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
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16_add16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 8; c -= 8) {
163 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
164 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
165
166
167 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
168 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
169 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
170 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
171 i0 += 8;
172
173
174 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
175
176
177 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
178 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
179 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
180 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
181 i1 += 8;
182
183
184 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
185
186 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
187 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
188
189 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
190 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
191 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
192 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
193 i2 += 8;
194
195
196 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
197
198
199 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
200 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
201 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
202 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
203 i3 += 8;
204
205
206 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
207
208 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
209 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
210
211 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
212 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
213 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
214 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
215 i4 += 8;
216
217
218 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
219
220
221 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
222 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
223 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
224 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
225 i5 += 8;
226
227
228 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
229
230 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
231 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
232
233 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
234 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
235 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
236 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
237 i6 += 8;
238
239
240 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
241
242
243 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
244 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
245 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
246 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
247 i7 += 8;
248
249
250 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
251
252 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
253 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
254
255 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
256 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
257 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
258 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
259 i8 += 8;
260
261
262 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
263
264
265 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
266 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
267 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
268 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
269 i9 += 8;
270
271
272 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
273
274 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
275 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
276
277 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
278 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
279 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
280 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
281 i10 += 8;
282
283
284 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
285
286
287 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
288 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
289 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
290 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
291 i11 += 8;
292
293
294 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
297 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
298
299 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
300 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
301 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
302 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
303 i12 += 8;
304
305
306 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
307
308
309 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
310 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
311 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
312 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
313 i13 += 8;
314
315
316 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
317
318 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
319 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
320
321 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
322 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
323 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
324 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
325 i14 += 8;
326
327
328 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
329
330
331 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
332 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
333 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
334 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
335 i15 += 8;
336
337
338 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
339
340 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
341 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
342
343 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
344 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
345 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
346 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
347 i16 += 8;
348
349
350 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
351
352
353 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
354 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
355 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
356 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
357 i17 += 8;
358
359
360 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
361
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
363 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
364
365 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
366 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
367 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
368 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
369 i18 += 8;
370
371
372 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
373
374
375 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
376 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
377 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
378 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
379 i19 += 8;
380
381
382 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
383
384 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
385 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
386
387 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
388 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
389 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
390 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
391 i20 += 8;
392
393
394 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
395
396
397 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
398 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
399 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
400 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
401 i21 += 8;
402
403
404 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
405
406 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
407 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
408
409 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
410 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
411 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
412 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
413 i22 += 8;
414
415
416 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
417
418
419 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
420 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
421 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
422 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
423 i23 += 8;
424
425
426 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
427
428 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
429 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
430
431 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
432 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
433 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
434 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
435 i24 += 8;
436
437
438 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
439
440 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
441 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
442
443 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
444
445 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
446 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
447
448 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
449 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
450 w = (const void*) ((const float*) w + 8);
451 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
452 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
453
454 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
455 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
456 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
457
458 vacc0123 = _mm_cvtps_epi32(vscaled0123);
459 vacc4567 = _mm_cvtps_epi32(vscaled4567);
460
461 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
462 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
463
464
465 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
466
467 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
468 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
469
470 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
471 output += 8;
472 }
473 if XNN_UNLIKELY(c != 0) {
474 {
475 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
476 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
477
478
479 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
480 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
481 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
482 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
483
484
485 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
486
487
488 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
489 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
490 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
491 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
492
493
494 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
495
496 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
497 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
498
499 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
500 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
501 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
502 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
503
504
505 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
506
507
508 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
509 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
510 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
511 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
512
513
514 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
518
519 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
520 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
521 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
522 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
523
524
525 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
526
527
528 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
529 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
530 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
531 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
532
533
534 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
535
536 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
537 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
538
539 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
540 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
541 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
542 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
543
544
545 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
546
547
548 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
549 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
550 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
551 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
552
553
554 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
555
556 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
557 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
558
559 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
560 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
561 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
562 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
563
564
565 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
566
567
568 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
569 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
570 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
571 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
572
573
574 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
575
576 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
577 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
578
579 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
580 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
581 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
582 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
583
584
585 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
586
587
588 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
589 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
590 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
591 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
592
593
594 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
595
596 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
597 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
598
599 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
600 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
601 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
602 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
603
604
605 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
606
607
608 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
609 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
610 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
611 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
612
613
614 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
615
616 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
617 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
618
619 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
620 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
621 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
622 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
623
624
625 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
626
627
628 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
629 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
630 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
631 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
632
633
634 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
635
636 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
637 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
638
639 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
640 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
641 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
642 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
643
644
645 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
646
647
648 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
649 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
650 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
651 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
652
653
654 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
655
656 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
657 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
658
659 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
660 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
661 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
662 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
663
664
665 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
666
667
668 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
669 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
670 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
671 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
672
673
674 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
675
676 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
677 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
678
679 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
680 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
681 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
682 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
683
684
685 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
686
687
688 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
689 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
690 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
691 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
692
693
694 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
695
696 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
697 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
698
699 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
700 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
701 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
702 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
703
704
705 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
706
707
708 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
709 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
710 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
711 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
712
713
714 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
715
716 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
717 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
718
719 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
720 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
721 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
722 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
723
724
725 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
726
727 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
728 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
729
730
731 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
732 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
733
734 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)));
735 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
736 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
737 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
738
739 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
740 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
741 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
742
743 vacc0123 = _mm_cvtps_epi32(vscaled0123);
744 vacc4567 = _mm_cvtps_epi32(vscaled4567);
745
746
747 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
748 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
749
750
751 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
752
753 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->sse4.output_min));
754
755 if (c & 4) {
756 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
757 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
758 output += 4;
759 }
760 if (c & 2) {
761 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
762 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
763 output += 2;
764 }
765 if (c & 1) {
766 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
767 output += 1;
768 }
769 }
770 }
771
772 output = (int8_t*) ((uintptr_t) output + output_increment);
773 } while (--output_width != 0);
774 }
775