IVGCVSW-3837 Add support for per-axis quantization to reference Convolution2d workload
Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I0ac08ba4864d48e6f64c4ac645dad8ea850be112
diff --git a/src/backends/backendsCommon/test/WorkloadDataValidation.cpp b/src/backends/backendsCommon/test/WorkloadDataValidation.cpp
index 9773914..70d00b3 100644
--- a/src/backends/backendsCommon/test/WorkloadDataValidation.cpp
+++ b/src/backends/backendsCommon/test/WorkloadDataValidation.cpp
@@ -605,15 +605,16 @@
const TensorShape weightShape{ cOutput, cInput, hInput, wInput };
const TensorShape biasShape { cOutput };
- constexpr DataType dataType = DataType::QuantisedAsymm8;
- constexpr DataType biasType = DataType::Signed32;
+ constexpr DataType inputType = DataType::QuantisedAsymm8;
+ constexpr DataType weightType = DataType::QuantizedSymm8PerAxis;
+ constexpr DataType biasType = DataType::Signed32;
constexpr float perTensorScale = 1.5f;
- const TensorInfo inputInfo (inputShape, dataType, perTensorScale);
- const TensorInfo outputInfo(outputShape, dataType, perTensorScale);
+ const TensorInfo inputInfo (inputShape, inputType, perTensorScale);
+ const TensorInfo outputInfo(outputShape, inputType, perTensorScale);
const std::vector<float> weightPerAxisScales = { 2.50f, 3.50f };
- const TensorInfo weightInfo(weightShape, dataType, weightPerAxisScales, 0);
+ const TensorInfo weightInfo(weightShape, weightType, weightPerAxisScales, 0);
Convolution2dQueueDescriptor queueDescriptor;
queueDescriptor.m_Parameters.m_BiasEnabled = true;
diff --git a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp
index 198904e..5fac09f 100644
--- a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.cpp
@@ -14,6 +14,7 @@
#include <backendsCommon/CpuTensorHandle.hpp>
+#include <backendsCommon/test/DataLayoutUtils.hpp>
#include <backendsCommon/test/TensorCopyUtils.hpp>
#include <backendsCommon/test/WorkloadTestUtils.hpp>
@@ -3035,6 +3036,98 @@
workloadFactory, memoryManager, 0.1f, 128, biasEnabled);
}
+LayerTestResult<uint8_t, 4> Convolution2dPerAxisQuantTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+ const armnn::DataLayout layout)
+{
+ using namespace armnn;
+
+ const DataType inputType = DataType::QuantisedAsymm8;
+ const DataType kernelType = DataType::QuantizedSymm8PerAxis;
+ const DataType biasType = DataType::Signed32;
+
+ TensorInfo inputInfo ({ 1, 3, 1, 2 }, inputType, 0.5f, 128);
+ TensorInfo outputInfo({ 1, 3, 1, 3 }, inputType, 1.0f, 128);
+
+ const std::vector<float> quantScales{ 0.5f, 0.75f, 1.0f };
+ constexpr unsigned int quantDimension = 0;
+
+ TensorInfo kernelInfo({ 3, 1, 1, 2 }, kernelType, quantScales, quantDimension);
+
+ const std::vector<float> biasQuantScales{ 0.25f, 0.375f, 0.5f };
+ TensorInfo biasInfo({ 3 }, biasType, biasQuantScales, quantDimension);
+
+ std::vector<uint8_t> inputData =
+ {
+ 138, 108, 138, 108, 138, 108
+ };
+
+ std::vector<int8_t> kernelData =
+ {
+ 1, 2, 1, 2, 1, 2
+ };
+
+ std::vector<int32_t> biasData =
+ {
+ 4, 4, 4
+ };
+
+ std::vector<uint8_t> expectedOutputData =
+ {
+ 121, 118, 115, 121, 118, 115, 121, 118, 115
+ };
+
+ if (layout == DataLayout::NCHW)
+ {
+ PermuteTensorNhwcToNchw(inputInfo, inputData);
+ PermuteTensorNhwcToNchw(kernelInfo, kernelData);
+ PermuteTensorNhwcToNchw(outputInfo, expectedOutputData);
+ }
+
+ Convolution2dDescriptor descriptor;
+ descriptor.m_StrideX = 1;
+ descriptor.m_StrideY = 1;
+ descriptor.m_PadLeft = 0;
+ descriptor.m_PadRight = 0;
+ descriptor.m_PadTop = 0;
+ descriptor.m_PadBottom = 0;
+ descriptor.m_BiasEnabled = true;
+ descriptor.m_DataLayout = layout;
+
+ std::unique_ptr<ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputInfo);
+ std::unique_ptr<ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputInfo);
+
+ WorkloadInfo workloadInfo;
+ ScopedCpuTensorHandle weightTensor(kernelInfo);
+ ScopedCpuTensorHandle biasTensor(biasInfo);
+
+ AllocateAndCopyDataToITensorHandle(&weightTensor, kernelData.data());
+ AllocateAndCopyDataToITensorHandle(&biasTensor, biasData.data());
+
+ Convolution2dQueueDescriptor queueDescriptor;
+ queueDescriptor.m_Parameters = descriptor;
+ queueDescriptor.m_Weight = &weightTensor;
+ queueDescriptor.m_Bias = &biasTensor;
+
+ AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get());
+ AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get());
+
+ std::unique_ptr<IWorkload> workload = workloadFactory.CreateConvolution2d(queueDescriptor, workloadInfo);
+ inputHandle->Allocate();
+ outputHandle->Allocate();
+
+ CopyDataToITensorHandle(inputHandle.get(), inputData.data());
+
+ ExecuteWorkload(*workload, memoryManager);
+
+ LayerTestResult<uint8_t, 4> ret(outputInfo);
+ CopyDataFromITensorHandle(ret.output.origin(), outputHandle.get());
+ ret.outputExpected = MakeTensor<uint8_t, 4>(outputInfo, expectedOutputData);
+
+ return ret;
+}
+
LayerTestResult<float,4> CompareConvolution2dTest(
armnn::IWorkloadFactory& workloadFactory,
const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
diff --git a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.hpp
index f5ff586..3aac975 100644
--- a/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.hpp
+++ b/src/backends/backendsCommon/test/layerTests/Conv2dTestImpl.hpp
@@ -111,6 +111,11 @@
const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
armnn::IWorkloadFactory& refWorkloadFactory);
+LayerTestResult<uint8_t, 4> Convolution2dPerAxisQuantTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+ const armnn::DataLayout layout);
+
//
// DepthwiseConvolution2d
//