diff options
Diffstat (limited to 'src/backends/reference')
4 files changed, 186 insertions, 11 deletions
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp index 9c1ce1e013..3bf83bd9be 100644 --- a/src/backends/reference/RefWorkloadFactory.cpp +++ b/src/backends/reference/RefWorkloadFactory.cpp @@ -154,7 +154,16 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateDepthwiseConvolution2d( std::unique_ptr<IWorkload> RefWorkloadFactory::CreateDetectionPostProcess( const armnn::DetectionPostProcessQueueDescriptor& descriptor, const armnn::WorkloadInfo& info) const { - return MakeWorkload<RefDetectionPostProcessFloat32Workload, RefDetectionPostProcessUint8Workload>(descriptor, info); + const DataType dataType = info.m_InputTensorInfos[0].GetDataType(); + switch (dataType) + { + case DataType::Float32: + return std::make_unique<RefDetectionPostProcessFloat32Workload>(descriptor, info); + case DataType::QuantisedAsymm8: + return std::make_unique<RefDetectionPostProcessUint8Workload>(descriptor, info); + default: + return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info); + } } std::unique_ptr<armnn::IWorkload> RefWorkloadFactory::CreateNormalization( diff --git a/src/backends/reference/test/RefDetectionPostProcessTests.cpp b/src/backends/reference/test/RefDetectionPostProcessTests.cpp index 39403f0284..a9faff70b1 100644 --- a/src/backends/reference/test/RefDetectionPostProcessTests.cpp +++ b/src/backends/reference/test/RefDetectionPostProcessTests.cpp @@ -74,8 +74,8 @@ void DetectionPostProcessTestImpl(bool useRegularNms, const std::vector<float>& const std::vector<float>& expectedNumDetections) { armnn::TensorInfo boxEncodingsInfo({ 1, 6, 4 }, armnn::DataType::Float32); - armnn::TensorInfo scoresInfo({ 1, 6, 4 }, armnn::DataType::Float32); - armnn::TensorInfo anchorsInfo({ 1, 6, 4 }, armnn::DataType::Float32); + armnn::TensorInfo scoresInfo({ 1, 6, 3 }, armnn::DataType::Float32); + armnn::TensorInfo anchorsInfo({ 6, 4 }, armnn::DataType::Float32); armnn::TensorInfo detectionBoxesInfo({ 1, 3, 4 }, armnn::DataType::Float32); armnn::TensorInfo detectionScoresInfo({ 1, 3 }, armnn::DataType::Float32); diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp index 802167a3a0..c89e586f03 100644 --- a/src/backends/reference/test/RefEndToEndTests.cpp +++ b/src/backends/reference/test/RefEndToEndTests.cpp @@ -4,6 +4,8 @@ // #include <backendsCommon/test/EndToEndTestImpl.hpp> + +#include <backendsCommon/test/DetectionPostProcessTestImpl.hpp> #include <backendsCommon/test/GatherEndToEndTestImpl.hpp> #include <backendsCommon/test/MergerTestImpl.hpp> #include <backendsCommon/test/ArithmeticTestImpl.hpp> @@ -453,4 +455,168 @@ BOOST_AUTO_TEST_CASE(RefGatherMultiDimUint8Test) GatherMultiDimEndToEnd<armnn::DataType::QuantisedAsymm8>(defaultBackends); } +BOOST_AUTO_TEST_CASE(RefDetectionPostProcessRegularNmsTest) +{ + std::vector<float> boxEncodings({ + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, -1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f + }); + std::vector<float> scores({ + 0.0f, 0.9f, 0.8f, + 0.0f, 0.75f, 0.72f, + 0.0f, 0.6f, 0.5f, + 0.0f, 0.93f, 0.95f, + 0.0f, 0.5f, 0.4f, + 0.0f, 0.3f, 0.2f + }); + std::vector<float> anchors({ + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 100.5f, 1.0f, 1.0f + }); + DetectionPostProcessRegularNmsEndToEnd<armnn::DataType::Float32>(defaultBackends, boxEncodings, scores, anchors); +} + +inline void QuantizeData(uint8_t* quant, const float* dequant, const TensorInfo& info) +{ + for (size_t i = 0; i < info.GetNumElements(); i++) + { + quant[i] = armnn::Quantize<uint8_t>(dequant[i], info.GetQuantizationScale(), info.GetQuantizationOffset()); + } +} + +BOOST_AUTO_TEST_CASE(RefDetectionPostProcessRegularNmsUint8Test) +{ + armnn::TensorInfo boxEncodingsInfo({ 1, 6, 4 }, armnn::DataType::Float32); + armnn::TensorInfo scoresInfo({ 1, 6, 3 }, armnn::DataType::Float32); + armnn::TensorInfo anchorsInfo({ 6, 4 }, armnn::DataType::Float32); + + boxEncodingsInfo.SetQuantizationScale(1.0f); + boxEncodingsInfo.SetQuantizationOffset(1); + scoresInfo.SetQuantizationScale(0.01f); + scoresInfo.SetQuantizationOffset(0); + anchorsInfo.SetQuantizationScale(0.5f); + anchorsInfo.SetQuantizationOffset(0); + + std::vector<float> boxEncodings({ + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, -1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f + }); + std::vector<float> scores({ + 0.0f, 0.9f, 0.8f, + 0.0f, 0.75f, 0.72f, + 0.0f, 0.6f, 0.5f, + 0.0f, 0.93f, 0.95f, + 0.0f, 0.5f, 0.4f, + 0.0f, 0.3f, 0.2f + }); + std::vector<float> anchors({ + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 100.5f, 1.0f, 1.0f + }); + + std::vector<uint8_t> qBoxEncodings(boxEncodings.size(), 0); + std::vector<uint8_t> qScores(scores.size(), 0); + std::vector<uint8_t> qAnchors(anchors.size(), 0); + QuantizeData(qBoxEncodings.data(), boxEncodings.data(), boxEncodingsInfo); + QuantizeData(qScores.data(), scores.data(), scoresInfo); + QuantizeData(qAnchors.data(), anchors.data(), anchorsInfo); + DetectionPostProcessRegularNmsEndToEnd<armnn::DataType::QuantisedAsymm8>(defaultBackends, qBoxEncodings, + qScores, qAnchors, + 1.0f, 1, 0.01f, 0, 0.5f, 0); +} + +BOOST_AUTO_TEST_CASE(RefDetectionPostProcessFastNmsTest) +{ + std::vector<float> boxEncodings({ + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, -1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f + }); + std::vector<float> scores({ + 0.0f, 0.9f, 0.8f, + 0.0f, 0.75f, 0.72f, + 0.0f, 0.6f, 0.5f, + 0.0f, 0.93f, 0.95f, + 0.0f, 0.5f, 0.4f, + 0.0f, 0.3f, 0.2f + }); + std::vector<float> anchors({ + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 100.5f, 1.0f, 1.0f + }); + DetectionPostProcessFastNmsEndToEnd<armnn::DataType::Float32>(defaultBackends, boxEncodings, scores, anchors); +} + +BOOST_AUTO_TEST_CASE(RefDetectionPostProcessFastNmsUint8Test) +{ + armnn::TensorInfo boxEncodingsInfo({ 1, 6, 4 }, armnn::DataType::Float32); + armnn::TensorInfo scoresInfo({ 1, 6, 3 }, armnn::DataType::Float32); + armnn::TensorInfo anchorsInfo({ 6, 4 }, armnn::DataType::Float32); + + boxEncodingsInfo.SetQuantizationScale(1.0f); + boxEncodingsInfo.SetQuantizationOffset(1); + scoresInfo.SetQuantizationScale(0.01f); + scoresInfo.SetQuantizationOffset(0); + anchorsInfo.SetQuantizationScale(0.5f); + anchorsInfo.SetQuantizationOffset(0); + + std::vector<float> boxEncodings({ + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, -1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f, + 0.0f, 1.0f, 0.0f, 0.0f, + 0.0f, 0.0f, 0.0f, 0.0f + }); + std::vector<float> scores({ + 0.0f, 0.9f, 0.8f, + 0.0f, 0.75f, 0.72f, + 0.0f, 0.6f, 0.5f, + 0.0f, 0.93f, 0.95f, + 0.0f, 0.5f, 0.4f, + 0.0f, 0.3f, 0.2f + }); + std::vector<float> anchors({ + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 0.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 10.5f, 1.0f, 1.0f, + 0.5f, 100.5f, 1.0f, 1.0f + }); + + std::vector<uint8_t> qBoxEncodings(boxEncodings.size(), 0); + std::vector<uint8_t> qScores(scores.size(), 0); + std::vector<uint8_t> qAnchors(anchors.size(), 0); + QuantizeData(qBoxEncodings.data(), boxEncodings.data(), boxEncodingsInfo); + QuantizeData(qScores.data(), scores.data(), scoresInfo); + QuantizeData(qAnchors.data(), anchors.data(), anchorsInfo); + DetectionPostProcessFastNmsEndToEnd<armnn::DataType::QuantisedAsymm8>(defaultBackends, qBoxEncodings, + qScores, qAnchors, + 1.0f, 1, 0.01f, 0, 0.5f, 0); +} + BOOST_AUTO_TEST_SUITE_END() diff --git a/src/backends/reference/workloads/DetectionPostProcess.cpp b/src/backends/reference/workloads/DetectionPostProcess.cpp index 958de8294b..2eb35f5ffa 100644 --- a/src/backends/reference/workloads/DetectionPostProcess.cpp +++ b/src/backends/reference/workloads/DetectionPostProcess.cpp @@ -105,15 +105,15 @@ void AllocateOutputData(unsigned int numOutput, unsigned int numSelected, const for (unsigned int i = 0; i < numOutput; ++i) { unsigned int boxIndex = i * 4; - unsigned int boxConorIndex = selectedBoxes[outputIndices[i]] * 4; if (i < numSelected) { + unsigned int boxCornorIndex = selectedBoxes[outputIndices[i]] * 4; detectionScores[i] = selectedScores[outputIndices[i]]; detectionClasses[i] = boost::numeric_cast<float>(selectedClasses[outputIndices[i]]); - detectionBoxes[boxIndex] = boxCorners[boxConorIndex]; - detectionBoxes[boxIndex + 1] = boxCorners[boxConorIndex + 1]; - detectionBoxes[boxIndex + 2] = boxCorners[boxConorIndex + 2]; - detectionBoxes[boxIndex + 3] = boxCorners[boxConorIndex + 3]; + detectionBoxes[boxIndex] = boxCorners[boxCornorIndex]; + detectionBoxes[boxIndex + 1] = boxCorners[boxCornorIndex + 1]; + detectionBoxes[boxIndex + 2] = boxCorners[boxCornorIndex + 2]; + detectionBoxes[boxIndex + 3] = boxCorners[boxCornorIndex + 3]; } else { @@ -125,7 +125,7 @@ void AllocateOutputData(unsigned int numOutput, unsigned int numSelected, const detectionBoxes[boxIndex + 3] = 0.0f; } } - numDetections[0] = boost::numeric_cast<float>(numOutput); + numDetections[0] = boost::numeric_cast<float>(numSelected); } } // anonymous namespace @@ -216,7 +216,7 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo, std::vector<unsigned int> outputIndices = GenerateRangeK(numSelected); TopKSort(numOutput, outputIndices.data(), selectedScoresAfterNms.data(), numSelected); - AllocateOutputData(numOutput, numSelected, boxCorners, outputIndices, + AllocateOutputData(detectionBoxesInfo.GetShape()[1], numOutput, boxCorners, outputIndices, selectedBoxesAfterNms, selectedClasses, selectedScoresAfterNms, detectionBoxes, detectionScores, detectionClasses, numDetections); } @@ -255,7 +255,7 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo, unsigned int numSelected = boost::numeric_cast<unsigned int>(selectedIndices.size()); unsigned int numOutput = std::min(desc.m_MaxDetections, numSelected); - AllocateOutputData(numOutput, numSelected, boxCorners, selectedIndices, + AllocateOutputData(detectionBoxesInfo.GetShape()[1], numOutput, boxCorners, selectedIndices, boxIndices, maxScoreClasses, maxScores, detectionBoxes, detectionScores, detectionClasses, numDetections); } |