commit 9f0693b41a33d4d17ef016d8a5490cc65a8cfb8a
author Aron Virginas-Tar <Aron.Virginas-Tar@arm.com> Wed Nov 06 14:32:30 2019 +0000
committer Aron Virginas-Tar <Aron.Virginas-Tar@arm.com> Wed Nov 06 15:28:34 2019 +0000
tree e19c9f330b90d3161e089b281be26aac0e44e7be
parent 3b959603ab1e3dc7b7b70798f357b3fe6c0e47c8

IVGCVSW-3841 Add support for per-axis quantization

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: Ife7fa63b8839465e8f9f8626f34ca8c0f4d12788

ConversionUtils.hpp

diff --git a/ConversionUtils.hpp b/ConversionUtils.hpp
index e4ac4a5..1975434 100644
--- a/ConversionUtils.hpp
+++ b/ConversionUtils.hpp
@@ -183,11 +183,12 @@
 
 inline bool IsOperandTypeSupportedForTensors(V1_2::OperandType type)
 {
-    return type == V1_2::OperandType::BOOL                ||
-           type == V1_2::OperandType::TENSOR_FLOAT16      ||
-           type == V1_2::OperandType::TENSOR_FLOAT32      ||
-           type == V1_2::OperandType::TENSOR_QUANT8_ASYMM ||
-           type == V1_2::OperandType::TENSOR_QUANT16_SYMM ||
+    return type == V1_2::OperandType::BOOL                           ||
+           type == V1_2::OperandType::TENSOR_FLOAT16                 ||
+           type == V1_2::OperandType::TENSOR_FLOAT32                 ||
+           type == V1_2::OperandType::TENSOR_QUANT8_ASYMM            ||
+           type == V1_2::OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL ||
+           type == V1_2::OperandType::TENSOR_QUANT16_SYMM            ||
            type == V1_2::OperandType::TENSOR_INT32;
 }
 
@@ -384,16 +385,37 @@
 // we accept some tolerance. We don't want ArmNN itself to accept these inconsistencies as it is up to the
 // user (us, in this case) to ensure they match.
 void SanitizeBiasQuantizationScale(armnn::TensorInfo& biasInfo,
-                                   const armnn::TensorInfo& weightInfo, const armnn::TensorInfo& inputInfo)
+                                   const armnn::TensorInfo& weightInfo,
+                                   const armnn::TensorInfo& inputInfo)
 {
-    const float expectedBiasScale = weightInfo.GetQuantizationScale() * inputInfo.GetQuantizationScale();
-    if (biasInfo.GetQuantizationScale() != expectedBiasScale)
+    if (weightInfo.HasPerAxisQuantization())
     {
-        boost::math::fpc::close_at_tolerance<float> comparer(boost::math::fpc::percent_tolerance(1.0f));
-        if (comparer(biasInfo.GetQuantizationScale(), expectedBiasScale))
+        // NOTE: Bias scale is always set to 0 for per-axis quantization and
+        // it needs to be calculated: scale[i] = input_scale * weight_scale[i]
+        auto UpdateBiasScaleValue = [&inputInfo](float biasScale) -> float
         {
-            ALOGW("Bias quantization scale has been modified to match input*weights");
-            biasInfo.SetQuantizationScale(expectedBiasScale);
+            return biasScale * inputInfo.GetQuantizationScale();
+        };
+
+        std::vector<float> biasScales(weightInfo.GetQuantizationScales());
+        std::transform(biasScales.begin(), biasScales.end(), biasScales.begin(), UpdateBiasScaleValue);
+
+        biasInfo.SetQuantizationScales(biasScales);
+        biasInfo.SetQuantizationDim(weightInfo.GetQuantizationDim());
+
+        ALOGV("Bias quantization params have been updated for per-axis quantization");
+    }
+    else
+    {
+        const float expectedBiasScale = weightInfo.GetQuantizationScale() * inputInfo.GetQuantizationScale();
+        if (biasInfo.GetQuantizationScale() != expectedBiasScale)
+        {
+            boost::math::fpc::close_at_tolerance<float> comparer(boost::math::fpc::percent_tolerance(1.0f));
+            if (comparer(biasInfo.GetQuantizationScale(), expectedBiasScale))
+            {
+                ALOGW("Bias quantization scale has been modified to match input * weights");
+                biasInfo.SetQuantizationScale(expectedBiasScale);
+            }
         }
     }
 }