Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p8c-minmax-fp32-sse41-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_25p8c__sse41_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 const int8_t* i9 = input[9];
90 assert(i9 != NULL);
91 if XNN_UNPREDICTABLE(i9 != zero) {
92 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
93 }
94 const int8_t* i10 = input[10];
95 assert(i10 != NULL);
96 if XNN_UNPREDICTABLE(i10 != zero) {
97 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
98 }
99 const int8_t* i11 = input[11];
100 assert(i11 != NULL);
101 if XNN_UNPREDICTABLE(i11 != zero) {
102 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
103 }
104 const int8_t* i12 = input[12];
105 assert(i12 != NULL);
106 if XNN_UNPREDICTABLE(i12 != zero) {
107 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
108 }
109 const int8_t* i13 = input[13];
110 assert(i13 != NULL);
111 if XNN_UNPREDICTABLE(i13 != zero) {
112 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
113 }
114 const int8_t* i14 = input[14];
115 assert(i14 != NULL);
116 if XNN_UNPREDICTABLE(i14 != zero) {
117 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
118 }
119 const int8_t* i15 = input[15];
120 assert(i15 != NULL);
121 if XNN_UNPREDICTABLE(i15 != zero) {
122 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
123 }
124 const int8_t* i16 = input[16];
125 assert(i16 != NULL);
126 if XNN_UNPREDICTABLE(i16 != zero) {
127 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
128 }
129 const int8_t* i17 = input[17];
130 assert(i17 != NULL);
131 if XNN_UNPREDICTABLE(i17 != zero) {
132 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
133 }
134 const int8_t* i18 = input[18];
135 assert(i18 != NULL);
136 if XNN_UNPREDICTABLE(i18 != zero) {
137 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
138 }
139 const int8_t* i19 = input[19];
140 assert(i19 != NULL);
141 if XNN_UNPREDICTABLE(i19 != zero) {
142 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
143 }
144 const int8_t* i20 = input[20];
145 assert(i20 != NULL);
146 if XNN_UNPREDICTABLE(i20 != zero) {
147 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
148 }
149 const int8_t* i21 = input[21];
150 assert(i21 != NULL);
151 if XNN_UNPREDICTABLE(i21 != zero) {
152 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
153 }
154 const int8_t* i22 = input[22];
155 assert(i22 != NULL);
156 if XNN_UNPREDICTABLE(i22 != zero) {
157 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
158 }
159 const int8_t* i23 = input[23];
160 assert(i23 != NULL);
161 if XNN_UNPREDICTABLE(i23 != zero) {
162 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
163 }
164 const int8_t* i24 = input[24];
165 assert(i24 != NULL);
166 if XNN_UNPREDICTABLE(i24 != zero) {
167 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
168 }
169 input = (const int8_t**) ((uintptr_t) input + input_stride);
170
171 size_t c = channels;
172 const void* w = weights;
173 for (; c >= 8; c -= 8) {
174 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
175 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
176
177
178 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
179 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
180 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
181 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
182 i0 += 8;
183
184
185 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
186
187
188 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
189 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
190 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
191 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
192 i1 += 8;
193
194
195 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
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 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
201 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
202 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
203 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
204 i2 += 8;
205
206
207 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
208
209
210 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
211 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
212 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
213 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
214 i3 += 8;
215
216
217 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
218
219 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
220 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
221
222 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
223 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
224 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
225 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
226 i4 += 8;
227
228
229 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
230
231
232 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
233 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
234 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
235 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
236 i5 += 8;
237
238
239 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
240
241 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
242 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
243
244 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
245 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
246 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
247 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
248 i6 += 8;
249
250
251 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
252
253
254 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
255 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
256 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
257 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
258 i7 += 8;
259
260
261 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
262
263 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
264 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
265
266 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
267 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
268 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
269 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
270 i8 += 8;
271
272
273 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
274
275
276 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
277 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
278 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
279 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
280 i9 += 8;
281
282
283 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
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 vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
289 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
290 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
291 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
292 i10 += 8;
293
294
295 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
296
297
298 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
299 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
300 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
301 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
302 i11 += 8;
303
304
305 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
306
307 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
308 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
309
310 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
311 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
312 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
313 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
314 i12 += 8;
315
316
317 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
318
319
320 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
321 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
322 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
323 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
324 i13 += 8;
325
326
327 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
328
329 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
330 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
331
332 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
333 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
334 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
335 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
336 i14 += 8;
337
338
339 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
340
341
342 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
343 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
344 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
345 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
346 i15 += 8;
347
348
349 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
350
351 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
352 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
353
354 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
355 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
356 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
357 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
358 i16 += 8;
359
360
361 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
362
363
364 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
365 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
366 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
367 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
368 i17 += 8;
369
370
371 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
375
376 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
377 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
378 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
379 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
380 i18 += 8;
381
382
383 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
384
385
386 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
387 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
388 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
389 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
390 i19 += 8;
391
392
393 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
394
395 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
396 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
397
398 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
399 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
400 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
401 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
402 i20 += 8;
403
404
405 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
406
407
408 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
409 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
410 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
411 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
412 i21 += 8;
413
414
415 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
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 vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
421 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
422 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
423 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
424 i22 += 8;
425
426
427 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
428
429
430 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
431 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
432 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
433 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
434 i23 += 8;
435
436
437 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
440 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
441
442 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
443 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
444 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
445 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
446 i24 += 8;
447
448
449 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
450
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
453
454 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
455
456 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
457 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
458
459 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
460 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
461
462 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
463 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
464
465 vacc0123 = _mm_cvtps_epi32(vscaled0123);
466 vacc4567 = _mm_cvtps_epi32(vscaled4567);
467
468 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
469
470
471 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
472
473 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
474
475 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
476 output += 8;
477 }
478 if XNN_UNLIKELY(c != 0) {
479 {
480 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
481 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
482
483
484 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
485 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
486 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
487 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
488
489
490 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
491
492
493 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
494 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
495 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
496 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
497
498
499 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
500
501 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
502 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
503
504 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
505 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
506 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
507 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
508
509
510 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
511
512
513 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
514 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
515 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
516 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
517
518
519 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
520
521 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
522 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
523
524 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
525 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
526 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
527 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
528
529
530 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
531
532
533 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
534 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
535 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
536 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
537
538
539 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
540
541 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
542 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
543
544 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
545 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
546 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
547 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
548
549
550 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
551
552
553 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
554 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
555 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
556 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
557
558
559 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
560
561 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
562 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
563
564 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
565 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
566 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
567 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
568
569
570 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
571
572
573 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
574 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
575 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
576 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
577
578
579 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
580
581 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
582 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
583
584 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
585 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
586 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
587 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
588
589
590 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
591
592
593 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
594 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
595 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
596 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
597
598
599 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
600
601 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
602 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
603
604 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
605 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
606 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
607 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
608
609
610 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
611
612
613 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
614 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
615 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
616 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
617
618
619 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
620
621 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
622 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
623
624 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
625 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
626 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
627 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
628
629
630 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
631
632
633 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
634 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
635 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
636 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
637
638
639 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
640
641 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
642 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
643
644 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
645 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
646 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
647 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
648
649
650 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
651
652
653 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
654 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
655 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
656 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
657
658
659 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
660
661 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
662 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
663
664 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
665 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
666 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
667 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
668
669
670 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
671
672
673 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
674 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
675 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
676 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
677
678
679 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
680
681 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
682 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
683
684 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
685 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
686 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
687 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
688
689
690 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
691
692
693 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
694 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
695 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
696 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
697
698
699 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
700
701 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
702 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
703
704 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
705 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
706 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
707 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
708
709
710 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
711
712
713 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
714 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
715 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
716 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
717
718
719 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
720
721 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
722 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
723
724 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
725 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
726 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
727 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
728
729
730 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
731
732 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
733 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
734
735
736 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
737 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
738
739 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
740 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
741
742 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
743 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
744
745 vacc0123 = _mm_cvtps_epi32(vscaled0123);
746 vacc4567 = _mm_cvtps_epi32(vscaled4567);
747
748
749 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
750
751
752 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
753
754 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
755
756 if (c & 4) {
757 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
758 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
759 output += 4;
760 }
761 if (c & 2) {
762 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
763 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
764 output += 2;
765 }
766 if (c & 1) {
767 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
768 output += 1;
769 }
770 }
771 }
772
773 output = (int8_t*) ((uintptr_t) output + output_increment);
774 } while (--output_width != 0);
775 }