xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/scale/neon/qasymm8_signed.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2021-2022 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/core/helpers/ScaleHelpers.h"
25 #include "src/cpu/kernels/scale/neon/list.h"
26 
27 namespace arm_compute
28 {
29 namespace
30 {
qasymm8_signed_neon_scale_bilinear(const ITensor * src,ITensor * dst,const ITensor * offsets,const ITensor * dx,const ITensor * dy,BorderMode border_mode,PixelValue constant_border_value,float sampling_offset,bool align_corners,const Window & window)31 void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
32                                         BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
33                                         bool align_corners, const Window &window)
34 {
35     // Data layout is NHWC
36     const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
37     const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
38 
39     const int32_t input_width  = src->info()->dimension(1);
40     const int32_t input_height = src->info()->dimension(2);
41 
42     // Compute the ratio between source and destination dimensions
43     const float scale_x = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), align_corners);
44     const float scale_y = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
45 
46     if(border_mode == BorderMode::CONSTANT)
47     {
48         const int32_t in_stride_y = src->info()->strides_in_bytes()[1];
49         const int32_t in_stride_z = src->info()->strides_in_bytes()[2];
50 
51         Window win_off;
52         win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
53         win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
54 
55         // Don't increment in X and Y direction for the input tensor
56         // A pointer to the start of this plane is needed as base for the precomputed offsets
57         Window win_in(window);
58         win_in.set(1, Window::Dimension(0, 0, 0));
59         win_in.set(2, Window::Dimension(0, 0, 0));
60 
61         for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
62         {
63             win_off.set(d, Window::Dimension(0, 0, 0));
64         }
65 
66         Iterator in(src, win_in);
67         Iterator out(dst, window);
68 
69         const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>());
70         execute_window_loop(window, [&](const Coordinates & id)
71         {
72             const int32_t index_h       = std::floor((id[2] + sampling_offset) * scale_y - sampling_offset);
73             const int32_t index_w       = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
74             const auto    dx_val        = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
75             const auto    dy_val        = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
76             const auto    pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
77 
78             const auto a00 = (0 <= index_w && index_w < input_width && 0 <= index_h && index_h < input_height) ?
79                              (*(pixel_row_ptr + index_w * in_stride_y + index_h * in_stride_z)) :
80                              const_border_value;
81             const auto a01 = (-1 <= index_w && index_w + 1 < input_width && 0 <= index_h && index_h < input_height) ?
82                              (*(pixel_row_ptr + (index_w + 1) * in_stride_y + index_h * in_stride_z)) :
83                              const_border_value;
84             const auto a10 = (0 <= index_w && index_w < input_width && -1 <= index_h && index_h < input_height - 1) ?
85                              (*(pixel_row_ptr + index_w * in_stride_y + (index_h + 1) * in_stride_z)) :
86                              const_border_value;
87             const auto a11 = (-1 <= index_w && index_w < input_width - 1 && -1 <= index_h && index_h < input_height - 1) ?
88                              (*(pixel_row_ptr + (index_w + 1) * in_stride_y + (index_h + 1) * in_stride_z)) :
89                              const_border_value;
90 
91             const float inp00                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
92             const float inp01                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
93             const float inp10                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
94             const float inp11                      = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
95             *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
96         },
97         in, out);
98     }
99     else if(border_mode == BorderMode::REPLICATE)
100     {
101         using FloatTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
102         using Int32TagType = typename wrapper::traits::neon_bitvector_tag_t<int32_t, wrapper::traits::BitWidth::W128>;
103 
104         const int     in_stride_x  = src->info()->strides_in_bytes()[1];
105         const int     in_stride_y  = src->info()->strides_in_bytes()[2];
106         const int     in_stride_b  = src->info()->strides_in_bytes()[3];
107         const int     out_stride_x = dst->info()->strides_in_bytes()[1];
108         const int     out_stride_y = dst->info()->strides_in_bytes()[2];
109         const int     out_stride_b = dst->info()->strides_in_bytes()[3];
110         const int     out_dim_ch   = dst->info()->dimension(0);
111         constexpr int step_cout    = 16;
112 
113         Window window_execution = window;
114         window_execution.set(Window::DimX, Window::Dimension(0, 1, 1));
115         Window win_in_out(window);
116         win_in_out.set(Window::DimY, Window::Dimension(0, 0, 0));
117         win_in_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
118         Iterator in(src, win_in_out);
119         Iterator out(dst, win_in_out);
120 
121         const int xo_start = window_execution[1].start();
122         const int xo_end   = window_execution[1].end();
123         const int xo_step  = window_execution[1].step();
124         const int yo_start = window_execution[2].start();
125         const int yo_end   = window_execution[2].end();
126         const int yo_step  = window_execution[2].step();
127         const int bo_start = window_execution[3].start();
128         const int bo_end   = window_execution[3].end();
129         const int bo_step  = window_execution[3].step();
130 
131         const float fp_coord_offset_y = sampling_offset * (scale_y - 1);
132         const float fp_coord_offset_x = sampling_offset * (scale_x - 1);
133 
134         const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
135         const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
136 
137         const float32x4_t vscale_in  = wrapper::vdup_n(iq_info.scale, FloatTagType{});
138         const int32x4_t   voffset_in = wrapper::vdup_n(iq_info.offset, Int32TagType{}); // Offsets will be Int32
139 
140         const float32x4_t invvscale_o = wrapper::vdup_n(1.f / oq_info.scale, FloatTagType{});
141         const float32x4_t voffset_o   = vdupq_n_f32(oq_info.offset);
142 
143         for(int bo = bo_start; bo < bo_end; bo += bo_step)
144         {
145             const int8_t *in_ptr  = reinterpret_cast<int8_t *>(in.ptr() + bo * in_stride_b);
146             int8_t       *out_ptr = reinterpret_cast<int8_t *>(out.ptr() + bo * out_stride_b);
147 
148             for(int yo = yo_start; yo < yo_end; yo += yo_step)
149             {
150                 // Floating-point coordinate
151                 const float yi_f = yo * scale_y + fp_coord_offset_y;
152                 // Integer coordinate
153                 const int yi = static_cast<int>(std::floor(yi_f));
154                 // Weight for the y coordinate
155                 const float a1 = (yi_f - static_cast<float>(yi));
156                 const float b1 = (1.f - a1);
157 
158                 const int yi0 = utility::clamp<int>(yi, 0, input_height - 1);
159                 const int yi1 = utility::clamp<int>(yi + 1, 0, input_height - 1);
160 
161                 const int8_t *in_ptr_yi0 = in_ptr + yi0 * in_stride_y;
162                 const int8_t *in_ptr_yi1 = in_ptr + yi1 * in_stride_y;
163 
164                 int8_t *out_ptr_yo = out_ptr + yo * out_stride_y;
165                 for(int xo = xo_start; xo < xo_end; xo += xo_step)
166                 {
167                     // Floating-point coordinate
168                     const float xi_f = xo * scale_x + fp_coord_offset_x;
169                     // Integer coordinate
170                     const int xi = static_cast<int>(std::floor(xi_f));
171                     // Weight for the x coordinate
172                     const float a = (xi_f - static_cast<float>(xi));
173                     const float b = (1.f - a);
174 
175                     const float s00_s = b * b1;
176                     const float s01_s = a * b1;
177                     const float s10_s = b * a1;
178                     const float s11_s = a * a1;
179 
180                     const auto s00 = wrapper::vdup_n(s00_s, FloatTagType{});
181                     const auto s01 = wrapper::vdup_n(s01_s, FloatTagType{});
182                     const auto s10 = wrapper::vdup_n(s10_s, FloatTagType{});
183                     const auto s11 = wrapper::vdup_n(s11_s, FloatTagType{});
184 
185                     const int xi0 = utility::clamp<int>(xi, 0, input_width - 1);
186                     const int xi1 = utility::clamp<int>(xi + 1, 0, input_width - 1);
187 
188                     const auto in_ptr_xi0_yi0 = in_ptr_yi0 + xi0 * in_stride_x;
189                     const auto in_ptr_xi1_yi0 = in_ptr_yi0 + xi1 * in_stride_x;
190                     const auto in_ptr_xi0_yi1 = in_ptr_yi1 + xi0 * in_stride_x;
191                     const auto in_ptr_xi1_yi1 = in_ptr_yi1 + xi1 * in_stride_x;
192 
193                     int8_t *out_ptr_xo_yo = out_ptr_yo + xo * out_stride_x;
194 
195                     int cout = 0;
196                     for(; cout <= (out_dim_ch - step_cout); cout += step_cout)
197                     {
198                         const auto in00 = wrapper::vloadq(in_ptr_xi0_yi0 + cout * sizeof(int8_t));
199                         const auto in01 = wrapper::vloadq(in_ptr_xi1_yi0 + cout * sizeof(int8_t));
200                         const auto in10 = wrapper::vloadq(in_ptr_xi0_yi1 + cout * sizeof(int8_t));
201                         const auto in11 = wrapper::vloadq(in_ptr_xi1_yi1 + cout * sizeof(int8_t));
202 
203                         const int16x8_t in00_low  = wrapper::vmovl(wrapper::vgetlow(in00));
204                         const int16x8_t in00_high = wrapper::vmovl(wrapper::vgethigh(in00));
205 
206                         const auto in00_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in00_low)), voffset_in)), vscale_in);
207                         const auto in00_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in00_low)), voffset_in)), vscale_in);
208                         const auto in00_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in00_high)), voffset_in)), vscale_in);
209                         const auto in00_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in00_high)), voffset_in)), vscale_in);
210 
211                         const int16x8_t in01_low  = wrapper::vmovl(wrapper::vgetlow(in01));
212                         const int16x8_t in01_high = wrapper::vmovl(wrapper::vgethigh(in01));
213 
214                         const auto in01_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in01_low)), voffset_in)), vscale_in);
215                         const auto in01_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in01_low)), voffset_in)), vscale_in);
216                         const auto in01_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in01_high)), voffset_in)), vscale_in);
217                         const auto in01_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in01_high)), voffset_in)), vscale_in);
218 
219                         const int16x8_t in10_low  = wrapper::vmovl(wrapper::vgetlow(in10));
220                         const int16x8_t in10_high = wrapper::vmovl(wrapper::vgethigh(in10));
221 
222                         const auto in10_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in10_low)), voffset_in)), vscale_in);
223                         const auto in10_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in10_low)), voffset_in)), vscale_in);
224                         const auto in10_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in10_high)), voffset_in)), vscale_in);
225                         const auto in10_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in10_high)), voffset_in)), vscale_in);
226 
227                         const int16x8_t in11_low  = wrapper::vmovl(wrapper::vgetlow(in11));
228                         const int16x8_t in11_high = wrapper::vmovl(wrapper::vgethigh(in11));
229 
230                         const auto in11_0 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in11_low)), voffset_in)), vscale_in);
231                         const auto in11_1 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in11_low)), voffset_in)), vscale_in);
232                         const auto in11_2 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgetlow(in11_high)), voffset_in)), vscale_in);
233                         const auto in11_3 = wrapper::vmul(wrapper::vcvt<float>(wrapper::vsub(wrapper::vmovl(wrapper::vgethigh(in11_high)), voffset_in)), vscale_in);
234 
235                         auto out_0 = wrapper::vmul(in00_0, s00);
236                         out_0      = wrapper::vmla(out_0, in01_0, s01);
237                         out_0      = wrapper::vmla(out_0, in10_0, s10);
238                         out_0      = wrapper::vmla(out_0, in11_0, s11);
239 
240                         auto out_1 = wrapper::vmul(in00_1, s00);
241                         out_1      = wrapper::vmla(out_1, in01_1, s01);
242                         out_1      = wrapper::vmla(out_1, in10_1, s10);
243                         out_1      = wrapper::vmla(out_1, in11_1, s11);
244 
245                         auto out_2 = wrapper::vmul(in00_2, s00);
246                         out_2      = wrapper::vmla(out_2, in01_2, s01);
247                         out_2      = wrapper::vmla(out_2, in10_2, s10);
248                         out_2      = wrapper::vmla(out_2, in11_2, s11);
249 
250                         auto out_3 = wrapper::vmul(in00_3, s00);
251                         out_3      = wrapper::vmla(out_3, in01_3, s01);
252                         out_3      = wrapper::vmla(out_3, in10_3, s10);
253                         out_3      = wrapper::vmla(out_3, in11_3, s11);
254 
255 #if defined(__aarch64__) && !defined(BARE_METAL)
256                         const auto out_0_int = wrapper::vcvta<int32_t>(wrapper::vmla(voffset_o, out_0, invvscale_o));
257                         const auto out_1_int = wrapper::vcvta<int32_t>(wrapper::vmla(voffset_o, out_1, invvscale_o));
258                         const auto out_2_int = wrapper::vcvta<int32_t>(wrapper::vmla(voffset_o, out_2, invvscale_o));
259                         const auto out_3_int = wrapper::vcvta<int32_t>(wrapper::vmla(voffset_o, out_3, invvscale_o));
260 #else  // defined(__aarch64__) && !defined(BARE_METAL)
261                         const auto out_0_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_0, invvscale_o));
262                         const auto out_1_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_1, invvscale_o));
263                         const auto out_2_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_2, invvscale_o));
264                         const auto out_3_int = wrapper::vcvt<int32_t>(wrapper::vmla(voffset_o, out_3, invvscale_o));
265 #endif // defined(__aarch64__) && !defined(BARE_METAL)
266                         const auto low_part  = wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_0_int), wrapper::vqmovn(out_1_int)));
267                         const auto high_part = wrapper::vqmovn(wrapper::vcombine(wrapper::vqmovn(out_2_int), wrapper::vqmovn(out_3_int)));
268                         const auto out       = wrapper::vcombine(low_part, high_part);
269 
270                         wrapper::vstore(out_ptr_xo_yo + cout * sizeof(int8_t), out);
271                     }
272 
273                     for(; cout < out_dim_ch; ++cout)
274                     {
275                         const int8_t in00 = *(in_ptr_xi0_yi0 + cout * sizeof(int8_t));
276                         const int8_t in01 = *(in_ptr_xi1_yi0 + cout * sizeof(int8_t));
277                         const int8_t in10 = *(in_ptr_xi0_yi1 + cout * sizeof(int8_t));
278                         const int8_t in11 = *(in_ptr_xi1_yi1 + cout * sizeof(int8_t));
279 
280                         const float in00_f = (static_cast<int32_t>(in00) - iq_info.offset) * iq_info.scale;
281                         const float in01_f = (static_cast<int32_t>(in01) - iq_info.offset) * iq_info.scale;
282                         const float in10_f = (static_cast<int32_t>(in10) - iq_info.offset) * iq_info.scale;
283                         const float in11_f = (static_cast<int32_t>(in11) - iq_info.offset) * iq_info.scale;
284 
285                         float out = in00_f * s00_s;
286                         out += in01_f * s01_s;
287                         out += in10_f * s10_s;
288                         out += in11_f * s11_s;
289 
290                         // Rounding modes of vector and scalar loops should match
291 #if defined(__aarch64__) && !defined(BARE_METAL)
292                         *(out_ptr_xo_yo + cout * sizeof(int8_t)) = quantize_qasymm8_signed(out, oq_info);
293 #else  // defined(__aarch64__) && !defined(BARE_METAL)
294                         *(out_ptr_xo_yo + cout * sizeof(int8_t)) = quantize_qasymm8_signed(out, oq_info, RoundingPolicy::TO_ZERO);
295 #endif // defined(__aarch64__) && !defined(BARE_METAL)
296                     }
297                 }
298             }
299         }
300     }
301     else
302     {
303         ARM_COMPUTE_ERROR("Not implemented");
304     }
305 }
306 }
307 namespace cpu
308 {
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)309 void qasymm8_signed_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
310                                InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
311                                bool align_corners, const Window &window)
312 {
313     if(policy == InterpolationPolicy::BILINEAR)
314     {
315         if(src->info()->quantization_info() == dst->info()->quantization_info() && border_mode == BorderMode::REPLICATE)
316         {
317             s8_neon_scale(src, dst, offsets, dx, dy, policy, border_mode, constant_border_value, sampling_offset, align_corners, window);
318         }
319         else
320         {
321             qasymm8_signed_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
322         }
323     }
324     else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
325     {
326         nearest_neon_scale<int8_t>(src, dst, offsets, sampling_offset, align_corners, window);
327     }
328 }
329 } // namespace cpu
330 } // namespace arm_compute