Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c8s8r-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_5f5m5l16c8s8r__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 > 5);
36
37
38 const __m128 vscale = _mm_set1_ps(params->fp32_scalar.scale);
39 XNN_FORCE_REALIZATION(vscale);
40
41 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
42 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
43 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
44 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47
48 do {
49 const void* w = weights;
50
51 // First pass to process 5 inputs.
52 {
53 int32_t* b = buffer;
54 const int8_t* i0 = input[0];
55 assert(i0 != NULL);
56 if XNN_UNPREDICTABLE(i0 != zero) {
57 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
58 }
59 const int8_t* i1 = input[1];
60 assert(i1 != NULL);
61 if XNN_UNPREDICTABLE(i1 != zero) {
62 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
63 }
64 const int8_t* i2 = input[2];
65 assert(i2 != NULL);
66 if XNN_UNPREDICTABLE(i2 != zero) {
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
68 }
69 const int8_t* i3 = input[3];
70 assert(i3 != NULL);
71 if XNN_UNPREDICTABLE(i3 != zero) {
72 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
73 }
74 const int8_t* i4 = input[4];
75 assert(i4 != NULL);
76 if XNN_UNPREDICTABLE(i4 != zero) {
77 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
78 }
79 input += 5;
80
81 size_t c = round_up_po2(channels, 8);
82
83 for (; c >= 16; c -= 16) {
84 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
85 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
86 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
87 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
88
89
90
91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
92 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
93 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
94 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
95 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
96 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
97 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
98 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
99 i0 += 16;
100
101
102 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
103 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
104
105
106 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
107 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
108 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
109 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
110 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
111 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
112 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
113 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
114 i1 += 16;
115
116
117 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
118 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
119
120 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
121 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
122 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
123 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
124
125 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
126 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
127 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
128 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
129 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
130 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
131 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
132 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
133 i2 += 16;
134
135
136 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
137 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
138
139
140 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
141 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
142 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
143 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
144 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
145 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
146 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
147 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
148 i3 += 16;
149
150
151 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
152 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
153
154 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
155 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
156 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
157 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
158
159 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
160 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
161 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
162 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
163 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
164 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
165 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
166 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
167 i4 += 16;
168
169
170 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
171 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
172
173 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
174 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
175 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
176 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
177
178 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
179
180 _mm_storeu_si128((__m128i*) b, vacc0123);
181 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
182 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
183 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
184 b += 16;
185 }
186
187 if XNN_UNLIKELY(c != 0) {
188 do {
189 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
190 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
191
192
193 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
194 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
195 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 * sizeof(int8_t)));
196 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
197 i0 += 8;
198
199
200 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
201
202
203 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
204 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
205 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 8 * sizeof(int8_t)));
206 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
207 i1 += 8;
208
209
210 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
214
215 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
216 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
217 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 16 * sizeof(int8_t)));
218 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
219 i2 += 8;
220
221
222 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
223
224
225 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
226 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
227 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 24 * sizeof(int8_t)));
228 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
229 i3 += 8;
230
231
232 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
233
234 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
235 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
236
237 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
238 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
239 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 32 * sizeof(int8_t)));
240 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
241 i4 += 8;
242
243
244 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
245
246 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
247 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
248
249 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
250
251 _mm_storeu_si128((__m128i*) (b), vacc0123);
252 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
253 b += 8;
254 c -= 8;
255 } while (c != 0);
256 }
257 }
258
259 // Middle pass to process 5 inputs in each iteration.
260 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
261 int32_t* b = buffer;
262 const int8_t* i0 = input[0];
263 assert(i0 != NULL);
264 if XNN_UNPREDICTABLE(i0 != zero) {
265 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
266 }
267 const int8_t* i1 = input[1];
268 assert(i1 != NULL);
269 if XNN_UNPREDICTABLE(i1 != zero) {
270 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
271 }
272 const int8_t* i2 = input[2];
273 assert(i2 != NULL);
274 if XNN_UNPREDICTABLE(i2 != zero) {
275 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
276 }
277 const int8_t* i3 = input[3];
278 assert(i3 != NULL);
279 if XNN_UNPREDICTABLE(i3 != zero) {
280 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
281 }
282 const int8_t* i4 = input[4];
283 assert(i4 != NULL);
284 if XNN_UNPREDICTABLE(i4 != zero) {
285 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
286 }
287 input += 5;
288
289 size_t c = round_up_po2(channels, 8);
290
291 for (; c >= 16; c -= 16) {
292 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
293 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
294 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
295 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
296
297
298
299 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
300 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
301 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
302 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
303 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
304 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
305 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
306 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
307 i0 += 16;
308
309
310 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
311 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
312
313
314 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
315 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
316 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
317 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
318 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
319 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
320 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
321 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
322 i1 += 16;
323
324
325 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
326 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
327
328 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
329 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
330 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
331 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
332
333 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
334 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
335 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
336 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
337 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
338 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
339 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
340 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
341 i2 += 16;
342
343
344 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
345 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
346
347
348 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
349 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
350 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
351 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
352 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
353 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
354 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
355 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
356 i3 += 16;
357
358
359 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
360 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
361
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
363 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
364 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
365 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
366
367 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
368 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
369 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
370 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
371 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
372 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
373 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
374 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
375 i4 += 16;
376
377
378 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
379 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
380
381 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
382 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
383 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
384 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
385
386 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
387
388 _mm_storeu_si128((__m128i*) (b), vacc0123);
389 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
390 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
391 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
392 b += 16;
393 }
394
395 if XNN_UNLIKELY(c != 0) {
396 do {
397 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
398 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
399
400
401 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
402 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
403 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
404 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
405 i0 += 8;
406
407
408 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
409
410
411 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
412 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
413 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
414 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
415 i1 += 8;
416
417
418 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
419
420 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
421 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
422
423 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
424 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
425 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
426 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
427 i2 += 8;
428
429
430 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
431
432
433 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
434 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
435 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
436 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
437 i3 += 8;
438
439
440 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
441
442 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
443 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
444
445 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
446 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
447 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
448 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
449 i4 += 8;
450
451
452 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
453
454 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
455 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
456
457 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
458
459 _mm_storeu_si128((__m128i*) (b), vacc0123);
460 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
461 b += 8;
462 c -= 8;
463 } while (c != 0);
464 }
465 }
466
467 // Last pass to process up to 5 inputs.
468 {
469 const int32_t* b = buffer;
470 const int8_t* i0 = input[0];
471 assert(i0 != NULL);
472 if XNN_UNPREDICTABLE(i0 != zero) {
473 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
474 }
475 const int8_t* i1 = input[1];
476 assert(i1 != NULL);
477 if XNN_UNPREDICTABLE(i1 != zero) {
478 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
479 }
480 const int8_t* i2 = input[2];
481 assert(i2 != NULL);
482 if XNN_UNPREDICTABLE(i2 != zero) {
483 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
484 }
485 const int8_t* i3 = input[3];
486 assert(i3 != NULL);
487 if XNN_UNPREDICTABLE(i3 != zero) {
488 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
489 }
490 const int8_t* i4 = input[4];
491 assert(i4 != NULL);
492 if XNN_UNPREDICTABLE(i4 != zero) {
493 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
494 }
495
496 size_t c = channels;
497
498 for (; c >= 16; c -= 16) {
499 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
500 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
501 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
502 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
503 b += 16;
504
505
506 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
507 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
508 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
509 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
510 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
511 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
512 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
513 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
514 i0 += 16;
515
516
517 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
518 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
519
520
521 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
522 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
523 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
524 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
525 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
526 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
527 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
528 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
529 i1 += 16;
530
531
532 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
533 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
534
535 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
536 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
537 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
538 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
539
540 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
541 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
542 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
543 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
544 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
545 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
546 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
547 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
548 i2 += 16;
549
550
551 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
552 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
553
554
555 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
556 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
557 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
558 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
559 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
560 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
561 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
562 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
563 i3 += 16;
564
565
566 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
567 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
568
569 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
570 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
571 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
572 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
573
574 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
575 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
576 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
577 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
578 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
579 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
580 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
581 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
582 i4 += 16;
583
584
585 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
586 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
587
588 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
589 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
590 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
591 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
592
593 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
594
595 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
596 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
597 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
598 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
599
600 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
601 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
602 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
603 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
604
605 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
606 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
607 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
608 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
609
610 vacc0123 = _mm_cvtps_epi32(vscaled0123);
611 vacc4567 = _mm_cvtps_epi32(vscaled4567);
612 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
613 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
614
615 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
616 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
617
618
619 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
620
621 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
622
623 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
624 output += 16;
625 }
626
627 if XNN_UNLIKELY(c != 0) {
628 do {
629 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
630 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
631 b += 8;
632
633
634 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
635 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
636 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
637 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
638 i0 += 8;
639
640
641 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
642
643
644 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
645 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
646 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
647 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
648 i1 += 8;
649
650
651 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
652
653 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
654 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
655
656 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
657 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
658 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
659 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
660 i2 += 8;
661
662
663 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
664
665
666 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
667 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
668 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
669 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
670 i3 += 8;
671
672
673 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
674
675 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
676 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
677
678 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
679 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
680 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
681 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
682 i4 += 8;
683
684
685 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
686
687 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
688 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
689
690 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
691 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
692
693 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
694 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
695
696 w = (void*) ((uintptr_t) w + 40);
697
698 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
699 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
700
701 vacc0123 = _mm_cvtps_epi32(vscaled0123);
702 vacc4567 = _mm_cvtps_epi32(vscaled4567);
703
704 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
705
706
707 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
708
709 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
710
711 if XNN_LIKELY(c >= 8) {
712 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
713 output += 8;
714 c -= 8;
715 } else {
716 if (c & 4) {
717 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
718 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
719 output += 4;
720 }
721 if (c & 2) {
722 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
723 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
724 output += 2;
725 }
726 if (c & 1) {
727 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
728 output += 1;
729 }
730 c = 0;
731 }
732 } while (c != 0);
733 }
734 }
735
736 input = (const int8_t**) ((uintptr_t) input + input_stride);
737 output = (int8_t*) ((uintptr_t) output + output_increment);
738 } while (--output_width != 0);
739 }