IVGCVSW-3967 Avg_Pooling2d Fails on CL NHWC FP16

 * Enable fp_mixed_precision flag for the failing test case

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: If13552165eb6598a84d213b82847b56a8c5f2783

3 files changed, 22 insertions, 5 deletions

diff --git a/src/backends/aclCommon/ArmComputeTensorUtils.cpp b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
index fabe9a15bf..c7d250a706 100644
--- a/src/backends/aclCommon/ArmComputeTensorUtils.cpp
+++ b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
@@ -140,7 +140,8 @@ arm_compute::DataLayout ConvertDataLayout(armnn::DataLayout dataLayout)
     }
 }
 
-arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDescriptor& descriptor)
+arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDescriptor& descriptor,
+                                                              bool fpMixedPrecision)
 {
     using arm_compute::PoolingType;
     using arm_compute::DimensionRoundingType;
@@ -172,7 +173,7 @@ arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDes
 
     const Size2D poolSize(descriptor.m_PoolWidth, descriptor.m_PoolHeight);
 
-    return arm_compute::PoolingLayerInfo(poolingType, poolSize, padStrideInfo, excludePadding);
+    return arm_compute::PoolingLayerInfo(poolingType, poolSize, padStrideInfo, excludePadding, fpMixedPrecision);
 }
 
 arm_compute::NormalizationLayerInfo BuildArmComputeNormalizationLayerInfo(const NormalizationDescriptor& descriptor)
diff --git a/src/backends/aclCommon/ArmComputeTensorUtils.hpp b/src/backends/aclCommon/ArmComputeTensorUtils.hpp
index 34565fe425..c30de19557 100644
--- a/src/backends/aclCommon/ArmComputeTensorUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeTensorUtils.hpp
@@ -48,8 +48,11 @@ arm_compute::TensorInfo BuildArmComputeTensorInfo(const armnn::TensorInfo& tenso
 /// armnn::DataLayout.
 arm_compute::DataLayout ConvertDataLayout(armnn::DataLayout dataLayout);
 
-/// Utility function used to setup an arm_compute::PoolingLayerInfo object from an armnn::Pooling2dDescriptor.
-arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDescriptor& descriptor);
+/// Utility function used to setup an arm_compute::PoolingLayerInfo object from given
+/// armnn::Pooling2dDescriptor
+/// bool fpMixedPrecision
+arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDescriptor& descriptor,
+                                                              bool fpMixedPrecision = false);
 
 /// Utility function to setup an arm_compute::NormalizationLayerInfo object from an armnn::NormalizationDescriptor.
 arm_compute::NormalizationLayerInfo BuildArmComputeNormalizationLayerInfo(const NormalizationDescriptor& desc);
diff --git a/src/backends/cl/workloads/ClPooling2dWorkload.cpp b/src/backends/cl/workloads/ClPooling2dWorkload.cpp
index 607bc58f38..bf899977d6 100644
--- a/src/backends/cl/workloads/ClPooling2dWorkload.cpp
+++ b/src/backends/cl/workloads/ClPooling2dWorkload.cpp
@@ -40,7 +40,20 @@ ClPooling2dWorkload::ClPooling2dWorkload(
     input.info()->set_data_layout(aclDataLayout);
     output.info()->set_data_layout(aclDataLayout);
 
-    arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters);
+    // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
+    // enable fp_mixed_precision for the the FP16 cases that
+    // accumulation reaches a limit beyond which there is no more increment of the value
+    bool fpMixedPrecision = false;
+    if (input.info()->data_type() == arm_compute::DataType::F16
+        && m_Data.m_Parameters.m_PoolType == PoolingAlgorithm::Average
+        && m_Data.m_Parameters.m_PoolWidth >= 100
+        && input.info()->dimension(3) >= 5
+        && input.info()->dimension(2) * input.info()->dimension(1) >= 3000)
+    {
+        fpMixedPrecision = true;
+    }
+
+    arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters, fpMixedPrecision);
 
     // Run the layer.
     m_PoolingLayer.configure(&input, &output, layerInfo);