IVGCVSW-2202 Refactoring Arithmetic* names to Elementwise* names for workloads and workload functions
Change-Id: I6f3fce12a55f7d38ceafcdfcd6b5181bf56e2c09
diff --git a/src/armnn/test/CreateWorkload.hpp b/src/armnn/test/CreateWorkload.hpp
index b8ba72f..349c6a7 100644
--- a/src/armnn/test/CreateWorkload.hpp
+++ b/src/armnn/test/CreateWorkload.hpp
@@ -103,8 +103,8 @@
typename DescriptorType,
typename LayerType,
armnn::DataType DataType>
-std::unique_ptr<WorkloadType> CreateArithmeticWorkloadTest(armnn::IWorkloadFactory& factory,
- armnn::Graph& graph)
+std::unique_ptr<WorkloadType> CreateElementwiseWorkloadTest(armnn::IWorkloadFactory & factory,
+ armnn::Graph & graph)
{
// Creates the layer we're testing.
Layer* const layer = graph.AddLayer<LayerType>("layer");
diff --git a/src/backends/README.md b/src/backends/README.md
index ddd1bb6..60e4d0b 100644
--- a/src/backends/README.md
+++ b/src/backends/README.md
@@ -68,7 +68,7 @@
RefLayerSupport.cpp \
RefWorkloadFactory.cpp \
workloads/Activation.cpp \
- workloads/ArithmeticFunction.cpp \
+ workloads/ElementwiseFunction.cpp \
workloads/Broadcast.cpp \
...
diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp
index 8cef9d7..c5f685d 100644
--- a/src/backends/cl/test/ClCreateWorkloadTests.cpp
+++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp
@@ -55,15 +55,15 @@
typename DescriptorType,
typename LayerType,
armnn::DataType DataType>
-static void ClCreateArithmethicWorkloadTest()
+static void ClCreateElementwiseWorkloadTest()
{
Graph graph;
ClWorkloadFactory factory =
ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager());
- auto workload = CreateArithmeticWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
+ auto workload = CreateElementwiseWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
- // Checks that inputs/outputs are as we expect them (see definition of CreateArithmeticWorkloadTest).
+ // Checks that inputs/outputs are as we expect them (see definition of CreateElementwiseWorkloadTest).
DescriptorType queueDescriptor = workload->GetData();
auto inputHandle1 = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
auto inputHandle2 = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[1]);
@@ -75,7 +75,7 @@
BOOST_AUTO_TEST_CASE(CreateAdditionFloatWorkload)
{
- ClCreateArithmethicWorkloadTest<ClAdditionWorkload,
+ ClCreateElementwiseWorkloadTest<ClAdditionWorkload,
AdditionQueueDescriptor,
AdditionLayer,
armnn::DataType::Float32>();
@@ -83,7 +83,7 @@
BOOST_AUTO_TEST_CASE(CreateAdditionFloat16Workload)
{
- ClCreateArithmethicWorkloadTest<ClAdditionWorkload,
+ ClCreateElementwiseWorkloadTest<ClAdditionWorkload,
AdditionQueueDescriptor,
AdditionLayer,
armnn::DataType::Float16>();
@@ -91,7 +91,7 @@
BOOST_AUTO_TEST_CASE(CreateSubtractionFloatWorkload)
{
- ClCreateArithmethicWorkloadTest<ClSubtractionWorkload,
+ ClCreateElementwiseWorkloadTest<ClSubtractionWorkload,
SubtractionQueueDescriptor,
SubtractionLayer,
armnn::DataType::Float32>();
@@ -99,7 +99,7 @@
BOOST_AUTO_TEST_CASE(CreateSubtractionFloat16Workload)
{
- ClCreateArithmethicWorkloadTest<ClSubtractionWorkload,
+ ClCreateElementwiseWorkloadTest<ClSubtractionWorkload,
SubtractionQueueDescriptor,
SubtractionLayer,
armnn::DataType::Float16>();
@@ -107,7 +107,7 @@
BOOST_AUTO_TEST_CASE(CreateMultiplicationFloatWorkloadTest)
{
- ClCreateArithmethicWorkloadTest<ClMultiplicationWorkload,
+ ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload,
MultiplicationQueueDescriptor,
MultiplicationLayer,
armnn::DataType::Float32>();
@@ -115,7 +115,7 @@
BOOST_AUTO_TEST_CASE(CreateMultiplicationFloat16WorkloadTest)
{
- ClCreateArithmethicWorkloadTest<ClMultiplicationWorkload,
+ ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload,
MultiplicationQueueDescriptor,
MultiplicationLayer,
armnn::DataType::Float16>();
@@ -123,7 +123,7 @@
BOOST_AUTO_TEST_CASE(CreateMultiplicationUint8WorkloadTest)
{
- ClCreateArithmethicWorkloadTest<ClMultiplicationWorkload,
+ ClCreateElementwiseWorkloadTest<ClMultiplicationWorkload,
MultiplicationQueueDescriptor,
MultiplicationLayer,
armnn::DataType::QuantisedAsymm8>();
@@ -131,7 +131,7 @@
BOOST_AUTO_TEST_CASE(CreateDivisionFloatWorkloadTest)
{
- ClCreateArithmethicWorkloadTest<ClDivisionFloatWorkload,
+ ClCreateElementwiseWorkloadTest<ClDivisionFloatWorkload,
DivisionQueueDescriptor,
DivisionLayer,
armnn::DataType::Float32>();
@@ -139,7 +139,7 @@
BOOST_AUTO_TEST_CASE(CreateDivisionFloat16WorkloadTest)
{
- ClCreateArithmethicWorkloadTest<ClDivisionFloatWorkload,
+ ClCreateElementwiseWorkloadTest<ClDivisionFloatWorkload,
DivisionQueueDescriptor,
DivisionLayer,
armnn::DataType::Float16>();
diff --git a/src/backends/neon/test/NeonCreateWorkloadTests.cpp b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
index 1201253..dc6ec16 100644
--- a/src/backends/neon/test/NeonCreateWorkloadTests.cpp
+++ b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
@@ -87,13 +87,13 @@
typename DescriptorType,
typename LayerType,
armnn::DataType DataType>
-static void NeonCreateArithmethicWorkloadTest()
+static void NeonCreateElementwiseWorkloadTest()
{
Graph graph;
NeonWorkloadFactory factory =
NeonWorkloadFactoryHelper::GetFactory(NeonWorkloadFactoryHelper::GetMemoryManager());
- auto workload = CreateArithmeticWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
+ auto workload = CreateElementwiseWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
DescriptorType queueDescriptor = workload->GetData();
auto inputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
@@ -107,7 +107,7 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
BOOST_AUTO_TEST_CASE(CreateAdditionFloat16Workload)
{
- NeonCreateArithmethicWorkloadTest<NeonAdditionFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonAdditionFloatWorkload,
AdditionQueueDescriptor,
AdditionLayer,
DataType::Float16>();
@@ -116,7 +116,7 @@
BOOST_AUTO_TEST_CASE(CreateAdditionFloatWorkload)
{
- NeonCreateArithmethicWorkloadTest<NeonAdditionFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonAdditionFloatWorkload,
AdditionQueueDescriptor,
AdditionLayer,
DataType::Float32>();
@@ -125,7 +125,7 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
BOOST_AUTO_TEST_CASE(CreateSubtractionFloat16Workload)
{
- NeonCreateArithmethicWorkloadTest<NeonSubtractionFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonSubtractionFloatWorkload,
SubtractionQueueDescriptor,
SubtractionLayer,
DataType::Float16>();
@@ -134,7 +134,7 @@
BOOST_AUTO_TEST_CASE(CreateSubtractionFloatWorkload)
{
- NeonCreateArithmethicWorkloadTest<NeonSubtractionFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonSubtractionFloatWorkload,
SubtractionQueueDescriptor,
SubtractionLayer,
DataType::Float32>();
@@ -143,7 +143,7 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
BOOST_AUTO_TEST_CASE(CreateMultiplicationFloat16Workload)
{
- NeonCreateArithmethicWorkloadTest<NeonMultiplicationFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonMultiplicationFloatWorkload,
MultiplicationQueueDescriptor,
MultiplicationLayer,
DataType::Float16>();
@@ -152,7 +152,7 @@
BOOST_AUTO_TEST_CASE(CreateMultiplicationFloatWorkload)
{
- NeonCreateArithmethicWorkloadTest<NeonMultiplicationFloatWorkload,
+ NeonCreateElementwiseWorkloadTest<NeonMultiplicationFloatWorkload,
MultiplicationQueueDescriptor,
MultiplicationLayer,
DataType::Float32>();
diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk
index 7162d4a..66675bd 100644
--- a/src/backends/reference/backend.mk
+++ b/src/backends/reference/backend.mk
@@ -12,17 +12,16 @@
RefLayerSupport.cpp \
RefWorkloadFactory.cpp \
workloads/Activation.cpp \
- workloads/ArithmeticFunction.cpp \
workloads/BatchToSpaceNd.cpp \
workloads/Broadcast.cpp \
workloads/ConvImpl.cpp \
+ workloads/ElementwiseFunction.cpp \
workloads/FullyConnected.cpp \
workloads/Mean.cpp \
workloads/Pad.cpp \
workloads/Pooling2d.cpp \
workloads/RefActivationFloat32Workload.cpp \
workloads/RefActivationUint8Workload.cpp \
- workloads/RefArithmeticWorkload.cpp \
workloads/RefBaseConstantWorkload.cpp \
workloads/RefBatchNormalizationFloat32Workload.cpp \
workloads/RefBatchNormalizationUint8Workload.cpp \
@@ -36,6 +35,7 @@
workloads/RefConvolution2dUint8Workload.cpp \
workloads/RefDepthwiseConvolution2dFloat32Workload.cpp \
workloads/RefDepthwiseConvolution2dUint8Workload.cpp \
+ workloads/RefElementwiseWorkload.cpp \
workloads/RefFakeQuantizationFloat32Workload.cpp \
workloads/RefFloorFloat32Workload.cpp \
workloads/RefFullyConnectedFloat32Workload.cpp \
diff --git a/src/backends/reference/test/RefCreateWorkloadTests.cpp b/src/backends/reference/test/RefCreateWorkloadTests.cpp
index 47f9d0e..8621122 100644
--- a/src/backends/reference/test/RefCreateWorkloadTests.cpp
+++ b/src/backends/reference/test/RefCreateWorkloadTests.cpp
@@ -67,11 +67,12 @@
typename DescriptorType,
typename LayerType,
armnn::DataType DataType>
-static void RefCreateArithmethicWorkloadTest()
+static void RefCreateElementwiseWorkloadTest()
{
Graph graph;
RefWorkloadFactory factory;
- auto workload = CreateArithmeticWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
+ auto workload = CreateElementwiseWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(
+ factory, graph);
CheckInputsOutput(std::move(workload),
TensorInfo({ 2, 3 }, DataType),
@@ -81,66 +82,66 @@
BOOST_AUTO_TEST_CASE(CreateAdditionFloatWorkload)
{
- RefCreateArithmethicWorkloadTest<RefAdditionFloat32Workload,
- AdditionQueueDescriptor,
- AdditionLayer,
- armnn::DataType::Float32>();
+ RefCreateElementwiseWorkloadTest<RefAdditionFloat32Workload,
+ AdditionQueueDescriptor,
+ AdditionLayer,
+ armnn::DataType::Float32>();
}
BOOST_AUTO_TEST_CASE(CreateAdditionUint8Workload)
{
- RefCreateArithmethicWorkloadTest<RefAdditionUint8Workload,
- AdditionQueueDescriptor,
- AdditionLayer,
- armnn::DataType::QuantisedAsymm8>();
+ RefCreateElementwiseWorkloadTest<RefAdditionUint8Workload,
+ AdditionQueueDescriptor,
+ AdditionLayer,
+ armnn::DataType::QuantisedAsymm8>();
}
BOOST_AUTO_TEST_CASE(CreateSubtractionFloatWorkload)
{
- RefCreateArithmethicWorkloadTest<RefSubtractionFloat32Workload,
- SubtractionQueueDescriptor,
- SubtractionLayer,
- armnn::DataType::Float32>();
+ RefCreateElementwiseWorkloadTest<RefSubtractionFloat32Workload,
+ SubtractionQueueDescriptor,
+ SubtractionLayer,
+ armnn::DataType::Float32>();
}
BOOST_AUTO_TEST_CASE(CreateSubtractionUint8Workload)
{
- RefCreateArithmethicWorkloadTest<RefSubtractionUint8Workload,
- SubtractionQueueDescriptor,
- SubtractionLayer,
- armnn::DataType::QuantisedAsymm8>();
+ RefCreateElementwiseWorkloadTest<RefSubtractionUint8Workload,
+ SubtractionQueueDescriptor,
+ SubtractionLayer,
+ armnn::DataType::QuantisedAsymm8>();
}
BOOST_AUTO_TEST_CASE(CreateMultiplicationFloatWorkload)
{
- RefCreateArithmethicWorkloadTest<RefMultiplicationFloat32Workload,
- MultiplicationQueueDescriptor,
- MultiplicationLayer,
- armnn::DataType::Float32>();
+ RefCreateElementwiseWorkloadTest<RefMultiplicationFloat32Workload,
+ MultiplicationQueueDescriptor,
+ MultiplicationLayer,
+ armnn::DataType::Float32>();
}
BOOST_AUTO_TEST_CASE(CreateMultiplicationUint8Workload)
{
- RefCreateArithmethicWorkloadTest<RefMultiplicationUint8Workload,
- MultiplicationQueueDescriptor,
- MultiplicationLayer,
- armnn::DataType::QuantisedAsymm8>();
+ RefCreateElementwiseWorkloadTest<RefMultiplicationUint8Workload,
+ MultiplicationQueueDescriptor,
+ MultiplicationLayer,
+ armnn::DataType::QuantisedAsymm8>();
}
BOOST_AUTO_TEST_CASE(CreateDivisionFloatWorkload)
{
- RefCreateArithmethicWorkloadTest<RefDivisionFloat32Workload,
- DivisionQueueDescriptor,
- DivisionLayer,
- armnn::DataType::Float32>();
+ RefCreateElementwiseWorkloadTest<RefDivisionFloat32Workload,
+ DivisionQueueDescriptor,
+ DivisionLayer,
+ armnn::DataType::Float32>();
}
BOOST_AUTO_TEST_CASE(CreateDivisionUint8Workload)
{
- RefCreateArithmethicWorkloadTest<RefDivisionUint8Workload,
- DivisionQueueDescriptor,
- DivisionLayer,
- armnn::DataType::QuantisedAsymm8>();
+ RefCreateElementwiseWorkloadTest<RefDivisionUint8Workload,
+ DivisionQueueDescriptor,
+ DivisionLayer,
+ armnn::DataType::QuantisedAsymm8>();
}
template <typename BatchNormalizationWorkloadType, armnn::DataType DataType>
diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index 2d9ad92..86c5f90 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -6,8 +6,6 @@
list(APPEND armnnRefBackendWorkloads_sources
Activation.cpp
Activation.hpp
- ArithmeticFunction.cpp
- ArithmeticFunction.hpp
BatchNormImpl.hpp
BatchToSpaceNd.cpp
BatchToSpaceNd.hpp
@@ -15,6 +13,8 @@
Broadcast.hpp
ConvImpl.cpp
ConvImpl.hpp
+ ElementwiseFunction.cpp
+ ElementwiseFunction.hpp
FullyConnected.cpp
FullyConnected.hpp
Merger.hpp
@@ -26,8 +26,6 @@
RefActivationFloat32Workload.hpp
RefActivationUint8Workload.cpp
RefActivationUint8Workload.hpp
- RefArithmeticWorkload.cpp
- RefArithmeticWorkload.hpp
RefBaseConstantWorkload.cpp
RefBaseConstantWorkload.hpp
RefBatchNormalizationFloat32Workload.cpp
@@ -50,6 +48,8 @@
RefConvolution2dFloat32Workload.hpp
RefConvolution2dUint8Workload.cpp
RefConvolution2dUint8Workload.hpp
+ RefElementwiseWorkload.cpp
+ RefElementwiseWorkload.hpp
RefDepthwiseConvolution2dFloat32Workload.cpp
RefDepthwiseConvolution2dFloat32Workload.hpp
RefDepthwiseConvolution2dUint8Workload.cpp
diff --git a/src/backends/reference/workloads/ArithmeticFunction.cpp b/src/backends/reference/workloads/ElementwiseFunction.cpp
similarity index 64%
rename from src/backends/reference/workloads/ArithmeticFunction.cpp
rename to src/backends/reference/workloads/ElementwiseFunction.cpp
index fede138..bea3d2f 100644
--- a/src/backends/reference/workloads/ArithmeticFunction.cpp
+++ b/src/backends/reference/workloads/ElementwiseFunction.cpp
@@ -3,7 +3,7 @@
// SPDX-License-Identifier: MIT
//
-#include "ArithmeticFunction.hpp"
+#include "ElementwiseFunction.hpp"
#include "Broadcast.hpp"
#include <functional>
@@ -11,7 +11,7 @@
{
template <typename Functor>
-ArithmeticFunction<Functor>::ArithmeticFunction(const TensorShape& inShape0,
+ElementwiseFunction<Functor>::ElementwiseFunction(const TensorShape& inShape0,
const TensorShape& inShape1,
const TensorShape& outShape,
const float* inData0,
@@ -23,7 +23,7 @@
} //namespace armnn
-template struct armnn::ArithmeticFunction<std::plus<float>>;
-template struct armnn::ArithmeticFunction<std::minus<float>>;
-template struct armnn::ArithmeticFunction<std::multiplies<float>>;
-template struct armnn::ArithmeticFunction<std::divides<float>>;
+template struct armnn::ElementwiseFunction<std::plus<float>>;
+template struct armnn::ElementwiseFunction<std::minus<float>>;
+template struct armnn::ElementwiseFunction<std::multiplies<float>>;
+template struct armnn::ElementwiseFunction<std::divides<float>>;
diff --git a/src/backends/reference/workloads/ArithmeticFunction.hpp b/src/backends/reference/workloads/ElementwiseFunction.hpp
similarity index 84%
rename from src/backends/reference/workloads/ArithmeticFunction.hpp
rename to src/backends/reference/workloads/ElementwiseFunction.hpp
index eafb644..5011616 100644
--- a/src/backends/reference/workloads/ArithmeticFunction.hpp
+++ b/src/backends/reference/workloads/ElementwiseFunction.hpp
@@ -11,9 +11,9 @@
{
template <typename Functor>
-struct ArithmeticFunction
+struct ElementwiseFunction
{
- ArithmeticFunction(const TensorShape& inShape0,
+ ElementwiseFunction(const TensorShape& inShape0,
const TensorShape& inShape1,
const TensorShape& outShape,
const float* inData0,
diff --git a/src/backends/reference/workloads/RefArithmeticWorkload.cpp b/src/backends/reference/workloads/RefArithmeticWorkload.cpp
deleted file mode 100644
index 6c39fa1..0000000
--- a/src/backends/reference/workloads/RefArithmeticWorkload.cpp
+++ /dev/null
@@ -1,69 +0,0 @@
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "RefArithmeticWorkload.hpp"
-#include "ArithmeticFunction.hpp"
-#include "RefWorkloadUtils.hpp"
-#include "Profiling.hpp"
-#include <vector>
-
-namespace armnn
-{
-
-template <typename ParentDescriptor, typename Functor>
-void BaseFloat32ArithmeticWorkload<ParentDescriptor, Functor>::ExecuteImpl(const char * debugString) const
-{
- ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, debugString);
-
- auto data = Float32Workload<ParentDescriptor>::GetData();
- const TensorShape& inShape0 = GetTensorInfo(data.m_Inputs[0]).GetShape();
- const TensorShape& inShape1 = GetTensorInfo(data.m_Inputs[1]).GetShape();
- const TensorShape& outShape = GetTensorInfo(data.m_Outputs[0]).GetShape();
-
- const float* inData0 = GetInputTensorDataFloat(0, data);
- const float* inData1 = GetInputTensorDataFloat(1, data);
- float* outData = GetOutputTensorDataFloat(0, data);
-
- ArithmeticFunction<Functor>(inShape0, inShape1, outShape, inData0, inData1, outData);
-}
-
-template <typename ParentDescriptor, typename Functor>
-void BaseUint8ArithmeticWorkload<ParentDescriptor, Functor>::ExecuteImpl(const char * debugString) const
-{
- ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, debugString);
-
- auto data = Uint8Workload<ParentDescriptor>::GetData();
- const TensorInfo& inputInfo0 = GetTensorInfo(data.m_Inputs[0]);
- const TensorInfo& inputInfo1 = GetTensorInfo(data.m_Inputs[1]);
- const TensorInfo& outputInfo = GetTensorInfo(data.m_Outputs[0]);
-
- auto dequant0 = Dequantize(GetInputTensorDataU8(0, data), inputInfo0);
- auto dequant1 = Dequantize(GetInputTensorDataU8(1, data), inputInfo1);
-
- std::vector<float> results(outputInfo.GetNumElements());
-
- ArithmeticFunction<Functor>(inputInfo0.GetShape(),
- inputInfo1.GetShape(),
- outputInfo.GetShape(),
- dequant0.data(),
- dequant1.data(),
- results.data());
-
- Quantize(GetOutputTensorDataU8(0, data), results.data(), outputInfo);
-}
-
-}
-
-template class armnn::BaseFloat32ArithmeticWorkload<armnn::AdditionQueueDescriptor, std::plus<float>>;
-template class armnn::BaseUint8ArithmeticWorkload<armnn::AdditionQueueDescriptor, std::plus<float>>;
-
-template class armnn::BaseFloat32ArithmeticWorkload<armnn::SubtractionQueueDescriptor, std::minus<float>>;
-template class armnn::BaseUint8ArithmeticWorkload<armnn::SubtractionQueueDescriptor, std::minus<float>>;
-
-template class armnn::BaseFloat32ArithmeticWorkload<armnn::MultiplicationQueueDescriptor, std::multiplies<float>>;
-template class armnn::BaseUint8ArithmeticWorkload<armnn::MultiplicationQueueDescriptor, std::multiplies<float>>;
-
-template class armnn::BaseFloat32ArithmeticWorkload<armnn::DivisionQueueDescriptor, std::divides<float>>;
-template class armnn::BaseUint8ArithmeticWorkload<armnn::DivisionQueueDescriptor, std::divides<float>>;
diff --git a/src/backends/reference/workloads/RefElementwiseWorkload.cpp b/src/backends/reference/workloads/RefElementwiseWorkload.cpp
new file mode 100644
index 0000000..8e312a7
--- /dev/null
+++ b/src/backends/reference/workloads/RefElementwiseWorkload.cpp
@@ -0,0 +1,69 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefElementwiseWorkload.hpp"
+#include "ElementwiseFunction.hpp"
+#include "RefWorkloadUtils.hpp"
+#include "Profiling.hpp"
+#include <vector>
+
+namespace armnn
+{
+
+template <typename ParentDescriptor, typename Functor>
+void BaseFloat32ElementwiseWorkload<ParentDescriptor, Functor>::ExecuteImpl(const char * debugString) const
+{
+ ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, debugString);
+
+ auto data = Float32Workload<ParentDescriptor>::GetData();
+ const TensorShape& inShape0 = GetTensorInfo(data.m_Inputs[0]).GetShape();
+ const TensorShape& inShape1 = GetTensorInfo(data.m_Inputs[1]).GetShape();
+ const TensorShape& outShape = GetTensorInfo(data.m_Outputs[0]).GetShape();
+
+ const float* inData0 = GetInputTensorDataFloat(0, data);
+ const float* inData1 = GetInputTensorDataFloat(1, data);
+ float* outData = GetOutputTensorDataFloat(0, data);
+
+ ElementwiseFunction<Functor>(inShape0, inShape1, outShape, inData0, inData1, outData);
+}
+
+template <typename ParentDescriptor, typename Functor>
+void BaseUint8ElementwiseWorkload<ParentDescriptor, Functor>::ExecuteImpl(const char * debugString) const
+{
+ ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, debugString);
+
+ auto data = Uint8Workload<ParentDescriptor>::GetData();
+ const TensorInfo& inputInfo0 = GetTensorInfo(data.m_Inputs[0]);
+ const TensorInfo& inputInfo1 = GetTensorInfo(data.m_Inputs[1]);
+ const TensorInfo& outputInfo = GetTensorInfo(data.m_Outputs[0]);
+
+ auto dequant0 = Dequantize(GetInputTensorDataU8(0, data), inputInfo0);
+ auto dequant1 = Dequantize(GetInputTensorDataU8(1, data), inputInfo1);
+
+ std::vector<float> results(outputInfo.GetNumElements());
+
+ ElementwiseFunction<Functor>(inputInfo0.GetShape(),
+ inputInfo1.GetShape(),
+ outputInfo.GetShape(),
+ dequant0.data(),
+ dequant1.data(),
+ results.data());
+
+ Quantize(GetOutputTensorDataU8(0, data), results.data(), outputInfo);
+}
+
+}
+
+template class armnn::BaseFloat32ElementwiseWorkload<armnn::AdditionQueueDescriptor, std::plus<float>>;
+template class armnn::BaseUint8ElementwiseWorkload<armnn::AdditionQueueDescriptor, std::plus<float>>;
+
+template class armnn::BaseFloat32ElementwiseWorkload<armnn::SubtractionQueueDescriptor, std::minus<float>>;
+template class armnn::BaseUint8ElementwiseWorkload<armnn::SubtractionQueueDescriptor, std::minus<float>>;
+
+template class armnn::BaseFloat32ElementwiseWorkload<armnn::MultiplicationQueueDescriptor, std::multiplies<float>>;
+template class armnn::BaseUint8ElementwiseWorkload<armnn::MultiplicationQueueDescriptor, std::multiplies<float>>;
+
+template class armnn::BaseFloat32ElementwiseWorkload<armnn::DivisionQueueDescriptor, std::divides<float>>;
+template class armnn::BaseUint8ElementwiseWorkload<armnn::DivisionQueueDescriptor, std::divides<float>>;
diff --git a/src/backends/reference/workloads/RefArithmeticWorkload.hpp b/src/backends/reference/workloads/RefElementwiseWorkload.hpp
similarity index 71%
rename from src/backends/reference/workloads/RefArithmeticWorkload.hpp
rename to src/backends/reference/workloads/RefElementwiseWorkload.hpp
index 7560617..156613a 100644
--- a/src/backends/reference/workloads/RefArithmeticWorkload.hpp
+++ b/src/backends/reference/workloads/RefElementwiseWorkload.hpp
@@ -17,13 +17,13 @@
typename armnn::DataType DataType,
typename ParentDescriptor,
typename armnn::StringMapping::Id DebugString>
-class RefArithmeticWorkload
+class RefElementwiseWorkload
{
// Needs specialization. The default is empty on purpose.
};
template <typename ParentDescriptor, typename Functor>
-class BaseFloat32ArithmeticWorkload : public Float32Workload<ParentDescriptor>
+class BaseFloat32ElementwiseWorkload : public Float32Workload<ParentDescriptor>
{
public:
using Float32Workload<ParentDescriptor>::Float32Workload;
@@ -33,21 +33,21 @@
template <typename Functor,
typename ParentDescriptor,
typename armnn::StringMapping::Id DebugString>
-class RefArithmeticWorkload<Functor, armnn::DataType::Float32, ParentDescriptor, DebugString>
- : public BaseFloat32ArithmeticWorkload<ParentDescriptor, Functor>
+class RefElementwiseWorkload<Functor, armnn::DataType::Float32, ParentDescriptor, DebugString>
+ : public BaseFloat32ElementwiseWorkload<ParentDescriptor, Functor>
{
public:
- using BaseFloat32ArithmeticWorkload<ParentDescriptor, Functor>::BaseFloat32ArithmeticWorkload;
+ using BaseFloat32ElementwiseWorkload<ParentDescriptor, Functor>::BaseFloat32ElementwiseWorkload;
virtual void Execute() const override
{
- using Parent = BaseFloat32ArithmeticWorkload<ParentDescriptor, Functor>;
+ using Parent = BaseFloat32ElementwiseWorkload<ParentDescriptor, Functor>;
Parent::ExecuteImpl(StringMapping::Instance().Get(DebugString));
}
};
template <typename ParentDescriptor, typename Functor>
-class BaseUint8ArithmeticWorkload : public Uint8Workload<ParentDescriptor>
+class BaseUint8ElementwiseWorkload : public Uint8Workload<ParentDescriptor>
{
public:
using Uint8Workload<ParentDescriptor>::Uint8Workload;
@@ -57,64 +57,64 @@
template <typename Functor,
typename ParentDescriptor,
typename armnn::StringMapping::Id DebugString>
-class RefArithmeticWorkload<Functor, armnn::DataType::QuantisedAsymm8, ParentDescriptor, DebugString>
- : public BaseUint8ArithmeticWorkload<ParentDescriptor, Functor>
+class RefElementwiseWorkload<Functor, armnn::DataType::QuantisedAsymm8, ParentDescriptor, DebugString>
+ : public BaseUint8ElementwiseWorkload<ParentDescriptor, Functor>
{
public:
- using BaseUint8ArithmeticWorkload<ParentDescriptor, Functor>::BaseUint8ArithmeticWorkload;
+ using BaseUint8ElementwiseWorkload<ParentDescriptor, Functor>::BaseUint8ElementwiseWorkload;
virtual void Execute() const override
{
- using Parent = BaseUint8ArithmeticWorkload<ParentDescriptor, Functor>;
+ using Parent = BaseUint8ElementwiseWorkload<ParentDescriptor, Functor>;
Parent::ExecuteImpl(StringMapping::Instance().Get(DebugString));
}
};
using RefAdditionFloat32Workload =
- RefArithmeticWorkload<std::plus<float>,
+ RefElementwiseWorkload<std::plus<float>,
DataType::Float32,
AdditionQueueDescriptor,
StringMapping::RefAdditionWorkload_Execute>;
using RefAdditionUint8Workload =
- RefArithmeticWorkload<std::plus<float>,
+ RefElementwiseWorkload<std::plus<float>,
DataType::QuantisedAsymm8,
AdditionQueueDescriptor,
StringMapping::RefAdditionWorkload_Execute>;
using RefSubtractionFloat32Workload =
- RefArithmeticWorkload<std::minus<float>,
+ RefElementwiseWorkload<std::minus<float>,
DataType::Float32,
SubtractionQueueDescriptor,
StringMapping::RefSubtractionWorkload_Execute>;
using RefSubtractionUint8Workload =
- RefArithmeticWorkload<std::minus<float>,
+ RefElementwiseWorkload<std::minus<float>,
DataType::QuantisedAsymm8,
SubtractionQueueDescriptor,
StringMapping::RefSubtractionWorkload_Execute>;
using RefMultiplicationFloat32Workload =
- RefArithmeticWorkload<std::multiplies<float>,
+ RefElementwiseWorkload<std::multiplies<float>,
DataType::Float32,
MultiplicationQueueDescriptor,
StringMapping::RefMultiplicationWorkload_Execute>;
using RefMultiplicationUint8Workload =
- RefArithmeticWorkload<std::multiplies<float>,
+ RefElementwiseWorkload<std::multiplies<float>,
DataType::QuantisedAsymm8,
MultiplicationQueueDescriptor,
StringMapping::RefMultiplicationWorkload_Execute>;
using RefDivisionFloat32Workload =
- RefArithmeticWorkload<std::divides<float>,
+ RefElementwiseWorkload<std::divides<float>,
DataType::Float32,
DivisionQueueDescriptor,
StringMapping::RefDivisionWorkload_Execute>;
using RefDivisionUint8Workload =
- RefArithmeticWorkload<std::divides<float>,
+ RefElementwiseWorkload<std::divides<float>,
DataType::QuantisedAsymm8,
DivisionQueueDescriptor,
StringMapping::RefDivisionWorkload_Execute>;
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index 20e9a9f..86d8624 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -6,8 +6,8 @@
#pragma once
#include "RefConstantUint8Workload.hpp"
-#include "ArithmeticFunction.hpp"
-#include "RefArithmeticWorkload.hpp"
+#include "ElementwiseFunction.hpp"
+#include "RefElementwiseWorkload.hpp"
#include "ConvImpl.hpp"
#include "RefBaseConstantWorkload.hpp"
#include "RefConvolution2dUint8Workload.hpp"