From: Kevin May <kevin.may@arm.com>
Date: Wed, 29 Apr 2020 16:05:05 +0000 (+0100)
Subject: IVGCVSW-4749 Fix CpuAcc Hal 1.3 FullyConnected Failures
X-Git-Tag: submit/tizen/20200730.023729~87
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=be7e35c5994c379e165fb1b7053301afb5d52e04;p=platform%2Fupstream%2Farmnn.git

IVGCVSW-4749 Fix CpuAcc Hal 1.3 FullyConnected Failures

* Use make_unique instead of MakeWorkloadHelper for FullyConnected Workload
* Add Neon Workload Tests for QAsymmU8 and QAsymmS8

Signed-off-by: Kevin May <kevin.may@arm.com>
Change-Id: Ie1359af8b0a0bec1e058e20fc789c2520ebe64e8
---

diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 237808c..56ee9a7 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -274,8 +274,7 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFloor(const FloorQueueDesc
 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateFullyConnected(
     const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) const
 {
-    return MakeWorkloadHelper<NeonFullyConnectedWorkload, NeonFullyConnectedWorkload>(
-        descriptor, info, m_MemoryManager->GetIntraLayerManager());
+    return std::make_unique<NeonFullyConnectedWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
 }
 
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateGather(const armnn::GatherQueueDescriptor& descriptor,
diff --git a/src/backends/neon/test/NeonCreateWorkloadTests.cpp b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
index 447bad1..a89602d 100644
--- a/src/backends/neon/test/NeonCreateWorkloadTests.cpp
+++ b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
@@ -325,8 +325,12 @@ static void NeonCreateFullyConnectedWorkloadTest()
     FullyConnectedQueueDescriptor queueDescriptor = workload->GetData();
     auto inputHandle  = PolymorphicDowncast<IAclTensorHandle*>(queueDescriptor.m_Inputs[0]);
     auto outputHandle = PolymorphicDowncast<IAclTensorHandle*>(queueDescriptor.m_Outputs[0]);
-    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 1, 4, 5}, DataType)));
-    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 7}, DataType)));
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateFullyConnectedWorkloadTest).
+    float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0;
+    float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0;
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 1, 4, 5}, DataType, inputsQScale)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 7}, DataType, outputQScale)));
 }
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -341,6 +345,16 @@ BOOST_AUTO_TEST_CASE(CreateFullyConnectedFloatWorkload)
     NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::Float32>();
 }
 
+BOOST_AUTO_TEST_CASE(CreateFullyConnectedQAsymmU8Workload)
+{
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::QAsymmU8>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateFullyConnectedQAsymmS8Workload)
+{
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::QAsymmS8>();
+}
+
 template <typename NormalizationWorkloadType, typename armnn::DataType DataType>
 static void NeonCreateNormalizationWorkloadTest(DataLayout dataLayout)
 {