Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-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_6f6m7l8c8s8r__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 > 6);
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 6 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 const int8_t* i5 = input[5];
80 assert(i5 != NULL);
81 if XNN_UNPREDICTABLE(i5 != zero) {
82 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
83 }
84 input += 6;
85
86 size_t c = round_up_po2(channels, 8);
87
88 for (; c >= 8; c -= 8) {
89 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
90
91
92 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
93 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
94 i0 += 8;
95
96 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
97
98 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
99 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
100 i1 += 8;
101
102 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
103
104 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
105 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
106 i2 += 8;
107
108 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
109
110 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
111 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
112 i3 += 8;
113
114 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
115
116 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
117 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
118 i4 += 8;
119
120 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
121
122 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
123 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
124 i5 += 8;
125
126 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
127
128 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
129
130 _mm256_storeu_si256((__m256i*) b, vacc01234567);
131 b += 8;
132 }
133
134 assert(c == 0);
135 }
136
137 // Middle pass to process 6 inputs in each iteration.
138 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
139 int32_t* b = buffer;
140 const int8_t* i0 = input[0];
141 assert(i0 != NULL);
142 if XNN_UNPREDICTABLE(i0 != zero) {
143 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
144 }
145 const int8_t* i1 = input[1];
146 assert(i1 != NULL);
147 if XNN_UNPREDICTABLE(i1 != zero) {
148 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
149 }
150 const int8_t* i2 = input[2];
151 assert(i2 != NULL);
152 if XNN_UNPREDICTABLE(i2 != zero) {
153 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
154 }
155 const int8_t* i3 = input[3];
156 assert(i3 != NULL);
157 if XNN_UNPREDICTABLE(i3 != zero) {
158 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
159 }
160 const int8_t* i4 = input[4];
161 assert(i4 != NULL);
162 if XNN_UNPREDICTABLE(i4 != zero) {
163 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
164 }
165 const int8_t* i5 = input[5];
166 assert(i5 != NULL);
167 if XNN_UNPREDICTABLE(i5 != zero) {
168 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
169 }
170 input += 6;
171
172 size_t c = round_up_po2(channels, 8);
173
174 for (; c >= 8; c -= 8) {
175 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
176
177
178 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
179 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
180 i0 += 8;
181
182 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
183
184 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
185 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
186 i1 += 8;
187
188 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
189
190 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
191 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
192 i2 += 8;
193
194 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
195
196 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
197 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
198 i3 += 8;
199
200 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
201
202 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
203 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
204 i4 += 8;
205
206 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
207
208 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
209 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
210 i5 += 8;
211
212 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
213
214 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
215
216 _mm256_storeu_si256((__m256i*) b, vacc01234567);
217 b += 8;
218 }
219
220 assert(c == 0);
221 }
222
223 // Last pass to process up to 7 inputs.
224 {
225 const int32_t* b = buffer;
226 const int8_t* i0 = input[0];
227 assert(i0 != NULL);
228 if XNN_UNPREDICTABLE(i0 != zero) {
229 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
230 }
231 const int8_t* i1 = input[1];
232 assert(i1 != NULL);
233 if XNN_UNPREDICTABLE(i1 != zero) {
234 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
235 }
236 const int8_t* i2 = input[2];
237 assert(i2 != NULL);
238 if XNN_UNPREDICTABLE(i2 != zero) {
239 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
240 }
241 const int8_t* i3 = input[3];
242 assert(i3 != NULL);
243 if XNN_UNPREDICTABLE(i3 != zero) {
244 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
245 }
246 const int8_t* i4 = input[4];
247 assert(i4 != NULL);
248 if XNN_UNPREDICTABLE(i4 != zero) {
249 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
250 }
251 const int8_t* i5 = input[5];
252 assert(i5 != NULL);
253 if XNN_UNPREDICTABLE(i5 != zero) {
254 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
255 }
256 const int8_t* i6 = input[6];
257 assert(i6 != NULL);
258 if XNN_UNPREDICTABLE(i6 != zero) {
259 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
260 }
261
262 size_t c = channels;
263
264 for (; c >= 8; c -= 8) {
265 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
266 b += 8;
267
268
269 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
270 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
271 i0 += 8;
272
273 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
274
275 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
276 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
277 i1 += 8;
278
279 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
280
281 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
282 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
283 i2 += 8;
284
285 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
286
287 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
288 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
289 i3 += 8;
290
291 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
292
293 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
294 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
295 i4 += 8;
296
297 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
298
299 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
300 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
301 i5 += 8;
302
303 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
304
305 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
306 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
307 i6 += 8;
308
309 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
310
311 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
312
313 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
314
315 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
316
317 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
318
319 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
320
321 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
322
323 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
324
325 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
326
327 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
328 output += 8;
329 }
330
331 if XNN_UNLIKELY(c != 0) {
332 {
333 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) b);
334 b += 8;
335
336 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
337 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t))));
338
339 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
340
341 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
342 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t))));
343
344 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
345
346 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
347 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t))));
348
349 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
350
351 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
352 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t))));
353
354 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
355
356 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
357 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
358
359 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
360
361 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
362 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t))));
363
364 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
365
366 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
367 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t))));
368
369 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
370
371 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
372 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
373 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
374 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
375
376 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
377
378 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
379
380 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
381
382 if (c & 4) {
383 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
384 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
385 output += 4;
386 }
387 if (c & 2) {
388 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
389 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
390 output += 2;
391 }
392 if (c & 1) {
393 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
394 output += 1;
395 }
396 }
397 }
398 }
399
400 input = (const int8_t**) ((uintptr_t) input + input_stride);
401 output = (int8_t*) ((uintptr_t) output + output_increment);
402 } while (--output_width != 0);
403 }