Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l8c8s8r-minmax-fp32-sse41-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_5f5m5l8c8s8r__sse41_mul16_add16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 5);
36
37
38 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
39 XNN_FORCE_REALIZATION(vscale);
40
41 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
42 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 5 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 input += 5;
80
81 size_t c = round_up_po2(channels, 8);
82
83 for (; c >= 8; c -= 8) {
84 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
85 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
86
87
88
89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
91 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
92 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
93 i0 += 8;
94
95
96 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
97
98
99 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
100 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
101 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
102 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
103 i1 += 8;
104
105
106 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
107
108 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
109 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
110
111 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
112 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
113 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
114 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
115 i2 += 8;
116
117
118 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
119
120
121 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
122 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
123 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
124 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
125 i3 += 8;
126
127
128 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
129
130 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
131 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
132
133 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
134 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
135 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
136 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
137 i4 += 8;
138
139
140 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
141
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
144
145 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
146
147 _mm_storeu_si128((__m128i*) b, vacc0123);
148 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
149 b += 8;
150 }
151
152 assert(c == 0);
153 }
154
155 // Middle pass to process 5 inputs in each iteration.
156 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
157 int32_t* b = buffer;
158 const int8_t* i0 = input[0];
159 assert(i0 != NULL);
160 if XNN_UNPREDICTABLE(i0 != zero) {
161 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
162 }
163 const int8_t* i1 = input[1];
164 assert(i1 != NULL);
165 if XNN_UNPREDICTABLE(i1 != zero) {
166 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
167 }
168 const int8_t* i2 = input[2];
169 assert(i2 != NULL);
170 if XNN_UNPREDICTABLE(i2 != zero) {
171 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
172 }
173 const int8_t* i3 = input[3];
174 assert(i3 != NULL);
175 if XNN_UNPREDICTABLE(i3 != zero) {
176 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
177 }
178 const int8_t* i4 = input[4];
179 assert(i4 != NULL);
180 if XNN_UNPREDICTABLE(i4 != zero) {
181 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
182 }
183 input += 5;
184
185 size_t c = round_up_po2(channels, 8);
186
187 for (; c >= 8; c -= 8) {
188 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
189 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
190
191
192
193 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
194 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
195 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
196 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
197 i0 += 8;
198
199
200 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
201
202
203 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
204 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
205 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
206 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
207 i1 += 8;
208
209
210 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
214
215 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
216 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
217 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
218 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
219 i2 += 8;
220
221
222 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
223
224
225 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
226 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
227 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
228 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
229 i3 += 8;
230
231
232 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
233
234 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
235 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
236
237 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
238 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
239 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
240 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
241 i4 += 8;
242
243
244 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
245
246 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
247 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
248
249 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
250
251 _mm_storeu_si128((__m128i*) (b), vacc0123);
252 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
253 b += 8;
254 }
255
256 assert(c == 0);
257 }
258
259 // Last pass to process up to 5 inputs.
260 {
261 const int32_t* b = buffer;
262 const int8_t* i0 = input[0];
263 assert(i0 != NULL);
264 if XNN_UNPREDICTABLE(i0 != zero) {
265 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
266 }
267 const int8_t* i1 = input[1];
268 assert(i1 != NULL);
269 if XNN_UNPREDICTABLE(i1 != zero) {
270 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
271 }
272 const int8_t* i2 = input[2];
273 assert(i2 != NULL);
274 if XNN_UNPREDICTABLE(i2 != zero) {
275 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
276 }
277 const int8_t* i3 = input[3];
278 assert(i3 != NULL);
279 if XNN_UNPREDICTABLE(i3 != zero) {
280 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
281 }
282 const int8_t* i4 = input[4];
283 assert(i4 != NULL);
284 if XNN_UNPREDICTABLE(i4 != zero) {
285 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
286 }
287
288 size_t c = channels;
289
290 for (; c >= 8; c -= 8) {
291 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
292 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
293 b += 8;
294
295
296 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
297 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
298 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
299 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
300 i0 += 8;
301
302
303 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
304
305
306 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
307 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
308 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
309 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
310 i1 += 8;
311
312
313 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
314
315 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
316 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
317
318 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
319 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
320 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
321 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
322 i2 += 8;
323
324
325 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
326
327
328 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
329 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
330 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
331 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
332 i3 += 8;
333
334
335 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
336
337 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
338 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
339
340 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
341 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
342 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
343 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
344 i4 += 8;
345
346
347 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
348
349 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
350 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
351
352 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
353
354 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
355 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
356
357 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
358 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
359
360 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
361 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
362
363 vacc0123 = _mm_cvtps_epi32(vscaled0123);
364 vacc4567 = _mm_cvtps_epi32(vscaled4567);
365
366 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
367
368
369 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
370
371 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
372
373 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
374 output += 8;
375 }
376
377 if XNN_UNLIKELY(c != 0) {
378 {
379 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
380 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
381 b += 8;
382
383
384 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
385 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
386 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
387 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
388
389
390 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
391
392
393 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
394 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
395 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
396 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
397
398
399 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
403
404 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
405 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
406 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
407 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
408
409
410 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
411
412
413 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
414 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
415 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
416 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
417
418
419 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
420
421 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
422 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
423
424 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
425 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
426 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
427 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
428
429
430 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
431
432 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
433 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
434
435 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
436 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
437
438 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
439 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
440
441
442 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
443 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
444
445 vacc0123 = _mm_cvtps_epi32(vscaled0123);
446 vacc4567 = _mm_cvtps_epi32(vscaled4567);
447
448 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
449
450
451 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
452
453 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
454
455 if (c & 4) {
456 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
457 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
458 output += 4;
459 }
460 if (c & 2) {
461 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
462 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
463 output += 2;
464 }
465 if (c & 1) {
466 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
467 output += 1;
468 }
469 }
470 }
471 }
472
473 input = (const int8_t**) ((uintptr_t) input + input_stride);
474 output = (int8_t*) ((uintptr_t) output + output_increment);
475 } while (--output_width != 0);
476 }