091fbf --- /dev/null +++ b/src/backends/neon/workloads/NeonReduceWorkload.hpp @@ -0,0 +1,30 @@ +// +// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +#include + +namespace armnn +{ + +arm_compute::Status NeonReduceWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const ReduceDescriptor& desc); + +class NeonReduceWorkload : public BaseWorkload +{ +public: + NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info); + + void Execute() const override; + +private: + mutable arm_compute::NEReductionOperation m_Layer; +}; + +} //namespace armnn diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp index 949100d50a..4eb526a04d 100644 --- a/src/backends/neon/workloads/NeonWorkloads.hpp +++ b/src/backends/neon/workloads/NeonWorkloads.hpp @@ -49,6 +49,7 @@ #include "NeonQuantizedLstmWorkload.hpp" #include "NeonQuantizeWorkload.hpp" #include "NeonRankWorkload.hpp" +#include "NeonReduceWorkload.hpp" #include "NeonReshapeWorkload.hpp" #include "NeonResizeWorkload.hpp" #include "NeonRsqrtWorkload.hpp" -- cgit v1.2.1