diff options
author | Aron Virginas-Tar <Aron.Virginas-Tar@arm.com> | 2019-11-04 15:00:19 +0000 |
---|---|---|
committer | Áron Virginás-Tar <aron.virginas-tar@arm.com> | 2019-11-05 12:36:31 +0000 |
commit | b67f95796b03c1a5d3b3fa6a40f6baf85bdd09cb (patch) | |
tree | d96468d22624886f9495dbaf8b381f232d347b8b /src/backends | |
parent | 13b653fd246b31279ae6cb42261671462b52d1be (diff) | |
download | armnn-b67f95796b03c1a5d3b3fa6a40f6baf85bdd09cb.tar.gz |
IVGCVSW-3836 Add support for Int32 per-axis scales
* Added ScaledInt32PerAxisDecoder implementation
* Added new case for Signed32 in MakeDecoder that returns a
ScaledInt32PerAxisDecoder if the tensor info has multiple
quantization scales
Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I8b3c11091644da993044d2a0fe2aba6b06b5af56
Diffstat (limited to 'src/backends')
-rw-r--r-- | src/backends/reference/workloads/BaseIterator.hpp | 23 | ||||
-rw-r--r-- | src/backends/reference/workloads/Decoders.hpp | 56 |
2 files changed, 71 insertions, 8 deletions
diff --git a/src/backends/reference/workloads/BaseIterator.hpp b/src/backends/reference/workloads/BaseIterator.hpp index 9fe3f15f9b..50475312a5 100644 --- a/src/backends/reference/workloads/BaseIterator.hpp +++ b/src/backends/reference/workloads/BaseIterator.hpp @@ -451,4 +451,25 @@ private: std::vector<float> m_Scale; }; -} //namespace armnn
\ No newline at end of file +class ScaledInt32PerAxisDecoder : public PerAxisIterator<const int32_t, Decoder<float>> +{ +public: + ScaledInt32PerAxisDecoder(const int32_t* data, const std::vector<float>& scales, unsigned int axisFactor) + : PerAxisIterator(data, axisFactor), m_Scales(scales) {} + + float Get() const override + { + return armnn::Dequantize(*m_Iterator, m_Scales[m_AxisIndex], 0); + } + + // Get scale of the current value + float GetScale() const + { + return m_Scales[m_AxisIndex]; + } + +private: + std::vector<float> m_Scales; +}; + +} // namespace armnn diff --git a/src/backends/reference/workloads/Decoders.hpp b/src/backends/reference/workloads/Decoders.hpp index dd2b28a50f..dcd498cb63 100644 --- a/src/backends/reference/workloads/Decoders.hpp +++ b/src/backends/reference/workloads/Decoders.hpp @@ -14,6 +14,54 @@ namespace armnn { +namespace +{ + +inline std::unique_ptr<Decoder<float>> MakeSigned32PerAxisDecoder(const TensorInfo& info, const void* data) +{ + auto params = armnnUtils::GetPerAxisParams(info); + return std::make_unique<ScaledInt32PerAxisDecoder>( + static_cast<const int32_t*>(data), + params.second, + params.first); +} + +inline std::unique_ptr<Decoder<float>> MakeSigned32Decoder(const TensorInfo& info, const void* data) +{ + if(info.HasMultipleQuantizationScales()) + { + // NOTE: If we have multiple quantization scales, we create a ScaledInt32PerAxisDecoder. + // This will be used to decode per-axis quantized convolution biases. + return MakeSigned32PerAxisDecoder(info, data); + } + else + { + if (info.GetQuantizationDim().has_value()) + { + // NOTE: Even though we only have a single quantization scale, if the quantization + // dimension is set, the tensor has per-axis quantization and we need to create a + // ScaledInt32PerAxisDecoder + return MakeSigned32PerAxisDecoder(info, data); + } + + const float scale = info.GetQuantizationScale(); + if (scale == 0.f) + { + // NOTE:: If no quantization scale is set, we create an Int32Decoder, which simply + // casts the int value to float. This will be used for any INT32 data other than + // convolution biases. + return std::make_unique<Int32Decoder>(static_cast<const int32_t*>(data)); + } + + // NOTE: If we only have a single (non-zero) quantization scale and no quantization + // dimension is specified, we need to create a ScaledInt32Decoder. This will be used + // to decode per-tensor quantized convolution biases. + return std::make_unique<ScaledInt32Decoder>(static_cast<const int32_t*>(data), scale); + } +} + +} // anonymous namespace + template<typename T> inline std::unique_ptr<Decoder<T>> MakeDecoder(const TensorInfo& info, const void* data = nullptr); @@ -54,13 +102,7 @@ inline std::unique_ptr<Decoder<float>> MakeDecoder(const TensorInfo& info, const } case DataType::Signed32: { - const float scale = info.GetQuantizationScale(); - if (scale == 0.f) - { - return std::make_unique<Int32Decoder>(static_cast<const int32_t*>(data)); - } - // NOTE: ScaledInt32Decoder is used for quantized convolution biases - return std::make_unique<ScaledInt32Decoder>(static_cast<const int32_t*>(data), scale); + return MakeSigned32Decoder(info, data); } default: { |