xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/pool3d/neon/quantized.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 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 #ifndef SRC_CORE_NEON_KERNELS_POOL3D_QUANTIZED_H
25 #define SRC_CORE_NEON_KERNELS_POOL3D_QUANTIZED_H
26 
27 #include "arm_compute/core/ITensor.h"
28 #include "arm_compute/core/Types.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30 #include "src/core/helpers/PoolingHelpers.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 
33 namespace arm_compute
34 {
35 namespace cpu
36 {
37 template <typename T>
avg_poolingMxNxD_q8_neon_ndhwc(const ITensor * src,ITensor * dst0,Pooling3dLayerInfo & pool_info,const Window & window_out,const int window_step_x)38 void avg_poolingMxNxD_q8_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window_out,
39                                     const int window_step_x)
40 
41 {
42     using q8x8_t  = typename wrapper::traits::neon_vector<T, 8>::type;
43     using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type;
44     using q16_t   = typename wrapper::traits::promote_t<T>;
45     using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type;
46     using q32_t   = typename wrapper::traits::promote_t<q16_t>;
47     using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type;
48 
49     int pool_stride_x = static_cast<int>(pool_info.stride.width);
50     int pool_stride_y = static_cast<int>(pool_info.stride.height);
51     int pool_stride_z = static_cast<int>(pool_info.stride.depth);
52 
53     const int pool_size_x = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.width;
54     const int pool_size_y = pool_info.is_global_pooling ? src->info()->tensor_shape().z() : pool_info.pool_size.height;
55     const int pool_size_z = pool_info.is_global_pooling ? src->info()->tensor_shape()[3] : pool_info.pool_size.depth;
56 
57     const int pool_pad_top    = static_cast<int>(pool_info.padding.top);
58     const int pool_pad_bottom = static_cast<int>(pool_info.padding.bottom);
59     const int pool_pad_left   = static_cast<int>(pool_info.padding.left);
60     const int pool_pad_right  = static_cast<int>(pool_info.padding.right);
61     const int pool_pad_front  = static_cast<int>(pool_info.padding.front);
62     const int pool_pad_back   = static_cast<int>(pool_info.padding.back);
63 
64     const int upper_bound_w = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_right);
65     const int upper_bound_h = src->info()->dimension(2) + (pool_info.exclude_padding ? 0 : pool_pad_bottom);
66     const int upper_bound_d = src->info()->dimension(3) + (pool_info.exclude_padding ? 0 : pool_pad_back);
67 
68     const int input_dim_c = src->info()->dimension(0);
69     const int input_dim_w = src->info()->dimension(1);
70     const int input_dim_h = src->info()->dimension(2);
71     const int input_dim_d = src->info()->dimension(3);
72 
73     const int y_stride = static_cast<int>(src->info()->strides_in_bytes().y());
74     const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z());
75     const int w_stride = static_cast<int>(src->info()->strides_in_bytes()[3]);
76     const int n_stride = static_cast<int>(src->info()->strides_in_bytes()[4]);
77 
78     const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
79 
80     const int window_end_x   = input_dim_c;
81     const int window_start_x = 0;
82 
83     Iterator out(dst0, window_out);
84 
85     const float32x4_t             half_scale_v = vdupq_n_f32(0.5f);
86     const UniformQuantizationInfo src_qinfo    = src->info()->quantization_info().uniform();
87     const UniformQuantizationInfo dst_qinfo    = dst0->info()->quantization_info().uniform();
88 
89     const float quant_rescale = dst_qinfo.scale / src_qinfo.scale;
90     // "new_offset" doesn't have to consider the "half_scale_v" in its computation
91     // With a requantization performed in a single step there won't be uncertainties introduced
92     const int32_t new_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / quant_rescale);
93 
94     execute_window_loop(window_out, [&](const Coordinates & id)
95     {
96         // Computing the theoretical input starting/ending points
97         const int in_idx_width  = static_cast<int>(id.y()) * pool_stride_x - pool_pad_left;
98         const int in_idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top;
99         const int in_idx_depth  = static_cast<int>(id[3]) * pool_stride_z - pool_pad_front;
100 
101         const int pool_start_x = std::max(0, -in_idx_width);
102         const int pool_end_x_t = std::min(input_dim_w + pool_pad_left - in_idx_width, pool_size_x);
103         const int pool_start_y = std::max(0, -in_idx_height);
104         const int pool_end_y_t = std::min(input_dim_h + pool_pad_top - in_idx_height, pool_size_y);
105 
106         const int pool_start_z = std::max(0, -in_idx_depth);
107         const int pool_end_z_t = std::min(input_dim_d + pool_pad_front - in_idx_depth, pool_size_z);
108 
109         // The end of width to consider in calculation should exclude PAD_X, PAD_Y and PAD_Z
110         const int pool_end_x = std::min(pool_end_x_t, input_dim_w - in_idx_width);
111         const int pool_end_y = std::min(pool_end_y_t, input_dim_h - in_idx_height);
112         const int pool_end_z = std::min(pool_end_z_t, input_dim_d - in_idx_depth);
113 
114         // Calculate scale
115         const float scale = calculate_avg_scale_pool3d(pool_info.exclude_padding, id, pool_size_x, pool_size_y, pool_size_z, upper_bound_w, upper_bound_h, upper_bound_d, pool_pad_left,
116                                                        pool_pad_top, pool_pad_front, pool_stride_x, pool_stride_y, pool_stride_z);
117 
118         const uint8_t *in_ptr_n = in_ptr_start + id[4] * n_stride;
119 
120         int x_off = window_start_x;
121 
122         for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) // C
123         {
124             q32x4_t vres1 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{});
125             q32x4_t vres2 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{});
126             q32x4_t vres3 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{});
127             q32x4_t vres4 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{});
128 
129             // Perform pooling
130             for(int z = pool_start_z; z < pool_end_z; ++z)
131             {
132                 const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride;
133                 for(int y = pool_start_y; y < pool_end_y; ++y)
134                 {
135                     const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride;
136                     for(int x = pool_start_x; x < pool_end_x; ++x)
137                     {
138                         const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride;
139                         const q8x16_t  data     = wrapper::vloadq(reinterpret_cast<const T *>(in_ptr_x) + x_off);
140 
141                         const q16x8_t data_q16  = wrapper::vmovl(wrapper::vgetlow(data));
142                         const q16x8_t data2_q16 = wrapper::vmovl(wrapper::vgethigh(data));
143                         vres1                   = wrapper::vadd(vres1, wrapper::vmovl(wrapper::vgetlow(data_q16)));
144                         vres2                   = wrapper::vadd(vres2, wrapper::vmovl(wrapper::vgethigh(data_q16)));
145                         vres3                   = wrapper::vadd(vres3, wrapper::vmovl(wrapper::vgetlow(data2_q16)));
146                         vres4                   = wrapper::vadd(vres4, wrapper::vmovl(wrapper::vgethigh(data2_q16)));
147                     }
148                 }
149             }
150 
151             if(src_qinfo != dst_qinfo)
152             {
153                 const float32x4x4_t vres =
154                 {
155                     {
156                         vcvtq_f32_q32(vres1),
157                         vcvtq_f32_q32(vres2),
158                         vcvtq_f32_q32(vres3),
159                         vcvtq_f32_q32(vres4),
160                     }
161                 };
162                 const auto requantized_dst = vrequantize_pooling_with_scale<q8x16_t>(vres, quant_rescale, scale, new_offset);
163                 // Store result
164                 wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, wrapper::vgetlow(requantized_dst));
165                 wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off + 8, wrapper::vgethigh(requantized_dst));
166             }
167             else
168             {
169                 const float32x4_t scale_v = vdupq_n_f32(scale);
170                 // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero
171                 vres1 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v));
172                 vres2 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v));
173                 vres3 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v));
174                 vres4 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres4), scale_v));
175 
176                 const q8x8_t res1 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres1), wrapper::vmovn(vres2)));
177                 const q8x8_t res2 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres3), wrapper::vmovn(vres4)));
178                 // Store result
179                 wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, res1);
180                 wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off + 8, res2);
181             }
182         }
183 
184         // Left-overs loop
185         for(; x_off < window_end_x; ++x_off)
186         {
187             q32_t res = static_cast<q32_t>(0.f);
188 
189             // Perform pooling
190             for(int z = pool_start_z; z < pool_end_z; ++z)
191             {
192                 const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride;
193                 for(int y = pool_start_y; y < pool_end_y; ++y)
194                 {
195                     const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride;
196                     for(int x = pool_start_x; x < pool_end_x; ++x)
197                     {
198                         const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride;
199                         const T        data     = *(reinterpret_cast<const T *>(in_ptr_x) + x_off);
200                         res += data;
201                     }
202                 }
203             }
204 
205             if(src_qinfo != dst_qinfo)
206             {
207                 const float res_f           = static_cast<float>(res);
208                 const float new_scale       = quant_rescale / scale;
209                 const auto  requantized_dst = quantize<T>(res_f, UniformQuantizationInfo(new_scale, new_offset));
210 
211                 // Store result
212                 *(reinterpret_cast<T *>(out.ptr()) + x_off) = requantized_dst;
213             }
214             else
215             {
216                 // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero
217                 res = static_cast<T>(0.5f + static_cast<float>(res) * scale);
218 
219                 // Store result
220                 *(reinterpret_cast<T *>(out.ptr()) + x_off) = res;
221             }
222         }
223     },
224     out);
225 }
226 
227 template <typename T>
max_poolingMxNxD_q8_neon_ndhwc(const ITensor * src,ITensor * dst0,Pooling3dLayerInfo & pool_info,const Window & window_out,const int window_step_x)228 void max_poolingMxNxD_q8_neon_ndhwc(const ITensor *src, ITensor *dst0, Pooling3dLayerInfo &pool_info, const Window &window_out,
229                                     const int window_step_x)
230 
231 {
232     using q8x8_t  = typename wrapper::traits::neon_vector<T, 8>::type;
233     using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type;
234 
235     const int window_half_step_x = window_step_x / 2;
236 
237     int pool_stride_x = static_cast<int>(pool_info.stride.width);
238     int pool_stride_y = static_cast<int>(pool_info.stride.height);
239     int pool_stride_z = static_cast<int>(pool_info.stride.depth);
240 
241     const int pool_size_x = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.width;
242     const int pool_size_y = pool_info.is_global_pooling ? src->info()->tensor_shape().z() : pool_info.pool_size.height;
243     const int pool_size_z = pool_info.is_global_pooling ? src->info()->tensor_shape()[3] : pool_info.pool_size.depth;
244 
245     const int pool_pad_top   = static_cast<int>(pool_info.padding.top);
246     const int pool_pad_left  = static_cast<int>(pool_info.padding.left);
247     const int pool_pad_front = static_cast<int>(pool_info.padding.front);
248 
249     const int input_dim_c = src->info()->dimension(0);
250     const int input_dim_w = src->info()->dimension(1);
251     const int input_dim_h = src->info()->dimension(2);
252     const int input_dim_d = src->info()->dimension(3);
253 
254     const int y_stride = static_cast<int>(src->info()->strides_in_bytes().y());
255     const int z_stride = static_cast<int>(src->info()->strides_in_bytes().z());
256     const int w_stride = static_cast<int>(src->info()->strides_in_bytes()[3]);
257     const int n_stride = static_cast<int>(src->info()->strides_in_bytes()[4]);
258 
259     const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
260 
261     const int window_end_x   = input_dim_c;
262     const int window_start_x = 0;
263 
264     Iterator out(dst0, window_out);
265 
266     const UniformQuantizationInfo src_qinfo = src->info()->quantization_info().uniform();
267     const UniformQuantizationInfo dst_qinfo = dst0->info()->quantization_info().uniform();
268 
269     const float                   requant_scale  = dst_qinfo.scale / src_qinfo.scale;
270     const int32_t                 requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale);
271     const UniformQuantizationInfo requant_qinfo  = UniformQuantizationInfo(requant_scale, requant_offset);
272 
273     execute_window_loop(window_out, [&](const Coordinates & id)
274     {
275         // Computing the theoretical input starting/ending points
276         const int in_idx_width  = static_cast<int>(id.y()) * pool_stride_x - pool_pad_left;
277         const int in_idx_height = static_cast<int>(id.z()) * pool_stride_y - pool_pad_top;
278         const int in_idx_depth  = static_cast<int>(id[3]) * pool_stride_z - pool_pad_front;
279 
280         const int pool_start_x = std::max(0, -in_idx_width);
281         const int pool_end_x_t = std::min(input_dim_w + pool_pad_left - in_idx_width, pool_size_x);
282         const int pool_start_y = std::max(0, -in_idx_height);
283         const int pool_end_y_t = std::min(input_dim_h + pool_pad_top - in_idx_height, pool_size_y);
284 
285         const int pool_start_z = std::max(0, -in_idx_depth);
286         const int pool_end_z_t = std::min(input_dim_d + pool_pad_front - in_idx_depth, pool_size_z);
287 
288         // The end of width to consider in calculation should exclude PAD_X, PAD_Y and PAD_Z
289         const int pool_end_x = std::min(pool_end_x_t, input_dim_w - in_idx_width);
290         const int pool_end_y = std::min(pool_end_y_t, input_dim_h - in_idx_height);
291         const int pool_end_z = std::min(pool_end_z_t, input_dim_d - in_idx_depth);
292 
293         const uint8_t *in_ptr_n = in_ptr_start + id[4] * n_stride;
294 
295         int x_off = window_start_x;
296 
297         for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) // C
298         {
299             q8x16_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_128_tag{});
300 
301             // Perform pooling
302             for(int z = pool_start_z; z < pool_end_z; ++z)
303             {
304                 const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride;
305                 for(int y = pool_start_y; y < pool_end_y; ++y)
306                 {
307                     const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride;
308                     for(int x = pool_start_x; x < pool_end_x; ++x)
309                     {
310                         const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride;
311                         const q8x16_t  data     = wrapper::vloadq(reinterpret_cast<const T *>(in_ptr_x) + x_off);
312 
313                         vres = wrapper::vmax(vres, data);
314                     }
315                 }
316             }
317 
318             // Store result
319             wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off, (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(vres), wrapper::vgethigh(vres),
320                             requant_qinfo) :
321                             vres);
322         }
323 
324         // Leftovers using half the window step
325         for(; x_off <= (window_end_x - window_half_step_x); x_off += window_half_step_x)
326         {
327             q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{});
328 
329             // Perform pooling
330             for(int z = pool_start_z; z < pool_end_z; ++z)
331             {
332                 const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride;
333                 for(int y = pool_start_y; y < pool_end_y; ++y)
334                 {
335                     const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride;
336                     for(int x = pool_start_x; x < pool_end_x; ++x)
337                     {
338                         const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride;
339                         const q8x8_t   data     = wrapper::vload(reinterpret_cast<const T *>(in_ptr_x) + x_off);
340 
341                         vres = wrapper::vmax(vres, data);
342                     }
343                 }
344             }
345 
346             // Store result
347             wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x_off,
348                             (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t>(vres, requant_qinfo) : vres);
349         }
350 
351         // Left-overs loop
352         for(; x_off < window_end_x; ++x_off)
353         {
354             T res = std::numeric_limits<T>::min();
355 
356             for(int z = pool_start_z; z < pool_end_z; ++z)
357             {
358                 const uint8_t *in_ptr_z = in_ptr_n + (z + in_idx_depth) * w_stride;
359                 for(int y = pool_start_y; y < pool_end_y; ++y)
360                 {
361                     const uint8_t *in_ptr_y = in_ptr_z + (y + in_idx_height) * z_stride;
362                     for(int x = pool_start_x; x < pool_end_x; ++x)
363                     {
364                         const uint8_t *in_ptr_x = in_ptr_y + (x + in_idx_width) * y_stride;
365                         const T        data     = *(reinterpret_cast<const T *>(in_ptr_x) + x_off);
366 
367                         res = std::max(res, data);
368                     }
369                 }
370             }
371 
372             // Store result
373             if(src_qinfo != dst_qinfo)
374             {
375                 const float res_f                           = static_cast<float>(res);
376                 *(reinterpret_cast<T *>(out.ptr()) + x_off) = quantize<T>(res_f, requant_qinfo);
377             }
378             else
379             {
380                 *(reinterpret_cast<T *>(out.ptr()) + x_off) = res;
381             }
382         }
383     },
384     out);
385 }
386 
387 } // namespace cpu
388 } // namespace arm_compute
389 
390 #endif // SRC_CORE_NEON_KERNELS_POOL3D_QUANTIZED_H