commit 21fc28bb68d7a6232eea3a0e8cfd800d95346ea4
author Aron Virginas-Tar <Aron.Virginas-Tar@arm.com> Tue Nov 26 14:04:54 2019 +0000
committer Aron Virginas-Tar <Aron.Virginas-Tar@arm.com> Tue Nov 26 15:00:28 2019 +0000
tree 5eca96ac31e3f8c62da3ad7742834180a0f7eeac
parent 9b6d4246804d39e5b7570992d63e31a044442385

IVGCVSW-3866 Add support for per-channel QSymm8 Convolution2d to Neon backend

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: If7d366b2894050c53a651016165f34be1a2ce7a6

src/armnn/CompatibleTypes.hpp

diff --git a/src/armnn/CompatibleTypes.hpp b/src/armnn/CompatibleTypes.hpp
index 4e31aed..06aa064 100644
--- a/src/armnn/CompatibleTypes.hpp
+++ b/src/armnn/CompatibleTypes.hpp
@@ -36,6 +36,12 @@
 }
 
 template<>
+inline bool CompatibleTypes<int8_t>(DataType dataType)
+{
+    return dataType == DataType::QuantisedSymm8 || dataType == DataType::QuantizedSymm8PerAxis;
+}
+
+template<>
 inline bool CompatibleTypes<int16_t>(DataType dataType)
 {
     return dataType == DataType::QuantisedSymm16;

src/backends/neon/NeonLayerSupport.cpp

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index ed0f41a..4474b12 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -274,11 +274,6 @@
                                                 const Optional<TensorInfo>& biases,
                                                 Optional<std::string&> reasonIfUnsupported) const
 {
-    if (weights.HasPerAxisQuantization())
-    {
-        return false;
-    }
-
     // Multiplier > 1.0f currently not supported in ACL
     if ((input.GetQuantizationScale() * weights.GetQuantizationScale()) / output.GetQuantizationScale() > 1.0f)
     {

src/backends/neon/test/NeonLayerTests.cpp

diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 26c5536..d74a4c6 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -110,6 +110,9 @@
                      false,
                      armnn::DataLayout::NCHW)
 
+ARMNN_AUTO_TEST_CASE(Convolution2dPerAxisQuantTestNchw, Convolution2dPerAxisQuantTest, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(Convolution2dPerAxisQuantTestNhwc, Convolution2dPerAxisQuantTest, DataLayout::NHWC);
+
 // DepthToSpace
 ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_1, DepthToSpaceTest1<DataType::Float32>, DataLayout::NCHW);
 ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_2, DepthToSpaceTest2<DataType::Float32>, DataLayout::NCHW);

src/backends/neon/workloads/NeonWorkloadUtils.hpp

diff --git a/src/backends/neon/workloads/NeonWorkloadUtils.hpp b/src/backends/neon/workloads/NeonWorkloadUtils.hpp
index f63946e..e9edc89 100644
--- a/src/backends/neon/workloads/NeonWorkloadUtils.hpp
+++ b/src/backends/neon/workloads/NeonWorkloadUtils.hpp
@@ -46,6 +46,9 @@
         case DataType::QuantisedAsymm8:
             CopyArmComputeTensorData(tensor, handle->GetConstTensor<uint8_t>());
             break;
+        case DataType::QuantizedSymm8PerAxis:
+            CopyArmComputeTensorData(tensor, handle->GetConstTensor<int8_t>());
+            break;
         case DataType::Signed32:
             CopyArmComputeTensorData(tensor, handle->GetConstTensor<int32_t>());
             break;