Fix Intermittent Neon™ ReduceMean QASYMM8 Mismatch

- Dividing scale by number of elements causes accuracy loss due to limitations in float datatype and truncation to int
- Adds rounding after division on aarch64 to negate this.

Resolves: [COMPMID-5839]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I54ef0f7e56f39da1fa5f30378f551b5ca419a61d
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/492456
Tested-by: bsgcomp <bsgcomp@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9110
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 7 insertions, 1 deletions

diff --git a/tests/validation/fixtures/ReduceMeanFixture.h b/tests/validation/fixtures/ReduceMeanFixture.h
index 304630e9f5..13354eedc1 100644
--- a/tests/validation/fixtures/ReduceMeanFixture.h
+++ b/tests/validation/fixtures/ReduceMeanFixture.h
@@ -124,7 +124,13 @@ protected:
         {
             TensorShape output_shape = i == 0 ? src_shape : out.shape();
             output_shape.set(axis[i], 1);
-            out = reference::reduction_operation<T, T>(i == 0 ? src : out, output_shape, axis[i], ReductionOperation::MEAN_SUM, quantization_info_output);
+            bool is_opencl = false;
+
+#ifdef ARM_COMPUTE_OPENCL_ENABLED
+            is_opencl = std::is_same<CLTensor, TensorType>::value; // Round down to zero on opencl to match kernel
+#endif /* ARM_COMPUTE_OPENCL_ENABLED */
+            out = reference::reduction_operation<T, T>(i == 0 ? src : out, output_shape, axis[i], ReductionOperation::MEAN_SUM, quantization_info_output, is_opencl ? RoundingPolicy::TO_ZERO : RoundingPolicy::TO_NEAREST_UP);
+
         }
 
         if(!keep_dims)