Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-minmax-fp32-sse2-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 <emmintrin.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_6f6m7l8c8s8r__sse2_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 > 6);
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_epi16(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 6 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 const int8_t* i5 = input[5];
80 assert(i5 != NULL);
81 if XNN_UNPREDICTABLE(i5 != zero) {
82 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
83 }
84 input += 6;
85
86 size_t c = round_up_po2(channels, 8);
87
88 for (; c >= 8; c -= 8) {
89 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
90 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
91
92
93
94 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
95 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
96 i0 += 8;
97
98 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
99 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
100
101 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
102
103
104 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
105 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
106 i1 += 8;
107
108 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
109 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
110
111 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
112
113 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
114 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
115 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
116
117 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
118 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
119 i2 += 8;
120
121 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
122 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
123
124 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
125
126
127 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
128 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
129 i3 += 8;
130
131 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
132 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
133
134 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
135
136 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
137 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
138 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
139
140 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
141 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
142 i4 += 8;
143
144 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
145 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
146
147 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
148
149
150 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
151 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
152 i5 += 8;
153
154 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
155 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
156
157 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
158
159 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
160 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
161 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
162
163 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t));
164
165 _mm_storeu_si128((__m128i*) b, vacc0123);
166 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
167 b += 8;
168 }
169
170 assert(c == 0);
171 }
172
173 // Middle pass to process 6 inputs in each iteration.
174 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
175 int32_t* b = buffer;
176 const int8_t* i0 = input[0];
177 assert(i0 != NULL);
178 if XNN_UNPREDICTABLE(i0 != zero) {
179 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
180 }
181 const int8_t* i1 = input[1];
182 assert(i1 != NULL);
183 if XNN_UNPREDICTABLE(i1 != zero) {
184 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
185 }
186 const int8_t* i2 = input[2];
187 assert(i2 != NULL);
188 if XNN_UNPREDICTABLE(i2 != zero) {
189 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
190 }
191 const int8_t* i3 = input[3];
192 assert(i3 != NULL);
193 if XNN_UNPREDICTABLE(i3 != zero) {
194 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
195 }
196 const int8_t* i4 = input[4];
197 assert(i4 != NULL);
198 if XNN_UNPREDICTABLE(i4 != zero) {
199 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
200 }
201 const int8_t* i5 = input[5];
202 assert(i5 != NULL);
203 if XNN_UNPREDICTABLE(i5 != zero) {
204 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
205 }
206 input += 6;
207
208 size_t c = round_up_po2(channels, 8);
209
210 for (; c >= 8; c -= 8) {
211 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
212 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
213
214
215
216 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
217 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
218 i0 += 8;
219
220 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
221 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
222
223 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
224
225
226 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
227 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
228 i1 += 8;
229
230 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
231 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
232
233 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
234
235 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
238
239 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
240 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
241 i2 += 8;
242
243 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
244 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
245
246 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
247
248
249 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
250 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
251 i3 += 8;
252
253 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
254 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
255
256 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
257
258 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
259 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
260 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
261
262 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
263 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
264 i4 += 8;
265
266 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
267 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
268
269 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
270
271
272 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
273 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
274 i5 += 8;
275
276 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
277 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
278
279 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
280
281 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
282 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
283 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
284
285 w = (const void*) ((uintptr_t) w + 48 * sizeof(int8_t));
286
287 _mm_storeu_si128((__m128i*) (b), vacc0123);
288 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
289 b += 8;
290 }
291
292 assert(c == 0);
293 }
294
295 // Last pass to process up to 7 inputs.
296 {
297 const int32_t* b = buffer;
298 const int8_t* i0 = input[0];
299 assert(i0 != NULL);
300 if XNN_UNPREDICTABLE(i0 != zero) {
301 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
302 }
303 const int8_t* i1 = input[1];
304 assert(i1 != NULL);
305 if XNN_UNPREDICTABLE(i1 != zero) {
306 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
307 }
308 const int8_t* i2 = input[2];
309 assert(i2 != NULL);
310 if XNN_UNPREDICTABLE(i2 != zero) {
311 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
312 }
313 const int8_t* i3 = input[3];
314 assert(i3 != NULL);
315 if XNN_UNPREDICTABLE(i3 != zero) {
316 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
317 }
318 const int8_t* i4 = input[4];
319 assert(i4 != NULL);
320 if XNN_UNPREDICTABLE(i4 != zero) {
321 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
322 }
323 const int8_t* i5 = input[5];
324 assert(i5 != NULL);
325 if XNN_UNPREDICTABLE(i5 != zero) {
326 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
327 }
328 const int8_t* i6 = input[6];
329 assert(i6 != NULL);
330 if XNN_UNPREDICTABLE(i6 != zero) {
331 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
332 }
333
334 size_t c = channels;
335
336 for (; c >= 8; c -= 8) {
337 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
338 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
339 b += 8;
340
341
342 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
343 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
344 i0 += 8;
345
346 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
347 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
348
349 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
350
351
352 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
353 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
354 i1 += 8;
355
356 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
357 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
358
359 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
360
361 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
363 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
364
365 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
366 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
367 i2 += 8;
368
369 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
370 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
371
372 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
373
374
375 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
376 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
377 i3 += 8;
378
379 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
380 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
381
382 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
383
384 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
385 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
386 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
387
388 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
389 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
390 i4 += 8;
391
392 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
393 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
394
395 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
396
397
398 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
399 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
400 i5 += 8;
401
402 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
403 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
404
405 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
406
407 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
408 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
409 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
410
411 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
412 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
413 i6 += 8;
414
415 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
416 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
417
418 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
419
420 const __m128i vsignprod6x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
421 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod6x01234567));
422 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod6x01234567));
423
424 w = (const void*) ((uintptr_t) w + 56 * sizeof(int8_t));
425
426 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
427 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
428
429 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
430 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
431
432 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
433 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
434
435 vacc0123 = _mm_cvtps_epi32(vscaled0123);
436 vacc4567 = _mm_cvtps_epi32(vscaled4567);
437
438 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
439
440 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
441
442 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
443
444
445 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
446 output += 8;
447 }
448
449 if XNN_UNLIKELY(c != 0) {
450 {
451 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
452 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
453 b += 8;
454
455
456 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
457 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
458
459 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
460 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
461
462 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
463
464
465 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
466 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
467
468 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
469 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
470
471 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
472
473 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
474 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
475 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
476
477 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
478 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
479
480 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
481 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
482
483 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
484
485
486 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
487 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
488
489 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
490 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
491
492 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
493
494 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
496 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
497
498 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
499 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
500
501 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
502 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
503
504 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
505
506
507 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
508 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
509
510 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
511 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
512
513 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
514
515 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
518
519 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
520 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
521
522 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
523 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
524
525 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
526
527 const __m128i vsignprod6x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
528 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod6x01234567));
529 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod6x01234567));
530
531 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
532 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
533
534 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
535 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
536
537
538 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
539 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
540
541 vacc0123 = _mm_cvtps_epi32(vscaled0123);
542 vacc4567 = _mm_cvtps_epi32(vscaled4567);
543
544 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
545
546 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
547
548 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
549
550
551 if (c & 4) {
552 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
553 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
554 output += 4;
555 }
556 if (c & 2) {
557 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
558 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
559 output += 2;
560 }
561 if (c & 1) {
562 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
563 output += 1;
564 }
565 }
566 }
567 }
568
569 input = (const int8_t**) ((uintptr_t) input + input_stride);
570 output = (int8_t*) ((uintptr_t) output + output_increment);
571 } while (--output_width != 0);
572 }