Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-minmax-fp32-sse41-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul16.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 <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_6f6m7l8c8s8r__sse41_mul16_add16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 6);
36
37
38 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
39 XNN_FORCE_REALIZATION(vscale);
40
41 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);
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 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
90 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
91
92
93
94 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
95 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
96 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
97 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
98 i0 += 8;
99
100
101 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
102
103
104 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
105 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
106 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
107 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
108 i1 += 8;
109
110
111 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
112
113 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
114 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
115
116 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
117 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
118 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
119 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
120 i2 += 8;
121
122
123 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
124
125
126 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
127 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
128 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
129 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
130 i3 += 8;
131
132
133 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
134
135 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
136 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
137
138 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
139 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
140 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
141 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
142 i4 += 8;
143
144
145 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
146
147
148 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
149 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
150 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
152 i5 += 8;
153
154
155 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
156
157 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
158 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
159
160 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
161
162 _mm_storeu_si128((__m128i*) b, vacc0123);
163 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
164 b += 8;
165 }
166
167 assert(c == 0);
168 }
169
170 // Middle pass to process 6 inputs in each iteration.
171 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
172 int32_t* b = buffer;
173 const int8_t* i0 = input[0];
174 assert(i0 != NULL);
175 if XNN_UNPREDICTABLE(i0 != zero) {
176 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
177 }
178 const int8_t* i1 = input[1];
179 assert(i1 != NULL);
180 if XNN_UNPREDICTABLE(i1 != zero) {
181 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
182 }
183 const int8_t* i2 = input[2];
184 assert(i2 != NULL);
185 if XNN_UNPREDICTABLE(i2 != zero) {
186 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
187 }
188 const int8_t* i3 = input[3];
189 assert(i3 != NULL);
190 if XNN_UNPREDICTABLE(i3 != zero) {
191 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
192 }
193 const int8_t* i4 = input[4];
194 assert(i4 != NULL);
195 if XNN_UNPREDICTABLE(i4 != zero) {
196 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
197 }
198 const int8_t* i5 = input[5];
199 assert(i5 != NULL);
200 if XNN_UNPREDICTABLE(i5 != zero) {
201 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
202 }
203 input += 6;
204
205 size_t c = round_up_po2(channels, 8);
206
207 for (; c >= 8; c -= 8) {
208 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
209 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
210
211
212
213 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
214 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
215 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
216 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
217 i0 += 8;
218
219
220 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
221
222
223 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
224 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
225 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
226 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
227 i1 += 8;
228
229
230 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
231
232 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
233 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
234
235 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
236 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
237 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
238 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
239 i2 += 8;
240
241
242 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
243
244
245 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
246 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
247 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
248 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
249 i3 += 8;
250
251
252 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
253
254 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
255 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
256
257 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
258 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
259 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
260 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
261 i4 += 8;
262
263
264 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
265
266
267 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
268 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
269 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
270 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
271 i5 += 8;
272
273
274 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
275
276 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
277 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
278
279 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
280
281 _mm_storeu_si128((__m128i*) (b), vacc0123);
282 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
283 b += 8;
284 }
285
286 assert(c == 0);
287 }
288
289 // Last pass to process up to 7 inputs.
290 {
291 const int32_t* b = buffer;
292 const int8_t* i0 = input[0];
293 assert(i0 != NULL);
294 if XNN_UNPREDICTABLE(i0 != zero) {
295 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
296 }
297 const int8_t* i1 = input[1];
298 assert(i1 != NULL);
299 if XNN_UNPREDICTABLE(i1 != zero) {
300 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
301 }
302 const int8_t* i2 = input[2];
303 assert(i2 != NULL);
304 if XNN_UNPREDICTABLE(i2 != zero) {
305 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
306 }
307 const int8_t* i3 = input[3];
308 assert(i3 != NULL);
309 if XNN_UNPREDICTABLE(i3 != zero) {
310 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
311 }
312 const int8_t* i4 = input[4];
313 assert(i4 != NULL);
314 if XNN_UNPREDICTABLE(i4 != zero) {
315 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
316 }
317 const int8_t* i5 = input[5];
318 assert(i5 != NULL);
319 if XNN_UNPREDICTABLE(i5 != zero) {
320 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
321 }
322 const int8_t* i6 = input[6];
323 assert(i6 != NULL);
324 if XNN_UNPREDICTABLE(i6 != zero) {
325 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
326 }
327
328 size_t c = channels;
329
330 for (; c >= 8; c -= 8) {
331 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
332 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
333 b += 8;
334
335
336 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
337 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
338 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
339 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
340 i0 += 8;
341
342
343 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
344
345
346 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
347 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
348 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
349 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
350 i1 += 8;
351
352
353 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
354
355 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
356 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
357
358 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
359 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
360 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
361 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
362 i2 += 8;
363
364
365 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
366
367
368 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
369 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
370 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
371 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
372 i3 += 8;
373
374
375 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
376
377 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
378 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
379
380 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
381 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
382 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
383 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
384 i4 += 8;
385
386
387 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
388
389
390 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
391 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
392 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
393 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
394 i5 += 8;
395
396
397 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
398
399 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
400 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
401
402 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
403 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
404 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
405 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
406 i6 += 8;
407
408
409 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
410
411 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
412 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
413
414 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
415
416 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
417 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
418
419 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
420 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
421
422 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
423 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
424
425 vacc0123 = _mm_cvtps_epi32(vscaled0123);
426 vacc4567 = _mm_cvtps_epi32(vscaled4567);
427
428 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
429
430
431 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
432
433 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
434
435 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
436 output += 8;
437 }
438
439 if XNN_UNLIKELY(c != 0) {
440 {
441 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
442 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
443 b += 8;
444
445
446 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
447 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
448 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
449 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
450
451
452 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
453
454
455 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
456 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
457 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
458 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
459
460
461 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
462
463 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
464 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
465
466 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
467 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
468 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
469 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
470
471
472 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
473
474
475 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
476 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
477 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
478 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
479
480
481 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
482
483 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
484 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
485
486 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
487 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
488 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
489 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
490
491
492 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
493
494
495 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
496 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
497 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
498 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
499
500
501 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
502
503 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
504 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
505
506 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
507 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
508 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
509 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
510
511
512 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
513
514 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
515 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
516
517 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
518 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
519
520 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
521 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
522
523
524 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
525 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
526
527 vacc0123 = _mm_cvtps_epi32(vscaled0123);
528 vacc4567 = _mm_cvtps_epi32(vscaled4567);
529
530 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
531
532
533 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
534
535 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
536
537 if (c & 4) {
538 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
539 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
540 output += 4;
541 }
542 if (c & 2) {
543 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
544 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
545 output += 2;
546 }
547 if (c & 1) {
548 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
549 output += 1;
550 }
551 }
552 }
553 }
554
555 input = (const int8_t**) ((uintptr_t) input + input_stride);
556 output = (int8_t*) ((uintptr_t) output + output_increment);
557 } while (--output_width != 0);
558 }