Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p8c-minmax-fp32-avx-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p8c__avx_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35
36 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
39 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42
43 do {
44 const int8_t* i0 = input[0];
45 assert(i0 != NULL);
46 if XNN_UNPREDICTABLE(i0 != zero) {
47 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
48 }
49 const int8_t* i1 = input[1];
50 assert(i1 != NULL);
51 if XNN_UNPREDICTABLE(i1 != zero) {
52 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
53 }
54 const int8_t* i2 = input[2];
55 assert(i2 != NULL);
56 if XNN_UNPREDICTABLE(i2 != zero) {
57 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
58 }
59 const int8_t* i3 = input[3];
60 assert(i3 != NULL);
61 if XNN_UNPREDICTABLE(i3 != zero) {
62 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
63 }
64 const int8_t* i4 = input[4];
65 assert(i4 != NULL);
66 if XNN_UNPREDICTABLE(i4 != zero) {
67 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
68 }
69 const int8_t* i5 = input[5];
70 assert(i5 != NULL);
71 if XNN_UNPREDICTABLE(i5 != zero) {
72 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
73 }
74 const int8_t* i6 = input[6];
75 assert(i6 != NULL);
76 if XNN_UNPREDICTABLE(i6 != zero) {
77 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
78 }
79 const int8_t* i7 = input[7];
80 assert(i7 != NULL);
81 if XNN_UNPREDICTABLE(i7 != zero) {
82 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
83 }
84 const int8_t* i8 = input[8];
85 assert(i8 != NULL);
86 if XNN_UNPREDICTABLE(i8 != zero) {
87 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
88 }
89 input = (const int8_t**) ((uintptr_t) input + input_stride);
90
91 size_t c = channels;
92 const void* w = weights;
93 for (; c >= 8; c -= 8) {
94 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
95 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
96
97
98 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
99 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
100 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
101 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
102 i0 += 8;
103
104
105 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
106
107
108 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
109 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
110 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
111 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
112 i1 += 8;
113
114
115 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
116
117 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
118 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
119
120 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
121 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
122 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
123 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
124 i2 += 8;
125
126
127 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
128
129
130 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
131 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
132 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
133 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
134 i3 += 8;
135
136
137 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
138
139 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
140 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
141
142 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
143 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
144 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
145 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
146 i4 += 8;
147
148
149 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
150
151
152 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
153 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
154 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
155 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
156 i5 += 8;
157
158
159 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
160
161 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
162 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
163
164 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
165 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
166 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
167 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
168 i6 += 8;
169
170
171 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
172
173
174 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
175 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
176 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
177 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
178 i7 += 8;
179
180
181 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
182
183 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
184 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
185
186 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
187 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
188 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
189 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
190 i8 += 8;
191
192
193 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
194
195 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
196 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
197
198 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
199
200 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
201 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
202
203 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
204 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
205
206 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
207 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
208
209 vacc0123 = _mm_cvtps_epi32(vscaled0123);
210 vacc4567 = _mm_cvtps_epi32(vscaled4567);
211
212 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
213
214
215 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
216
217 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
218
219 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
220 output += 8;
221 }
222 if XNN_UNLIKELY(c != 0) {
223 {
224 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
225 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
226
227
228 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
229 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
230 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
231 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
232
233
234 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
235
236
237 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
238 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
239 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
240 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
241
242
243 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
244
245 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
246 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
247
248 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
249 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
250 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
251 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
252
253
254 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
255
256
257 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
258 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
259 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
260 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
261
262
263 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
264
265 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
266 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
267
268 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
269 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
270 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
271 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
272
273
274 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
275
276
277 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
278 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
279 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
280 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
281
282
283 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
284
285 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
286 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
287
288 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
289 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
290 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
291 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
292
293
294 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
295
296
297 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
298 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
299 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
300 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
301
302
303 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
304
305 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
306 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
307
308 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
309 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
310 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
311 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
312
313
314 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
315
316 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
317 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
318
319
320 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
321 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
322
323 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
324 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
325
326 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
327 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
328
329 vacc0123 = _mm_cvtps_epi32(vscaled0123);
330 vacc4567 = _mm_cvtps_epi32(vscaled4567);
331
332
333 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
334
335
336 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
337
338 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
339
340 if (c & 4) {
341 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
342 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
343 output += 4;
344 }
345 if (c & 2) {
346 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
347 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
348 output += 2;
349 }
350 if (c & 1) {
351 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
352 output += 1;
353 }
354 }
355 }
356
357 output = (int8_t*) ((uintptr_t) output + output_increment);
358 } while (--output_width != 0);
359 }