Add optimization for global pooling in pooling_layer.cl

- Simplify the implementation when the pooling size has the same spatial
  dimensions of the input tensor
- Rework the heuristic for F32/F16
- Add test for validating the global pooling path
- Fix compare_dimensions in validation. The validation fails because we have different
  number of dimensions for NCHW and NHWC (e.g. 1,1,2,1(NCHW) -> 2,1,1,1(NHWC)

Change-Id: Iba680cb30bf2a5d0952265a4cc9794f368549ca5
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5510
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 3 insertions, 1 deletions

diff --git a/src/core/gpu/cl/kernels/ClPoolingKernel.cpp b/src/core/gpu/cl/kernels/ClPoolingKernel.cpp
index 78243402bf..a432877a1d 100644
--- a/src/core/gpu/cl/kernels/ClPoolingKernel.cpp
+++ b/src/core/gpu/cl/kernels/ClPoolingKernel.cpp
@@ -173,9 +173,11 @@ std::tuple<Status, Window, ClPoolingConfig> validate_and_configure_window(ITenso
         }
         case DataLayout::NHWC:
         {
+            const size_t vec_size = dst->data_type() == DataType::F32 ? 2 : 4;
+
             // Initialize border size
             border_size                       = BorderSize();
-            num_elems_processed_per_iteration = adjust_vec_size(4, dst->dimension(0));
+            num_elems_processed_per_iteration = adjust_vec_size(vec_size, dst->dimension(0));
             win                               = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration));
             break;
         }