xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuScaleKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2016-2022 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 #include "src/cpu/kernels/CpuScaleKernel.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Window.h"
28 #include "src/core/common/Registrars.h"
29 #include "src/core/helpers/ScaleHelpers.h"
30 #include "src/core/helpers/WindowHelpers.h"
31 #include "src/cpu/kernels/scale/neon/list.h"
32 #include "src/cpu/kernels/scale/sve/list.h"
33 #include "support/Rounding.h"
34 
35 #include <arm_neon.h>
36 #include <map>
37 
38 namespace arm_compute
39 {
40 namespace cpu
41 {
42 namespace kernels
43 {
44 namespace
45 {
46 static const std::vector<CpuScaleKernel::ScaleKernel> available_kernels =
47 {
48     {
49         "sve_fp16_scale",
50         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0202() 51         {
52             return data.dt == DataType::F16 && data.isa.sve && data.isa.fp16 && data.interpolation_policy != InterpolationPolicy::BILINEAR;
53         },
54         REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_scale)
55     },
56     {
57         "sve_fp32_scale",
58         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0302() 59         {
60             return data.dt == DataType::F32 && data.isa.sve && data.interpolation_policy != InterpolationPolicy::BILINEAR;
61         },
62         REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_scale)
63     },
64     {
65         "sve_qu8_scale",
66         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0402() 67         {
68             return data.dt == DataType::QASYMM8 && data.isa.sve && data.interpolation_policy != InterpolationPolicy::BILINEAR;
69         },
70         REGISTER_QASYMM8_SVE(arm_compute::cpu::qasymm8_sve_scale)
71     },
72     {
73         "sve_qs8_scale",
74         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0502() 75         {
76             return data.dt == DataType::QASYMM8_SIGNED && data.isa.sve && data.interpolation_policy != InterpolationPolicy::BILINEAR;
77         },
78         REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::qasymm8_signed_sve_scale)
79     },
80     {
81         "sve_u8_scale",
82         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0602() 83         {
84             return data.dt == DataType::U8 && data.isa.sve && data.interpolation_policy != InterpolationPolicy::BILINEAR;
85         },
86         REGISTER_INTEGER_SVE(arm_compute::cpu::u8_sve_scale)
87     },
88     {
89         "sve_s16_scale",
90         [](const ScaleKernelDataTypeISASelectorData & data)
__anone5d1f65b0702() 91         {
92             return data.dt == DataType::S16 && data.isa.sve && data.interpolation_policy != InterpolationPolicy::BILINEAR;
93         },
94         REGISTER_INTEGER_SVE(arm_compute::cpu::s16_sve_scale)
95     },
96     {
97         "neon_fp16_scale",
__anone5d1f65b0802() 98         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.fp16; },
99         REGISTER_FP16_NEON(arm_compute::cpu::common_neon_scale<float16_t>)
100     },
101     {
102         "neon_fp32_scale",
__anone5d1f65b0902() 103         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::F32; },
104         REGISTER_FP32_NEON(arm_compute::cpu::common_neon_scale<float>)
105     },
106     {
107         "neon_qu8_scale",
__anone5d1f65b0a02() 108         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8; },
109         REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_neon_scale)
110     },
111     {
112         "neon_qs8_scale",
__anone5d1f65b0b02() 113         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
114         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_neon_scale)
115     },
116     {
117         "neon_u8_scale",
__anone5d1f65b0c02() 118         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::U8; },
119         REGISTER_INTEGER_NEON(arm_compute::cpu::u8_neon_scale)
120     },
121     {
122         "neon_s8_scale",
__anone5d1f65b0d02() 123         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::S8; },
124         REGISTER_INTEGER_NEON(arm_compute::cpu::s8_neon_scale)
125     },
126     {
127         "neon_s16_scale",
__anone5d1f65b0e02() 128         [](const ScaleKernelDataTypeISASelectorData & data) { return data.dt == DataType::S16; },
129         REGISTER_INTEGER_NEON(arm_compute::cpu::s16_neon_scale)
130     },
131 };
132 
validate_arguments(const ITensorInfo * src,const ITensorInfo * dx,const ITensorInfo * dy,const ITensorInfo * offsets,ITensorInfo * dst,const ScaleKernelInfo & info)133 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy,
134                           const ITensorInfo *offsets, ITensorInfo *dst, const ScaleKernelInfo &info)
135 {
136     const auto *uk = CpuScaleKernel::get_implementation(ScaleKernelDataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_isa(), info.interpolation_policy });
137 
138     ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
139 
140     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
141     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
142     ARM_COMPUTE_RETURN_ERROR_ON(dst == src);
143     ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
144     ARM_COMPUTE_UNUSED(info.constant_border_value);
145     ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.use_padding, "Padding is not supported");
146 
147     const DataLayout data_layout   = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
148     const auto       width_index   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
149     const auto       height_index  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
150     const auto       output_width  = dst->dimension(width_index);
151     const auto       output_height = dst->dimension(height_index);
152     ARM_COMPUTE_RETURN_ERROR_ON(output_width == 0);
153     ARM_COMPUTE_RETURN_ERROR_ON(output_height == 0);
154 
155     ARM_COMPUTE_RETURN_ERROR_ON((src->data_type() == DataType::S8) && (data_layout != DataLayout::NHWC || info.interpolation_policy != InterpolationPolicy::BILINEAR
156                                                                        || info.border_mode != BorderMode::REPLICATE));
157 
158     if(info.interpolation_policy == InterpolationPolicy::NEAREST_NEIGHBOR && offsets != nullptr)
159     {
160         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32);
161     }
162 
163     if(info.interpolation_policy == InterpolationPolicy::BILINEAR && offsets != nullptr)
164     {
165         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32);
166         if(dx != nullptr && dy != nullptr)
167         {
168             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dx, 1, DataType::F32);
169             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dy, 1, DataType::F32);
170         }
171     }
172 
173     ARM_COMPUTE_RETURN_ERROR_ON(info.align_corners && !scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy));
174 
175     if(info.interpolation_policy == InterpolationPolicy::AREA)
176     {
177         ARM_COMPUTE_RETURN_ERROR_ON(data_layout != DataLayout::NCHW);
178         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
179     }
180 
181     return Status{};
182 }
183 } // namespace
184 
configure(const ITensorInfo * src,const ITensorInfo * dx,const ITensorInfo * dy,const ITensorInfo * offsets,ITensorInfo * dst,const ScaleKernelInfo & info)185 void CpuScaleKernel::configure(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets,
186                                ITensorInfo *dst, const ScaleKernelInfo &info)
187 {
188     ARM_COMPUTE_UNUSED(dx, dy, offsets);
189     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
190     // Perform validation step
191     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
192                                                   dx,
193                                                   dy,
194                                                   offsets,
195                                                   dst,
196                                                   info));
197 
198     const auto *uk = CpuScaleKernel::get_implementation(ScaleKernelDataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_isa(), info.interpolation_policy });
199     ARM_COMPUTE_ERROR_ON_NULLPTR(uk);
200 
201     _run_method = uk->ukernel;
202     _name       = std::string("CpuScaleKernel").append("/").append(uk->name).append("_").append(string_from_interpolation_policy(info.interpolation_policy));
203 
204     // Get data layout and width/height indices
205     _data_layout         = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
206     const int idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
207     const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
208 
209     _policy                = info.interpolation_policy;
210     _border_mode           = info.border_mode;
211     _constant_border_value = info.constant_border_value;
212     _align_corners         = info.align_corners;
213 
214     if(info.sampling_policy == SamplingPolicy::CENTER)
215     {
216         _sampling_offset = 0.5f;
217     }
218 
219     // Compute the ratio between source width/height and destination width/height
220     const auto wr = scale_utils::calculate_resize_ratio(src->dimension(idx_width), dst->dimension(idx_width), _align_corners);
221     const auto hr = scale_utils::calculate_resize_ratio(src->dimension(idx_height), dst->dimension(idx_height), _align_corners);
222 
223     // Area interpolation behaves as Nearest Neighbour in case of up-sampling
224     _policy = (_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : _policy;
225 
226     if(_border_mode == BorderMode::UNDEFINED)
227     {
228         _border_mode           = BorderMode::CONSTANT;
229         _constant_border_value = PixelValue();
230     }
231 
232 #ifdef ENABLE_NCHW_KERNELS
233     // Configure scale function to run
234     if(_data_layout == DataLayout::NCHW)
235     {
236         std::string function_to_call("scale_");
237         function_to_call += string_from_data_type(src->data_type()) + "_";
238         function_to_call += string_from_data_layout(_data_layout) + "_";
239         function_to_call += string_from_interpolation_policy(_policy);
240 
241         static std::map<std::string, ScaleFunctionPtr> map_function =
242         {
243             { "scale_U8_NCHW_AREA_CONSTANT", &CpuScaleKernel::scale_area_nchw_u8 },
244 
245             { "scale_U8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<uint8_t> },
246             { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
247 
248             { "scale_QASYMM8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<uint8_t> },
249             { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
250 
251             { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<int8_t> },
252             { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int8_t> },
253 
254             { "scale_S16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<int16_t> },
255             { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int16_t> },
256 
257 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
258             { "scale_F16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float16_t> },
259             { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float16_t> },
260 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
261 
262             { "scale_F32_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float> },
263             { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float> },
264         };
265         auto it = map_function.find(function_to_call);
266         if(it != map_function.end())
267         {
268             _func = it->second;
269         }
270     }
271 #endif // ENABLE_NCHW_KERNELS
272 
273     // Configure window
274     Window win = calculate_max_window(*dst, Steps());
275     ICpuKernel::configure(win);
276 }
277 
278 #ifdef ENABLE_NCHW_KERNELS
279 template <typename T>
scale_nearest_nchw(const ITensor * src,ITensor * dst,const ITensor * dx,const ITensor * dy,const ITensor * offsets,const Window & window)280 void CpuScaleKernel::scale_nearest_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
281 {
282     ARM_COMPUTE_UNUSED(dx, dy);
283     const size_t in_stride_x = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
284 
285     // Compute the ratio between source height and destination height
286     const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
287 
288     // Don't increment in X and Y direction for the input tensor
289     // A pointer to the start of this plane is needed as base for the precomputed offsets
290     Window win_in(window);
291     win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
292     win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
293 
294     // Set offsets window
295     Window win_off;
296     win_off.set(Window::DimX, window[Window::DimX]);
297     win_off.set(Window::DimY, window[Window::DimY]);
298     for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
299     {
300         win_off.set(d, Window::Dimension(0, 0, 0));
301     }
302 
303     // Create iterators
304     Iterator src_i(src, win_in);
305     Iterator dst_i(dst, window);
306     Iterator offsets_i(offsets, win_off);
307     execute_window_loop(window, [&](const Coordinates & id)
308     {
309         const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets_i.ptr());
310         const auto in_yi       = static_cast<int32_t>(_align_corners ? utils::rounding::round_half_away_from_zero((id.y() + _sampling_offset) * hr) : std::floor((
311                                                           id.y() + _sampling_offset)
312                                                       * hr));
313         const int32_t offset_row            = in_yi * in_stride_x;
314         *reinterpret_cast<T *>(dst_i.ptr()) = *(reinterpret_cast<const T *>(src_i.ptr()) + offsets_ptr[0] + offset_row);
315     },
316     src_i, offsets_i, dst_i);
317 }
318 
319 template <typename T>
scale_bilinear_nchw(const ITensor * src,ITensor * dst,const ITensor * dx,const ITensor * dy,const ITensor * offsets,const Window & window)320 void CpuScaleKernel::scale_bilinear_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
321 {
322     // Compute the ratio between source height and destination height
323     const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
324     Window     win_off;
325     win_off.set(Window::DimX, window.x());
326     win_off.set(Window::DimY, window.y());
327 
328     // Don't increment in X and Y direction for the input tensor
329     // A pointer to the start of this plane is needed as base for the precomputed offsets
330     Window win_in(window);
331     win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
332     win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
333 
334     for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
335     {
336         win_off.set(d, Window::Dimension(0, 0, 0));
337     }
338 
339     Iterator src_i(src, win_in);
340     Iterator dst_i(dst, window);
341     Iterator offsets_i(offsets, win_off);
342     Iterator dx_i(dx, win_off);
343     Iterator dy_i(dy, win_off);
344 
345     const int32_t in_dim_w    = src->info()->dimension(0);
346     const int32_t in_dim_h    = src->info()->dimension(1);
347     const int32_t in_stride_w = in_dim_w + src->info()->padding().left + src->info()->padding().right;
348 
349     if(_border_mode == BorderMode::CONSTANT)
350     {
351 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
352         using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
353 #else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
354         using ConstType = T;
355 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
356         const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
357         execute_window_loop(window, [&](const Coordinates & id)
358         {
359             const int32_t index_h       = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
360             const auto    index_w       = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
361             const auto    dx_val        = *(reinterpret_cast<const float *>(dx_i.ptr()));
362             const auto    dy_val        = *(reinterpret_cast<const float *>(dy_i.ptr()));
363             const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
364 
365             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;
366             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;
367             const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h
368                               && index_h < in_dim_h - 1) ?
369                              (*(pixel_row_ptr + index_w + index_h * in_stride_w + in_stride_w)) :
370                              const_border_value;
371             const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h
372                               && index_h < in_dim_h - 1) ?
373                              (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w + in_stride_w)) :
374                              const_border_value;
375 
376             *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
377         },
378         src_i, offsets_i, dx_i, dy_i, dst_i);
379     }
380     else if(_border_mode == BorderMode::REPLICATE)
381     {
382         execute_window_loop(window, [&](const Coordinates & id)
383         {
384             const int  index_h       = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
385             const auto index_w       = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
386             const auto dx_val        = *(reinterpret_cast<const float *>(dx_i.ptr()));
387             const auto dy_val        = *(reinterpret_cast<const float *>(dy_i.ptr()));
388             const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
389 
390             auto clamped_x  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
391             auto clamped_x1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
392             auto clamped_y  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
393             auto clamped_y1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
394 
395             const auto a00 = *(pixel_row_ptr + clamped_x + clamped_y * in_stride_w);
396             const auto a01 = *(pixel_row_ptr + clamped_x1 + clamped_y * in_stride_w);
397             const auto a10 = *(pixel_row_ptr + clamped_x + clamped_y1 * in_stride_w);
398             const auto a11 = *(pixel_row_ptr + clamped_x1 + clamped_y1 * in_stride_w);
399 
400             *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
401         },
402         src_i, offsets_i, dx_i, dy_i, dst_i);
403     }
404     else
405     {
406         ARM_COMPUTE_ERROR("Not implemented");
407     }
408 }
409 
scale_area_nchw_u8(const ITensor * src,ITensor * dst,const ITensor * dx,const ITensor * dy,const ITensor * offsets,const Window & window)410 void CpuScaleKernel::scale_area_nchw_u8(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
411 {
412     ARM_COMPUTE_UNUSED(dx, dy, offsets);
413     using namespace scale_helpers;
414 
415     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
416 
417     // Don't increment in width/height/channels for the input tensor
418     // A pointer to the start of this plane is needed as base for the precomputed offsets
419     Window win_in(window);
420     win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
421     win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
422     win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
423 
424     Iterator src_i(src, win_in);
425     Iterator dst_i(dst, window);
426 
427     const auto   wr        = scale_utils::calculate_resize_ratio(src->info()->dimension(0), dst->info()->dimension(0), _align_corners);
428     const auto   hr        = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
429     const auto   w         = src->info()->dimension(0);
430     const auto   h         = src->info()->dimension(1);
431     const size_t in_stride = src->info()->strides_in_bytes()[1];
432 
433     execute_window_loop(window, [&](const Coordinates & id)
434     {
435         const auto in_ptr = reinterpret_cast<const uint8_t *>(src_i.ptr());
436 
437         uint8x8_t tmp0 = vdup_n_u8(0);
438         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x(), id.y()), tmp0, 0);
439         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 1, id.y()), tmp0, 1);
440         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 2, id.y()), tmp0, 2);
441         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 3, id.y()), tmp0, 3);
442         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 4, id.y()), tmp0, 4);
443         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 5, id.y()), tmp0, 5);
444         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 6, id.y()), tmp0, 6);
445         tmp0           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 7, id.y()), tmp0, 7);
446 
447         uint8x8_t tmp1 = vdup_n_u8(0);
448         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 8, id.y()), tmp1, 0);
449         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 9, id.y()), tmp1, 1);
450         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 10, id.y()), tmp1, 2);
451         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 11, id.y()), tmp1, 3);
452         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 12, id.y()), tmp1, 4);
453         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 13, id.y()), tmp1, 5);
454         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 14, id.y()), tmp1, 6);
455         tmp1           = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 15, id.y()), tmp1, 7);
456 
457         vst1q_u8(dst_i.ptr(), vcombine_u8(tmp0, tmp1));
458     },
459     src_i, dst_i);
460 }
461 
462 template <typename T>
scale_bilinear_qasymm(const ITensor * src,ITensor * dst,const ITensor * dx,const ITensor * dy,const ITensor * offsets,const Window & window)463 void CpuScaleKernel::scale_bilinear_qasymm(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
464 {
465     // Get data layout and width/height indices
466     const int idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
467     const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
468 
469     // Compute the ratio between source height and destination height
470     const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), _align_corners);
471     Window     win_off;
472     win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
473     win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
474 
475     // Don't increment in X and Y direction for the input tensor
476     // A pointer to the start of this plane is needed as base for the precomputed offsets
477     Window win_in(window);
478     win_in.set(idx_width, Window::Dimension(0, 0, 0));
479     win_in.set(idx_height, Window::Dimension(0, 0, 0));
480 
481     for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
482     {
483         win_off.set(d, Window::Dimension(0, 0, 0));
484     }
485 
486     Iterator src_i(src, win_in);
487     Iterator dst_i(dst, window);
488 
489     const int32_t in_dim_w = src->info()->dimension(idx_width);
490     const int32_t in_dim_h = src->info()->dimension(idx_height);
491     const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
492     const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
493 
494     const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
495     const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
496 
497     if(_border_mode == BorderMode::CONSTANT)
498     {
499 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
500         using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
501 #else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
502         using ConstType = T;
503 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
504         const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
505         execute_window_loop(window, [&](const Coordinates & id)
506         {
507             const int32_t index_h       = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
508             const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
509             const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
510             const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
511             const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
512 
513             const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
514                              (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
515                              const_border_value;
516             const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
517                              (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
518                              const_border_value;
519             const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
520                              (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
521                              const_border_value;
522             const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
523                              (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
524                              const_border_value;
525 
526             const float inp00                   = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
527             const float inp01                   = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
528             const float inp10                   = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
529             const float inp11                   = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
530             *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
531         },
532         src_i, dst_i);
533     }
534     else if(_border_mode == BorderMode::REPLICATE)
535     {
536         execute_window_loop(window, [&](const Coordinates & id)
537         {
538             const int     index_h       = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
539             const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
540             const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
541             const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
542             const auto    pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
543 
544             auto clamped_w  = utility::clamp<int>(index_w, 0, in_dim_w - 1);
545             auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
546             auto clamped_h  = utility::clamp<int>(index_h, 0, in_dim_h - 1);
547             auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
548 
549             const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
550             const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
551             const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
552             const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
553 
554             const float inp00                   = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
555             const float inp01                   = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
556             const float inp10                   = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
557             const float inp11                   = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
558             *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
559         },
560         src_i, dst_i);
561     }
562     else
563     {
564         ARM_COMPUTE_ERROR("Not implemented");
565     }
566 }
567 #endif // ENABLE_NCHW_KERNELS
568 
validate(const ITensorInfo * input,const ITensorInfo * dx,const ITensorInfo * dy,const ITensorInfo * offsets,ITensorInfo * output,const ScaleKernelInfo & info)569 Status CpuScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
570                                 const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
571 {
572     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, dx, dy, offsets, output, info));
573     return Status{};
574 }
575 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)576 void CpuScaleKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
577 {
578     ARM_COMPUTE_UNUSED(info);
579     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
580     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
581     ARM_COMPUTE_ERROR_ON(_func == nullptr && _data_layout == DataLayout::NCHW);
582     ARM_COMPUTE_ERROR_ON(_run_method == nullptr && _data_layout == DataLayout::NHWC);
583 
584     const auto src     = tensors.get_const_tensor(TensorType::ACL_SRC);
585     auto       dst     = tensors.get_tensor(TensorType::ACL_DST);
586     const auto dx      = tensors.get_const_tensor(TensorType::ACL_INT_0);
587     const auto dy      = tensors.get_const_tensor(TensorType::ACL_INT_1);
588     const auto offsets = tensors.get_const_tensor(TensorType::ACL_INT_2);
589 
590     if(_data_layout == DataLayout::NCHW)
591     {
592         (this->*_func)(src, dst, dx, dy, offsets, window);
593     }
594     else
595     {
596         _run_method(src, dst, offsets, dx, dy, _policy, _border_mode, _constant_border_value, _sampling_offset, _align_corners, window);
597     }
598 }
599 
name() const600 const char *CpuScaleKernel::name() const
601 {
602     return _name.c_str();
603 }
604 
get_available_kernels()605 const std::vector<CpuScaleKernel::ScaleKernel> &CpuScaleKernel::get_available_kernels()
606 {
607     return available_kernels;
608 }
609 
610 } // namespace kernels
611 } // namespace cpu
612 } // namespace arm_compute
613