IVGCVSW-7752 DTS: Fix QuantizePerChannel tests
* Added validation for scale on all Quantized types
* Added Encoder for Per Axis UINT16 Symmetrical Quantized type
* Added error for Per Axis Asymmetrical Quantized type not supported
Signed-off-by: John Mcloughlin <john.mcloughlin@arm.com>
Change-Id: I433519ccacd71219a92bde2b81955d6abf9219c5
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index d2b14cd..f18c6bf 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -799,31 +799,46 @@
for (unsigned int i = 0; i < numOutputs; i++) {
OutputSlot& outputSlot = layer->GetOutputSlot(i);
TensorInfo info = outputSlot.GetTensorInfo();
- if (DataType::QAsymmU8 == info.GetDataType())
- {
- if (0.f == info.GetQuantizationScale())
- {
- noErrors = false;
- std::stringstream ss;
- ss << "output " << i << " of layer " << GetLayerTypeAsCString(layer->GetType())
- << " (" << layer->GetNameStr() << ") is of type"
- << " Quantized 8 bit but its scale parameter has not been set";
- ReportError(ss.str(), errMessages);
- }
- // Softmax under QuantisedAsymm8 must always be scale (1.0f/256.0f) and offset 0
- if ((info.GetQuantizationScale() != (1.0f / 256.0f) ||
- info.GetQuantizationOffset() != 0) &&
- layer->GetType() == armnn::LayerType::Softmax)
- {
- std::stringstream ss;
- ss << "Quantization parameters for Softmax layer (Scale: " <<
- info.GetQuantizationScale() << " and Offset: " << info.GetQuantizationOffset() <<
- ") are incorrect and have been updated to Scale: 0.00390625 and Offset: 0";
- ARMNN_LOG(warning) << ss.str();
- info.SetQuantizationScale((1.0f /256.0f));
- info.SetQuantizationOffset(0);
- outputSlot.SetTensorInfo(info);
- }
+ auto quantizationDataType = info.GetDataType();
+ auto quantizationScales = info.GetQuantizationScales();
+ // For any Quantized Tensor ensure scale(s) are set
+ switch(quantizationDataType) {
+ case DataType::QAsymmU8:
+ case DataType::QSymmS16:
+ case DataType::QSymmS8:
+ case DataType::QAsymmS8:
+ if ((quantizationDataType == DataType::QAsymmU8 || quantizationDataType == DataType::QAsymmS8)
+ && info.HasPerAxisQuantization()) {
+ throw InvalidArgumentException("Per Axis Quantization is not supported in "
+ "Asymmetric Quantization Datatype.");
+ }
+ if ((!info.HasPerAxisQuantization() && info.GetQuantizationScale() == 0.f)
+ || (info.HasPerAxisQuantization() && (quantizationScales.end() !=
+ std::find(quantizationScales.begin(), quantizationScales.end(), 0.f)))) {
+ noErrors = false;
+ std::stringstream ss;
+ ss << "output " << i << " of layer " << GetLayerTypeAsCString(layer->GetType())
+ << " (" << layer->GetNameStr() << ") is of type"
+ << " Quantized value but the scale parameter has not been set";
+ ReportError(ss.str(), errMessages);
+ }
+ // Softmax under QuantisedAsymm8 must always be scale (1.0f/256.0f) and offset 0
+ if (!info.HasPerAxisQuantization() && quantizationDataType == DataType::QAsymmU8 &&
+ (info.GetQuantizationScale() != (1.0f / 256.0f) ||
+ info.GetQuantizationOffset() != 0) &&
+ layer->GetType() == armnn::LayerType::Softmax) {
+ std::stringstream ss;
+ ss << "Quantization parameters for Softmax layer (Scale: " <<
+ info.GetQuantizationScale() << " and Offset: " << info.GetQuantizationOffset() <<
+ ") are incorrect and have been updated to Scale: 0.00390625 and Offset: 0";
+ ARMNN_LOG(warning) << ss.str();
+ info.SetQuantizationScale((1.0f / 256.0f));
+ info.SetQuantizationOffset(0);
+ outputSlot.SetTensorInfo(info);
+ }
+ break;
+ default:
+ break;
}
}
return noErrors;
diff --git a/src/backends/reference/workloads/BaseIterator.hpp b/src/backends/reference/workloads/BaseIterator.hpp
index 2d27951..1665c1f 100644
--- a/src/backends/reference/workloads/BaseIterator.hpp
+++ b/src/backends/reference/workloads/BaseIterator.hpp
@@ -1,5 +1,5 @@
//
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
@@ -896,4 +896,31 @@
std::vector<float> m_Scales;
};
+class QSymm16PerAxisEncoder : public PerAxisIterator<int16_t, Encoder<float>>
+{
+public:
+ QSymm16PerAxisEncoder(int16_t* data, const std::vector<float>& scale,
+ unsigned int axisFactor, unsigned int axisDimensionality)
+ : PerAxisIterator(data, axisFactor, axisDimensionality), m_Scale(scale) {}
+
+ void Set(float right)
+ {
+ *m_Iterator = armnn::Quantize<int16_t>(right, m_Scale[m_AxisIndex], 0);
+ }
+
+ float Get() const
+ {
+ return armnn::Dequantize(*m_Iterator, m_Scale[m_AxisIndex], 0);
+ }
+
+ // Get scale of the current value
+ float GetScale() const
+ {
+ return m_Scale[m_AxisIndex];
+ }
+
+private:
+ std::vector<float> m_Scale;
+};
+
} // namespace armnn
diff --git a/src/backends/reference/workloads/Encoders.hpp b/src/backends/reference/workloads/Encoders.hpp
index d6d6114..8a70237 100644
--- a/src/backends/reference/workloads/Encoders.hpp
+++ b/src/backends/reference/workloads/Encoders.hpp
@@ -1,5 +1,5 @@
//
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2017-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
@@ -56,10 +56,24 @@
}
case armnn::DataType::QSymmS16:
{
- return std::make_unique<QSymm16Encoder>(
- static_cast<int16_t*>(data),
- info.GetQuantizationScale(),
- info.GetQuantizationOffset());
+ if (info.HasPerAxisQuantization())
+ {
+ unsigned int axis = info.GetQuantizationDim().value();
+ auto axisDimensionality = info.GetShape()[axis];
+ std::pair<unsigned int, std::vector<float>> params = armnnUtils::GetPerAxisParams(info);
+ return std::make_unique<QSymm16PerAxisEncoder>(
+ static_cast<int16_t*>(data),
+ params.second,
+ params.first,
+ axisDimensionality);
+ }
+ else
+ {
+ return std::make_unique<QSymm16Encoder>(
+ static_cast<int16_t *>(data),
+ info.GetQuantizationScale(),
+ info.GetQuantizationOffset());
+ }
}
case armnn::DataType::Signed32:
{