Fix MeanStdDevNormalizationLayer reference outputting nan for FP16

- Bring the epsilon up to 1e-3 for FP16 (both backends) since it was causing the reference's variance being negative and its square root being NaN
- Bring the epsilon up to 1e-7 for FP16 NEON test for the same problem on the NEON kernel
- Adjust the CL kernel's vec_size when input tensor's width < 16 and use macros agnostic of vector size for sum reduction
- Add previously mismatching tensor shapes

Resolve COMPMID-4354

Change-Id: I823c871aacb72326f90c86b24cb16c3e2d4bd15e
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5630
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 8 insertions, 21 deletions

diff --git a/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp b/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp
index 7dc34f17b5..9f98b67582 100644
--- a/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp
+++ b/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp
@@ -54,22 +54,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, f
     }
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    if(output != nullptr)
-    {
-        ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-        // Output auto inizialitation if not yet initialized
-        auto_init_if_empty(*output, *input);
-    }
-
-    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
-
-    // This kernel doesn't need padding
-    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
-    return std::make_pair(Status{}, win);
-}
 } // namespace
 
 CLMeanStdDevNormalizationKernel::CLMeanStdDevNormalizationKernel()
@@ -90,10 +74,15 @@ void CLMeanStdDevNormalizationKernel::configure(const CLCompileContext &compile_
 
     ARM_COMPUTE_ERROR_THROW_ON(CLMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon));
 
+    if(output != nullptr)
+    {
+        auto_init_if_empty(*output->info(), *input->info());
+    }
+
     _input  = input;
     _output = output;
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->info()->element_size(), input->info()->dimension(0));
 
     // Set build options
     CLBuildOptions build_opts;
@@ -107,9 +96,8 @@ void CLMeanStdDevNormalizationKernel::configure(const CLCompileContext &compile_
     _kernel = create_kernel(compile_context, "mean_stddev_normalization", build_opts.options());
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), (_run_in_place) ? nullptr : output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
+    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    ICLKernel::configure_internal(win);
 
     // Set config_id for enabling LWS tuning
     _config_id = "mean_stddev_normalization_layer_";
@@ -123,7 +111,6 @@ void CLMeanStdDevNormalizationKernel::configure(const CLCompileContext &compile_
 Status CLMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first);
     return Status{};
 }