Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c8s8r-minmax-fp32-sse41-mul16.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(
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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
106 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
107 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
108 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
109
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
111 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
112 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
113 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
114 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
115 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
116 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
117 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
118 i1 += 16;
119
120
121 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
122 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
123
124 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
125 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
126 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
127 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
128
129 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
130 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
131 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
132 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
133 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
134 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
135 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
136 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
137 i2 += 16;
138
139
140 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
141 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
142
143 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
144 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
145 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
146 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
147
148 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
149 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
150 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
151 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
152 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
153 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
154 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
155 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
156 i3 += 16;
157
158
159 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
160 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
161
162 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
163 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
164 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
165 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
166
167 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
168 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
169 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
170 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
171 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
172 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
173 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
174 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
175 i4 += 16;
176
177
178 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
179 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
180
181 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
183 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
184 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
185
186 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t));
187
188 _mm_storeu_si128((__m128i*) b, vacc0123);
189 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
190 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
191 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
192 b += 16;
193 }
194
195 if XNN_UNLIKELY(c != 0) {
196 do {
197 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
198 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
199
200
201 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
202 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
203 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 * sizeof(int8_t)));
204 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
205 i0 += 8;
206
207
208 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
209
210 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
211 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
212
213 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
214 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
215 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 8 * sizeof(int8_t)));
216 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
217 i1 += 8;
218
219
220 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
221
222 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
223 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
224
225 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
226 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
227 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 16 * sizeof(int8_t)));
228 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
229 i2 += 8;
230
231
232 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
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 vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
238 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
239 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 24 * sizeof(int8_t)));
240 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
241 i3 += 8;
242
243
244 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
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 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
250 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
251 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + sizeof(int32_t) * 8 + 32 * sizeof(int8_t)));
252 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
253 i4 += 8;
254
255
256 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
257
258 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
259 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
260
261 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t));
262
263 _mm_storeu_si128((__m128i*) (b), vacc0123);
264 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
265 b += 8;
266 c -= 8;
267 } while (c != 0);
268 }
269 }
270
271 // Middle pass to process 5 inputs in each iteration.
272 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
273 int32_t* b = buffer;
274 const int8_t* i0 = input[0];
275 assert(i0 != NULL);
276 if XNN_UNPREDICTABLE(i0 != zero) {
277 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
278 }
279 const int8_t* i1 = input[1];
280 assert(i1 != NULL);
281 if XNN_UNPREDICTABLE(i1 != zero) {
282 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
283 }
284 const int8_t* i2 = input[2];
285 assert(i2 != NULL);
286 if XNN_UNPREDICTABLE(i2 != zero) {
287 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
288 }
289 const int8_t* i3 = input[3];
290 assert(i3 != NULL);
291 if XNN_UNPREDICTABLE(i3 != zero) {
292 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
293 }
294 const int8_t* i4 = input[4];
295 assert(i4 != NULL);
296 if XNN_UNPREDICTABLE(i4 != zero) {
297 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
298 }
299 input += 5;
300
301 size_t c = round_up_po2(channels, 8);
302
303 for (; c >= 16; c -= 16) {
304 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
305 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
306 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
307 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
308
309
310
311 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
312 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
313 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
314 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
315 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
316 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
317 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
318 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
319 i0 += 16;
320
321
322 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
323 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
324
325 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
326 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
327 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
328 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
329
330 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
331 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
332 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
333 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
334 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
335 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
336 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
337 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
338 i1 += 16;
339
340
341 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
342 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
343
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
345 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
346 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
347 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
348
349 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
350 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
351 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
352 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
353 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
354 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
355 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
356 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
357 i2 += 16;
358
359
360 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
361 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
362
363 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
364 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
365 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
366 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
367
368 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
369 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
370 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
371 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
372 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
373 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
374 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
375 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
376 i3 += 16;
377
378
379 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
380 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
381
382 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
383 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
384 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
385 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
386
387 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
388 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
389 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
390 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
391 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
392 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
393 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
394 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
395 i4 += 16;
396
397
398 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
399 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
400
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
403 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
404 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
405
406 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
407
408 _mm_storeu_si128((__m128i*) (b), vacc0123);
409 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
410 _mm_storeu_si128((__m128i*) (b + 8), vacc89AB);
411 _mm_storeu_si128((__m128i*) (b + 12), vaccCDEF);
412 b += 16;
413 }
414
415 if XNN_UNLIKELY(c != 0) {
416 do {
417 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
418 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
419
420
421 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
422 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
423 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
424 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
425 i0 += 8;
426
427
428 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
429
430 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
431 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
432
433 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
434 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
435 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
436 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
437 i1 += 8;
438
439
440 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
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 vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
446 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
447 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
448 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
449 i2 += 8;
450
451
452 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
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 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
458 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
459 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
460 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
461 i3 += 8;
462
463
464 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
465
466 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
467 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
468
469 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
470 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
471 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
472 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
473 i4 += 8;
474
475
476 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
477
478 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
479 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
480
481 w = (const void*) ((uintptr_t) w + 40 * sizeof(int8_t));
482
483 _mm_storeu_si128((__m128i*) (b), vacc0123);
484 _mm_storeu_si128((__m128i*) (b + 4), vacc4567);
485 b += 8;
486 c -= 8;
487 } while (c != 0);
488 }
489 }
490
491 // Last pass to process up to 5 inputs.
492 {
493 const int32_t* b = buffer;
494 const int8_t* i0 = input[0];
495 assert(i0 != NULL);
496 if XNN_UNPREDICTABLE(i0 != zero) {
497 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
498 }
499 const int8_t* i1 = input[1];
500 assert(i1 != NULL);
501 if XNN_UNPREDICTABLE(i1 != zero) {
502 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
503 }
504 const int8_t* i2 = input[2];
505 assert(i2 != NULL);
506 if XNN_UNPREDICTABLE(i2 != zero) {
507 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
508 }
509 const int8_t* i3 = input[3];
510 assert(i3 != NULL);
511 if XNN_UNPREDICTABLE(i3 != zero) {
512 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
513 }
514 const int8_t* i4 = input[4];
515 assert(i4 != NULL);
516 if XNN_UNPREDICTABLE(i4 != zero) {
517 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
518 }
519
520 size_t c = channels;
521
522 for (; c >= 16; c -= 16) {
523 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
524 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
525 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) (b + 8));
526 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) (b + 12));
527 b += 16;
528
529
530 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
531 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
532 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 0 * sizeof(int8_t)));
533 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
534 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
535 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
536 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
537 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
538 i0 += 16;
539
540
541 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
542 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
545 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
546 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
547 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
548
549 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
550 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
551 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
552 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
553 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
554 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
555 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
556 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
557 i1 += 16;
558
559
560 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
561 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
562
563 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
564 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
565 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
566 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
567
568 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
569 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
570 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
571 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
572 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
573 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
574 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40 * sizeof(int8_t)));
575 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
576 i2 += 16;
577
578
579 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
580 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
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 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
585 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
586
587 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
588 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
589 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
590 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
591 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
592 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
593 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56 * sizeof(int8_t)));
594 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
595 i3 += 16;
596
597
598 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
599 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
600
601 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
602 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
603 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
604 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
605
606 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
607 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
608 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64 * sizeof(int8_t)));
609 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
610 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
611 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
612 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72 * sizeof(int8_t)));
613 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
614 i4 += 16;
615
616
617 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
618 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
619
620 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
621 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
622 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
623 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
624
625 w = (const void*) ((uintptr_t) w + 80 * sizeof(int8_t));
626
627 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
628 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
629 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
630 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
631
632 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
633 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
634 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
635 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
636
637 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
638 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
639 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
640 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
641
642 vacc0123 = _mm_cvtps_epi32(vscaled0123);
643 vacc4567 = _mm_cvtps_epi32(vscaled4567);
644 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
645 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
646
647 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
648 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
649
650
651 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
652
653 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
654
655 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
656 output += 16;
657 }
658
659 if XNN_UNLIKELY(c != 0) {
660 do {
661 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) b);
662 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) (b + 4));
663 b += 8;
664
665
666 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
667 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
668 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) w);
669 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
670 i0 += 8;
671
672
673 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
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 vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
679 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
680 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
681 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
682 i1 += 8;
683
684
685 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
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 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
691 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
692 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
693 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
694 i2 += 8;
695
696
697 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
698
699 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
700 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
701
702 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
703 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
704 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
705 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
706 i3 += 8;
707
708
709 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
710
711 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
712 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
713
714 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
715 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
716 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
717 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
718 i4 += 8;
719
720
721 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
722
723 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
724 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
725
726 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
727 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
728
729 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
730 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
731
732 w = (void*) ((uintptr_t) w + 40);
733
734 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
735 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
736
737 vacc0123 = _mm_cvtps_epi32(vscaled0123);
738 vacc4567 = _mm_cvtps_epi32(vscaled4567);
739
740 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
741
742
743 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
744
745 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
746
747 if XNN_LIKELY(c >= 8) {
748 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
749 output += 8;
750 c -= 8;
751 } else {
752 if (c & 4) {
753 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
754 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
755 output += 4;
756 }
757 if (c & 2) {
758 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
759 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
760 output += 2;
761 }
762 if (c & 1) {
763 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
764 output += 1;
765 }
766 c = 0;
767 }
768 } while (c != 0);
769 }
770 }
771
772 input = (const int8_t**) ((uintptr_t) input + input_stride);
773 output = (int8_t*) ((uintptr_t) output + output_increment);
774 } while (--output_width != 0);
775 }