COMPMID-1379: Created WindowIterator and TensorAccessor

- WindowIterator: used to iterate over arbitrary positions of a window. (More flexible than execute_window_loop which only can iterate over entire dimensions)
- TensorAccessor: RSH's code uses pointers to specialised types and strides in element sizes, this helps interfacing with their code.

Change-Id: I8ded8758d345668804873409f949b8cec694d289
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139082
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

6 files changed, 514 insertions, 5 deletions

diff --git a/arm_compute/core/Window.h b/arm_compute/core/Window.h
index 6f172ecebf..73c8d4385b 100644
--- a/arm_compute/core/Window.h
+++ b/arm_compute/core/Window.h
@@ -224,7 +224,13 @@ public:
      * @return The number of iterations
      */
     constexpr size_t num_iterations(size_t dimension) const;
-
+    /** Return the total number of iterations needed to iterate through the entire window
+     *
+     * @return Number of total iterations
+     */
+    size_t num_iterations_total() const;
+    /** Return the shape of the window in number of steps */
+    TensorShape shape() const;
     /** Split a window into a set of sub windows along a given dimension
      *
      * For example to split a window into 3 sub-windows along the Y axis, you would have to do:<br/>
diff --git a/arm_compute/core/Window.inl b/arm_compute/core/Window.inl
index 8401227eee..c6fc8848aa 100644
--- a/arm_compute/core/Window.inl
+++ b/arm_compute/core/Window.inl
@@ -247,4 +247,24 @@ inline void Window::use_tensor_dimensions(const TensorShape &shape, size_t first
         set(n, Window::Dimension(0, std::max(shape[n], static_cast<size_t>(1))));
     }
 }
+
+inline TensorShape Window::shape() const
+{
+    TensorShape shape;
+    for(size_t d = 0; d < TensorShape::num_max_dimensions; ++d)
+    {
+        shape.set(d, (_dims[d].end() - _dims[d].start()) / _dims[d].step());
+    }
+    return shape;
+}
+
+inline size_t Window::num_iterations_total() const
+{
+    size_t total = 1;
+    for(size_t d = 0; d < Coordinates::num_max_dimensions; ++d)
+    {
+        total *= num_iterations(d);
+    }
+    return total;
+}
 }
diff --git a/arm_compute/core/WindowIterator.h b/arm_compute/core/WindowIterator.h
new file mode 100644
index 0000000000..13e9973506
--- /dev/null
+++ b/arm_compute/core/WindowIterator.h
@@ -0,0 +1,292 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_WINDOW_ITERATOR_H__
+#define __ARM_COMPUTE_WINDOW_ITERATOR_H__
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Window.h"
+
+//FIXME: Delete the "PRINTF" before the release. In the meantime it's probably going to be useful to debug
+//#define PRINTF printf
+#define PRINTF(...)
+
+namespace arm_compute
+{
+/** Convert an offset in window steps into absolute coordinates.
+ *
+ * @param[in] w      Window @p offset is related to.
+ * @param[in] offset Offset inside the window expressed in number of window steps.
+ *
+ * @return Absolute coordinates.
+ */
+inline Coordinates convert_window_coord_to_position(const Window &w, const Coordinates &offset)
+{
+    Coordinates position;
+    for(unsigned int i = 0; i < Coordinates::num_max_dimensions; ++i)
+    {
+        position.set(i, w[i].start() + offset[i] * w[i].step());
+    }
+    return position;
+}
+
+/** Tensor accessors to make it easier to interface with arm_gemm */
+template <typename T>
+class TensorAccessor
+{
+public:
+    /** Constructor:
+     *
+     * @param[in] tensor Source tensor, must be allocated.
+     */
+    TensorAccessor(const ITensor &tensor)
+        : _first(tensor.ptr_to_element(Coordinates())), _strides(tensor.info()->strides_in_bytes())
+    {
+    }
+    /** Get the stride of the dimension dim expressed in number of Ts.
+     *
+     * @param[in] dim Dimension of the wanted stride.
+     *
+     * @return Stride in number of Ts.
+     */
+    inline size_t stride(size_t dim) const
+    {
+        return _strides[dim] / sizeof(T);
+    }
+
+    /** Returns a pointer to the element at coordinates (x,y,z,w)
+     *
+     * @param[in] x X coordinates
+     * @param[in] y (optional) Y coordinates
+     * @param[in] z (optional) Z coordinates
+     * @param[in] w (optional) W coordinates
+     */
+    inline T *get_ptr(unsigned int x, unsigned int y = 0, unsigned int z = 0, unsigned int w = 0)
+    {
+        return reinterpret_cast<T *>(_first + x * _strides[0] + y * _strides[1] + z * _strides[2] + w * _strides[3]);
+    }
+
+    /** Returns a pointer to the element at coordinates (x,y,z,w)
+     *
+     * @param[in] x X coordinates
+     * @param[in] y (optional) Y coordinates
+     * @param[in] z (optional) Z coordinates
+     * @param[in] w (optional) W coordinates
+     */
+    inline T *operator()(unsigned int x, unsigned int y = 0, unsigned int z = 0, unsigned int w = 0)
+    {
+        return get_ptr(x, y, z, w);
+    }
+
+private:
+    uint8_t       *_first;   /**< Pointer to the first element of the tensor.*/
+    const Strides &_strides; /**< Strides in bytes of the tensor */
+};
+
+/** Iterate over a portion of a Window */
+template <typename L>
+class WindowIterator
+{
+public:
+    /** Construct a WindowIterator object
+     *
+     * @param[in] w               Window to use for the iteration
+     * @param[in] start           Where to start iterating from (In Window coordinates)
+     * @param[in] end             Where to stop iterating (In Window coordinates).
+     * @param[in] lambda_function Lambda function to call for every iteration between start and end. (It will be called last for end - 1)
+     */
+    WindowIterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
+        : _lambda_function(std::move(lambda_function)),
+          _position(convert_window_coord_to_position(w, start)),
+          _end(convert_window_coord_to_position(w, end)),
+          _w(w)
+    {
+    }
+    /** Iterate over the lowest 3 dimensions of the window.
+     *
+     * @param[in] on_new_row_size Callback to be called before lambda_function every time the width of the row processed changes.
+     */
+    template <typename M>
+    void iterate_3D(M &&on_new_row_size)
+    {
+        while(_end.z() != _position.z())
+        {
+            PRINTF("New slice %d\n", _position.z());
+            iterate_2D_internal(on_new_row_size, _w.x().end() - _w.x().step(), _w.y().end() - _w.y().step());
+            _position[2] += _w.z().step();
+            _position[1] = _w.y().start();
+            _position[0] = _w.x().start();
+        }
+        // Left over:
+        PRINTF("Left over slice\n");
+        iterate_2D(on_new_row_size);
+    }
+
+    /** Iterate over the lowest 2 dimensions of the window.
+     *
+     * @param[in] on_new_row_size Callback to be called before lambda_function every time the width of the row processed changes.
+     */
+    template <typename M>
+    void iterate_2D(M &&on_new_row_size)
+    {
+        iterate_2D_internal(on_new_row_size, _end.x(), _end.y());
+    }
+
+    /** Change the step used for the iteration.
+     *
+     * @note Does not affect the start and end points.
+     *
+     * @param[in] dim  Dimension to change
+     * @param[in] step New step to use for the given dimension.
+     */
+    inline void set_step(size_t dim, int step)
+    {
+        _w.set_dimension_step(dim, step);
+    }
+
+    /** Returns the coordinates in absolute coordinates of the end position
+         *
+         * @return End position coordinates.
+         */
+    const Coordinates &end_position() const
+    {
+        return _end;
+    }
+
+private:
+    template <typename M>
+    void iterate_2D_internal(M &&on_new_row_size, int end_x, int end_y)
+    {
+        //Is there more than one row to process ?
+        if(end_y == _position.y())
+        {
+            // Single row:
+            PRINTF("Partial row only\n");
+            // Both start and end belong to the same row:
+            iterate_over_dim0(end_x + _w.x().step(), on_new_row_size);
+        }
+        else
+        {
+            // Do we start from the beginning of the row ?
+            if(_w.x().start() != _position.x())
+            {
+                //Start in the middle of a row: process left-over X
+                PRINTF("Partial row first\n");
+                iterate_over_dim0(_w.x().end(), on_new_row_size);
+                _position[1] += _w.y().step();
+            }
+
+            //Middle rows
+            bool no_leftover = end_x + _w.x().step() == _w.x().end();
+            if(no_leftover)
+            {
+                PRINTF("no left over\n");
+                //Switch to full row size:
+                on_new_row_size(_w[0].start(), _w.x().end());
+                // Shouldn't be possible to reach that point and not have at least one entire row to process
+                ARM_COMPUTE_ERROR_ON(_w.y().end() == _position.y());
+                // No leftover: all the rows lefts to process are full width:
+                iterate_over_dim1(_w.y().end());
+            }
+            else
+            {
+                PRINTF("with left over\n");
+                // Are there full rows to process ?
+                if(_position[1] != end_y)
+                {
+                    PRINTF("full rows\n");
+                    //Switch to full row size:
+                    on_new_row_size(_w[0].start(), _w.x().end());
+                    iterate_over_dim1(_w.y().end() - _w.y().step());
+                }
+
+                PRINTF("Final leftover\n");
+                //Leftover end x
+                _position[0] = _w.x().start();
+                iterate_over_dim0(end_x + _w.x().step(), on_new_row_size);
+            }
+        }
+    }
+
+    /** Process full rows below 'end'
+     *
+     * @param[in] end Y position to stop at.
+     */
+    void iterate_over_dim1(int end)
+    {
+        for(; _position[1] != end; _position[1] += _w[1].step())
+        {
+            _position[0] = _w[0].start();
+            iterate_over_dim0(_w[0].end());
+        }
+    }
+
+    /** Process elements of a given row up to 'end'
+     *
+     * @param[in] end             X position to stop at.
+     * @param[in] on_new_row_size Callback to call before starting iterating
+     */
+    template <typename M>
+    void iterate_over_dim0(int end, M &&on_new_row_size)
+    {
+        on_new_row_size(_position.x(), end);
+        iterate_over_dim0(end);
+    }
+
+    /** Process elements of a given row up to 'end'
+     *
+     * @param[in] end X position to stop at.
+     */
+    void iterate_over_dim0(int end)
+    {
+        PRINTF("X [%d, %d, %d]\n", _position.x(), end, _w[0].step());
+        // Both start and end belong to the same row:
+        ARM_COMPUTE_ERROR_ON(_position[0] > end);
+        for(; _position.x() < end; _position[0] += _w[0].step())
+        {
+            _lambda_function(_position);
+        }
+    }
+
+    L           _lambda_function; /**< Function to call for each iteration */
+    Coordinates _position;        /**< Absolute coordinates of the current position */
+    Coordinates _end;             /**< Absolute coordinates of the point after the last iteration */
+    Window      _w;               /**< Window to iterate over */
+};
+
+/** Create a WindowIterator object
+ *
+ * @param[in] w               Window to use for the iteration
+ * @param[in] start           Where to start iterating from (In Window coordinates)
+ * @param[in] end             Where to stop iterating (In Window coordinates).
+ * @param[in] lambda_function Lambda function to call for every iteration between start and end. (It will be called last for end - 1)
+ *
+ * @return A WindowIterator object.
+ */
+template <typename L>
+WindowIterator<L> create_window_iterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
+{
+    return WindowIterator<L>(w, start, end, std::move(lambda_function));
+}
+}
+#endif /*__ARM_COMPUTE_WINDOW_ITERATOR_H__*/
diff --git a/tests/framework/datasets/InitializerListDataset.h b/tests/framework/datasets/InitializerListDataset.h
index f90e0b747a..ec1550df4d 100644
--- a/tests/framework/datasets/InitializerListDataset.h
+++ b/tests/framework/datasets/InitializerListDataset.h
@@ -25,7 +25,7 @@
 #define ARM_COMPUTE_TEST_DATASET_LIST
 
 #include "Dataset.h"
-#include "support/ToolchainSupport.h"
+#include "utils/TypePrinter.h"
 
 #include <initializer_list>
 #include <string>
@@ -84,8 +84,7 @@ public:
          */
         std::string description() const
         {
-            using support::cpp11::to_string;
-            return _name + "=" + to_string(*_iterator);
+            return _name + "=" + arm_compute::to_string(*_iterator);
         }
 
         /** Get the current value.
diff --git a/tests/validation/UNIT/WindowIterator.cpp b/tests/validation/UNIT/WindowIterator.cpp
new file mode 100644
index 0000000000..0c1ab12fe7
--- /dev/null
+++ b/tests/validation/UNIT/WindowIterator.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/WindowIterator.h"
+#include "tests/Utils.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "utils/TypePrinter.h"
+
+#include <stdexcept>
+
+using namespace arm_compute;
+using namespace arm_compute::test;
+using namespace arm_compute::test::validation;
+
+TEST_SUITE(UNIT)
+TEST_SUITE(WindowIterator)
+
+template <typename Dim, typename... Dims>
+Window create_window(Dim &&dim0, Dims &&... dims)
+{
+    Window win;
+    const std::array < Dim, 1 + sizeof...(Dims) > dimensions{ { dim0, std::forward<Dims>(dims)... } };
+    for(size_t i = 0; i < dimensions.size(); i++)
+    {
+        win.set(i, dimensions[i]);
+    }
+    return win;
+}
+
+template <typename T>
+std::vector<T> create_vector(std::initializer_list<T> list_objs)
+{
+    std::vector<T> vec_objs;
+    for(auto it : list_objs)
+    {
+        vec_objs.push_back(it);
+    }
+    return vec_objs;
+}
+
+DATA_TEST_CASE(WholeWindow, framework::DatasetMode::ALL, zip(framework::dataset::make("Window", { create_window(Window::Dimension(0, 1)),
+                                                                                                  create_window(Window::Dimension(1, 5, 2), Window::Dimension(3, 5)),
+                                                                                                  create_window(Window::Dimension(4, 16, 4), Window::Dimension(3, 13, 5), Window::Dimension(1, 3, 2))
+                                                                                                }),
+                                                             framework::dataset::make("Expected", { create_vector({ Coordinates(0, 0) }),
+                                                                                                    create_vector({ Coordinates(1, 3), Coordinates(3, 3), Coordinates(1, 4), Coordinates(3, 4) }),
+                                                                                                    create_vector({ Coordinates(4, 3, 1), Coordinates(8, 3, 1), Coordinates(12, 3, 1), Coordinates(4, 8, 1), Coordinates(8, 8, 1), Coordinates(12, 8, 1) })
+                                                                                                  })),
+               window, expected)
+{
+    unsigned int i            = 0;
+    int          row_size     = 0;
+    TensorShape  window_shape = window.shape();
+    Coordinates  start_offset = index2coords(window_shape, 0);
+    Coordinates  end_offset   = index2coords(window_shape, window.num_iterations_total() - 1);
+    auto window_iterator      = create_window_iterator(window, start_offset, end_offset, [&](const Coordinates & id)
+    {
+        ARM_COMPUTE_EXPECT_EQUAL(row_size, (window[0].end() - window[0].start()), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_ASSERT(i < expected.size());
+        Coordinates expected_coords = expected[i++];
+        //Set number of dimensions to the maximum (To match the number of dimensions used by the id passed to the lambda function)
+        expected_coords.set_num_dimensions(Coordinates::num_max_dimensions);
+        ARM_COMPUTE_EXPECT_EQUAL(id, expected_coords, framework::LogLevel::ERRORS);
+    });
+    window_iterator.iterate_3D([&](int start, int end)
+    {
+        ARM_COMPUTE_EXPECT_EQUAL(window[0].start(), start, framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT_EQUAL(window[0].end(), end, framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(end > start, framework::LogLevel::ERRORS);
+        row_size = end - start;
+    });
+    ARM_COMPUTE_EXPECT_EQUAL(i, expected.size(), framework::LogLevel::ERRORS);
+}
+
+DATA_TEST_CASE(PartialWindow2D, framework::DatasetMode::ALL, zip(zip(zip(combine(framework::dataset::make("Window",
+                                                                                                          create_window(Window::Dimension(4, 20, 4), Window::Dimension(3, 18, 5), Window::Dimension(1, 2, 1))),
+                                                                                 framework::dataset::make("Start", { 0, 1, 3, 2, 4 })),
+                                                                         framework::dataset::make("End", { 0, 2, 5, 8, 7 })),
+                                                                     framework::dataset::make("RowSize",
+{
+    create_vector({ 4 }),
+    create_vector({ 8, 8 }),
+    create_vector({ 4, 8, 8 }),
+    create_vector({ 8, 8, 16, 16, 16, 16, 4 }),
+    create_vector({ 16, 16, 16, 16 }),
+})),
+framework::dataset::make("Expected", { create_vector({ Coordinates(4, 3, 1) }), create_vector({ Coordinates(8, 3, 1), Coordinates(12, 3, 1) }), create_vector({ Coordinates(16, 3, 1), Coordinates(4, 8, 1), Coordinates(8, 8, 1) }), create_vector({ Coordinates(12, 3, 1), Coordinates(16, 3, 1), Coordinates(4, 8, 1), Coordinates(8, 8, 1), Coordinates(12, 8, 1), Coordinates(16, 8, 1), Coordinates(4, 13, 1) }), create_vector({ Coordinates(4, 8, 1), Coordinates(8, 8, 1), Coordinates(12, 8, 1), Coordinates(16, 8, 1) }) })),
+window, start, end, expected_row_size, expected)
+{
+    unsigned int i            = 0;
+    int          row_size     = 0;
+    TensorShape  window_shape = window.shape();
+    Coordinates  start_offset = index2coords(window_shape, start);
+    Coordinates  end_offset   = index2coords(window_shape, end);
+    auto window_iterator      = create_window_iterator(window, start_offset, end_offset, [&](const Coordinates & id)
+    {
+        ARM_COMPUTE_ASSERT(i < expected.size());
+        ARM_COMPUTE_EXPECT_EQUAL(expected_row_size[i], row_size, framework::LogLevel::ERRORS);
+        Coordinates expected_coords = expected[i++];
+        //Set number of dimensions to the maximum (To match the number of dimensions used by the id passed to the lambda function)
+        expected_coords.set_num_dimensions(Coordinates::num_max_dimensions);
+        ARM_COMPUTE_EXPECT_EQUAL(id, expected_coords, framework::LogLevel::ERRORS);
+    });
+    window_iterator.iterate_3D([&](int start, int end)
+    {
+        ARM_COMPUTE_EXPECT(start >= window[0].start(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(end <= window[0].end(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(end > start, framework::LogLevel::ERRORS);
+        row_size = end - start;
+    });
+    ARM_COMPUTE_EXPECT_EQUAL(i, expected.size(), framework::LogLevel::ERRORS);
+}
+
+TEST_SUITE_END()
+TEST_SUITE_END()
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index 49c07938bd..9bd8c2a667 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -60,7 +60,6 @@ std::string to_string_if_not_null(T *arg)
         return to_string(*arg);
     }
 }
-
 /** Formatted output of the Dimensions type.
  *
  * @param[out] os         Output stream.
@@ -1627,6 +1626,48 @@ inline std::string to_string(const Termination &termination)
     return str.str();
 }
 
+/** Formatted output of a vector of objects.
+ *
+ * @param[out] os   Output stream
+ * @param[in]  args Vector of objects to print
+ *
+ * @return Modified output stream.
+ */
+template <typename T>
+inline ::std::ostream &operator<<(::std::ostream &os, const std::vector<T> &args)
+{
+    os << "[";
+    bool first = true;
+    for(auto &arg : args)
+    {
+        if(first)
+        {
+            first = false;
+        }
+        else
+        {
+            os << ", ";
+        }
+        os << arg;
+    }
+    os << "]";
+    return os;
+}
+
+/** Formatted output of a vector of objects.
+ *
+ * @param[in] args Vector of objects to print
+ *
+ * @return String representing args.
+ */
+template <typename T>
+std::string to_string(const std::vector<T> &args)
+{
+    std::stringstream str;
+    str << args;
+    return str.str();
+}
+
 /** Formatted output of the WinogradInfo type. */
 inline ::std::ostream &operator<<(::std::ostream &os, const WinogradInfo &info)
 {
@@ -1644,5 +1685,18 @@ inline std::string to_string(const WinogradInfo &type)
     str << type;
     return str.str();
 }
+
+/** Fallback method: try to use std::to_string:
+ *
+ * @param[in] val Value to convert to string
+ *
+ * @return String representing val.
+ */
+template <typename T>
+inline std::string to_string(const T &val)
+{
+    return support::cpp11::to_string(val);
+}
+
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_TEST_TYPE_PRINTER_H__ */