Compute Library
 21.02
NEScaleKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2021 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  */
25 
30 #include "src/core/CPP/Validate.h"
39 #include "support/Rounding.h"
40 #include <arm_neon.h>
41 #include <map>
42 
43 namespace arm_compute
44 {
45 namespace
46 {
47 struct ScaleSelectorData
48 {
50 };
52 using ScaleKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *,
53  InterpolationPolicy, BorderMode, PixelValue, float, bool, const Window &)>::type;
54 struct ScaleKernel
55 {
56  const char *name;
57  const ScaleSelectorPtr is_selected;
58  ScaleKernelPtr ukernel;
59 };
60 
61 static const ScaleKernel available_kernels[] =
62 {
63 #if defined(__ARM_FEATURE_SVE)
64  {
65  "fp16_sve_scale",
66  [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
68  },
69  {
70  "f32_sve_scale",
71  [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
73  },
74  {
75  "qasymm8_sve_scale",
76  [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
78  },
79  {
80  "qasymm8_signed_sve_scale",
81  [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
83  },
84  {
85  "u8_sve_scale",
86  [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
88  },
89  {
90  "s16_sve_scale",
91  [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
93  },
94 #else /* !defined(__ARM_FEATURE_SVE) */
95 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
96  {
97  "common_neon_scale",
98  [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
99  REGISTER_FP16_NEON(arm_compute::cpu::common_neon_scale<float16_t>)
100  },
101 #endif /* !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
102  {
103  "common_neon_scale",
104  [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
105  REGISTER_FP32_NEON(arm_compute::cpu::common_neon_scale<float>)
106  },
107  {
108  "qasymm8_neon_scale",
109  [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
111  },
112  {
113  "qasymm8_signed_neon_scale",
114  [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
116  },
117  {
118  "common_neon_scale",
119  [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
120  REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale<uint8_t>)
121  },
122  {
123  "common_neon_scale",
124  [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
125  REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale<int16_t>)
126  },
127 #endif /* !defined(__ARM_FEATURE_SVE) */
128 };
129 
130 /** Micro-kernel selector
131  *
132  * @param[in] data Selection data passed to help pick the appropriate micro-kernel
133  *
134  * @return A matching micro-kernel else nullptr
135  */
136 const ScaleKernel *get_implementation(const ScaleSelectorData &data)
137 {
138  for(const auto &uk : available_kernels)
139  {
140  if(uk.is_selected(data))
141  {
142  return &uk;
143  }
144  }
145  return nullptr;
146 }
147 
148 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
149  const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
150 {
151  const auto *uk = get_implementation(ScaleSelectorData{ input->data_type() });
152  ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
153 
156  ARM_COMPUTE_RETURN_ERROR_ON(output == input);
157  ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
158  ARM_COMPUTE_UNUSED(info.constant_border_value);
159  ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.use_padding, "Padding is not supported");
160 
161  const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : info.data_layout;
162  const auto width_index = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
163  const auto height_index = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
164  const auto output_width = output->dimension(width_index);
165  const auto output_height = output->dimension(height_index);
166  ARM_COMPUTE_RETURN_ERROR_ON(output_width == 0);
167  ARM_COMPUTE_RETURN_ERROR_ON(output_height == 0);
168 
169  if(info.interpolation_policy == InterpolationPolicy::NEAREST_NEIGHBOR)
170  {
172  }
173 
174  if(info.interpolation_policy == InterpolationPolicy::BILINEAR)
175  {
179  }
180 
181  ARM_COMPUTE_RETURN_ERROR_ON(info.align_corners && !scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy));
182 
183  if(info.interpolation_policy == InterpolationPolicy::AREA)
184  {
187  }
188 
189  return Status{};
190 }
191 } // namespace
192 
194  : _func(nullptr), _offsets(nullptr), _dx(nullptr), _dy(nullptr), _input(nullptr), _output(nullptr), _policy(), _border_mode(), _constant_border_value(PixelValue()), _sampling_offset(0),
195  _align_corners(false), _data_layout(DataLayout::UNKNOWN)
196 {
197 }
198 
199 void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets,
200  ITensor *output, const ScaleKernelInfo &info)
201 {
202  ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
203  // Perform validation step
205  dx != nullptr ? dx->info() : nullptr,
206  dy != nullptr ? dy->info() : nullptr,
207  offsets != nullptr ? offsets->info() : nullptr,
208  output->info(),
209  info));
210 
211  // Get data layout and width/height indices
212  _data_layout = info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : info.data_layout;
215 
216  _input = input;
217  _output = output;
218  _offsets = offsets;
219  _dx = dx;
220  _dy = dy;
221  _policy = info.interpolation_policy;
222  _border_mode = info.border_mode;
223  _constant_border_value = info.constant_border_value;
224  _align_corners = info.align_corners;
225 
227  {
228  _sampling_offset = 0.5f;
229  }
230 
231  // Compute the ratio between source width/height and destination width/height
232  const auto wr = scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), _align_corners);
233  const auto hr = scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), _align_corners);
234 
235  // Area interpolation behaves as Nearest Neighbour in case of up-sampling
236  _policy = (_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : _policy;
237 
238  if(_border_mode == BorderMode::UNDEFINED)
239  {
240  _border_mode = BorderMode::CONSTANT;
241  _constant_border_value = PixelValue();
242  }
243 
244  // Configure scale function to run
245  if(_data_layout == DataLayout::NCHW)
246  {
247  std::string function_to_call("scale_");
248  function_to_call += string_from_data_type(_input->info()->data_type()) + "_";
249  function_to_call += string_from_data_layout(_data_layout) + "_";
250  function_to_call += string_from_interpolation_policy(_policy);
251 
252  static std::map<std::string, ScaleFunctionPtr> map_function =
253  {
254  { "scale_U8_NCHW_AREA_CONSTANT", &NEScaleKernel::scale_area_nchw_u8 },
255 
256  { "scale_U8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<uint8_t> },
257  { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
258 
259  { "scale_QASYMM8_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<uint8_t> },
260  { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<uint8_t> },
261 
262  { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_qasymm<int8_t> },
263  { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<int8_t> },
264 
265  { "scale_S16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<int16_t> },
266  { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<int16_t> },
267 
268 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
269  { "scale_F16_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float16_t> },
270  { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<float16_t> },
271 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
272 
273  { "scale_F32_NCHW_BILINEAR", &NEScaleKernel::scale_bilinear_nchw<float> },
274  { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &NEScaleKernel::scale_nearest_nchw<float> },
275  };
276  auto it = map_function.find(function_to_call);
277  if(it != map_function.end())
278  {
279  _func = it->second;
280  }
281  }
282 
283  // Configure window
284  Window win = calculate_max_window(*output->info(), Steps());
285  Coordinates coord;
286  coord.set_num_dimensions(output->info()->num_dimensions());
287  output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
288  INEKernel::configure(win);
289 }
290 
291 template <typename T>
292 void NEScaleKernel::scale_nearest_nchw(const Window &window)
293 {
294  const size_t in_stride_x = _input->info()->dimension(0) + _input->info()->padding().left + _input->info()->padding().right;
295 
296  // Compute the ratio between source height and destination height
297  const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
298 
299  // Don't increment in X and Y direction for the input tensor
300  // A pointer to the start of this plane is needed as base for the precomputed offsets
301  Window win_in(window);
302  win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
303  win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
304 
305  // Set offsets window
306  Window win_off;
307  win_off.set(Window::DimX, window[Window::DimX]);
308  win_off.set(Window::DimY, window[Window::DimY]);
309  for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
310  {
311  win_off.set(d, Window::Dimension(0, 0, 0));
312  }
313 
314  // Create iterators
315  Iterator in(_input, win_in);
316  Iterator out(_output, window);
317  Iterator offsets(_offsets, win_off);
318  execute_window_loop(window, [&](const Coordinates & id)
319  {
320  const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr());
321  const auto in_yi = static_cast<int32_t>(_align_corners ? utils::rounding::round_half_away_from_zero((id.y() + _sampling_offset) * hr) : std::floor((id.y() + _sampling_offset) * hr));
322  const int32_t offset_row = in_yi * in_stride_x;
323  *reinterpret_cast<T *>(out.ptr()) = *(reinterpret_cast<const T *>(in.ptr()) + offsets_ptr[0] + offset_row);
324  },
325  in, offsets, out);
326 }
327 
328 template <typename T>
329 void NEScaleKernel::scale_bilinear_nchw(const Window &window)
330 {
331  // Compute the ratio between source height and destination height
332  const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
333  Window win_off;
334  win_off.set(Window::DimX, window.x());
335  win_off.set(Window::DimY, window.y());
336 
337  // Don't increment in X and Y direction for the input tensor
338  // A pointer to the start of this plane is needed as base for the precomputed offsets
339  Window win_in(window);
340  win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
341  win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
342 
343  for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
344  {
345  win_off.set(d, Window::Dimension(0, 0, 0));
346  }
347 
348  Iterator in(_input, win_in);
349  Iterator out(_output, window);
350  Iterator offsets(_offsets, win_off);
351  Iterator dx(_dx, win_off);
352  Iterator dy(_dy, win_off);
353 
354  const int32_t in_dim_w = _input->info()->dimension(0);
355  const int32_t in_dim_h = _input->info()->dimension(1);
356  const int32_t in_stride_w = in_dim_w + _input->info()->padding().left + _input->info()->padding().right;
357 
358  if(_border_mode == BorderMode::CONSTANT)
359  {
360 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
361  using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
362 #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
363  using ConstType = T;
364 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
365  const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
366  execute_window_loop(window, [&](const Coordinates & id)
367  {
368  const int32_t index_h = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
369  const auto index_w = *(reinterpret_cast<const int32_t *>(offsets.ptr()));
370  const auto dx_val = *(reinterpret_cast<const float *>(dx.ptr()));
371  const auto dy_val = *(reinterpret_cast<const float *>(dy.ptr()));
372  const auto pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
373 
374  const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + index_h * in_stride_w)) : const_border_value;
375  const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w)) : const_border_value;
376  const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h
377  && index_h < in_dim_h - 1) ?
378  (*(pixel_row_ptr + index_w + index_h * in_stride_w + in_stride_w)) :
379  const_border_value;
380  const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h
381  && index_h < in_dim_h - 1) ?
382  (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w + in_stride_w)) :
383  const_border_value;
384 
385  *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
386  },
387  in, offsets, dx, dy, out);
388  }
389  else if(_border_mode == BorderMode::REPLICATE)
390  {
391  execute_window_loop(window, [&](const Coordinates & id)
392  {
393  const int index_h = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
394  const auto index_w = *(reinterpret_cast<const int32_t *>(offsets.ptr()));
395  const auto dx_val = *(reinterpret_cast<const float *>(dx.ptr()));
396  const auto dy_val = *(reinterpret_cast<const float *>(dy.ptr()));
397  const auto pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
398 
399  auto clamped_x = utility::clamp<int>(index_w, 0, in_dim_w - 1);
400  auto clamped_x1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
401  auto clamped_y = utility::clamp<int>(index_h, 0, in_dim_h - 1);
402  auto clamped_y1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
403 
404  const auto a00 = *(pixel_row_ptr + clamped_x + clamped_y * in_stride_w);
405  const auto a01 = *(pixel_row_ptr + clamped_x1 + clamped_y * in_stride_w);
406  const auto a10 = *(pixel_row_ptr + clamped_x + clamped_y1 * in_stride_w);
407  const auto a11 = *(pixel_row_ptr + clamped_x1 + clamped_y1 * in_stride_w);
408 
409  *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
410  },
411  in, offsets, dx, dy, out);
412  }
413  else
414  {
415  ARM_COMPUTE_ERROR("Not implemented");
416  }
417 }
418 
419 void NEScaleKernel::scale_area_nchw_u8(const Window &window)
420 {
421  using namespace scale_helpers;
422 
424 
425  // Don't increment in width/height/channels for the input tensor
426  // A pointer to the start of this plane is needed as base for the precomputed offsets
427  Window win_in(window);
428  win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
429  win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
430  win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
431 
432  Iterator in(_input, win_in);
433  Iterator out(_output, window);
434 
435  const auto wr = scale_utils::calculate_resize_ratio(_input->info()->dimension(0), _output->info()->dimension(0), _align_corners);
436  const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(1), _output->info()->dimension(1), _align_corners);
437  const auto w = _input->info()->dimension(0);
438  const auto h = _input->info()->dimension(1);
439  const size_t in_stride = _input->info()->strides_in_bytes()[1];
440 
441  execute_window_loop(window, [&](const Coordinates & id)
442  {
443  const auto in_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
444 
445  uint8x8_t tmp0 = vdup_n_u8(0);
446  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x(), id.y()), tmp0, 0);
447  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 1, id.y()), tmp0, 1);
448  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 2, id.y()), tmp0, 2);
449  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 3, id.y()), tmp0, 3);
450  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 4, id.y()), tmp0, 4);
451  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 5, id.y()), tmp0, 5);
452  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 6, id.y()), tmp0, 6);
453  tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 7, id.y()), tmp0, 7);
454 
455  uint8x8_t tmp1 = vdup_n_u8(0);
456  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 8, id.y()), tmp1, 0);
457  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 9, id.y()), tmp1, 1);
458  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 10, id.y()), tmp1, 2);
459  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 11, id.y()), tmp1, 3);
460  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 12, id.y()), tmp1, 4);
461  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 13, id.y()), tmp1, 5);
462  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 14, id.y()), tmp1, 6);
463  tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 15, id.y()), tmp1, 7);
464 
465  vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1));
466  },
467  in, out);
468 }
469 
470 template <typename T>
471 void NEScaleKernel::scale_bilinear_qasymm(const Window &window)
472 {
473  // Get data layout and width/height indices
476 
477  // Compute the ratio between source height and destination height
478  const auto hr = scale_utils::calculate_resize_ratio(_input->info()->dimension(idx_height), _output->info()->dimension(idx_height), _align_corners);
479  Window win_off;
480  win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
481  win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
482 
483  // Don't increment in X and Y direction for the input tensor
484  // A pointer to the start of this plane is needed as base for the precomputed offsets
485  Window win_in(window);
486  win_in.set(idx_width, Window::Dimension(0, 0, 0));
487  win_in.set(idx_height, Window::Dimension(0, 0, 0));
488 
489  for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d)
490  {
491  win_off.set(d, Window::Dimension(0, 0, 0));
492  }
493 
494  Iterator in(_input, win_in);
495  Iterator out(_output, window);
496 
497  const int32_t in_dim_w = _input->info()->dimension(idx_width);
498  const int32_t in_dim_h = _input->info()->dimension(idx_height);
499  const int32_t stride_w = _input->info()->strides_in_bytes()[idx_width];
500  const int32_t stride_h = _input->info()->strides_in_bytes()[idx_height];
501 
502  const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
503  const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
504 
505  if(_border_mode == BorderMode::CONSTANT)
506  {
507 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
508  using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
509 #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
510  using ConstType = T;
511 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
512  const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
513  execute_window_loop(window, [&](const Coordinates & id)
514  {
515  const int32_t index_h = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
516  const int32_t index_w = *(reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
517  const auto dx_val = *(reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
518  const auto dy_val = *(reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
519  const auto pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
520 
521  const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
522  (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
523  const_border_value;
524  const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
525  (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
526  const_border_value;
527  const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
528  (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
529  const_border_value;
530  const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
531  (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
532  const_border_value;
533 
534  const float inp00 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
535  const float inp01 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
536  const float inp10 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
537  const float inp11 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
538  *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
539  },
540  in, out);
541  }
542  else if(_border_mode == BorderMode::REPLICATE)
543  {
544  execute_window_loop(window, [&](const Coordinates & id)
545  {
546  const int index_h = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
547  const int32_t index_w = *(reinterpret_cast<const int32_t *>(_offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
548  const auto dx_val = *(reinterpret_cast<const float *>(_dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
549  const auto dy_val = *(reinterpret_cast<const float *>(_dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
550  const auto pixel_row_ptr = reinterpret_cast<const T *>(in.ptr());
551 
552  auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
553  auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
554  auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
555  auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
556 
557  const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
558  const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
559  const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
560  const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
561 
562  const float inp00 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
563  const float inp01 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
564  const float inp10 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
565  const float inp11 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
566  *reinterpret_cast<T *>(out.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
567  },
568  in, out);
569  }
570  else
571  {
572  ARM_COMPUTE_ERROR("Not implemented");
573  }
574 }
575 
577  const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
578 {
579  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, dx, dy, offsets, output, info));
580  return Status{};
581 }
582 
583 void NEScaleKernel::run(const Window &window, const ThreadInfo &info)
584 {
585  ARM_COMPUTE_UNUSED(info);
588  ARM_COMPUTE_ERROR_ON(_func == nullptr && _data_layout == DataLayout::NCHW);
589 
590  if(_data_layout == DataLayout::NCHW)
591  {
592  (this->*_func)(window);
593  }
594  else
595  {
596  const auto *uk = get_implementation(ScaleSelectorData{ _input->info()->data_type() });
597  uk->ukernel(_input, _output, _offsets, _dx, _dy, _policy, _border_mode, _constant_border_value, _sampling_offset, _align_corners, window);
598  }
599 }
600 } // namespace arm_compute
BorderMode
Methods available to handle borders.
Definition: Types.h:265
void s16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
BorderMode border_mode
Border mode policy.
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
InterpolationPolicy
Interpolation method.
Definition: Types.h:392
SimpleTensor< float > w
Definition: DFT.cpp:156
DataType dt
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
void qasymm8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
InterpolationPolicy interpolation_policy
Interpolation type to use.
#define REGISTER_FP16_NEON(func_name)
Definition: Registrars.h:42
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
1 channel, 1 U8 per channel
#define REGISTER_FP32_NEON(func_name)
Definition: Registrars.h:52
void get(uint8_t &v) const
Interpret the pixel value as a U8.
Definition: PixelValue.h:241
void qasymm8_signed_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
half_float::half half
16-bit floating point type
Definition: Types.h:46
1 channel, 1 F32 per channel
Output values are defined by bilinear interpolation between the pixels.
#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
#define REGISTER_QASYMM8_SIGNED_NEON(func_name)
Definition: Registrars.h:62
const DataLayout data_layout
Definition: Im2Col.cpp:151
bool align_corners
Align corners of input and output.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
void fp16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
Quantization info when assuming per layer quantization.
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
T round_half_away_from_zero(T value)
Round floating-point value with half value rounding away from zero.
Definition: Rounding.h:106
Status class.
Definition: Error.h:52
Output values are defined to match the source pixel whose center is nearest to the sample position...
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
decltype(strategy::transforms) typedef type
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
virtual void set_valid_region(const ValidRegion &valid_region)=0
Set the valid region of the tensor.
1 channel, 1 F16 per channel
Samples are taken at pixel center.
bool is_align_corners_allowed_sampling_policy(SamplingPolicy sampling_policy)
Returns if aligned corners are allowed for the given sampling policy.
Definition: ScaleUtils.h:52
#define REGISTER_INTEGER_NEON(func_name)
Definition: Registrars.h:92
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
1 channel, 1 S32 per channel
static Status validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
Static function to check if given info will lead to a valid configuration of NEScaleKernel.
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
void u8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
const ScaleSelectorPtr is_selected
#define REGISTER_QASYMM8_NEON(func_name)
Definition: Registrars.h:72
SamplingPolicy sampling_policy
Sampling policy used by the interpolation.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
uint8_t pixel_area_c1u8_clamp(const uint8_t *first_pixel_ptr, size_t stride, size_t width, size_t height, float wr, float hr, int x, int y)
Return the pixel at (x,y) using area interpolation by clamping when out of borders.
Definition: ScaleHelpers.h:283
Coordinates of an item.
Definition: Coordinates.h:37
void qasymm8_signed_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
const std::string & string_from_interpolation_policy(InterpolationPolicy policy)
Translates a given interpolation policy to a string.
Definition: Utils.cpp:212
void fp32_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
Samples are taken at pixel top left corner.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
Definition: Helpers.inl:139
void configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets, ITensor *output, const ScaleKernelInfo &info)
Initialise the kernel&#39;s inputs, output and interpolation policy.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
virtual PaddingSize padding() const =0
Padding of tensor.
unsigned int left
left of the border
Definition: Types.h:378
unsigned int right
right of the border
Definition: Types.h:376
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
1 channel, 1 S16 per channel
Output values are determined by averaging the source pixels whose areas fall under the area of the de...
Num samples, channels, height, width.
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
const std::string & string_from_data_layout(DataLayout dl)
Convert a data layout identity into a string.
Definition: Utils.cpp:123
Information about executing thread and CPU.
Definition: CPPTypes.h:235
const char * name
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
Borders are left undefined.
Pixels outside the image are assumed to have the same value as the closest image pixel.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
constexpr const Dimension & y() const
Alias to access the second dimension of the window.
Definition: Window.h:154
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
static float dequantize(QUANTIZED_TYPE value, const UniformQuantizationInfo &qinfo)
Dequantize a value given a 8-bit asymmetric quantization scheme.
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
PixelValue constant_border_value
Constant value to use for constant border mode policy.
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
Definition: Helpers.inl:77
float calculate_resize_ratio(size_t input_size, size_t output_size, bool align_corners=false)
Returns resize ratio between input and output with consideration of aligned corners.
Definition: ScaleUtils.cpp:27
void set_num_dimensions(size_t num_dimensions)
Set number of dimensions.
Definition: Dimensions.h:149
ScaleKernelPtr ukernel
quantized, asymmetric fixed-point 8-bit number signed
Includes all wrapper headers at once.
Container for valid region of a window.
Definition: Types.h:188
size_t get_data_layout_dimension_index(const DataLayout data_layout, const DataLayoutDimension data_layout_dimension)
Get the index of the given dimension.
Definition: Helpers.inl:193
DataLayout data_layout
Data layout to use.
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
float delta_bilinear(float a00, float a01, float a10, float a11, float dx_val, float dy_val)
Computes bilinear interpolation using the top-left, top-right, bottom-left, bottom-right pixels and t...
Definition: ScaleHelpers.h:343
DataType
Available data types.
Definition: Types.h:77
DataLayout
[DataLayout enum definition]
Definition: Types.h:120
void qasymm8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, bool align_corners, const Window &window)
Definition: qasymm8.cpp:135
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
NEScaleKernel()
Default constructor.
virtual DataLayout data_layout() const =0
Get the data layout of the tensor.
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:145