COMPMID-1981: fixed Nightly DepthConcatenate QASMM8 mismatches

Change-Id: Ibd72e03ffa6aa50543b3c410d6167231fdd16990
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/662
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com>

1 files changed, 13 insertions, 10 deletions

diff --git a/tests/validation/reference/DepthConcatenateLayer.cpp b/tests/validation/reference/DepthConcatenateLayer.cpp
index 7775614171..f0dab3baac 100644
--- a/tests/validation/reference/DepthConcatenateLayer.cpp
+++ b/tests/validation/reference/DepthConcatenateLayer.cpp
@@ -51,20 +51,23 @@ SimpleTensor<T> depthconcatenate_layer(const std::vector<SimpleTensor<T>> &srcs,
     const int depth_out    = dst.shape().z();
     const int out_stride_z = width_out * height_out;
     const int batches      = dst.shape().total_size_upper(3);
-
     if(srcs[0].data_type() == DataType::QASYMM8 && srcs[0].quantization_info() != dst.quantization_info())
     {
-        // input tensors can have smaller width and height than the output, so for each output's slice we need to requantize 0 (as this is the value
-        // used in NEFillBorderKernel by NEDepthConcatenateLayer) using the corresponding quantization info for that particular slice/input tensor.
-        int slice = 0;
-        for(const auto &src : srcs)
+        for(int b = 0; b < batches; ++b)
         {
-            auto ptr_slice = static_cast<T *>(dst(Coordinates(0, 0, slice)));
-            std::transform(ptr_slice, ptr_slice + dst.num_elements() / depth_out, ptr_slice, [src, dst](T t)
+            // input tensors can have smaller width and height than the output, so for each output's slice we need to requantize 0 (as this is the value
+            // used in NEFillBorderKernel by NEDepthConcatenateLayer) using the corresponding quantization info for that particular slice/input tensor.
+            int slice = 0;
+            for(const auto &src : srcs)
             {
-                return dst.quantization_info().quantize(src.quantization_info().dequantize(0), RoundingPolicy::TO_NEAREST_UP);
-            });
-            slice += src.shape().z();
+                auto       ptr_slice = static_cast<T *>(dst(Coordinates(0, 0, slice, b)));
+                const auto num_elems_in_slice((dst.num_elements() / depth_out) * src.shape().z());
+                std::transform(ptr_slice, ptr_slice + num_elems_in_slice, ptr_slice, [src, dst](T t)
+                {
+                    return dst.quantization_info().quantize(src.quantization_info().dequantize(0), RoundingPolicy::TO_NEAREST_UP);
+                });
+                slice += src.shape().z();
+            }
         }
     }
     else