From 671a11e1c8e1e4db7bcae9ce97b0c97ebcb97464 Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Fri, 6 Jul 2018 15:11:36 +0100 Subject: 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 Reviewed-by: Georgios Pinitas --- arm_compute/core/Window.h | 8 +- arm_compute/core/Window.inl | 20 +++ arm_compute/core/WindowIterator.h | 292 ++++++++++++++++++++++++++++++++++++++ 3 files changed, 319 insertions(+), 1 deletion(-) create mode 100644 arm_compute/core/WindowIterator.h (limited to 'arm_compute/core') 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:
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(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 +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(_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 +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 + 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 + 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 + 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 + 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 +WindowIterator create_window_iterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function) +{ + return WindowIterator(w, start, end, std::move(lambda_function)); +} +} +#endif /*__ARM_COMPUTE_WINDOW_ITERATOR_H__*/ -- cgit v1.2.1