Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c8s8r-minmax-rndnu-neon-mul8-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-neon-mul8.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_dwconv_minmax_rndnu_ukernel_5f5m5l16c8s8r__neon_mul8_ld128(
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 const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
38 const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
39 const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
40 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
41 const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
42 const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
43 do {
44 const void* w = weights;
45
46 // First pass to process 5 inputs.
47 {
48 int32_t* b = buffer;
49 const int8_t* i0 = input[0];
50 assert(i0 != NULL);
51 if XNN_UNPREDICTABLE(i0 != zero) {
52 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
53 }
54 const int8_t* i1 = input[1];
55 assert(i1 != NULL);
56 if XNN_UNPREDICTABLE(i1 != zero) {
57 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
58 }
59 const int8_t* i2 = input[2];
60 assert(i2 != NULL);
61 if XNN_UNPREDICTABLE(i2 != zero) {
62 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
63 }
64 const int8_t* i3 = input[3];
65 assert(i3 != NULL);
66 if XNN_UNPREDICTABLE(i3 != zero) {
67 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
68 }
69 const int8_t* i4 = input[4];
70 assert(i4 != NULL);
71 if XNN_UNPREDICTABLE(i4 != zero) {
72 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
73 }
74 input += 5;
75
76 size_t c = round_up_po2(channels, 8);
77
78 for (; c >= 16; c -= 16) {
79 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
80 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
81 int32x4_t vacc89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
82 int32x4_t vaccCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
83
84 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
85 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
86
87 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
88 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
89
90 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
91 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
92 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
93 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
94
95 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
96 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
97
98 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
99 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
100
101 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
102 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
103 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
104 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
105
106 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
107 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108
109 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
110 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
111
112 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
113 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
114 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
115 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
116
117 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
118 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119
120 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
121 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
122
123 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
124 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
125 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
126 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
127
128 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
129 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
130
131 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
132 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
133
134 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
135 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
136 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
137 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
138
139 vst1q_s32(b, vacc0123); b += 4;
140 vst1q_s32(b, vacc4567); b += 4;
141 vst1q_s32(b, vacc89AB); b += 4;
142 vst1q_s32(b, vaccCDEF); b += 4;
143 }
144
145 if XNN_UNLIKELY(c != 0) {
146 do {
147 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
148 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
149
150 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
151 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
152
153 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
154
155 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
156 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
157
158 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
159 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
160
161 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
162
163 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
164 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
165
166 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
167 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
168
169 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
170
171 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
172 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
173
174 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
175 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
176
177 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
178
179 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
180 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
181
182 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
183 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
184
185 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
186
187 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
188 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
189
190 vst1q_s32(b, vacc0123); b += 4;
191 vst1q_s32(b, vacc4567); b += 4;
192 c -= 8;
193 } while (c != 0);
194 }
195 }
196
197 // Middle pass to process 5 inputs in each iteration.
198 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
199 int32_t* b = buffer;
200
201 const int8_t* i0 = input[0];
202 assert(i0 != NULL);
203 if XNN_UNPREDICTABLE(i0 != zero) {
204 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
205 }
206 const int8_t* i1 = input[1];
207 assert(i1 != NULL);
208 if XNN_UNPREDICTABLE(i1 != zero) {
209 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
210 }
211 const int8_t* i2 = input[2];
212 assert(i2 != NULL);
213 if XNN_UNPREDICTABLE(i2 != zero) {
214 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
215 }
216 const int8_t* i3 = input[3];
217 assert(i3 != NULL);
218 if XNN_UNPREDICTABLE(i3 != zero) {
219 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
220 }
221 const int8_t* i4 = input[4];
222 assert(i4 != NULL);
223 if XNN_UNPREDICTABLE(i4 != zero) {
224 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
225 }
226 input += 5;
227
228 size_t c = round_up_po2(channels, 8);
229
230 for (; c >= 16; c -= 16) {
231 int32x4_t vacc0123 = vld1q_s32(b);
232 int32x4_t vacc4567 = vld1q_s32(b + 4);
233 int32x4_t vacc89AB = vld1q_s32(b + 8);
234 int32x4_t vaccCDEF = vld1q_s32(b + 12);
235
236 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
237 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
238
239 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
240 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
241
242 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
243 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
244 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
245 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
246
247 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
248 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
249
250 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
251 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
252
253 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
254 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
255 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
256 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
257
258 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
259 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
260
261 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
262 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
263
264 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
265 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
266 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
267 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
268
269 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
270 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
271
272 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
273 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
274
275 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
276 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
277 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
278 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
279
280 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
281 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
282
283 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
284 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
285
286 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
287 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
288 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
289 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
290
291 vst1q_s32(b, vacc0123); b += 4;
292 vst1q_s32(b, vacc4567); b += 4;
293 vst1q_s32(b, vacc89AB); b += 4;
294 vst1q_s32(b, vaccCDEF); b += 4;
295 }
296
297 if XNN_UNLIKELY(c != 0) {
298 do {
299 int32x4_t vacc0123 = vld1q_s32(b);
300 int32x4_t vacc4567 = vld1q_s32(b + 4);
301
302 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
303 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
304
305 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
306
307 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
308 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
309
310 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
311 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
312
313 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
314
315 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
316 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
317
318 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
319 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
320
321 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
322
323 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
324 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
325
326 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
327 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
328
329 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
330
331 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
332 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
333
334 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
335 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
336
337 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
338
339 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
340 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
341
342 vst1q_s32(b, vacc0123); b += 4;
343 vst1q_s32(b, vacc4567); b += 4;
344 c -= 8;
345 } while (c != 0);
346 }
347 }
348
349 // Last pass to process up to 5 inputs.
350 {
351 const int32_t* b = buffer;
352 const int8_t* i0 = input[0];
353 assert(i0 != NULL);
354 if XNN_UNPREDICTABLE(i0 != zero) {
355 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
356 }
357 const int8_t* i1 = input[1];
358 assert(i1 != NULL);
359 if XNN_UNPREDICTABLE(i1 != zero) {
360 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
361 }
362 const int8_t* i2 = input[2];
363 assert(i2 != NULL);
364 if XNN_UNPREDICTABLE(i2 != zero) {
365 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
366 }
367 const int8_t* i3 = input[3];
368 assert(i3 != NULL);
369 if XNN_UNPREDICTABLE(i3 != zero) {
370 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
371 }
372 const int8_t* i4 = input[4];
373 assert(i4 != NULL);
374 if XNN_UNPREDICTABLE(i4 != zero) {
375 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
376 }
377
378 size_t c = channels;
379 for (; c >= 16; c -= 16) {
380 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
381 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
382 int32x4_t vacc89AB = vld1q_s32(b); b += 4;
383 int32x4_t vaccCDEF = vld1q_s32(b); b += 4;
384
385 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
386 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
387
388 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
389 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
390
391 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
392 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
393 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
394 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
395
396 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
397 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
398
399 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
400 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
401
402 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
403 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
404 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
405 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
406
407 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
408 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
409
410 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
411 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
412
413 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
414 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
415 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
416 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
417
418 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
419 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
420
421 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
422 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
423
424 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
425 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
426 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
427 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
428
429 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
430 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
431
432 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
433 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
434
435 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
436 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
437 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
438 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
439
440 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
441 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
442 vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
443 vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
444
445 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
446 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
447 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
448 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
449
450 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
451 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
452 vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
453 vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
454
455 #if XNN_ARCH_ARM64
456 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
457 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
458
459 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
460 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
461
462 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
463 #else // !XNN_ARCH_ARM64
464 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
465 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
466
467 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
468 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
469
470 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
471 #endif // !XNN_ARCH_ARM64
472
473 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
474
475 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
476
477 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
478 }
479
480
481 if XNN_UNLIKELY(c != 0) {
482 do {
483 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
484 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
485
486 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
487 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
488
489 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
490
491 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
492 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
493 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
494 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
495
496 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
497
498 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
499 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
500 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
501 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
502
503 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
504
505 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
506 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
507 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
508 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
509
510 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
511
512 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
513 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
514 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
515 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
516
517 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
518
519 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
520 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
521
522 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
523 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
524
525 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
526 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
527
528 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
529 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
530
531 #if XNN_ARCH_ARM64
532 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
533 #else
534 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
535 #endif
536 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
537
538 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
539 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
540 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
541
542 if XNN_LIKELY(c >= 8) {
543 vst1_s8(output, vout01234567); output += 8;
544 c -= 8;
545 } else {
546 if (c & 4) {
547 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
548 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
549 }
550 if (c & 2) {
551 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
552 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
553 }
554 if (c & 1) {
555 vst1_lane_s8(output, vout01234567, 0); output += 1;
556 }
557 c = 0;
558 }
559
560 } while (c != 0);
561 }
562 }
563
564 input = (const int8_t**) ((uintptr_t) input + input_stride);
565 output = (int8_t*) ((uintptr_t) output + output_increment);
566 } while (--output_width != 0);
567 }