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>

17 files changed, 116 insertions, 18 deletions

diff --git a/arm_compute/core/CL/kernels/CLMeanStdDevKernel.h b/arm_compute/core/CL/kernels/CLMeanStdDevKernel.h
index 9f30f76e1b..2c09590dab 100644
--- a/arm_compute/core/CL/kernels/CLMeanStdDevKernel.h
+++ b/arm_compute/core/CL/kernels/CLMeanStdDevKernel.h
@@ -63,12 +63,15 @@ public:
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
 
+    BorderSize border_size() const override;
+
 private:
     const ICLImage *_input;
     float          *_mean;
     float          *_stddev;
     cl::Buffer     *_global_sum;
     cl::Buffer     *_global_sum_squared;
+    BorderSize      _border_size;
 };
 }
 #endif /* __ARM_COMPUTE_CLMEANSTDDEVKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h b/arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h
index 83407ccb7d..9c72b20d58 100644
--- a/arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h
+++ b/arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h
@@ -64,6 +64,8 @@ public:
     // Inherited methods overridden:
     void run(const Window &window) override;
 
+    BorderSize border_size() const override;
+
 private:
     const IImage *_input;
     float        *_mean;
@@ -71,6 +73,7 @@ private:
     uint64_t     *_global_sum;
     uint64_t     *_global_sum_squared;
     std::mutex    _mtx;
+    BorderSize    _border_size;
 };
 }
 #endif /* __ARM_COMPUTE_NEMEANSTDDEVKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLMeanStdDev.h b/arm_compute/runtime/CL/functions/CLMeanStdDev.h
index e33bcdd779..7622138236 100644
--- a/arm_compute/runtime/CL/functions/CLMeanStdDev.h
+++ b/arm_compute/runtime/CL/functions/CLMeanStdDev.h
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_CLMEANSTDDEV_H__
 
 #include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLMeanStdDevKernel.h"
 #include "arm_compute/runtime/IFunction.h"
 
@@ -38,17 +39,18 @@ public:
     CLMeanStdDev();
     /** Initialise the kernel's inputs and outputs.
      *
-     * @param[in]  input  Input image. Data types supported: U8.
-     * @param[out] mean   Output average pixel value.
-     * @param[out] stddev (Optional)Output standard deviation of pixel values.
+     * @param[in, out] input  Input image. Data types supported: U8. (Written to only for border filling)
+     * @param[out]     mean   Output average pixel value.
+     * @param[out]     stddev (Optional)Output standard deviation of pixel values.
      */
-    void configure(const ICLImage *input, float *mean, float *stddev = nullptr);
+    void configure(ICLImage *input, float *mean, float *stddev = nullptr);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
     CLMeanStdDevKernel _mean_stddev_kernel; /**< Kernel that standard deviation calculation. */
+    CLFillBorderKernel _fill_border_kernel; /**< Kernel that fills the border with zeroes. */
     cl::Buffer         _global_sum;         /**< Variable that holds the global sum among calls in order to ease reduction */
     cl::Buffer         _global_sum_squared; /**< Variable that holds the global sum of squared values among calls in order to ease reduction */
 };
diff --git a/arm_compute/runtime/NEON/functions/NEMeanStdDev.h b/arm_compute/runtime/NEON/functions/NEMeanStdDev.h
index 3770b2a270..e8bf8dfa3d 100644
--- a/arm_compute/runtime/NEON/functions/NEMeanStdDev.h
+++ b/arm_compute/runtime/NEON/functions/NEMeanStdDev.h
@@ -24,6 +24,7 @@
 #ifndef __ARM_COMPUTE_NEMEANSTDDEV_H__
 #define __ARM_COMPUTE_NEMEANSTDDEV_H__
 
+#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
 #include "arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/IFunction.h"
@@ -44,17 +45,18 @@ public:
     NEMeanStdDev();
     /** Initialise the kernel's inputs and outputs.
      *
-     * @param[in]  input  Input image. Data type supported: U8.
-     * @param[out] mean   Output average pixel value.
-     * @param[out] stddev (Optional) Output standard deviation of pixel values.
+     * @param[in, out] input  Input image. Data types supported: U8. (Written to only for border filling)
+     * @param[out]     mean   Output average pixel value.
+     * @param[out]     stddev (Optional) Output standard deviation of pixel values.
      */
-    void configure(const IImage *input, float *mean, float *stddev = nullptr);
+    void configure(IImage *input, float *mean, float *stddev = nullptr);
 
     // Inherited methods overridden:
     void run() override;
 
 private:
     NEMeanStdDevKernel _mean_stddev_kernel; /**< Kernel that standard deviation calculation. */
+    NEFillBorderKernel _fill_border_kernel; /**< Kernel that fills tensor's borders with zeroes. */
     uint64_t           _global_sum;         /**< Variable that holds the global sum among calls in order to ease reduction */
     uint64_t           _global_sum_squared; /**< Variable that holds the global sum of squared values among calls in order to ease reduction */
 };
diff --git a/src/core/CL/kernels/CLMeanStdDevKernel.cpp b/src/core/CL/kernels/CLMeanStdDevKernel.cpp
index b0b748f466..1b70d7513f 100644
--- a/src/core/CL/kernels/CLMeanStdDevKernel.cpp
+++ b/src/core/CL/kernels/CLMeanStdDevKernel.cpp
@@ -40,10 +40,15 @@
 using namespace arm_compute;
 
 CLMeanStdDevKernel::CLMeanStdDevKernel()
-    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr)
+    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _border_size(0)
 {
 }
 
+BorderSize CLMeanStdDevKernel::border_size() const
+{
+    return _border_size;
+}
+
 void CLMeanStdDevKernel::configure(const ICLImage *input, float *mean, cl::Buffer *global_sum, float *stddev, cl::Buffer *global_sum_squared)
 {
     ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input);
@@ -83,6 +88,9 @@ void CLMeanStdDevKernel::configure(const ICLImage *input, float *mean, cl::Buffe
     constexpr unsigned int num_elems_processed_per_iteration_x = 8;
     const unsigned int     num_elems_processed_per_iteration_y = input->info()->dimension(1);
 
+    _border_size = BorderSize(std::max(static_cast<int>(num_elems_processed_per_iteration_x) - static_cast<int>(input->info()->dimension(0)),
+                                       static_cast<int>(input->info()->dimension(0) % num_elems_processed_per_iteration_x)));
+
     Window                win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
     AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
     update_window_and_padding(win, input_access);
diff --git a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
index 4616203d66..1eb7e45e36 100644
--- a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
+++ b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
@@ -85,10 +85,15 @@ std::pair<uint64x1_t, uint64x1_t> accumulate(const Window &window, Iterator &ite
 } // namespace
 
 NEMeanStdDevKernel::NEMeanStdDevKernel()
-    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx()
+    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx(), _border_size(0)
 {
 }
 
+BorderSize NEMeanStdDevKernel::border_size() const
+{
+    return _border_size;
+}
+
 void NEMeanStdDevKernel::configure(const IImage *input, float *mean, uint64_t *global_sum, float *stddev, uint64_t *global_sum_squared)
 {
     ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input);
@@ -105,6 +110,9 @@ void NEMeanStdDevKernel::configure(const IImage *input, float *mean, uint64_t *g
 
     constexpr unsigned int num_elems_processed_per_iteration = 16;
 
+    _border_size = BorderSize(std::max(static_cast<int>(num_elems_processed_per_iteration) - static_cast<int>(input->info()->dimension(0)),
+                                       static_cast<int>(input->info()->dimension(0) % num_elems_processed_per_iteration)));
+
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
 
diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
index 79fcba1dfb..176e3d688e 100644
--- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
+++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
@@ -224,7 +224,7 @@ void NELogits1DMaxKernel::configure(const ITensor *input, ITensor *output)
 
     _input       = input;
     _output      = output;
-    _border_size = BorderSize(0, input_width % num_elems_processed_per_iteration, 0, 0);
+    _border_size = BorderSize(0, num_elems_processed_per_iteration - (input_width % num_elems_processed_per_iteration), 0, 0);
 
     // Configure kernel window
     constexpr unsigned int num_elems_written_per_row = 1;
diff --git a/src/runtime/CL/functions/CLMeanStdDev.cpp b/src/runtime/CL/functions/CLMeanStdDev.cpp
index 56ba146790..838f7e73d2 100644
--- a/src/runtime/CL/functions/CLMeanStdDev.cpp
+++ b/src/runtime/CL/functions/CLMeanStdDev.cpp
@@ -23,19 +23,19 @@
  */
 #include "arm_compute/runtime/CL/functions/CLMeanStdDev.h"
 
-#include "arm_compute/core/CL/kernels/CLMeanStdDevKernel.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 
 using namespace arm_compute;
 
 CLMeanStdDev::CLMeanStdDev()
     : _mean_stddev_kernel(),
+      _fill_border_kernel(),
       _global_sum(),
       _global_sum_squared()
 {
 }
 
-void CLMeanStdDev::configure(const ICLImage *input, float *mean, float *stddev)
+void CLMeanStdDev::configure(ICLImage *input, float *mean, float *stddev)
 {
     _global_sum = cl::Buffer(CLScheduler::get().context(), CL_MEM_ALLOC_HOST_PTR | CL_MEM_READ_WRITE, sizeof(cl_ulong));
 
@@ -45,9 +45,11 @@ void CLMeanStdDev::configure(const ICLImage *input, float *mean, float *stddev)
     }
 
     _mean_stddev_kernel.configure(input, mean, &_global_sum, stddev, &_global_sum_squared);
+    _fill_border_kernel.configure(input, _mean_stddev_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<uint8_t>(0)));
 }
 
 void CLMeanStdDev::run()
 {
+    CLScheduler::get().enqueue(_fill_border_kernel);
     CLScheduler::get().enqueue(_mean_stddev_kernel);
 }
diff --git a/src/runtime/NEON/functions/NEMeanStdDev.cpp b/src/runtime/NEON/functions/NEMeanStdDev.cpp
index 47143f5e5b..ab8e72bf1d 100644
--- a/src/runtime/NEON/functions/NEMeanStdDev.cpp
+++ b/src/runtime/NEON/functions/NEMeanStdDev.cpp
@@ -23,19 +23,19 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEMeanStdDev.h"
 
-#include "arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
 using namespace arm_compute;
 
 NEMeanStdDev::NEMeanStdDev()
-    : _mean_stddev_kernel(), _global_sum(0), _global_sum_squared(0)
+    : _mean_stddev_kernel(), _fill_border_kernel(), _global_sum(0), _global_sum_squared(0)
 {
 }
 
-void NEMeanStdDev::configure(const IImage *input, float *mean, float *stddev)
+void NEMeanStdDev::configure(IImage *input, float *mean, float *stddev)
 {
     _mean_stddev_kernel.configure(input, mean, &_global_sum, stddev, &_global_sum_squared);
+    _fill_border_kernel.configure(input, _mean_stddev_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<uint8_t>(0)));
 }
 
 void NEMeanStdDev::run()
@@ -43,5 +43,6 @@ void NEMeanStdDev::run()
     _global_sum         = 0;
     _global_sum_squared = 0;
 
+    _fill_border_kernel.run(_fill_border_kernel.window());
     NEScheduler::get().schedule(&_mean_stddev_kernel, Window::DimY);
 }
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();