Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-8f8m9l16c8s8r-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_8f8m9l16c8s8r__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 > 8);
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 8 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 const int8_t* i5 = input[5];
75 assert(i5 != NULL);
76 if XNN_UNPREDICTABLE(i5 != zero) {
77 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
78 }
79 const int8_t* i6 = input[6];
80 assert(i6 != NULL);
81 if XNN_UNPREDICTABLE(i6 != zero) {
82 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
83 }
84 const int8_t* i7 = input[7];
85 assert(i7 != NULL);
86 if XNN_UNPREDICTABLE(i7 != zero) {
87 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
88 }
89 input += 8;
90
91 size_t c = round_up_po2(channels, 8);
92
93 for (; c >= 16; c -= 16) {
94 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
95 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
96 int32x4_t vacc89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
97 int32x4_t vaccCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
98
99 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
100 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101
102 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
103 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
104
105 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
106 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
107 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
108 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
109
110 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
111 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
112
113 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
114 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
115
116 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
117 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
118 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
119 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
120
121 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
122 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
123
124 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
125 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
126
127 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
128 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
129 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
130 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
131
132 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
133 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
134
135 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
136 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
137
138 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
139 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
140 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
141 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
142
143 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
144 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
145
146 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
147 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
148
149 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
150 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
151 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
152 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
153
154 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
155 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
156
157 vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
158 vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
159
160 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
161 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
162 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
163 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
164
165 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
166 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
167
168 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
169 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
170
171 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
172 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
173 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
174 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
175
176 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
177 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
178
179 vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
180 vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
181
182 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
183 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
184 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
185 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
186
187 vst1q_s32(b, vacc0123); b += 4;
188 vst1q_s32(b, vacc4567); b += 4;
189 vst1q_s32(b, vacc89AB); b += 4;
190 vst1q_s32(b, vaccCDEF); b += 4;
191 }
192
193 if XNN_UNLIKELY(c != 0) {
194 do {
195 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
196 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
197
198 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
199 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
200
201 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
202
203 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
204 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
205
206 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
207 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
208
209 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
210
211 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
212 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
213
214 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
215 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
216
217 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
218
219 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
220 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
221
222 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
223 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
224
225 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
226
227 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
228 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
229
230 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
231 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
232
233 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
234
235 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
236 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
237
238 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
239 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
240
241 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
242
243 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
244 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
245
246 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
247 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
248
249 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
250
251 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
252 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
253
254 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
255 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
256
257 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
258
259 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
260 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
261
262 vst1q_s32(b, vacc0123); b += 4;
263 vst1q_s32(b, vacc4567); b += 4;
264 c -= 8;
265 } while (c != 0);
266 }
267 }
268
269 // Middle pass to process 8 inputs in each iteration.
270 for (size_t ks = kernel_size - 8; ks > 9; ks -= 8) {
271 int32_t* b = buffer;
272
273 const int8_t* i0 = input[0];
274 assert(i0 != NULL);
275 if XNN_UNPREDICTABLE(i0 != zero) {
276 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
277 }
278 const int8_t* i1 = input[1];
279 assert(i1 != NULL);
280 if XNN_UNPREDICTABLE(i1 != zero) {
281 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
282 }
283 const int8_t* i2 = input[2];
284 assert(i2 != NULL);
285 if XNN_UNPREDICTABLE(i2 != zero) {
286 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
287 }
288 const int8_t* i3 = input[3];
289 assert(i3 != NULL);
290 if XNN_UNPREDICTABLE(i3 != zero) {
291 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
292 }
293 const int8_t* i4 = input[4];
294 assert(i4 != NULL);
295 if XNN_UNPREDICTABLE(i4 != zero) {
296 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
297 }
298 const int8_t* i5 = input[5];
299 assert(i5 != NULL);
300 if XNN_UNPREDICTABLE(i5 != zero) {
301 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
302 }
303 const int8_t* i6 = input[6];
304 assert(i6 != NULL);
305 if XNN_UNPREDICTABLE(i6 != zero) {
306 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
307 }
308 const int8_t* i7 = input[7];
309 assert(i7 != NULL);
310 if XNN_UNPREDICTABLE(i7 != zero) {
311 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
312 }
313 input += 8;
314
315 size_t c = round_up_po2(channels, 8);
316
317 for (; c >= 16; c -= 16) {
318 int32x4_t vacc0123 = vld1q_s32(b);
319 int32x4_t vacc4567 = vld1q_s32(b + 4);
320 int32x4_t vacc89AB = vld1q_s32(b + 8);
321 int32x4_t vaccCDEF = vld1q_s32(b + 12);
322
323 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
324 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
325
326 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
327 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
328
329 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
330 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
331 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
332 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
333
334 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
335 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
336
337 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
338 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
339
340 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
341 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
342 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
343 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
344
345 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
346 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
347
348 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
349 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
350
351 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
352 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
353 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
354 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
355
356 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
357 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
358
359 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
360 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
361
362 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
363 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
364 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
365 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
366
367 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
368 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
369
370 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
371 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
372
373 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
374 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
375 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
376 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
377
378 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
379 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
380
381 vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
382 vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
383
384 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
385 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
386 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
387 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
388
389 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
390 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
391
392 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
393 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
394
395 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
396 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
397 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
398 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
399
400 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
401 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
402
403 vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
404 vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
405
406 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
407 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
408 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
409 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
410
411 vst1q_s32(b, vacc0123); b += 4;
412 vst1q_s32(b, vacc4567); b += 4;
413 vst1q_s32(b, vacc89AB); b += 4;
414 vst1q_s32(b, vaccCDEF); b += 4;
415 }
416
417 if XNN_UNLIKELY(c != 0) {
418 do {
419 int32x4_t vacc0123 = vld1q_s32(b);
420 int32x4_t vacc4567 = vld1q_s32(b + 4);
421
422 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
423 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
424
425 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
426
427 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
428 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
429
430 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
431 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
432
433 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
434
435 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
436 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
437
438 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
439 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
440
441 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
442
443 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
444 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
445
446 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
447 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
448
449 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
450
451 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
452 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
453
454 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
455 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
456
457 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
458
459 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
460 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
461
462 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
463 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
464
465 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
466
467 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
468 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
469
470 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
471 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
472
473 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
474
475 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
476 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
477
478 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
479 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
480
481 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
482
483 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
484 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
485
486 vst1q_s32(b, vacc0123); b += 4;
487 vst1q_s32(b, vacc4567); b += 4;
488 c -= 8;
489 } while (c != 0);
490 }
491 }
492
493 // Last pass to process up to 9 inputs.
494 {
495 const int32_t* b = buffer;
496 const int8_t* i0 = input[0];
497 assert(i0 != NULL);
498 if XNN_UNPREDICTABLE(i0 != zero) {
499 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
500 }
501 const int8_t* i1 = input[1];
502 assert(i1 != NULL);
503 if XNN_UNPREDICTABLE(i1 != zero) {
504 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
505 }
506 const int8_t* i2 = input[2];
507 assert(i2 != NULL);
508 if XNN_UNPREDICTABLE(i2 != zero) {
509 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
510 }
511 const int8_t* i3 = input[3];
512 assert(i3 != NULL);
513 if XNN_UNPREDICTABLE(i3 != zero) {
514 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
515 }
516 const int8_t* i4 = input[4];
517 assert(i4 != NULL);
518 if XNN_UNPREDICTABLE(i4 != zero) {
519 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
520 }
521 const int8_t* i5 = input[5];
522 assert(i5 != NULL);
523 if XNN_UNPREDICTABLE(i5 != zero) {
524 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
525 }
526 const int8_t* i6 = input[6];
527 assert(i6 != NULL);
528 if XNN_UNPREDICTABLE(i6 != zero) {
529 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
530 }
531 const int8_t* i7 = input[7];
532 assert(i7 != NULL);
533 if XNN_UNPREDICTABLE(i7 != zero) {
534 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
535 }
536 const int8_t* i8 = input[8];
537 assert(i8 != NULL);
538 if XNN_UNPREDICTABLE(i8 != zero) {
539 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
540 }
541
542 size_t c = channels;
543 for (; c >= 16; c -= 16) {
544 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
545 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
546 int32x4_t vacc89AB = vld1q_s32(b); b += 4;
547 int32x4_t vaccCDEF = vld1q_s32(b); b += 4;
548
549 const int8x16_t vi0x0123456789ABCDEF = vld1q_s8(i0); i0 += 16;
550 const int8x16_t vk0x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
551
552 int16x8_t vprod01234567 = vmull_s8(vget_low_s8(vi0x0123456789ABCDEF), vget_low_s8(vk0x0123456789ABCDEF));
553 int16x8_t vprod89ABCDEF = vmull_s8(vget_high_s8(vi0x0123456789ABCDEF), vget_high_s8(vk0x0123456789ABCDEF));
554
555 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
556 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
557 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
558 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
559
560 const int8x16_t vi1x0123456789ABCDEF = vld1q_s8(i1); i1 += 16;
561 const int8x16_t vk1x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
562
563 vprod01234567 = vmull_s8(vget_low_s8(vi1x0123456789ABCDEF), vget_low_s8(vk1x0123456789ABCDEF));
564 vprod89ABCDEF = vmull_s8(vget_high_s8(vi1x0123456789ABCDEF), vget_high_s8(vk1x0123456789ABCDEF));
565
566 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
567 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
568 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
569 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
570
571 const int8x16_t vi2x0123456789ABCDEF = vld1q_s8(i2); i2 += 16;
572 const int8x16_t vk2x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
573
574 vprod01234567 = vmull_s8(vget_low_s8(vi2x0123456789ABCDEF), vget_low_s8(vk2x0123456789ABCDEF));
575 vprod89ABCDEF = vmull_s8(vget_high_s8(vi2x0123456789ABCDEF), vget_high_s8(vk2x0123456789ABCDEF));
576
577 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
578 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
579 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
580 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
581
582 const int8x16_t vi3x0123456789ABCDEF = vld1q_s8(i3); i3 += 16;
583 const int8x16_t vk3x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
584
585 vprod01234567 = vmull_s8(vget_low_s8(vi3x0123456789ABCDEF), vget_low_s8(vk3x0123456789ABCDEF));
586 vprod89ABCDEF = vmull_s8(vget_high_s8(vi3x0123456789ABCDEF), vget_high_s8(vk3x0123456789ABCDEF));
587
588 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
589 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
590 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
591 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
592
593 const int8x16_t vi4x0123456789ABCDEF = vld1q_s8(i4); i4 += 16;
594 const int8x16_t vk4x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
595
596 vprod01234567 = vmull_s8(vget_low_s8(vi4x0123456789ABCDEF), vget_low_s8(vk4x0123456789ABCDEF));
597 vprod89ABCDEF = vmull_s8(vget_high_s8(vi4x0123456789ABCDEF), vget_high_s8(vk4x0123456789ABCDEF));
598
599 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
600 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
601 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
602 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
603
604 const int8x16_t vi5x0123456789ABCDEF = vld1q_s8(i5); i5 += 16;
605 const int8x16_t vk5x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
606
607 vprod01234567 = vmull_s8(vget_low_s8(vi5x0123456789ABCDEF), vget_low_s8(vk5x0123456789ABCDEF));
608 vprod89ABCDEF = vmull_s8(vget_high_s8(vi5x0123456789ABCDEF), vget_high_s8(vk5x0123456789ABCDEF));
609
610 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
611 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
612 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
613 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
614
615 const int8x16_t vi6x0123456789ABCDEF = vld1q_s8(i6); i6 += 16;
616 const int8x16_t vk6x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
617
618 vprod01234567 = vmull_s8(vget_low_s8(vi6x0123456789ABCDEF), vget_low_s8(vk6x0123456789ABCDEF));
619 vprod89ABCDEF = vmull_s8(vget_high_s8(vi6x0123456789ABCDEF), vget_high_s8(vk6x0123456789ABCDEF));
620
621 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
622 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
623 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
624 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
625
626 const int8x16_t vi7x0123456789ABCDEF = vld1q_s8(i7); i7 += 16;
627 const int8x16_t vk7x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
628
629 vprod01234567 = vmull_s8(vget_low_s8(vi7x0123456789ABCDEF), vget_low_s8(vk7x0123456789ABCDEF));
630 vprod89ABCDEF = vmull_s8(vget_high_s8(vi7x0123456789ABCDEF), vget_high_s8(vk7x0123456789ABCDEF));
631
632 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
633 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
634 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
635 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
636
637 const int8x16_t vi8x0123456789ABCDEF = vld1q_s8(i8); i8 += 16;
638 const int8x16_t vk8x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
639
640 vprod01234567 = vmull_s8(vget_low_s8(vi8x0123456789ABCDEF), vget_low_s8(vk8x0123456789ABCDEF));
641 vprod89ABCDEF = vmull_s8(vget_high_s8(vi8x0123456789ABCDEF), vget_high_s8(vk8x0123456789ABCDEF));
642
643 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
644 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
645 vacc89AB = vaddw_s16(vacc89AB, vget_low_s16(vprod89ABCDEF));
646 vaccCDEF = vaddw_s16(vaccCDEF, vget_high_s16(vprod89ABCDEF));
647
648 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
649 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
650 vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
651 vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
652
653 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
654 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
655 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
656 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
657
658 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
659 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
660 vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
661 vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
662
663 #if XNN_ARCH_ARM64
664 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
665 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
666
667 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
668 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
669
670 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
671 #else // !XNN_ARCH_ARM64
672 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
673 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
674
675 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
676 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
677
678 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
679 #endif // !XNN_ARCH_ARM64
680
681 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
682
683 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
684
685 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
686 }
687
688
689 if XNN_UNLIKELY(c != 0) {
690 do {
691 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
692 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
693
694 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8;
695 const int8x8_t vk0x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
696
697 int16x8_t vprod01234567 = vmull_s8(vi0x01234567, vk0x01234567);
698
699 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
700 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
701 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8;
702 const int8x8_t vk1x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
703
704 vprod01234567 = vmull_s8(vi1x01234567, vk1x01234567);
705
706 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
707 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
708 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8;
709 const int8x8_t vk2x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
710
711 vprod01234567 = vmull_s8(vi2x01234567, vk2x01234567);
712
713 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
714 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
715 const int8x8_t vi3x01234567 = vld1_s8(i3); i3 += 8;
716 const int8x8_t vk3x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
717
718 vprod01234567 = vmull_s8(vi3x01234567, vk3x01234567);
719
720 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
721 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
722 const int8x8_t vi4x01234567 = vld1_s8(i4); i4 += 8;
723 const int8x8_t vk4x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
724
725 vprod01234567 = vmull_s8(vi4x01234567, vk4x01234567);
726
727 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
728 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
729 const int8x8_t vi5x01234567 = vld1_s8(i5); i5 += 8;
730 const int8x8_t vk5x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
731
732 vprod01234567 = vmull_s8(vi5x01234567, vk5x01234567);
733
734 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
735 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
736 const int8x8_t vi6x01234567 = vld1_s8(i6); i6 += 8;
737 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
738
739 vprod01234567 = vmull_s8(vi6x01234567, vk6x01234567);
740
741 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
742 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
743 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8;
744 const int8x8_t vk7x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
745
746 vprod01234567 = vmull_s8(vi7x01234567, vk7x01234567);
747
748 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
749 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
750 const int8x8_t vi8x01234567 = vld1_s8(i8); i8 += 8;
751 const int8x8_t vk8x01234567 = vld1_s8(w); w = (const int8_t*) w + 8;
752
753 vprod01234567 = vmull_s8(vi8x01234567, vk8x01234567);
754
755 vacc0123 = vaddw_s16(vacc0123, vget_low_s16(vprod01234567));
756 vacc4567 = vaddw_s16(vacc4567, vget_high_s16(vprod01234567));
757
758 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
759 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
760
761 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
762 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
763
764 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
765 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
766
767 #if XNN_ARCH_ARM64
768 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
769 #else
770 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
771 #endif
772 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
773
774 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
775 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
776 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
777
778 if XNN_LIKELY(c >= 8) {
779 vst1_s8(output, vout01234567); output += 8;
780 c -= 8;
781 } else {
782 if (c & 4) {
783 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
784 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
785 }
786 if (c & 2) {
787 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
788 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
789 }
790 if (c & 1) {
791 vst1_lane_s8(output, vout01234567, 0); output += 1;
792 }
793 c = 0;
794 }
795
796 } while (c != 0);
797 }
798 }
799
800 input = (const int8_t**) ((uintptr_t) input + input_stride);
801 output = (int8_t*) ((uintptr_t) output + output_increment);
802 } while (--output_width != 0);
803 }