Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-minmax-fp32-sse41-mul16.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(
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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
104 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
105
106 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
107 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
108 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
109 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
110 i1 += 8;
111
112
113 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
114
115 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
116 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
117
118 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
119 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
120 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
121 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
122 i2 += 8;
123
124
125 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
126
127 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
128 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
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_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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
152 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
153
154 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
155 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
156 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
157 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
158 i5 += 8;
159
160
161 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
162
163 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
164 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
165
166 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
167
168 _mm_storeu_si128((__m128i*) b, vacc0123);
169 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
170 b += 8;
171 }
172
173 assert(c == 0);
174 }
175
176 // Middle pass to process 6 inputs in each iteration.
177 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
178 int32_t* b = buffer;
179 const int8_t* i0 = input[0];
180 assert(i0 != NULL);
181 if XNN_UNPREDICTABLE(i0 != zero) {
182 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
183 }
184 const int8_t* i1 = input[1];
185 assert(i1 != NULL);
186 if XNN_UNPREDICTABLE(i1 != zero) {
187 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
188 }
189 const int8_t* i2 = input[2];
190 assert(i2 != NULL);
191 if XNN_UNPREDICTABLE(i2 != zero) {
192 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
193 }
194 const int8_t* i3 = input[3];
195 assert(i3 != NULL);
196 if XNN_UNPREDICTABLE(i3 != zero) {
197 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
198 }
199 const int8_t* i4 = input[4];
200 assert(i4 != NULL);
201 if XNN_UNPREDICTABLE(i4 != zero) {
202 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
203 }
204 const int8_t* i5 = input[5];
205 assert(i5 != NULL);
206 if XNN_UNPREDICTABLE(i5 != zero) {
207 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
208 }
209 input += 6;
210
211 size_t c = round_up_po2(channels, 8);
212
213 for (; c >= 8; c -= 8) {
214 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
215 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
216
217
218
219 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
220 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
221 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
222 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
223 i0 += 8;
224
225
226 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
227
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
230
231 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
232 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
233 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
234 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
235 i1 += 8;
236
237
238 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
242
243 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
244 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
245 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
246 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
247 i2 += 8;
248
249
250 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
251
252 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
253 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
254
255 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
256 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
257 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
258 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
259 i3 += 8;
260
261
262 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
263
264 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
265 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
266
267 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
268 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
269 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
270 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
271 i4 += 8;
272
273
274 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
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 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
280 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
281 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
282 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
283 i5 += 8;
284
285
286 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
287
288 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
289 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
290
291 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
292
293 _mm_storeu_si128((__m128i*) (b), vacc0123);
294 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
295 b += 8;
296 }
297
298 assert(c == 0);
299 }
300
301 // Last pass to process up to 7 inputs.
302 {
303 const int32_t* b = buffer;
304 const int8_t* i0 = input[0];
305 assert(i0 != NULL);
306 if XNN_UNPREDICTABLE(i0 != zero) {
307 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
308 }
309 const int8_t* i1 = input[1];
310 assert(i1 != NULL);
311 if XNN_UNPREDICTABLE(i1 != zero) {
312 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
313 }
314 const int8_t* i2 = input[2];
315 assert(i2 != NULL);
316 if XNN_UNPREDICTABLE(i2 != zero) {
317 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
318 }
319 const int8_t* i3 = input[3];
320 assert(i3 != NULL);
321 if XNN_UNPREDICTABLE(i3 != zero) {
322 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
323 }
324 const int8_t* i4 = input[4];
325 assert(i4 != NULL);
326 if XNN_UNPREDICTABLE(i4 != zero) {
327 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
328 }
329 const int8_t* i5 = input[5];
330 assert(i5 != NULL);
331 if XNN_UNPREDICTABLE(i5 != zero) {
332 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
333 }
334 const int8_t* i6 = input[6];
335 assert(i6 != NULL);
336 if XNN_UNPREDICTABLE(i6 != zero) {
337 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
338 }
339
340 size_t c = channels;
341
342 for (; c >= 8; c -= 8) {
343 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
344 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
345 b += 8;
346
347
348 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
349 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
350 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
351 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
352 i0 += 8;
353
354
355 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
356
357 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
358 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
359
360 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
361 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
362 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
363 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
364 i1 += 8;
365
366
367 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
368
369 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
370 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
371
372 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
373 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
374 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
375 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
376 i2 += 8;
377
378
379 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
380
381 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
382 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
383
384 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
385 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
386 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
387 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
388 i3 += 8;
389
390
391 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
392
393 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
394 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
395
396 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
397 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
398 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
399 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
400 i4 += 8;
401
402
403 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
404
405 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
406 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
407
408 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
409 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
410 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
411 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
412 i5 += 8;
413
414
415 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
416
417 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
418 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
419
420 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
421 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
422 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
423 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
424 i6 += 8;
425
426
427 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
428
429 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
430 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
431
432 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
433
434 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
435 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
436
437 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
438 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
439
440 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
441 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
442
443 vacc0123 = _mm_cvtps_epi32(vscaled0123);
444 vacc4567 = _mm_cvtps_epi32(vscaled4567);
445
446 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
447
448
449 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
450
451 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
452
453 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
454 output += 8;
455 }
456
457 if XNN_UNLIKELY(c != 0) {
458 {
459 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
460 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
461 b += 8;
462
463
464 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
465 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
466 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
467 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
468
469
470 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
471
472 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
473 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
474
475 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
476 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
477 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
478 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
479
480
481 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
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 vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
487 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
488 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
489 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
490
491
492 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
493
494 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
495 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
496
497 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
498 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
499 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
500 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
501
502
503 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
504
505 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
506 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
507
508 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
509 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
510 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
511 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
512
513
514 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
518
519 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
520 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
521 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
522 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
523
524
525 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
526
527 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
528 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
529
530 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
531 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
532 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
533 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
534
535
536 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
537
538 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
539 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
540
541 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
542 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
543
544 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
545 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
546
547
548 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
549 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
550
551 vacc0123 = _mm_cvtps_epi32(vscaled0123);
552 vacc4567 = _mm_cvtps_epi32(vscaled4567);
553
554 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
555
556
557 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
558
559 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
560
561 if (c & 4) {
562 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
563 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
564 output += 4;
565 }
566 if (c & 2) {
567 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
568 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
569 output += 2;
570 }
571 if (c & 1) {
572 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
573 output += 1;
574 }
575 }
576 }
577 }
578
579 input = (const int8_t**) ((uintptr_t) input + input_stride);
580 output = (int8_t*) ((uintptr_t) output + output_increment);
581 } while (--output_width != 0);
582 }