Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l8c8s8r-minmax-fp32-avx2-mul32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-avx2-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l8c8s8r__avx2_mul32(
21 size_t channels,
22 size_t output_width,
23 const int8_t** input,
24 const void* weights,
25 int8_t* output,
26 intptr_t input_stride,
27 size_t output_increment,
28 size_t input_offset,
29 const int8_t* zero,
30 size_t kernel_size,
31 int32_t* buffer,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36 assert(kernel_size > 5);
37
38
39 const __m256 vscale = _mm256_set1_ps(params->fp32_scalar.scale);
40 XNN_FORCE_REALIZATION(vscale);
41 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
42 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 5 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 input += 5;
80
81 size_t c = round_up_po2(channels, 8);
82
83 for (; c >= 8; c -= 8) {
84 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
85
86
87 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
88 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
89 i0 += 8;
90
91 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
92
93 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
94 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
95 i1 += 8;
96
97 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
98
99 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
100 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
101 i2 += 8;
102
103 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
104
105 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
106 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
107 i3 += 8;
108
109 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
110
111 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
112 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
113 i4 += 8;
114
115 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
116
117 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
118
119 _mm256_storeu_si256((__m256i*) b, vacc01234567);
120 b += 8;
121 }
122
123 assert(c == 0);
124 }
125
126 // Middle pass to process 5 inputs in each iteration.
127 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
128 int32_t* b = buffer;
129 const int8_t* i0 = input[0];
130 assert(i0 != NULL);
131 if XNN_UNPREDICTABLE(i0 != zero) {
132 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
133 }
134 const int8_t* i1 = input[1];
135 assert(i1 != NULL);
136 if XNN_UNPREDICTABLE(i1 != zero) {
137 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
138 }
139 const int8_t* i2 = input[2];
140 assert(i2 != NULL);
141 if XNN_UNPREDICTABLE(i2 != zero) {
142 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
143 }
144 const int8_t* i3 = input[3];
145 assert(i3 != NULL);
146 if XNN_UNPREDICTABLE(i3 != zero) {
147 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
148 }
149 const int8_t* i4 = input[4];
150 assert(i4 != NULL);
151 if XNN_UNPREDICTABLE(i4 != zero) {
152 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
153 }
154 input += 5;
155
156 size_t c = round_up_po2(channels, 8);
157
158 for (; c >= 8; c -= 8) {
159 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
160
161
162 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
163 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
164 i0 += 8;
165
166 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
167
168 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
169 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
170 i1 += 8;
171
172 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
173
174 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
175 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
176 i2 += 8;
177
178 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
179
180 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
181 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
182 i3 += 8;
183
184 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
185
186 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
187 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
188 i4 += 8;
189
190 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
191
192 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
193
194 _mm256_storeu_si256((__m256i*) b, vacc01234567);
195 b += 8;
196 }
197
198 assert(c == 0);
199 }
200
201 // Last pass to process up to 5 inputs.
202 {
203 const int32_t* b = buffer;
204 const int8_t* i0 = input[0];
205 assert(i0 != NULL);
206 if XNN_UNPREDICTABLE(i0 != zero) {
207 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
208 }
209 const int8_t* i1 = input[1];
210 assert(i1 != NULL);
211 if XNN_UNPREDICTABLE(i1 != zero) {
212 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
213 }
214 const int8_t* i2 = input[2];
215 assert(i2 != NULL);
216 if XNN_UNPREDICTABLE(i2 != zero) {
217 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
218 }
219 const int8_t* i3 = input[3];
220 assert(i3 != NULL);
221 if XNN_UNPREDICTABLE(i3 != zero) {
222 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
223 }
224 const int8_t* i4 = input[4];
225 assert(i4 != NULL);
226 if XNN_UNPREDICTABLE(i4 != zero) {
227 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
228 }
229
230 size_t c = channels;
231
232 for (; c >= 8; c -= 8) {
233 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
234 b += 8;
235
236
237 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
238 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
239 i0 += 8;
240
241 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
242
243 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
244 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
245 i1 += 8;
246
247 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
248
249 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
250 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
251 i2 += 8;
252
253 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
254
255 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
256 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
257 i3 += 8;
258
259 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
260
261 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
262 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
263 i4 += 8;
264
265 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
266
267 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
268
269 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
270
271 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
272
273 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
274
275 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
276
277 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
278
279 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
280
281 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
282
283 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
284 output += 8;
285 }
286
287 if XNN_UNLIKELY(c != 0) {
288 {
289 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
290 b += 8;
291
292 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
293 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
294
295 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
296
297 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
298 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
299
300 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
301
302 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
303 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
304
305 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
306
307 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
308 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
309
310 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
311
312 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
313 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
314
315 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
316
317 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
318 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
319 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
320 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
321
322 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
323
324 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
325
326 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
327
328 if (c & 4) {
329 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
330 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
331 output += 4;
332 }
333 if (c & 2) {
334 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
335 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
336 output += 2;
337 }
338 if (c & 1) {
339 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
340 output += 1;
341 }
342 }
343 }
344 }
345
346 input = (const int8_t**) ((uintptr_t) input + input_stride);
347 output = (int8_t*) ((uintptr_t) output + output_increment);
348 } while (--output_width != 0);
349 }