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 #include <xnnpack/unaligned.h>
21
22
xnn_qc8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul16_add16(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const int8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const int8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const int8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const int8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const int8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const int8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 const int8_t* i9 = input[9];
85 assert(i9 != NULL);
86 if XNN_UNPREDICTABLE(i9 != zero) {
87 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
88 }
89 const int8_t* i10 = input[10];
90 assert(i10 != NULL);
91 if XNN_UNPREDICTABLE(i10 != zero) {
92 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
93 }
94 const int8_t* i11 = input[11];
95 assert(i11 != NULL);
96 if XNN_UNPREDICTABLE(i11 != zero) {
97 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
98 }
99 const int8_t* i12 = input[12];
100 assert(i12 != NULL);
101 if XNN_UNPREDICTABLE(i12 != zero) {
102 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
103 }
104 const int8_t* i13 = input[13];
105 assert(i13 != NULL);
106 if XNN_UNPREDICTABLE(i13 != zero) {
107 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
108 }
109 const int8_t* i14 = input[14];
110 assert(i14 != NULL);
111 if XNN_UNPREDICTABLE(i14 != zero) {
112 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
113 }
114 const int8_t* i15 = input[15];
115 assert(i15 != NULL);
116 if XNN_UNPREDICTABLE(i15 != zero) {
117 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
118 }
119 const int8_t* i16 = input[16];
120 assert(i16 != NULL);
121 if XNN_UNPREDICTABLE(i16 != zero) {
122 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
123 }
124 const int8_t* i17 = input[17];
125 assert(i17 != NULL);
126 if XNN_UNPREDICTABLE(i17 != zero) {
127 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
128 }
129 const int8_t* i18 = input[18];
130 assert(i18 != NULL);
131 if XNN_UNPREDICTABLE(i18 != zero) {
132 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
133 }
134 const int8_t* i19 = input[19];
135 assert(i19 != NULL);
136 if XNN_UNPREDICTABLE(i19 != zero) {
137 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
138 }
139 const int8_t* i20 = input[20];
140 assert(i20 != NULL);
141 if XNN_UNPREDICTABLE(i20 != zero) {
142 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
143 }
144 const int8_t* i21 = input[21];
145 assert(i21 != NULL);
146 if XNN_UNPREDICTABLE(i21 != zero) {
147 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
148 }
149 const int8_t* i22 = input[22];
150 assert(i22 != NULL);
151 if XNN_UNPREDICTABLE(i22 != zero) {
152 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
153 }
154 const int8_t* i23 = input[23];
155 assert(i23 != NULL);
156 if XNN_UNPREDICTABLE(i23 != zero) {
157 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
158 }
159 const int8_t* i24 = input[24];
160 assert(i24 != NULL);
161 if XNN_UNPREDICTABLE(i24 != zero) {
162 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
163 }
164 input = (const int8_t**) ((uintptr_t) input + input_stride);
165
166 size_t c = channels;
167 const void* w = weights;
168 for (; c >= 8; c -= 8) {
169 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
170 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
171
172
173 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
174 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
175 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
176 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
177 i0 += 8;
178
179
180 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
181
182
183 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
184 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
185 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
186 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
187 i1 += 8;
188
189
190 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
194
195 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
196 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
197 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
198 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
199 i2 += 8;
200
201
202 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
203
204
205 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
206 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
207 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
208 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
209 i3 += 8;
210
211
212 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
213
214 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
215 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
216
217 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
218 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
219 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
220 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
221 i4 += 8;
222
223
224 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
225
226
227 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
228 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
229 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
230 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
231 i5 += 8;
232
233
234 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
235
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
238
239 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
240 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
241 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
242 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
243 i6 += 8;
244
245
246 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
247
248
249 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
250 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
251 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
252 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
253 i7 += 8;
254
255
256 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
257
258 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
259 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
260
261 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
262 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
263 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
264 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
265 i8 += 8;
266
267
268 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
269
270
271 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
272 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
273 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
274 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
275 i9 += 8;
276
277
278 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
279
280 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
281 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
282
283 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
284 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
285 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
286 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
287 i10 += 8;
288
289
290 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
291
292
293 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
294 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
295 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
296 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
297 i11 += 8;
298
299
300 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
301
302 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
303 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
304
305 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
306 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
307 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
308 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
309 i12 += 8;
310
311
312 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
313
314
315 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
316 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
317 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
318 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
319 i13 += 8;
320
321
322 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
323
324 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
325 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
326
327 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
328 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
329 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
330 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
331 i14 += 8;
332
333
334 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
335
336
337 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
338 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
339 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
340 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
341 i15 += 8;
342
343
344 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
345
346 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
347 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
348
349 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
350 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
351 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
352 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
353 i16 += 8;
354
355
356 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
357
358
359 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
360 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
361 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
362 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
363 i17 += 8;
364
365
366 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
367
368 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
369 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
370
371 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
372 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
373 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
374 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
375 i18 += 8;
376
377
378 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
379
380
381 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
382 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
383 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
384 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
385 i19 += 8;
386
387
388 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
392
393 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
394 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
395 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
396 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
397 i20 += 8;
398
399
400 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
401
402
403 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
404 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
405 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
406 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
407 i21 += 8;
408
409
410 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
411
412 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
413 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
414
415 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
416 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
417 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
418 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
419 i22 += 8;
420
421
422 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
423
424
425 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
426 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
427 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
428 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
429 i23 += 8;
430
431
432 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
433
434 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
435 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
436
437 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
438 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
439 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
440 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
441 i24 += 8;
442
443
444 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
445
446 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
447 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
448
449 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
450
451 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
452 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
453
454 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
455 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
456 w = (const void*) ((const float*) w + 8);
457 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
458 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
459
460 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
461 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
462 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
463
464 vacc0123 = _mm_cvtps_epi32(vscaled0123);
465 vacc4567 = _mm_cvtps_epi32(vscaled4567);
466
467 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
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 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
474 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
475
476 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
477 output += 8;
478 }
479 if XNN_UNLIKELY(c != 0) {
480 {
481 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
482 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
483
484
485 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
486 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
487 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
488 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
489
490
491 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
492
493
494 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
495 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
496 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
497 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
498
499
500 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
501
502 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
503 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
504
505 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
506 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
507 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
508 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
509
510
511 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
512
513
514 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
515 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
516 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
517 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
518
519
520 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
521
522 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
523 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
524
525 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
526 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
527 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
528 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
529
530
531 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
532
533
534 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
535 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
536 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
537 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
538
539
540 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
541
542 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
543 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
544
545 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
546 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
547 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
548 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
549
550
551 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
552
553
554 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
555 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
556 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
557 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
558
559
560 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
561
562 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
563 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
564
565 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
566 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
567 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
568 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
569
570
571 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
572
573
574 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
575 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
576 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
577 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
578
579
580 vprod01234567 = _mm_macc_epi16(vxi9x01234567, vxk9x01234567, vprod01234567);
581
582 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
583 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
584
585 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
586 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
587 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
588 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
589
590
591 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
592
593
594 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
595 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
596 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
597 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
598
599
600 vprod01234567 = _mm_macc_epi16(vxi11x01234567, vxk11x01234567, vprod01234567);
601
602 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
603 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
604
605 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
606 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
607 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
608 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
609
610
611 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
612
613
614 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
615 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
616 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
617 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
618
619
620 vprod01234567 = _mm_macc_epi16(vxi13x01234567, vxk13x01234567, vprod01234567);
621
622 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
623 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
624
625 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
626 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
627 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
628 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
629
630
631 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
632
633
634 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
635 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
636 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
637 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
638
639
640 vprod01234567 = _mm_macc_epi16(vxi15x01234567, vxk15x01234567, vprod01234567);
641
642 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
643 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
644
645 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
646 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
647 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
648 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
649
650
651 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
652
653
654 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
655 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
656 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
657 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
658
659
660 vprod01234567 = _mm_macc_epi16(vxi17x01234567, vxk17x01234567, vprod01234567);
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 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
666 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
667 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
668 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
669
670
671 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
672
673
674 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
675 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
676 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
677 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
678
679
680 vprod01234567 = _mm_macc_epi16(vxi19x01234567, vxk19x01234567, vprod01234567);
681
682 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
683 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
684
685 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
686 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
687 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
688 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
689
690
691 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
692
693
694 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
695 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
696 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
697 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
698
699
700 vprod01234567 = _mm_macc_epi16(vxi21x01234567, vxk21x01234567, vprod01234567);
701
702 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
703 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
704
705 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
706 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
707 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
708 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
709
710
711 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
712
713
714 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
715 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
716 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
717 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
718
719
720 vprod01234567 = _mm_macc_epi16(vxi23x01234567, vxk23x01234567, vprod01234567);
721
722 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
723 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
724
725 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
726 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
727 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
728 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
729
730
731 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
732
733 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
734 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
735
736
737 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
738 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
739
740 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)));
741 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
742 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
743 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
744
745 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
746 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
747 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
748
749 vacc0123 = _mm_cvtps_epi32(vscaled0123);
750 vacc4567 = _mm_cvtps_epi32(vscaled4567);
751
752
753 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
754 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
755
756
757 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
758
759 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
760
761 if (c & 4) {
762 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
763 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
764 output += 4;
765 }
766 if (c & 2) {
767 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
768 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
769 output += 2;
770 }
771 if (c & 1) {
772 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
773 output += 1;
774 }
775 }
776 }
777
778 output = (int8_t*) ((uintptr_t) output + output_increment);
779 } while (--output_width != 0);
780 }
781