COMPMID-417 NEON/CL MeanStdDev bugfix using FillBorderKernel

Change-Id: Ic48ba7f69783d0e1e80611264e2bc67d1732436e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/81293
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

8 files changed, 71 insertions, 2 deletions

diff --git a/tests/AssetsLibrary.h b/tests/AssetsLibrary.h
index 3dd30e7629..18ffd773c8 100644
--- a/tests/AssetsLibrary.h
+++ b/tests/AssetsLibrary.h
@@ -150,6 +150,15 @@ public:
      */
     RawTensor get(const std::string &name, Format format, Channel channel);
 
+    /** Puts garbage values all around the tensor for testing purposes
+     *
+     * @param[in, out] tensor       To be filled tensor.
+     * @param[in]      distribution Distribution used to fill the tensor's surroundings.
+     * @param[in]      seed_offset  The offset will be added to the global seed before initialising the random generator.
+     */
+    template <typename T, typename D>
+    void fill_borders_with_garbage(T &&tensor, D &&distribution, std::random_device::result_type seed_offset) const;
+
     /** Fills the specified @p tensor with random values drawn from @p
      * distribution.
      *
@@ -348,6 +357,32 @@ private:
 };
 
 template <typename T, typename D>
+void AssetsLibrary::fill_borders_with_garbage(T &&tensor, D &&distribution, std::random_device::result_type seed_offset) const
+{
+    const PaddingSize padding_size = tensor.padding();
+
+    Window window;
+    window.set(0, Window::Dimension(-padding_size.left, tensor.shape()[0] + padding_size.right, 1));
+    window.set(1, Window::Dimension(-padding_size.top, tensor.shape()[1] + padding_size.bottom, 1));
+
+    std::mt19937 gen(_seed);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        TensorShape shape = tensor.shape();
+
+        // If outside of valid region
+        if(id.x() < 0 || id.x() >= static_cast<int>(shape.x()) || id.y() < 0 || id.y() >= static_cast<int>(shape.y()))
+        {
+            using ResultType         = typename std::remove_reference<D>::type::result_type;
+            const ResultType value   = distribution(gen);
+            void *const      out_ptr = tensor(id);
+            store_value_with_data_type(out_ptr, value, tensor.data_type());
+        }
+    });
+}
+
+template <typename T, typename D>
 void AssetsLibrary::fill(T &&tensor, D &&distribution, std::random_device::result_type seed_offset) const
 {
     Window window;
@@ -366,6 +401,8 @@ void AssetsLibrary::fill(T &&tensor, D &&distribution, std::random_device::resul
         void *const      out_ptr = tensor(id);
         store_value_with_data_type(out_ptr, value, tensor.data_type());
     });
+
+    fill_borders_with_garbage(tensor, distribution, seed_offset);
 }
 
 template <typename D>
diff --git a/tests/CL/CLAccessor.h b/tests/CL/CLAccessor.h
index 260da035c5..b1d7a078b1 100644
--- a/tests/CL/CLAccessor.h
+++ b/tests/CL/CLAccessor.h
@@ -59,6 +59,7 @@ public:
     DataType    data_type() const override;
     int         num_channels() const override;
     int         num_elements() const override;
+    PaddingSize padding() const override;
     int         fixed_point_position() const override;
     const void *operator()(const Coordinates &coord) const override;
     void *operator()(const Coordinates &coord) override;
@@ -113,6 +114,11 @@ inline int CLAccessor::num_elements() const
     return _tensor.info()->tensor_shape().total_size();
 }
 
+inline PaddingSize CLAccessor::padding() const
+{
+    return _tensor.info()->padding();
+}
+
 inline int CLAccessor::fixed_point_position() const
 {
     return _tensor.info()->fixed_point_position();
diff --git a/tests/IAccessor.h b/tests/IAccessor.h
index 0517981df5..ef06e9e9da 100644
--- a/tests/IAccessor.h
+++ b/tests/IAccessor.h
@@ -61,6 +61,9 @@ public:
     /** Number of elements of the tensor. */
     virtual int num_elements() const = 0;
 
+    /** Available padding around the tensor. */
+    virtual PaddingSize padding() const = 0;
+
     /** Number of bits for the fractional part. */
     virtual int fixed_point_position() const = 0;
 
diff --git a/tests/NEON/Accessor.h b/tests/NEON/Accessor.h
index 5949b350a0..c379018d39 100644
--- a/tests/NEON/Accessor.h
+++ b/tests/NEON/Accessor.h
@@ -53,6 +53,7 @@ public:
     DataType    data_type() const override;
     int         num_channels() const override;
     int         num_elements() const override;
+    PaddingSize padding() const override;
     int         fixed_point_position() const override;
     const void *operator()(const Coordinates &coord) const override;
     void *operator()(const Coordinates &coord) override;
@@ -101,6 +102,11 @@ inline int Accessor::num_elements() const
     return _tensor.info()->tensor_shape().total_size();
 }
 
+inline PaddingSize Accessor::padding() const
+{
+    return _tensor.info()->padding();
+}
+
 inline int Accessor::fixed_point_position() const
 {
     return _tensor.info()->fixed_point_position();
diff --git a/tests/RawTensor.cpp b/tests/RawTensor.cpp
index 402b5f3d0b..1d400a58d7 100644
--- a/tests/RawTensor.cpp
+++ b/tests/RawTensor.cpp
@@ -146,6 +146,11 @@ int RawTensor::num_elements() const
     return _shape.total_size();
 }
 
+PaddingSize RawTensor::padding() const
+{
+    return PaddingSize(0);
+}
+
 const RawTensor::BufferType *RawTensor::data() const
 {
     return _buffer.get();
diff --git a/tests/RawTensor.h b/tests/RawTensor.h
index 2480917c38..9d65e4f319 100644
--- a/tests/RawTensor.h
+++ b/tests/RawTensor.h
@@ -114,6 +114,9 @@ public:
     /** Number of elements of the tensor. */
     int num_elements() const;
 
+    /** Available padding around the tensor. */
+    PaddingSize padding() const;
+
     /** The number of bits for the fractional part of the fixed point numbers. */
     int fixed_point_position() const;
 
diff --git a/tests/validation/NEON/MeanStdDev.cpp b/tests/validation/NEON/MeanStdDev.cpp
index 5fcd81ec95..d39435f2b3 100644
--- a/tests/validation/NEON/MeanStdDev.cpp
+++ b/tests/validation/NEON/MeanStdDev.cpp
@@ -134,8 +134,8 @@ BOOST_DATA_TEST_CASE(RunLarge, Large2DShapes(), shape)
     std::pair<float, float> ref_output = Reference::compute_reference_mean_and_standard_deviation(shape);
 
     // Validate output
-    validate(output.first, ref_output.first);
-    validate(output.second, ref_output.second, 0.f, 0.001f);
+    validate(output.first, ref_output.first, 0.f, 0.0001f);
+    validate(output.second, ref_output.second, 0.f, 0.01f);
 }
 
 BOOST_AUTO_TEST_SUITE_END()
diff --git a/tests/validation_new/SimpleTensor.h b/tests/validation_new/SimpleTensor.h
index 6392e38e25..61d6f1cd04 100644
--- a/tests/validation_new/SimpleTensor.h
+++ b/tests/validation_new/SimpleTensor.h
@@ -127,6 +127,9 @@ public:
     /** Number of elements of the tensor. */
     int num_elements() const override;
 
+    /** Available padding around the tensor. */
+    PaddingSize padding() const override;
+
     /** The number of bits for the fractional part of the fixed point numbers. */
     int fixed_point_position() const override;
 
@@ -292,6 +295,12 @@ int SimpleTensor<T>::num_elements() const
 }
 
 template <typename T>
+PaddingSize SimpleTensor<T>::padding() const
+{
+    return PaddingSize(0);
+}
+
+template <typename T>
 const T *SimpleTensor<T>::data() const
 {
     return _buffer.get();