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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20
21
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])22 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul16_add16(
23 size_t channels,
24 size_t output_width,
25 const int8_t** input,
26 const void* weights,
27 int8_t* output,
28 size_t input_stride,
29 size_t output_increment,
30 size_t input_offset,
31 const int8_t* zero,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 for (; c >= 8; c -= 8) {
168 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
169 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
170
171
172 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
173 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
174 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
175 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
176 i0 += 8;
177
178
179 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
180
181
182 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
183 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
184 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
185 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
186 i1 += 8;
187
188
189 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
193
194 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
195 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
196 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
197 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
198 i2 += 8;
199
200
201 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
202
203
204 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
205 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
206 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
207 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
208 i3 += 8;
209
210
211 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
212
213 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
214 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
215
216 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
217 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
218 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
219 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
220 i4 += 8;
221
222
223 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
224
225
226 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
227 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
228 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
229 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
230 i5 += 8;
231
232
233 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
234
235 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
236 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
237
238 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
239 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
240 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
241 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
242 i6 += 8;
243
244
245 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
246
247
248 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
249 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
250 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
251 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
252 i7 += 8;
253
254
255 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
256
257 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
258 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
259
260 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
261 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
262 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
263 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
264 i8 += 8;
265
266
267 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
268
269
270 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
271 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
272 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
273 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
274 i9 += 8;
275
276
277 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
278
279 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
280 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
281
282 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
283 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
284 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
285 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
286 i10 += 8;
287
288
289 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
290
291
292 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
293 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
294 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
295 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
296 i11 += 8;
297
298
299 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
300
301 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
302 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
303
304 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
305 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
306 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
307 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
308 i12 += 8;
309
310
311 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
312
313
314 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
315 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
316 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
317 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
318 i13 += 8;
319
320
321 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
322
323 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
324 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
325
326 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
327 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
328 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
329 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
330 i14 += 8;
331
332
333 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
334
335
336 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
337 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
338 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
339 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
340 i15 += 8;
341
342
343 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
347
348 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
349 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
350 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
351 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
352 i16 += 8;
353
354
355 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
356
357
358 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
359 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
360 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
361 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
362 i17 += 8;
363
364
365 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
366
367 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
368 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
369
370 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
371 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
372 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
373 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
374 i18 += 8;
375
376
377 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
378
379
380 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
381 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
382 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
383 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
384 i19 += 8;
385
386
387 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
388
389 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
390 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
391
392 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
393 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
394 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
395 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
396 i20 += 8;
397
398
399 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
400
401
402 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
403 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
404 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
405 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
406 i21 += 8;
407
408
409 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
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 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
415 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
416 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
417 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
418 i22 += 8;
419
420
421 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
422
423
424 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
425 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
426 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
427 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
428 i23 += 8;
429
430
431 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
432
433 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
434 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
435
436 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
437 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
438 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
439 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
440 i24 += 8;
441
442
443 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
444
445 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
446 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
447
448 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
449
450 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
451 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
452
453 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
454 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
455 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
456
457 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
458 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
459 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
460
461 vacc0123 = _mm_cvtps_epi32(vscaled0123);
462 vacc4567 = _mm_cvtps_epi32(vscaled4567);
463
464 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
465 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
466
467
468 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
469
470 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
471 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
472
473 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
474 output += 8;
475 }
476 if XNN_UNLIKELY(c != 0) {
477 {
478 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
479 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
480
481
482 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
483 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
484 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
485 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
486
487
488 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
489
490
491 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
492 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
493 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
494 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
495
496
497 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
498
499 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
500 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
501
502 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
503 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
504 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
505 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
506
507
508 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
509
510
511 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
512 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
513 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
514 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
515
516
517 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
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 vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
523 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
524 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
525 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
526
527
528 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
529
530
531 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
532 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
533 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
534 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
535
536
537 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
538
539 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
540 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
541
542 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
543 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
544 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
545 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
546
547
548 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
549
550
551 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
552 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
553 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
554 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
555
556
557 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
558
559 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
560 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
561
562 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
563 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
564 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
565 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
566
567
568 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
569
570
571 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
572 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
573 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
574 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
575
576
577 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
578
579 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
580 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
581
582 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
583 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
584 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
585 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
586
587
588 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
589
590
591 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
592 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
593 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
594 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
595
596
597 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
598
599 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
600 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
601
602 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
603 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
604 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
605 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
606
607
608 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
609
610
611 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
612 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
613 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
614 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
615
616
617 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
618
619 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
620 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
621
622 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
623 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
624 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
625 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
626
627
628 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
629
630
631 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
632 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
633 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
634 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
635
636
637 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
638
639 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
640 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
641
642 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
643 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
644 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
645 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
646
647
648 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
649
650
651 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
652 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
653 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
654 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
655
656
657 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
658
659 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
660 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
661
662 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
663 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
664 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
665 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
666
667
668 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
669
670
671 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
672 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
673 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
674 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
675
676
677 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
678
679 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
680 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
681
682 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
683 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
684 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
685 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
686
687
688 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
689
690
691 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
692 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
693 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
694 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
695
696
697 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
698
699 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
700 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
701
702 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
703 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
704 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
705 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
706
707
708 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
709
710
711 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
712 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
713 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
714 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
715
716
717 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
718
719 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
720 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
721
722 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
723 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
724 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
725 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
726
727
728 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
729
730 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
731 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
732
733
734 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
735 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
736
737 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
738 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
739 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
740
741 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
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 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
750 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
751
752
753 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
754
755 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
756
757 if (c & 4) {
758 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
759 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
760 output += 4;
761 }
762 if (c & 2) {
763 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
764 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
765 output += 2;
766 }
767 if (c & 1) {
768 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
769 output += 1;
770 }
771 }
772 }
773
774 output = (int8_t*) ((uintptr_t) output + output_increment);
775 } while (--output_width != 0);
776 }
777