Compute Library
 21.02
WindowIterator.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_WINDOW_ITERATOR_H
25 #define ARM_COMPUTE_WINDOW_ITERATOR_H
27 #include "arm_compute/core/Error.h"
30 
31 //FIXME: Delete the "ARM_COMPUTE_PRINTF" before the release. In the meantime it's probably going to be useful to debug
32 //#define ARM_COMPUTE_PRINTF printf
33 #define ARM_COMPUTE_PRINTF(...)
34 
35 namespace arm_compute
36 {
37 /** Convert an offset in window steps into absolute coordinates.
38  *
39  * @param[in] w Window @p offset is related to.
40  * @param[in] offset Offset inside the window expressed in number of window steps.
41  *
42  * @return Absolute coordinates.
43  */
45 {
46  Coordinates position;
47  for(unsigned int i = 0; i < Coordinates::num_max_dimensions; ++i)
48  {
49  position.set(i, w[i].start() + offset[i] * w[i].step());
50  }
51  return position;
52 }
53 
54 /** Tensor accessors to make it easier to interface with arm_gemm */
55 template <typename T>
57 {
58 public:
59  /** Constructor:
60  *
61  * @param[in] tensor Source tensor, must be allocated.
62  */
63  TensorAccessor(const ITensor &tensor)
64  : _first(tensor.ptr_to_element(Coordinates())), _strides(tensor.info()->strides_in_bytes())
65  {
66  }
67  /** Get the stride of the dimension dim expressed in number of Ts.
68  *
69  * @param[in] dim Dimension of the wanted stride.
70  *
71  * @return Stride in number of Ts.
72  */
73  inline size_t stride(size_t dim) const
74  {
75  ARM_COMPUTE_ERROR_ON(_strides[dim] % sizeof(T) != 0);
76  return _strides[dim] / sizeof(T);
77  }
78 
79  /** Manually set the stride of a dimension
80  *
81  * @param[in] dim Dimension of the stride to set.
82  * @param[in] size Value to set the stride to (in bytes).
83  */
84  void set_stride(size_t dim, size_t size)
85  {
86  _strides[dim] = size;
87  }
88 
89  /** Manually set the strides
90  *
91  * @param[in] strides Strides to set
92  */
93  void set_strides(const Strides &strides)
94  {
95  _strides = strides;
96  }
97 
98  /** Returns a pointer to the element at coordinates (x,y,z,w)
99  *
100  * @param[in] x X coordinates
101  * @param[in] y (optional) Y coordinates
102  * @param[in] z (optional) Z coordinates
103  * @param[in] w (optional) W coordinates
104  */
105  inline T *get_ptr(unsigned int x, unsigned int y = 0, unsigned int z = 0, unsigned int w = 0)
106  {
107  return reinterpret_cast<T *>(_first + x * _strides[0] + y * _strides[1] + z * _strides[2] + w * _strides[3]);
108  }
109 
110  /** Returns a pointer to the element at coordinates (x,y,z,w)
111  *
112  * @param[in] x X coordinates
113  * @param[in] y (optional) Y coordinates
114  * @param[in] z (optional) Z coordinates
115  * @param[in] w (optional) W coordinates
116  */
117  inline T *operator()(unsigned int x, unsigned int y = 0, unsigned int z = 0, unsigned int w = 0)
118  {
119  return get_ptr(x, y, z, w);
120  }
121 
122  /** Returns a pointer to the first element of the tensor
123  *
124  * @return Pointer to the first element.
125  */
126  inline T *first_element()
127  {
128  return reinterpret_cast<T *>(_first);
129  }
130 
131  /** Returns a pointer to the first element of the tensor
132  *
133  * @return Pointer to the first element.
134  */
135  inline T *operator()()
136  {
137  return first_element();
138  }
139 
140 private:
141  uint8_t *_first; /**< Pointer to the first element of the tensor.*/
142  Strides _strides; /**< Strides in bytes of the tensor */
143 };
144 
145 /** Iterate over a portion of a Window */
146 template <typename L>
148 {
149 public:
150  /** Construct a WindowIterator object
151  *
152  * @param[in] w Window to use for the iteration
153  * @param[in] start Where to start iterating from (In Window coordinates)
154  * @param[in] end Where to stop iterating (In Window coordinates).
155  * @param[in] lambda_function Lambda function to call for every iteration between start and end. (It will be called last for end - 1)
156  */
157  WindowIterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
158  : _lambda_function(std::move(lambda_function)),
159  _position(convert_window_coord_to_position(w, start)),
160  _end(convert_window_coord_to_position(w, end)),
161  _w(w)
162  {
163  }
164  /** Iterate over the lowest 3 dimensions of the window.
165  *
166  * @param[in] on_new_row_size Callback to be called before lambda_function every time the width of the row processed changes.
167  */
168  template <typename M>
169  void iterate_3D(M &&on_new_row_size)
170  {
171  while(_end.z() != _position.z())
172  {
173  ARM_COMPUTE_PRINTF("New slice %d\n", _position.z());
174  iterate_2D_internal(on_new_row_size, _w.x().end() - _w.x().step(), _w.y().end() - _w.y().step());
175  _position[2] += _w.z().step();
176  _position[1] = _w.y().start();
177  _position[0] = _w.x().start();
178  }
179  // Left over:
180  ARM_COMPUTE_PRINTF("Left over slice\n");
181  iterate_2D(on_new_row_size);
182  }
183 
184  /** Iterate over the lowest 2 dimensions of the window.
185  *
186  * @param[in] on_new_row_size Callback to be called before lambda_function every time the width of the row processed changes.
187  */
188  template <typename M>
189  void iterate_2D(M &&on_new_row_size)
190  {
191  iterate_2D_internal(on_new_row_size, _end.x(), _end.y());
192  }
193 
194  /** Change the step used for the iteration.
195  *
196  * @note Does not affect the start and end points.
197  *
198  * @param[in] dim Dimension to change
199  * @param[in] step New step to use for the given dimension.
200  */
201  inline void set_step(size_t dim, int step)
202  {
203  _w.set_dimension_step(dim, step);
204  }
205 
206  /** Returns the coordinates in absolute coordinates of the end position
207  *
208  * @return End position coordinates.
209  */
210  const Coordinates &end_position() const
211  {
212  return _end;
213  }
214 
215 private:
216  template <typename M>
217  void iterate_2D_internal(M &&on_new_row_size, int end_x, int end_y)
218  {
219  //Is there more than one row to process ?
220  if(end_y == _position.y())
221  {
222  // Single row:
223  ARM_COMPUTE_PRINTF("Partial row only\n");
224  // Both start and end belong to the same row:
225  iterate_over_dim0(end_x + _w.x().step(), on_new_row_size);
226  }
227  else
228  {
229  // Do we start from the beginning of the row ?
230  if(_w.x().start() != _position.x())
231  {
232  //Start in the middle of a row: process left-over X
233  ARM_COMPUTE_PRINTF("Partial row first\n");
234  iterate_over_dim0(_w.x().end(), on_new_row_size);
235  _position[1] += _w.y().step();
236  }
237 
238  //Middle rows
239  bool no_leftover = end_x + _w.x().step() == _w.x().end();
240  if(no_leftover)
241  {
242  ARM_COMPUTE_PRINTF("no left over\n");
243  //Switch to full row size:
244  on_new_row_size(_w[0].start(), _w.x().end());
245  // Shouldn't be possible to reach that point and not have at least one entire row to process
246  ARM_COMPUTE_ERROR_ON(_w.y().end() == _position.y());
247  // No leftover: all the rows lefts to process are full width:
248  iterate_over_dim1(end_y + _w.y().step());
249  }
250  else
251  {
252  ARM_COMPUTE_PRINTF("with left over\n");
253  // Are there full rows to process ?
254  if(_position[1] != end_y)
255  {
256  ARM_COMPUTE_PRINTF("full rows\n");
257  //Switch to full row size:
258  on_new_row_size(_w[0].start(), _w.x().end());
259  iterate_over_dim1(end_y);
260  }
261 
262  ARM_COMPUTE_PRINTF("Final leftover\n");
263  //Leftover end x
264  _position[0] = _w.x().start();
265  iterate_over_dim0(end_x + _w.x().step(), on_new_row_size);
266  }
267  }
268  }
269 
270  /** Process full rows below 'end'
271  *
272  * @param[in] end Y position to stop at.
273  */
274  void iterate_over_dim1(int end)
275  {
276  for(; _position[1] != end; _position[1] += _w[1].step())
277  {
278  _position[0] = _w[0].start();
279  iterate_over_dim0(_w[0].end());
280  }
281  }
282 
283  /** Process elements of a given row up to 'end'
284  *
285  * @param[in] end X position to stop at.
286  * @param[in] on_new_row_size Callback to call before starting iterating
287  */
288  template <typename M>
289  void iterate_over_dim0(int end, M &&on_new_row_size)
290  {
291  on_new_row_size(_position.x(), end);
292  iterate_over_dim0(end);
293  }
294 
295  /** Process elements of a given row up to 'end'
296  *
297  * @param[in] end X position to stop at.
298  */
299  void iterate_over_dim0(int end)
300  {
301  ARM_COMPUTE_PRINTF("X [%d, %d, %d]\n", _position.x(), end, _w[0].step());
302  // Both start and end belong to the same row:
303  ARM_COMPUTE_ERROR_ON(_position[0] > end);
304  for(; _position.x() < end; _position[0] += _w[0].step())
305  {
306  _lambda_function(_position);
307  }
308  }
309 
310  L _lambda_function; /**< Function to call for each iteration */
311  Coordinates _position; /**< Absolute coordinates of the current position */
312  Coordinates _end; /**< Absolute coordinates of the point after the last iteration */
313  Window _w; /**< Window to iterate over */
314 };
315 
316 /** Create a WindowIterator object
317  *
318  * @param[in] w Window to use for the iteration
319  * @param[in] start Where to start iterating from (In Window coordinates)
320  * @param[in] end Where to stop iterating (In Window coordinates).
321  * @param[in] lambda_function Lambda function to call for every iteration between start and end. (It will be called last for end - 1)
322  *
323  * @return A WindowIterator object.
324  */
325 template <typename L>
326 WindowIterator<L> create_window_iterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
327 {
328  return WindowIterator<L>(w, start, end, std::move(lambda_function));
329 }
330 }
331 #endif /*ARM_COMPUTE_WINDOW_ITERATOR_H*/
unsigned int M
void set(size_t dimension, T value, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: Dimensions.h:76
__global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
Definition: helpers.h:846
SimpleTensor< float > w
Definition: DFT.cpp:156
size_t stride(size_t dim) const
Get the stride of the dimension dim expressed in number of Ts.
Iterate over a portion of a Window.
WindowIterator< L > create_window_iterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
Create a WindowIterator object.
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
T * operator()(unsigned int x, unsigned int y=0, unsigned int z=0, unsigned int w=0)
Returns a pointer to the element at coordinates (x,y,z,w)
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
void set_strides(const Strides &strides)
Manually set the strides.
WindowIterator(const Window &w, const Coordinates &start, const Coordinates &end, L &&lambda_function)
Construct a WindowIterator object.
Coordinates of an item.
Definition: Coordinates.h:37
Coordinates convert_window_coord_to_position(const Window &w, const Coordinates &offset)
Convert an offset in window steps into absolute coordinates.
void end(TokenStream &in, bool &valid)
Definition: MLGOParser.cpp:290
void iterate_2D(M &&on_new_row_size)
Iterate over the lowest 2 dimensions of the window.
const Coordinates & end_position() const
Returns the coordinates in absolute coordinates of the end position.
#define ARM_COMPUTE_PRINTF(...)
Strides of an item in bytes.
Definition: Strides.h:37
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
constexpr int step
Definition: fp32.cpp:35
void set_stride(size_t dim, size_t size)
Manually set the stride of a dimension.
T * operator()()
Returns a pointer to the first element of the tensor.
void set_step(size_t dim, int step)
Change the step used for the iteration.
TensorAccessor(const ITensor &tensor)
Constructor:
T * get_ptr(unsigned int x, unsigned int y=0, unsigned int z=0, unsigned int w=0)
Returns a pointer to the element at coordinates (x,y,z,w)
T * first_element()
Returns a pointer to the first element of the tensor.
static constexpr size_t num_max_dimensions
Number of dimensions the tensor has.
Definition: Dimensions.h:46
Tensor accessors to make it easier to interface with arm_gemm.
Describe a multidimensional execution window.
Definition: Window.h:39
void iterate_3D(M &&on_new_row_size)
Iterate over the lowest 3 dimensions of the window.