Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l8c8s8r-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_8f8m9l8c8s8r__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 > 8);
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 8 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 const int8_t* i6 = input[6];
85 assert(i6 != NULL);
86 if XNN_UNPREDICTABLE(i6 != zero) {
87 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
88 }
89 const int8_t* i7 = input[7];
90 assert(i7 != NULL);
91 if XNN_UNPREDICTABLE(i7 != zero) {
92 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
93 }
94 input += 8;
95
96 size_t c = round_up_po2(channels, 8);
97
98 for (; c >= 8; c -= 8) {
99 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
100 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
101
102
103
104 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
105 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
106 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
107 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
108 i0 += 8;
109
110
111 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
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 vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
117 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
118 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
119 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
120 i1 += 8;
121
122
123 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
124
125 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
126 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
127
128 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
129 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
130 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
131 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
132 i2 += 8;
133
134
135 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
136
137 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
138 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
139
140 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
141 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
142 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
143 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
144 i3 += 8;
145
146
147 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
148
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
151
152 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
153 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
154 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
155 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
156 i4 += 8;
157
158
159 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
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 vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
165 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
166 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
167 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
168 i5 += 8;
169
170
171 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
172
173 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
174 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
175
176 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
177 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
178 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
179 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
180 i6 += 8;
181
182
183 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
184
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
187
188 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
189 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
190 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
191 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
192 i7 += 8;
193
194
195 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
196
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
198 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
199
200 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
201
202 _mm_storeu_si128((__m128i*) b, vacc0123);
203 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
204 b += 8;
205 }
206
207 assert(c == 0);
208 }
209
210 // Middle pass to process 8 inputs in each iteration.
211 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
212 int32_t* b = buffer;
213 const int8_t* i0 = input[0];
214 assert(i0 != NULL);
215 if XNN_UNPREDICTABLE(i0 != zero) {
216 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
217 }
218 const int8_t* i1 = input[1];
219 assert(i1 != NULL);
220 if XNN_UNPREDICTABLE(i1 != zero) {
221 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
222 }
223 const int8_t* i2 = input[2];
224 assert(i2 != NULL);
225 if XNN_UNPREDICTABLE(i2 != zero) {
226 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
227 }
228 const int8_t* i3 = input[3];
229 assert(i3 != NULL);
230 if XNN_UNPREDICTABLE(i3 != zero) {
231 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
232 }
233 const int8_t* i4 = input[4];
234 assert(i4 != NULL);
235 if XNN_UNPREDICTABLE(i4 != zero) {
236 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
237 }
238 const int8_t* i5 = input[5];
239 assert(i5 != NULL);
240 if XNN_UNPREDICTABLE(i5 != zero) {
241 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
242 }
243 const int8_t* i6 = input[6];
244 assert(i6 != NULL);
245 if XNN_UNPREDICTABLE(i6 != zero) {
246 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
247 }
248 const int8_t* i7 = input[7];
249 assert(i7 != NULL);
250 if XNN_UNPREDICTABLE(i7 != zero) {
251 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
252 }
253 input += 8;
254
255 size_t c = round_up_po2(channels, 8);
256
257 for (; c >= 8; c -= 8) {
258 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
259 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
260
261
262
263 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
264 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
265 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
266 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
267 i0 += 8;
268
269
270 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
271
272 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
273 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
274
275 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
276 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
277 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
278 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
279 i1 += 8;
280
281
282 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
286
287 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
288 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
289 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
290 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
291 i2 += 8;
292
293
294 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
297 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
298
299 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
300 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
301 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
302 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
303 i3 += 8;
304
305
306 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
307
308 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
309 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
310
311 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
312 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
313 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
314 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
315 i4 += 8;
316
317
318 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
322
323 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
324 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
325 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
326 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
327 i5 += 8;
328
329
330 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
331
332 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
333 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
334
335 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
336 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
337 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
338 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
339 i6 += 8;
340
341
342 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
343
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
345 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
346
347 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
348 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
349 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
350 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
351 i7 += 8;
352
353
354 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
358
359 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
360
361 _mm_storeu_si128((__m128i*) (b), vacc0123);
362 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
363 b += 8;
364 }
365
366 assert(c == 0);
367 }
368
369 // Last pass to process up to 9 inputs.
370 {
371 const int32_t* b = buffer;
372 const int8_t* i0 = input[0];
373 assert(i0 != NULL);
374 if XNN_UNPREDICTABLE(i0 != zero) {
375 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
376 }
377 const int8_t* i1 = input[1];
378 assert(i1 != NULL);
379 if XNN_UNPREDICTABLE(i1 != zero) {
380 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
381 }
382 const int8_t* i2 = input[2];
383 assert(i2 != NULL);
384 if XNN_UNPREDICTABLE(i2 != zero) {
385 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
386 }
387 const int8_t* i3 = input[3];
388 assert(i3 != NULL);
389 if XNN_UNPREDICTABLE(i3 != zero) {
390 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
391 }
392 const int8_t* i4 = input[4];
393 assert(i4 != NULL);
394 if XNN_UNPREDICTABLE(i4 != zero) {
395 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
396 }
397 const int8_t* i5 = input[5];
398 assert(i5 != NULL);
399 if XNN_UNPREDICTABLE(i5 != zero) {
400 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
401 }
402 const int8_t* i6 = input[6];
403 assert(i6 != NULL);
404 if XNN_UNPREDICTABLE(i6 != zero) {
405 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
406 }
407 const int8_t* i7 = input[7];
408 assert(i7 != NULL);
409 if XNN_UNPREDICTABLE(i7 != zero) {
410 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
411 }
412 const int8_t* i8 = input[8];
413 assert(i8 != NULL);
414 if XNN_UNPREDICTABLE(i8 != zero) {
415 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
416 }
417
418 size_t c = channels;
419
420 for (; c >= 8; c -= 8) {
421 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
422 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
423 b += 8;
424
425
426 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
427 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
428 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
429 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
430 i0 += 8;
431
432
433 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
434
435 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
436 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
437
438 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
439 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
440 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
441 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
442 i1 += 8;
443
444
445 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
446
447 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
448 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
449
450 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
451 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
452 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
453 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
454 i2 += 8;
455
456
457 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
458
459 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
460 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
461
462 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
463 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
464 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
465 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
466 i3 += 8;
467
468
469 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
470
471 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
472 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
473
474 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
475 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
476 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
477 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
478 i4 += 8;
479
480
481 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
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 vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
487 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
488 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
489 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
490 i5 += 8;
491
492
493 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
496 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
497
498 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
499 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
500 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
501 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
502 i6 += 8;
503
504
505 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
506
507 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
508 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
509
510 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
511 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
512 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
513 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
514 i7 += 8;
515
516
517 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
518
519 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
520 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
521
522 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
523 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
524 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
525 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
526 i8 += 8;
527
528
529 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
530
531 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
532 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
533
534 w = (const void*) ((uintptr_t) w + 72 * sizeof(int8_t));
535
536 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
537 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
538
539 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
540 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
541
542 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
543 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
544
545 vacc0123 = _mm_cvtps_epi32(vscaled0123);
546 vacc4567 = _mm_cvtps_epi32(vscaled4567);
547
548 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
549
550
551 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
552
553 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
554
555 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
556 output += 8;
557 }
558
559 if XNN_UNLIKELY(c != 0) {
560 {
561 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
562 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
563 b += 8;
564
565
566 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
567 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
568 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
569 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
570
571
572 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
573
574 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
575 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
576
577 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
578 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
579 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
580 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
581
582
583 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
584
585 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
586 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
587
588 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
589 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
590 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
591 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
592
593
594 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
595
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
597 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
598
599 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
600 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
601 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
602 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
603
604
605 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
606
607 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
608 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
609
610 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
611 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
612 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
613 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
614
615
616 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
617
618 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
619 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
620
621 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
622 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
623 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
624 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
625
626
627 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
628
629 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
630 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
631
632 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
633 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
634 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
635 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
636
637
638 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
639
640 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
641 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
642
643 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
644 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
645 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
646 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
647
648
649 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
650
651 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
652 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
653
654 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
655 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
656 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
657 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
658
659
660 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
661
662 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
663 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
664
665 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
666 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
667
668 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
669 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
670
671
672 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
673 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
674
675 vacc0123 = _mm_cvtps_epi32(vscaled0123);
676 vacc4567 = _mm_cvtps_epi32(vscaled4567);
677
678 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
679
680
681 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
682
683 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
684
685 if (c & 4) {
686 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
687 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
688 output += 4;
689 }
690 if (c & 2) {
691 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
692 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
693 output += 2;
694 }
695 if (c & 1) {
696 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
697 output += 1;
698 }
699 }
700 }
701 }
702
703 input = (const int8_t**) ((uintptr_t) input + input_stride);
704 output = (int8_t*) ((uintptr_t) output + output_increment);
705 } while (--output_width != 0);
706 }