Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l8c8s8r-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_8f8m9l8c8s8r__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 > 8);
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 8 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 const int8_t* i6 = input[6];
85 assert(i6 != NULL);
86 if XNN_UNPREDICTABLE(i6 != zero) {
87 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
88 }
89 const int8_t* i7 = input[7];
90 assert(i7 != NULL);
91 if XNN_UNPREDICTABLE(i7 != zero) {
92 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
93 }
94 input += 8;
95
96 size_t c = round_up_po2(channels, 8);
97
98 for (; c >= 8; c -= 8) {
99 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
100 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
101
102
103
104 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
105 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
106 i0 += 8;
107
108 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
109 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
110
111 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
112
113
114 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
115 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
116 i1 += 8;
117
118 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
119 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
120
121 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
122
123 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
124 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
125 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
126
127 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
128 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
129 i2 += 8;
130
131 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
132 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
133
134 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
135
136
137 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
138 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
139 i3 += 8;
140
141 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
142 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
143
144 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
145
146 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
149
150 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
151 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
152 i4 += 8;
153
154 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
155 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
156
157 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
158
159
160 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
161 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
162 i5 += 8;
163
164 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
165 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
166
167 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
168
169 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
170 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
171 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
172
173 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
174 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
175 i6 += 8;
176
177 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
178 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
179
180 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
181
182
183 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
184 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
185 i7 += 8;
186
187 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
188 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
189
190 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
191
192 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
195
196 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
197
198 _mm_storeu_si128((__m128i*) b, vacc0123);
199 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
200 b += 8;
201 }
202
203 assert(c == 0);
204 }
205
206 // Middle pass to process 8 inputs in each iteration.
207 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
208 int32_t* b = buffer;
209 const int8_t* i0 = input[0];
210 assert(i0 != NULL);
211 if XNN_UNPREDICTABLE(i0 != zero) {
212 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
213 }
214 const int8_t* i1 = input[1];
215 assert(i1 != NULL);
216 if XNN_UNPREDICTABLE(i1 != zero) {
217 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
218 }
219 const int8_t* i2 = input[2];
220 assert(i2 != NULL);
221 if XNN_UNPREDICTABLE(i2 != zero) {
222 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
223 }
224 const int8_t* i3 = input[3];
225 assert(i3 != NULL);
226 if XNN_UNPREDICTABLE(i3 != zero) {
227 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
228 }
229 const int8_t* i4 = input[4];
230 assert(i4 != NULL);
231 if XNN_UNPREDICTABLE(i4 != zero) {
232 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
233 }
234 const int8_t* i5 = input[5];
235 assert(i5 != NULL);
236 if XNN_UNPREDICTABLE(i5 != zero) {
237 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
238 }
239 const int8_t* i6 = input[6];
240 assert(i6 != NULL);
241 if XNN_UNPREDICTABLE(i6 != zero) {
242 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
243 }
244 const int8_t* i7 = input[7];
245 assert(i7 != NULL);
246 if XNN_UNPREDICTABLE(i7 != zero) {
247 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
248 }
249 input += 8;
250
251 size_t c = round_up_po2(channels, 8);
252
253 for (; c >= 8; c -= 8) {
254 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
255 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
256
257
258
259 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
260 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
261 i0 += 8;
262
263 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
264 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
265
266 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
267
268
269 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
270 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
271 i1 += 8;
272
273 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
274 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
275
276 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
277
278 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
279 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
280 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
281
282 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
283 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
284 i2 += 8;
285
286 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
287 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
288
289 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
290
291
292 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
293 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
294 i3 += 8;
295
296 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
297 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
298
299 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
300
301 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
302 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
303 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
304
305 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
306 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
307 i4 += 8;
308
309 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
310 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
311
312 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
313
314
315 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
316 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
317 i5 += 8;
318
319 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
320 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
321
322 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
323
324 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
325 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
326 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
327
328 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
329 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
330 i6 += 8;
331
332 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
333 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
334
335 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
336
337
338 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
339 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
340 i7 += 8;
341
342 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
343 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
344
345 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
346
347 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
348 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
349 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
350
351 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
352
353 _mm_storeu_si128((__m128i*) (b), vacc0123);
354 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
355 b += 8;
356 }
357
358 assert(c == 0);
359 }
360
361 // Last pass to process up to 9 inputs.
362 {
363 const int32_t* b = buffer;
364 const int8_t* i0 = input[0];
365 assert(i0 != NULL);
366 if XNN_UNPREDICTABLE(i0 != zero) {
367 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
368 }
369 const int8_t* i1 = input[1];
370 assert(i1 != NULL);
371 if XNN_UNPREDICTABLE(i1 != zero) {
372 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
373 }
374 const int8_t* i2 = input[2];
375 assert(i2 != NULL);
376 if XNN_UNPREDICTABLE(i2 != zero) {
377 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
378 }
379 const int8_t* i3 = input[3];
380 assert(i3 != NULL);
381 if XNN_UNPREDICTABLE(i3 != zero) {
382 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
383 }
384 const int8_t* i4 = input[4];
385 assert(i4 != NULL);
386 if XNN_UNPREDICTABLE(i4 != zero) {
387 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
388 }
389 const int8_t* i5 = input[5];
390 assert(i5 != NULL);
391 if XNN_UNPREDICTABLE(i5 != zero) {
392 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
393 }
394 const int8_t* i6 = input[6];
395 assert(i6 != NULL);
396 if XNN_UNPREDICTABLE(i6 != zero) {
397 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
398 }
399 const int8_t* i7 = input[7];
400 assert(i7 != NULL);
401 if XNN_UNPREDICTABLE(i7 != zero) {
402 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
403 }
404 const int8_t* i8 = input[8];
405 assert(i8 != NULL);
406 if XNN_UNPREDICTABLE(i8 != zero) {
407 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
408 }
409
410 size_t c = channels;
411
412 for (; c >= 8; c -= 8) {
413 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
414 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
415 b += 8;
416
417
418 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
419 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
420 i0 += 8;
421
422 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
423 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
424
425 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
426
427
428 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
429 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
430 i1 += 8;
431
432 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
433 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
434
435 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
436
437 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
440
441 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
442 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
443 i2 += 8;
444
445 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
446 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
447
448 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
449
450
451 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
452 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
453 i3 += 8;
454
455 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
456 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
457
458 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
459
460 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
461 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
462 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
463
464 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
465 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
466 i4 += 8;
467
468 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
469 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
470
471 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
472
473
474 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
475 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
476 i5 += 8;
477
478 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
479 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
480
481 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
482
483 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
484 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
485 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
486
487 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
488 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
489 i6 += 8;
490
491 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
492 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
493
494 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
495
496
497 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
498 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
499 i7 += 8;
500
501 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
502 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
503
504 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
505
506 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
507 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
508 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
509
510 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
511 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
512 i8 += 8;
513
514 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
515 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
516
517 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
518
519 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
520 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
521 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
522
523 w = (const void*) ((uintptr_t) w + 72 * sizeof(int8_t));
524
525 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
526 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
527
528 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
529 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
530
531 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
532 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
533
534 vacc0123 = _mm_cvtps_epi32(vscaled0123);
535 vacc4567 = _mm_cvtps_epi32(vscaled4567);
536
537 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
538
539 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
540
541 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
542
543
544 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
545 output += 8;
546 }
547
548 if XNN_UNLIKELY(c != 0) {
549 {
550 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
551 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
552 b += 8;
553
554
555 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
556 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
557
558 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
559 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
560
561 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
562
563
564 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
565 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
566
567 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
568 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
569
570 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
571
572 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
573 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
574 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
575
576 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
577 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
578
579 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
580 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
581
582 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
583
584
585 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
586 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
587
588 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
589 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
590
591 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
592
593 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
594 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
595 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
596
597 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
598 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
599
600 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
601 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
602
603 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
604
605
606 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
607 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
608
609 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
610 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
611
612 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
613
614 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
615 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
616 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
617
618 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
619 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
620
621 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
622 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
623
624 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
625
626
627 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
628 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
629
630 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
631 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
632
633 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
634
635 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
636 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
637 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
638
639 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
640 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
641
642 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
643 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
644
645 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
646
647 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
648 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
649 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
650
651 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
652 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
653
654 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
655 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
656
657
658 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
659 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
660
661 vacc0123 = _mm_cvtps_epi32(vscaled0123);
662 vacc4567 = _mm_cvtps_epi32(vscaled4567);
663
664 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
665
666 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
667
668 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
669
670
671 if (c & 4) {
672 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
673 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
674 output += 4;
675 }
676 if (c & 2) {
677 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
678 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
679 output += 2;
680 }
681 if (c & 1) {
682 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
683 output += 1;
684 }
685 }
686 }
687 }
688
689 input = (const int8_t**) ((uintptr_t) input + input_stride);
690 output = (int8_t*) ((uintptr_t) output + output_increment);
691 } while (--output_width != 0);
692 }