Function CheckScaleSetOnQuantizedType

src/armnn/Network.cpp:802–843

}

bool CheckScaleSetOnQuantizedType(Layer* layer, Optional<std::vector<std::string>&> errMessages)
{
    bool noErrors = true;
    unsigned int numOutputs = layer->GetNumOutputSlots();
    for (unsigned int i = 0; i < numOutputs; i++) {
        OutputSlot& outputSlot = layer->GetOutputSlot(i);
        TensorInfo info = outputSlot.GetTensorInfo();
        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.");
                }
                // 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);
                    ReportError(ss.str(), errMessages);
                }
                break;
            default:
                break;
        }
    }
    return noErrors;
}

OptimizationResult AttemptBackendAssignment(BackendSettings& backendSettings,
                                            Graph& graph,