Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-25p8c-minmax-fp32-avx-mul16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/unaligned.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_25p8c__avx_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
34 XNN_FORCE_REALIZATION(vscale);
35
36 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);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
39 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42
43 do {
44 const int8_t* i0 = input[0];
45 assert(i0 != NULL);
46 if XNN_UNPREDICTABLE(i0 != zero) {
47 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
48 }
49 const int8_t* i1 = input[1];
50 assert(i1 != NULL);
51 if XNN_UNPREDICTABLE(i1 != zero) {
52 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
53 }
54 const int8_t* i2 = input[2];
55 assert(i2 != NULL);
56 if XNN_UNPREDICTABLE(i2 != zero) {
57 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
58 }
59 const int8_t* i3 = input[3];
60 assert(i3 != NULL);
61 if XNN_UNPREDICTABLE(i3 != zero) {
62 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
63 }
64 const int8_t* i4 = input[4];
65 assert(i4 != NULL);
66 if XNN_UNPREDICTABLE(i4 != zero) {
67 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
68 }
69 const int8_t* i5 = input[5];
70 assert(i5 != NULL);
71 if XNN_UNPREDICTABLE(i5 != zero) {
72 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
73 }
74 const int8_t* i6 = input[6];
75 assert(i6 != NULL);
76 if XNN_UNPREDICTABLE(i6 != zero) {
77 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
78 }
79 const int8_t* i7 = input[7];
80 assert(i7 != NULL);
81 if XNN_UNPREDICTABLE(i7 != zero) {
82 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
83 }
84 const int8_t* i8 = input[8];
85 assert(i8 != NULL);
86 if XNN_UNPREDICTABLE(i8 != zero) {
87 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
88 }
89 const int8_t* i9 = input[9];
90 assert(i9 != NULL);
91 if XNN_UNPREDICTABLE(i9 != zero) {
92 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
93 }
94 const int8_t* i10 = input[10];
95 assert(i10 != NULL);
96 if XNN_UNPREDICTABLE(i10 != zero) {
97 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
98 }
99 const int8_t* i11 = input[11];
100 assert(i11 != NULL);
101 if XNN_UNPREDICTABLE(i11 != zero) {
102 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
103 }
104 const int8_t* i12 = input[12];
105 assert(i12 != NULL);
106 if XNN_UNPREDICTABLE(i12 != zero) {
107 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
108 }
109 const int8_t* i13 = input[13];
110 assert(i13 != NULL);
111 if XNN_UNPREDICTABLE(i13 != zero) {
112 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
113 }
114 const int8_t* i14 = input[14];
115 assert(i14 != NULL);
116 if XNN_UNPREDICTABLE(i14 != zero) {
117 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
118 }
119 const int8_t* i15 = input[15];
120 assert(i15 != NULL);
121 if XNN_UNPREDICTABLE(i15 != zero) {
122 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
123 }
124 const int8_t* i16 = input[16];
125 assert(i16 != NULL);
126 if XNN_UNPREDICTABLE(i16 != zero) {
127 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
128 }
129 const int8_t* i17 = input[17];
130 assert(i17 != NULL);
131 if XNN_UNPREDICTABLE(i17 != zero) {
132 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
133 }
134 const int8_t* i18 = input[18];
135 assert(i18 != NULL);
136 if XNN_UNPREDICTABLE(i18 != zero) {
137 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
138 }
139 const int8_t* i19 = input[19];
140 assert(i19 != NULL);
141 if XNN_UNPREDICTABLE(i19 != zero) {
142 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
143 }
144 const int8_t* i20 = input[20];
145 assert(i20 != NULL);
146 if XNN_UNPREDICTABLE(i20 != zero) {
147 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
148 }
149 const int8_t* i21 = input[21];
150 assert(i21 != NULL);
151 if XNN_UNPREDICTABLE(i21 != zero) {
152 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
153 }
154 const int8_t* i22 = input[22];
155 assert(i22 != NULL);
156 if XNN_UNPREDICTABLE(i22 != zero) {
157 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
158 }
159 const int8_t* i23 = input[23];
160 assert(i23 != NULL);
161 if XNN_UNPREDICTABLE(i23 != zero) {
162 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
163 }
164 const int8_t* i24 = input[24];
165 assert(i24 != NULL);
166 if XNN_UNPREDICTABLE(i24 != zero) {
167 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
168 }
169 input = (const int8_t**) ((uintptr_t) input + input_stride);
170
171 size_t c = channels;
172 const void* w = weights;
173 for (; c >= 8; c -= 8) {
174 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
175 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
176
177
178 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
179 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
180 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
181 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
182 i0 += 8;
183
184
185 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
186
187 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
189
190 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
191 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
192 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
193 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
194 i1 += 8;
195
196
197 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
198
199 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
200 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
201
202 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
203 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
204 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
205 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
206 i2 += 8;
207
208
209 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
210
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
213
214 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
215 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
216 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
217 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
218 i3 += 8;
219
220
221 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
222
223 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
224 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
225
226 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
227 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
228 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
229 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
230 i4 += 8;
231
232
233 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
234
235 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
236 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
237
238 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
239 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
240 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
241 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
242 i5 += 8;
243
244
245 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
246
247 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
248 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
249
250 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
251 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
252 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
253 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
254 i6 += 8;
255
256
257 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
258
259 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
260 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
261
262 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
263 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
264 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
265 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
266 i7 += 8;
267
268
269 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
270
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
272 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
273
274 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
275 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
276 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
277 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
278 i8 += 8;
279
280
281 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
282
283 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
284 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
285
286 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
287 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
288 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
289 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
290 i9 += 8;
291
292
293 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
296 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
297
298 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
299 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
300 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
301 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
302 i10 += 8;
303
304
305 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
306
307 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
308 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
309
310 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
311 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
312 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
313 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
314 i11 += 8;
315
316
317 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
318
319 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
320 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
321
322 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
323 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
324 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
325 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
326 i12 += 8;
327
328
329 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
330
331 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
332 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
333
334 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
335 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
336 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
337 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
338 i13 += 8;
339
340
341 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
342
343 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
344 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
345
346 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
347 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
348 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
349 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
350 i14 += 8;
351
352
353 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
354
355 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
356 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
357
358 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
359 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
360 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
361 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
362 i15 += 8;
363
364
365 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
366
367 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
368 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
369
370 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
371 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
372 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
373 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
374 i16 += 8;
375
376
377 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
378
379 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
380 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
381
382 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
383 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
384 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
385 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
386 i17 += 8;
387
388
389 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
390
391 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
392 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
393
394 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
395 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
396 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
397 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
398 i18 += 8;
399
400
401 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
402
403 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
404 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
405
406 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
407 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
408 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
409 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
410 i19 += 8;
411
412
413 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
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 vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
419 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
420 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
421 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
422 i20 += 8;
423
424
425 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
426
427 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
428 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
429
430 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
431 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
432 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
433 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
434 i21 += 8;
435
436
437 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
440 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
441
442 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
443 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
444 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
445 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
446 i22 += 8;
447
448
449 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
450
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
453
454 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
455 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
456 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
457 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
458 i23 += 8;
459
460
461 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
462
463 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
464 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
465
466 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
467 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
468 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
469 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
470 i24 += 8;
471
472
473 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
474
475 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
476 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
477
478 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
479
480 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
481 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
482
483 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
484 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
485
486 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
487 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
488
489 vacc0123 = _mm_cvtps_epi32(vscaled0123);
490 vacc4567 = _mm_cvtps_epi32(vscaled4567);
491
492 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
493
494
495 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
496
497 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
498
499 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
500 output += 8;
501 }
502 if XNN_UNLIKELY(c != 0) {
503 {
504 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
505 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
506
507
508 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
509 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
510 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
511 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
512
513
514 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
518
519 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
520 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
521 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
522 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
523
524
525 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
526
527 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
528 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
529
530 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
531 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
532 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
533 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
534
535
536 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
537
538 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
539 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
540
541 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
542 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
543 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
544 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
545
546
547 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
548
549 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
550 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
551
552 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
553 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
554 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
555 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
556
557
558 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
559
560 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
561 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
562
563 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
564 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
565 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
566 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
567
568
569 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
570
571 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
572 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
573
574 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
575 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
576 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
577 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
578
579
580 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
581
582 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
583 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
584
585 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
586 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
587 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
588 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
589
590
591 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
592
593 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
594 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
595
596 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
597 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
598 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
599 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
600
601
602 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
603
604 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
605 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
606
607 const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
608 const __m128i vxi9x01234567 = _mm_cvtepi8_epi16(vi9x01234567);
609 const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
610 const __m128i vxk9x01234567 = _mm_cvtepi8_epi16(vk9x01234567);
611
612
613 vprod01234567 = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567);
614
615 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
616 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
617
618 const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
619 const __m128i vxi10x01234567 = _mm_cvtepi8_epi16(vi10x01234567);
620 const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
621 const __m128i vxk10x01234567 = _mm_cvtepi8_epi16(vk10x01234567);
622
623
624 vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
625
626 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
627 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
628
629 const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
630 const __m128i vxi11x01234567 = _mm_cvtepi8_epi16(vi11x01234567);
631 const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
632 const __m128i vxk11x01234567 = _mm_cvtepi8_epi16(vk11x01234567);
633
634
635 vprod01234567 = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567);
636
637 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
638 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
639
640 const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
641 const __m128i vxi12x01234567 = _mm_cvtepi8_epi16(vi12x01234567);
642 const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
643 const __m128i vxk12x01234567 = _mm_cvtepi8_epi16(vk12x01234567);
644
645
646 vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
647
648 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
649 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
650
651 const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
652 const __m128i vxi13x01234567 = _mm_cvtepi8_epi16(vi13x01234567);
653 const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
654 const __m128i vxk13x01234567 = _mm_cvtepi8_epi16(vk13x01234567);
655
656
657 vprod01234567 = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567);
658
659 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
660 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
661
662 const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
663 const __m128i vxi14x01234567 = _mm_cvtepi8_epi16(vi14x01234567);
664 const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
665 const __m128i vxk14x01234567 = _mm_cvtepi8_epi16(vk14x01234567);
666
667
668 vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
669
670 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
671 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
672
673 const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
674 const __m128i vxi15x01234567 = _mm_cvtepi8_epi16(vi15x01234567);
675 const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
676 const __m128i vxk15x01234567 = _mm_cvtepi8_epi16(vk15x01234567);
677
678
679 vprod01234567 = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567);
680
681 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
682 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
683
684 const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
685 const __m128i vxi16x01234567 = _mm_cvtepi8_epi16(vi16x01234567);
686 const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
687 const __m128i vxk16x01234567 = _mm_cvtepi8_epi16(vk16x01234567);
688
689
690 vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
691
692 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
693 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
694
695 const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
696 const __m128i vxi17x01234567 = _mm_cvtepi8_epi16(vi17x01234567);
697 const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
698 const __m128i vxk17x01234567 = _mm_cvtepi8_epi16(vk17x01234567);
699
700
701 vprod01234567 = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567);
702
703 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
704 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
705
706 const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
707 const __m128i vxi18x01234567 = _mm_cvtepi8_epi16(vi18x01234567);
708 const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
709 const __m128i vxk18x01234567 = _mm_cvtepi8_epi16(vk18x01234567);
710
711
712 vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
713
714 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
715 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
716
717 const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
718 const __m128i vxi19x01234567 = _mm_cvtepi8_epi16(vi19x01234567);
719 const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
720 const __m128i vxk19x01234567 = _mm_cvtepi8_epi16(vk19x01234567);
721
722
723 vprod01234567 = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567);
724
725 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
726 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
727
728 const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
729 const __m128i vxi20x01234567 = _mm_cvtepi8_epi16(vi20x01234567);
730 const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
731 const __m128i vxk20x01234567 = _mm_cvtepi8_epi16(vk20x01234567);
732
733
734 vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
735
736 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
737 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
738
739 const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
740 const __m128i vxi21x01234567 = _mm_cvtepi8_epi16(vi21x01234567);
741 const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
742 const __m128i vxk21x01234567 = _mm_cvtepi8_epi16(vk21x01234567);
743
744
745 vprod01234567 = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567);
746
747 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
748 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
749
750 const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
751 const __m128i vxi22x01234567 = _mm_cvtepi8_epi16(vi22x01234567);
752 const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
753 const __m128i vxk22x01234567 = _mm_cvtepi8_epi16(vk22x01234567);
754
755
756 vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
757
758 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
759 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
760
761 const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
762 const __m128i vxi23x01234567 = _mm_cvtepi8_epi16(vi23x01234567);
763 const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
764 const __m128i vxk23x01234567 = _mm_cvtepi8_epi16(vk23x01234567);
765
766
767 vprod01234567 = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567);
768
769 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
770 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
771
772 const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
773 const __m128i vxi24x01234567 = _mm_cvtepi8_epi16(vi24x01234567);
774 const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
775 const __m128i vxk24x01234567 = _mm_cvtepi8_epi16(vk24x01234567);
776
777
778 vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
779
780 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
781 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
782
783
784 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
785 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
786
787 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
788 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
789
790 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
791 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
792
793 vacc0123 = _mm_cvtps_epi32(vscaled0123);
794 vacc4567 = _mm_cvtps_epi32(vscaled4567);
795
796
797 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
798
799
800 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
801
802 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
803
804 if (c & 4) {
805 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
806 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
807 output += 4;
808 }
809 if (c & 2) {
810 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
811 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
812 output += 2;
813 }
814 if (c & 1) {
815 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
816 output += 1;
817 }
818 }
819 }
820
821 output = (int8_t*) ((uintptr_t) output + output_increment);
822 } while (--output_width != 0);
823 }