Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c8s8r-minmax-fp32-sse41-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include "xnnpack/dwconv.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_8f8m9l16c8s8r__sse41_mul16_add16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 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_epi8(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 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 >= 16; c -= 16) {
99 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
100 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
101 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
102 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
103
104
105
106 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
107 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
108 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
109 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
110 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
111 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
112 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
113 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
114 i0 += 16;
115
116
117 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
118 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
119
120
121 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
122 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
123 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
124 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
125 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
126 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
127 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
128 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
129 i1 += 16;
130
131
132 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
133 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
134
135 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
136 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
137 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
138 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
139
140 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
141 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
142 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
143 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
144 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
145 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
146 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
147 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
148 i2 += 16;
149
150
151 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
152 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
153
154
155 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
156 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
157 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
158 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
159 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
160 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
161 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
162 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
163 i3 += 16;
164
165
166 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
167 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
168
169 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
170 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
171 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
172 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
173
174 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
175 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
176 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
177 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
178 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
179 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
180 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
181 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
182 i4 += 16;
183
184
185 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
186 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
187
188
189 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
190 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
191 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
192 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
193 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
194 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
195 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
196 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
197 i5 += 16;
198
199
200 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
201 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
205 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
206 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
207
208 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
209 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
210 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
211 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
212 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
213 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
214 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
215 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
216 i6 += 16;
217
218
219 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
220 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
221
222
223 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
224 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
225 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
226 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
227 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
228 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
229 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
230 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
231 i7 += 16;
232
233
234 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
235 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
236
237 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
238 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
239 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
240 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
241
242 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t));
243
244 _mm_storeu_si128((__m128i*) b, vacc0123);
245 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
246 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
247 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
248 b += 16;
249 }
250
251 if XNN_UNLIKELY(c != 0) {
252 do {
253 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
254 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
255
256
257 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
258 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
259 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 * sizeof(int8_t)));
260 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
261 i0 += 8;
262
263
264 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
265
266
267 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
268 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
269 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 8 * sizeof(int8_t)));
270 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
271 i1 += 8;
272
273
274 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
275
276 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
277 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
278
279 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
280 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
281 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 16 * sizeof(int8_t)));
282 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
283 i2 += 8;
284
285
286 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
287
288
289 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
290 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
291 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 24 * sizeof(int8_t)));
292 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
293 i3 += 8;
294
295
296 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
300
301 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
302 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
303 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 32 * sizeof(int8_t)));
304 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
305 i4 += 8;
306
307
308 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
309
310
311 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
312 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
313 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 40 * sizeof(int8_t)));
314 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
315 i5 += 8;
316
317
318 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
322
323 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
324 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
325 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 48 * sizeof(int8_t)));
326 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
327 i6 += 8;
328
329
330 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
331
332
333 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
334 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
335 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 56 * sizeof(int8_t)));
336 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
337 i7 += 8;
338
339
340 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
341
342 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
343 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
344
345 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t));
346
347 _mm_storeu_si128((__m128i*) (b), vacc0123);
348 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
349 b += 8;
350 c -= 8;
351 } while (c != 0);
352 }
353 }
354
355 // Middle pass to process 8 inputs in each iteration.
356 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
357 int32_t* b = buffer;
358 const int8_t* i0 = input[0];
359 assert(i0 != NULL);
360 if XNN_UNPREDICTABLE(i0 != zero) {
361 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
362 }
363 const int8_t* i1 = input[1];
364 assert(i1 != NULL);
365 if XNN_UNPREDICTABLE(i1 != zero) {
366 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
367 }
368 const int8_t* i2 = input[2];
369 assert(i2 != NULL);
370 if XNN_UNPREDICTABLE(i2 != zero) {
371 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
372 }
373 const int8_t* i3 = input[3];
374 assert(i3 != NULL);
375 if XNN_UNPREDICTABLE(i3 != zero) {
376 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
377 }
378 const int8_t* i4 = input[4];
379 assert(i4 != NULL);
380 if XNN_UNPREDICTABLE(i4 != zero) {
381 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
382 }
383 const int8_t* i5 = input[5];
384 assert(i5 != NULL);
385 if XNN_UNPREDICTABLE(i5 != zero) {
386 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
387 }
388 const int8_t* i6 = input[6];
389 assert(i6 != NULL);
390 if XNN_UNPREDICTABLE(i6 != zero) {
391 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
392 }
393 const int8_t* i7 = input[7];
394 assert(i7 != NULL);
395 if XNN_UNPREDICTABLE(i7 != zero) {
396 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
397 }
398 input += 8;
399
400 size_t c = round_up_po2(channels, 8);
401
402 for (; c >= 16; c -= 16) {
403 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
404 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
405 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
406 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
407
408
409
410 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
411 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
412 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
413 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
414 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
415 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
416 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
417 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
418 i0 += 16;
419
420
421 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
422 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
423
424
425 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
426 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
427 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
428 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
429 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
430 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
431 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
432 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
433 i1 += 16;
434
435
436 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
437 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
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 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
442 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
443
444 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
445 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
446 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
447 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
448 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
449 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
450 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
451 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
452 i2 += 16;
453
454
455 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
456 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
457
458
459 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
460 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
461 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
462 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
463 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
464 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
465 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
466 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
467 i3 += 16;
468
469
470 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
471 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
472
473 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
474 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
475 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
476 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
477
478 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
479 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
480 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
481 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
482 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
483 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
484 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
485 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
486 i4 += 16;
487
488
489 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
490 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
491
492
493 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
494 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
495 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)));
496 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
497 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
498 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
499 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t)));
500 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
501 i5 += 16;
502
503
504 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
505 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
506
507 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
508 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
509 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
510 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
511
512 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
513 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
514 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)));
515 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
516 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
517 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
518 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t)));
519 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
520 i6 += 16;
521
522
523 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
524 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
525
526
527 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
528 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
529 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)));
530 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
531 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
532 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
533 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t)));
534 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
535 i7 += 16;
536
537
538 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
539 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
540
541 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
542 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
543 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
544 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
545
546 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
547
548 _mm_storeu_si128((__m128i*) (b), vacc0123);
549 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
550 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
551 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
552 b += 16;
553 }
554
555 if XNN_UNLIKELY(c != 0) {
556 do {
557 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
558 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
559
560
561 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
562 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
563 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
564 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
565 i0 += 8;
566
567
568 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
569
570
571 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
572 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
573 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
574 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
575 i1 += 8;
576
577
578 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
579
580 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
581 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
582
583 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
584 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
585 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
586 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
587 i2 += 8;
588
589
590 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
591
592
593 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
594 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
595 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
596 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
597 i3 += 8;
598
599
600 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
601
602 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
603 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
604
605 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
606 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
607 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
608 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
609 i4 += 8;
610
611
612 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
613
614
615 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
616 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
617 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
618 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
619 i5 += 8;
620
621
622 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
623
624 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
625 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
626
627 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
628 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
629 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
630 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
631 i6 += 8;
632
633
634 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
635
636
637 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
638 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
639 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
640 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
641 i7 += 8;
642
643
644 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
645
646 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
647 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
648
649 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
650
651 _mm_storeu_si128((__m128i*) (b), vacc0123);
652 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
653 b += 8;
654 c -= 8;
655 } while (c != 0);
656 }
657 }
658
659 // Last pass to process up to 9 inputs.
660 {
661 const int32_t* b = buffer;
662 const int8_t* i0 = input[0];
663 assert(i0 != NULL);
664 if XNN_UNPREDICTABLE(i0 != zero) {
665 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
666 }
667 const int8_t* i1 = input[1];
668 assert(i1 != NULL);
669 if XNN_UNPREDICTABLE(i1 != zero) {
670 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
671 }
672 const int8_t* i2 = input[2];
673 assert(i2 != NULL);
674 if XNN_UNPREDICTABLE(i2 != zero) {
675 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
676 }
677 const int8_t* i3 = input[3];
678 assert(i3 != NULL);
679 if XNN_UNPREDICTABLE(i3 != zero) {
680 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
681 }
682 const int8_t* i4 = input[4];
683 assert(i4 != NULL);
684 if XNN_UNPREDICTABLE(i4 != zero) {
685 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
686 }
687 const int8_t* i5 = input[5];
688 assert(i5 != NULL);
689 if XNN_UNPREDICTABLE(i5 != zero) {
690 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
691 }
692 const int8_t* i6 = input[6];
693 assert(i6 != NULL);
694 if XNN_UNPREDICTABLE(i6 != zero) {
695 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
696 }
697 const int8_t* i7 = input[7];
698 assert(i7 != NULL);
699 if XNN_UNPREDICTABLE(i7 != zero) {
700 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
701 }
702 const int8_t* i8 = input[8];
703 assert(i8 != NULL);
704 if XNN_UNPREDICTABLE(i8 != zero) {
705 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
706 }
707
708 size_t c = channels;
709
710 for (; c >= 16; c -= 16) {
711 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
712 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
713 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
714 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
715 b += 16;
716
717
718 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
719 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
720 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
721 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
722 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
723 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
724 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
725 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
726 i0 += 16;
727
728
729 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
730 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
731
732
733 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
734 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
735 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
736 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
737 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
738 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
739 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
740 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
741 i1 += 16;
742
743
744 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
745 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
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 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
750 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
751
752 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
753 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
754 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
755 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
756 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
757 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
758 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
759 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
760 i2 += 16;
761
762
763 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
764 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
765
766
767 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
768 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
769 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
770 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
771 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
772 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
773 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
774 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
775 i3 += 16;
776
777
778 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
779 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
780
781 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
782 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
783 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
784 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
785
786 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
787 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
788 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
789 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
790 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
791 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
792 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
793 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
794 i4 += 16;
795
796
797 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
798 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
799
800
801 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
802 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
803 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80 * sizeof(int8_t)));
804 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
805 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
806 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
807 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88 * sizeof(int8_t)));
808 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
809 i5 += 16;
810
811
812 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
813 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
814
815 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
816 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
817 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
818 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
819
820 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
821 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
822 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96 * sizeof(int8_t)));
823 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
824 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
825 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
826 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 104 * sizeof(int8_t)));
827 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
828 i6 += 16;
829
830
831 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
832 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
833
834
835 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
836 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
837 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 112 * sizeof(int8_t)));
838 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
839 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
840 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
841 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 120 * sizeof(int8_t)));
842 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
843 i7 += 16;
844
845
846 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
847 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
848
849 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
850 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
851 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
852 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
853
854 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
855 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
856 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 128 * sizeof(int8_t)));
857 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
858 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
859 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
860 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 136 * sizeof(int8_t)));
861 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
862 i8 += 16;
863
864
865 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
866 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
867
868 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
869 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
870 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
871 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
872
873 w = (const void*) ((uintptr_t) w + 144 * sizeof(int8_t));
874
875 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
876 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
877 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
878 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
879
880 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
881 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
882 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
883 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
884
885 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
886 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
887 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
888 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
889
890 vacc0123 = _mm_cvtps_epi32(vscaled0123);
891 vacc4567 = _mm_cvtps_epi32(vscaled4567);
892 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
893 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
894
895 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
896 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
897
898
899 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
900
901 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
902
903 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
904 output += 16;
905 }
906
907 if XNN_UNLIKELY(c != 0) {
908 do {
909 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
910 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
911 b += 8;
912
913
914 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
915 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
916 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
917 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
918 i0 += 8;
919
920
921 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
922
923
924 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
925 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
926 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
927 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
928 i1 += 8;
929
930
931 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
932
933 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
934 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
935
936 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
937 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
938 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
939 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
940 i2 += 8;
941
942
943 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
944
945
946 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
947 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
948 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
949 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
950 i3 += 8;
951
952
953 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
954
955 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
956 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
957
958 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
959 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
960 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
961 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
962 i4 += 8;
963
964
965 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
966
967
968 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
969 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
970 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
971 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
972 i5 += 8;
973
974
975 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
976
977 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
978 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
979
980 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
981 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
982 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
983 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
984 i6 += 8;
985
986
987 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
988
989
990 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
991 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
992 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
993 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
994 i7 += 8;
995
996
997 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
998
999 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1000 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1001
1002 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
1003 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
1004 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
1005 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
1006 i8 += 8;
1007
1008
1009 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
1010
1011 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
1012 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
1013
1014 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
1015 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
1016
1017 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
1018 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
1019
1020 w = (void*) ((uintptr_t) w + 72);
1021
1022 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
1023 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
1024
1025 vacc0123 = _mm_cvtps_epi32(vscaled0123);
1026 vacc4567 = _mm_cvtps_epi32(vscaled4567);
1027
1028 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
1029
1030
1031 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
1032
1033 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
1034
1035 if XNN_LIKELY(c >= 8) {
1036 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
1037 output += 8;
1038 c -= 8;
1039 } else {
1040 if (c & 4) {
1041 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
1042 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
1043 output += 4;
1044 }
1045 if (c & 2) {
1046 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
1047 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
1048 output += 2;
1049 }
1050 if (c & 1) {
1051 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
1052 output += 1;
1053 }
1054 c = 0;
1055 }
1056 } while (c != 0);
1057 }
1058 }
1059
1060 input = (const int8_t**) ((uintptr_t) input + input_stride);
1061 output = (int8_t*) ((uintptr_t) output + output_increment);
1062 } while (--output_width != 0);
1063 }