COMPMID-1800: (Nightly) Mismatches in SC9863 board for NEON FP16

Fixes for:

- ReduceMean, reduction on the X axis for FP16 with 8 elements was
performed only up to a certain point. The fix now takes into account the
number of elements of the vector and does as many reductions as
necessary.

- YOLOLayer, activation for FP16 has to be performed on 32 bits until
the FP16 approximations is fixed.

Change-Id: I75373f4edd37de476e6fe1a56de3ef386b65c619

1 files changed, 4 insertions, 1 deletions

diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
index b77219cd79..182e93d177 100644
--- a/src/core/NEON/kernels/NEReductionOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
@@ -154,7 +154,10 @@ struct RedOpX
         input);
 
         auto carry_addition = wrapper::vpadd(wrapper::vgethigh(vec_sum_value), wrapper::vgetlow(vec_sum_value));
-        carry_addition      = wrapper::vpadd(carry_addition, carry_addition);
+        for(int i = 0; i < S / 4; ++i)
+        {
+            carry_addition = wrapper::vpadd(carry_addition, carry_addition);
+        }
 
         auto res = wrapper::vgetlane(carry_addition, 0);
         if(op == ReductionOperation::MEAN_SUM)