Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l8c8s8r-minmax-fp32-sse41-mul16.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(
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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
99 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
100
101 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
102 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
103 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
104 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
105 i1 += 8;
106
107
108 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
109
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
112
113 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
114 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
115 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
116 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
117 i2 += 8;
118
119
120 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
121
122 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
123 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
124
125 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
126 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
127 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
128 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
129 i3 += 8;
130
131
132 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
133
134 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
135 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
136
137 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
138 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
139 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
140 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
141 i4 += 8;
142
143
144 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
145
146 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
147 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
148
149 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
150
151 _mm_storeu_si128((__m128i*) b, vacc0123);
152 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
153 b += 8;
154 }
155
156 assert(c == 0);
157 }
158
159 // Middle pass to process 5 inputs in each iteration.
160 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
161 int32_t* b = buffer;
162 const int8_t* i0 = input[0];
163 assert(i0 != NULL);
164 if XNN_UNPREDICTABLE(i0 != zero) {
165 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
166 }
167 const int8_t* i1 = input[1];
168 assert(i1 != NULL);
169 if XNN_UNPREDICTABLE(i1 != zero) {
170 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
171 }
172 const int8_t* i2 = input[2];
173 assert(i2 != NULL);
174 if XNN_UNPREDICTABLE(i2 != zero) {
175 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
176 }
177 const int8_t* i3 = input[3];
178 assert(i3 != NULL);
179 if XNN_UNPREDICTABLE(i3 != zero) {
180 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
181 }
182 const int8_t* i4 = input[4];
183 assert(i4 != NULL);
184 if XNN_UNPREDICTABLE(i4 != zero) {
185 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
186 }
187 input += 5;
188
189 size_t c = round_up_po2(channels, 8);
190
191 for (; c >= 8; c -= 8) {
192 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
193 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
194
195
196
197 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
198 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
199 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
200 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
201 i0 += 8;
202
203
204 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
205
206 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
207 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
208
209 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
210 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
211 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
212 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
213 i1 += 8;
214
215
216 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
217
218 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
219 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
220
221 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
222 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
223 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
224 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
225 i2 += 8;
226
227
228 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
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 vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
234 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
235 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
236 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
237 i3 += 8;
238
239
240 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
241
242 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
243 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
244
245 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
246 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
247 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
248 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
249 i4 += 8;
250
251
252 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
253
254 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
255 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
256
257 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
258
259 _mm_storeu_si128((__m128i*) (b), vacc0123);
260 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
261 b += 8;
262 }
263
264 assert(c == 0);
265 }
266
267 // Last pass to process up to 5 inputs.
268 {
269 const int32_t* b = buffer;
270 const int8_t* i0 = input[0];
271 assert(i0 != NULL);
272 if XNN_UNPREDICTABLE(i0 != zero) {
273 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
274 }
275 const int8_t* i1 = input[1];
276 assert(i1 != NULL);
277 if XNN_UNPREDICTABLE(i1 != zero) {
278 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
279 }
280 const int8_t* i2 = input[2];
281 assert(i2 != NULL);
282 if XNN_UNPREDICTABLE(i2 != zero) {
283 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
284 }
285 const int8_t* i3 = input[3];
286 assert(i3 != NULL);
287 if XNN_UNPREDICTABLE(i3 != zero) {
288 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
289 }
290 const int8_t* i4 = input[4];
291 assert(i4 != NULL);
292 if XNN_UNPREDICTABLE(i4 != zero) {
293 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
294 }
295
296 size_t c = channels;
297
298 for (; c >= 8; c -= 8) {
299 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
300 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
301 b += 8;
302
303
304 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
305 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
306 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
307 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
308 i0 += 8;
309
310
311 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
312
313 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
314 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
315
316 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
317 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
318 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
319 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
320 i1 += 8;
321
322
323 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
324
325 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
326 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
327
328 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
329 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
330 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
331 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
332 i2 += 8;
333
334
335 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
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 vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
341 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
342 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
343 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
344 i3 += 8;
345
346
347 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
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 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
353 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
354 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
355 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
356 i4 += 8;
357
358
359 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
360
361 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
362 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
363
364 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
365
366 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
367 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
368
369 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
370 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
371
372 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
373 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
374
375 vacc0123 = _mm_cvtps_epi32(vscaled0123);
376 vacc4567 = _mm_cvtps_epi32(vscaled4567);
377
378 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
379
380
381 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
382
383 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
384
385 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
386 output += 8;
387 }
388
389 if XNN_UNLIKELY(c != 0) {
390 {
391 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
392 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
393 b += 8;
394
395
396 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
397 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
398 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
399 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
400
401
402 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
403
404 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
405 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
406
407 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
408 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
409 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
410 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
411
412
413 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
414
415 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
416 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
417
418 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
419 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
420 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
421 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
422
423
424 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
425
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
428
429 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
430 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
431 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
432 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
433
434
435 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
436
437 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
438 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
439
440 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
441 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
442 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
443 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
444
445
446 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
447
448 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
449 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
450
451 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
452 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
453
454 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
455 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
456
457
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 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
465
466
467 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
468
469 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
470
471 if (c & 4) {
472 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
473 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
474 output += 4;
475 }
476 if (c & 2) {
477 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
478 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
479 output += 2;
480 }
481 if (c & 1) {
482 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
483 output += 1;
484 }
485 }
486 }
487 }
488
489 input = (const int8_t**) ((uintptr_t) input + input_stride);
490 output = (int8_t*) ((uintptr_t) output + output_increment);
491 } while (--output_width != 0);
492 }