1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "lite/backends/arm/math/gemv_arm_int8.h"
#include <arm_neon.h>
#include "lite/backends/arm/math/saturate.h"
namespace paddle {
namespace lite {
namespace arm {
namespace math {
template <typename dtype>
inline void write_gemv_out(const int* in,
dtype* out,
const float* scale,
const float* bias,
int size,
bool is_relu);
template <>
inline void write_gemv_out(const int* in,
float* out,
const float* scale,
const float* bias,
int size,
bool is_relu) {
int i = 0;
float32x4_t vzero = vdupq_n_f32(0.f);
for (; i < size - 7; i += 8) {
float32x4_t vout0 = bias ? vld1q_f32(bias) : vdupq_n_f32(0.f);
float32x4_t vout1 = bias ? vld1q_f32(bias + 4) : vdupq_n_f32(0.f);
int32x4_t vin0 = vld1q_s32(in);
int32x4_t vin1 = vld1q_s32(in + 4);
float32x4_t vscale0 = vld1q_f32(scale);
float32x4_t vscale1 = vld1q_f32(scale + 4);
float32x4_t vinf0 = vcvtq_f32_s32(vin0);
float32x4_t vinf1 = vcvtq_f32_s32(vin1);
vout0 = vmlaq_f32(vout0, vinf0, vscale0);
vout1 = vmlaq_f32(vout1, vinf1, vscale1);
if (is_relu) {
vout0 = vmaxq_f32(vout0, vzero);
vout1 = vmaxq_f32(vout1, vzero);
}
vst1q_f32(out, vout0);
vst1q_f32(out + 4, vout1);
bias += 8;
in += 8;
out += 8;
scale += 8;
}
for (; i < size; ++i) {
out[0] = *(in++) * *(scale)++;
out[0] += bias ? *(bias++) : 0.f;
out[0] = is_relu ? (out[0] > 0.f ? out[0] : 0.f) : out[0];
out++;
}
}
template <>
inline void write_gemv_out(const int* in,
signed char* out,
const float* scale,
const float* bias,
int size,
bool flag_relu) {
if (bias) {
for (int i = 0; i < size; ++i) {
out[0] =
saturate_cast<signed char>(roundf(*(in++) * *(scale++) + *(bias++)));
out[0] = out[0] < -127 ? -127 : out[0]; // -127 - 127
if (flag_relu) {
out[0] = out[0] > 0 ? out[0] : 0;
}
out++;
}
} else {
for (int i = 0; i < size; ++i) {
out[0] = saturate_cast<signed char>(roundf(*(in++) * *(scale++)));
out[0] = out[0] < -127 ? -127 : out[0]; // -127 - 127
if (flag_relu) {
out[0] = out[0] > 0 ? out[0] : 0;
}
out++;
}
}
}
template <typename dtype>
bool gemv_int8_oth(const int8_t* A,
const int8_t* x,
dtype* y,
bool transA,
int M,
int N,
const float* scale,
bool is_bias,
const float* bias,
bool is_relu) {
if (transA) {
LOG(ERROR) << "ERROR: sgemv, transA is not supported now";
return false;
}
dtype* data_out = y;
const int8_t* data_in = x;
const int8_t* weights_ptr = A;
int cnt = N >> 4;
int tail = N & 15;
#ifdef __aarch64__
int out_cnt = M >> 3;
#pragma omp parallel for
for (int j = 0; j < out_cnt; j++) {
int out_idx = j * 8;
dtype* out_ptr = data_out + out_idx;
const float* scale_ptr = scale + out_idx;
int ptr_out[8] = {0, 0, 0, 0, 0, 0, 0, 0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * out_idx);
const int8_t* ptr_w1 = ptr_w0 + N;
const int8_t* ptr_w2 = ptr_w1 + N;
const int8_t* ptr_w3 = ptr_w2 + N;
const int8_t* ptr_w4 = ptr_w3 + N;
const int8_t* ptr_w5 = ptr_w4 + N;
const int8_t* ptr_w6 = ptr_w5 + N;
const int8_t* ptr_w7 = ptr_w6 + N;
auto bias_ptr = is_bias ? bias + out_idx : nullptr;
int cnt_loop = cnt;
asm volatile(
"prfm pldl1keep, [%[in]] \n" /* preload din */
"prfm pldl1keep, [%[w0]] \n" /* preload w0 */
"prfm pldl1keep, [%[w1]] \n" /* preload w1 */
"prfm pldl1keep, [%[w2]] \n" /* preload w2 */
"prfm pldl1keep, [%[w3]] \n" /* preload w3 */
"prfm pldl1keep, [%[w4]] \n" /* preload w4 */
"prfm pldl1keep, [%[w5]] \n" /* preload w5 */
"prfm pldl1keep, [%[w6]] \n" /* preload w6 */
"prfm pldl1keep, [%[w7]] \n" /* preload w7 */
"movi v0.4s, #0 \n" /* set out0 to 0 */
"movi v1.4s, #0 \n" /* set out1 to 0 */
"movi v2.4s, #0 \n" /* set out2 to 0 */
"movi v3.4s, #0 \n" /* set out3 to 0 */
"movi v4.4s, #0 \n" /* set out4 to 0 */
"movi v5.4s, #0 \n" /* set out5 to 0 */
"movi v6.4s, #0 \n" /* set out6 to 0 */
"movi v7.4s, #0 \n" /* set out7 to 0 */
/* check main loop */
"cmp %w[cnt], #1 \n" /* check whether has main loop */
"blt 2f \n" /* jump to tail */
/* main loop */
"1: \n" /* main loop */
"ldr q8, [%[in]], #16 \n" /* load input, 16 int8 */
"ldr q9, [%[w0]], #16 \n" /* load w0, 16 int8 */
"ldr q10, [%[w1]], #16 \n" /* load w1, 16 int8 */
"ldr q11, [%[w2]], #16 \n" /* load w2, 16 int8 */
"ldr q12, [%[w3]], #16 \n" /* load w3, 16 int8 */
"ldr q13, [%[w4]], #16 \n" /* load w4, 16 int8 */
"ldr q14, [%[w5]], #16 \n" /* load w5, 16 int8 */
"ldr q15, [%[w6]], #16 \n" /* load w6, 16 int8 */
"ldr q16, [%[w7]], #16 \n" /* load w7, 16 int8 */
/* mul, lower 8 int8 * int8 = int16 */
"smull v18.8h, v8.8b, v9.8b \n" /* mul in * w0, low, 8 int8 */
"smull v19.8h, v8.8b, v10.8b\n" /* mul in * w1, low, 8 int8 */
"smull v20.8h, v8.8b, v11.8b\n" /* mul in * w2, low, 8 int8 */
"smull v21.8h, v8.8b, v12.8b\n" /* mul in * w3, low, 8 int8 */
"smull v22.8h, v8.8b, v13.8b\n" /* mul in * w4, low, 8 int8 */
"smull v23.8h, v8.8b, v14.8b\n" /* mul in * w5, low, 8 int8 */
"smull v24.8h, v8.8b, v15.8b\n" /* mul in * w6, low, 8 int8 */
"smull v25.8h, v8.8b, v16.8b\n" /* mul in * w7, low, 8 int8 */
/* mul, higher 8 int8 * int8 + int16 = int16 */
"smlal2 v18.8h,v8.16b,v9.16b \n" /* mul in * w0, high, 8 int8 */
"smlal2 v19.8h,v8.16b,v10.16b\n" /* mul in * w1, high, 8 int8 */
"smlal2 v20.8h,v8.16b,v11.16b\n" /* mul in * w2, high, 8 int8 */
"smlal2 v21.8h,v8.16b,v12.16b\n" /* mul in * w2, high, 8 int8 */
"smlal2 v22.8h,v8.16b,v13.16b\n" /* mul in * w2, high, 8 int8 */
"smlal2 v23.8h,v8.16b,v14.16b\n" /* mul in * w2, high, 8 int8 */
"smlal2 v24.8h,v8.16b,v15.16b\n" /* mul in * w2, high, 8 int8 */
"smlal2 v25.8h,v8.16b,v16.16b\n" /* mul in * w2, high, 8 int8 */
"subs %w[cnt], %w[cnt], #1 \n" /* sub main loop count */
/* add int16 to int32 */
"sadalp v0.4s, v18.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v1.4s, v19.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v2.4s, v20.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v3.4s, v21.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v4.4s, v22.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v5.4s, v23.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v6.4s, v24.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"sadalp v7.4s, v25.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"bne 1b \n" /* jump to main loop */
/* pair add to final result */
"2: \n" /* reduce to scale */
"addp v8.4s , v0.4s , v1.4s \n" /* pair add to 4 int32*/
"addp v9.4s , v2.4s , v3.4s \n" /* pair add to 4 int32*/
"addp v10.4s, v4.4s , v5.4s \n" /* pair add to 4 int32*/
"addp v11.4s, v6.4s , v7.4s \n" /* pair add to 4 int32*/
"addp v12.4s, v8.4s , v9.4s \n" /* pair add to 4 int32*/
"addp v13.4s, v10.4s, v11.4s \n" /* pair add to 4 int32*/
/* write to output */
"stp q12, q13, [%[out]] \n" /* save result */
: [in] "+r"(ptr_in),
[w0] "+r"(ptr_w0),
[w1] "+r"(ptr_w1),
[w2] "+r"(ptr_w2),
[w3] "+r"(ptr_w3),
[w4] "+r"(ptr_w4),
[w5] "+r"(ptr_w5),
[w6] "+r"(ptr_w6),
[w7] "+r"(ptr_w7),
[cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25");
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
ptr_out[1] += ptr_in[i] * ptr_w1[i];
ptr_out[2] += ptr_in[i] * ptr_w2[i];
ptr_out[3] += ptr_in[i] * ptr_w3[i];
ptr_out[4] += ptr_in[i] * ptr_w4[i];
ptr_out[5] += ptr_in[i] * ptr_w5[i];
ptr_out[6] += ptr_in[i] * ptr_w6[i];
ptr_out[7] += ptr_in[i] * ptr_w7[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 8, is_relu);
}
//! deal with remains
#pragma omp parallel for
for (int j = out_cnt * 8; j < M; j++) {
// int *ptr_out = data_out + j;
dtype* out_ptr = data_out + j;
const float* scale_ptr = scale + j;
int ptr_out[1] = {0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * j);
int cnt_loop = cnt;
auto bias_ptr = is_bias ? bias + j : nullptr;
asm volatile(
"prfm pldl1keep, [%[in]] \n" /* preload din */
"prfm pldl1keep, [%[w0]] \n" /* preload w0 */
"movi v0.4s, #0 \n" /* set out0 to 0 */
/* check main loop */
"cmp %w[cnt], #1 \n" /* check whether has main loop */
"blt 2f \n" /* jump to tail */
/* main loop */
"1: \n" /* main loop */
"ldr q8, [%[in]], #16 \n" /* load input, 16 int8 */
"ldr q9, [%[w0]], #16 \n" /* load w0, 16 int8 */
/* mul, lower 8 int8 * int8 = int16 */
"smull v18.8h, v8.8b, v9.8b \n" /* mul in * w0, low, 8 int8 */
"subs %w[cnt], %w[cnt], #1 \n" /* sub main loop count */
/* mul, higher 8 int8 * int8 + int16 = int16 */
"smlal2 v18.8h,v8.16b,v9.16b \n" /* mul in * w0, high, 8 int8 */
/* add int16 to int32 */
"sadalp v0.4s, v18.8h \n" /* pair acc, 8 int16 -> 4 int32 */
"bne 1b \n" /* jump to main loop */
/* pair add to final result */
"2: \n" /* reduce to scale */
"addv s8, v0.4s \n" /* reduction to out0 */
/* write to output */
"str s8, [%[out]] \n" /* save result */
: [in] "+r"(ptr_in), [w0] "+r"(ptr_w0), [cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc", "memory", "v0", "v8", "v9", "v18");
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 1, is_relu);
}
#else // __aarch64__
int out_cnt = M >> 2;
#pragma omp parallel for
for (int j = 0; j < out_cnt; j++) {
int out_idx = j * 4;
dtype* out_ptr = data_out + out_idx;
const float* scale_ptr = scale + out_idx;
int ptr_out[4] = {0, 0, 0, 0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * out_idx);
const int8_t* ptr_w1 = ptr_w0 + N;
const int8_t* ptr_w2 = ptr_w1 + N;
const int8_t* ptr_w3 = ptr_w2 + N;
int cnt_loop = cnt;
auto bias_ptr = is_bias ? bias + out_idx : nullptr;
asm volatile(
"pld [%[in]] @ preload cache line, input\n"
"pld [%[w0]] @ preload cache line, weights r0\n"
"pld [%[w1]] @ preload cache line, weights r1\n"
"pld [%[w2]] @ preload cache line, weights r2\n"
"pld [%[w3]] @ preload cache line, weights r3\n"
"vmov.u32 q0, #0 @ set q0 to 0\n"
"vmov.u32 q1, #0 @ set q1 to 0\n"
"vmov.u32 q2, #0 @ set q2 to 0\n"
"vmov.u32 q3, #0 @ set q3 to 0\n"
// "vld1.32 {d20-d21}, %[bias] @ load bias data"
"cmp %[cnt], #1 @ check whether has main loop\n"
"blt 2f @ jump to pair add\n"
/* main loop */
"1: @ main loop\n"
"vld1.8 {d8-d9}, [%[in]]! @ load input, q4\n"
"vld1.8 {d12-d13}, [%[w0]]! @ load weights r0, q6\n"
"vld1.8 {d14-d15}, [%[w1]]! @ load weights r1, q7\n"
"vld1.8 {d16-d17}, [%[w2]]! @ load weights r2, q8\n"
"vld1.8 {d18-d19}, [%[w3]]! @ load weights r3, q9\n"
/* mul, int8 * int8 = int16 */
"vmull.s8 q12, d8, d12 @ mul add\n"
"vmull.s8 q13, d8, d14 @ mul add\n"
"vmull.s8 q14, d8, d16 @ mul add\n"
"vmull.s8 q15, d8, d18 @ mul add\n"
/* mla, int8 * int8 + int16 = int16 */
"vmlal.s8 q12, d9, d13 @ mul add\n"
"vmlal.s8 q13, d9, d15 @ mul add\n"
"vmlal.s8 q14, d9, d17 @ mul add\n"
"vmlal.s8 q15, d9, d19 @ mul add\n"
/* pacc, int16 + int32 = int32 */
"vpadal.s16 q0, q12 @ pair acc\n"
"vpadal.s16 q1, q13 @ pair acc\n"
"vpadal.s16 q2, q14 @ pair acc\n"
"vpadal.s16 q3, q15 @ pair acc\n"
"subs %[cnt], #1 @ sub loop count \n"
/* check loop end */
"bne 1b @ jump to main loop\n"
/* pair add to final result */
"2: @ pair add \n"
"vpadd.s32 d8, d0, d1 @ pair add, first step\n"
"vpadd.s32 d9, d2, d3 @ pair add, first step\n"
"vpadd.s32 d10, d4, d5 @ pair add, first step\n"
"vpadd.s32 d11, d6, d7 @ pair add, first step\n"
"vpadd.s32 d0, d8, d9 @ pair add, second step\n"
"vpadd.s32 d1, d10, d11 @ pair add, second step\n"
/* write output */
"vst1.32 {d0-d1}, [%[out]] @ save result\n"
: [in] "+r"(ptr_in),
[w0] "+r"(ptr_w0),
[w1] "+r"(ptr_w1),
[w2] "+r"(ptr_w2),
[w3] "+r"(ptr_w3),
[cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc",
"memory",
"q0",
"q1",
"q2",
"q3",
"q4",
"q5",
"q6",
"q7",
"q8",
"q9",
"q12",
"q13",
"q14",
"q15");
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
ptr_out[1] += ptr_in[i] * ptr_w1[i];
ptr_out[2] += ptr_in[i] * ptr_w2[i];
ptr_out[3] += ptr_in[i] * ptr_w3[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 4, is_relu);
}
//! deal with remains
#pragma omp parallel for
for (int j = out_cnt * 4; j < M; j++) {
dtype* out_ptr = data_out + j;
const float* scale_ptr = scale + j;
int ptr_out[1] = {0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * j);
int cnt_loop = cnt;
auto bias_ptr = is_bias ? bias + j : nullptr;
asm volatile(
"pld [%[in]] @ preload cache line, input\n"
"pld [%[w0]] @ preload cache line, weights r0\n"
"vmov.u32 q0, #0 @ set q0 to 0\n"
"cmp %[cnt], #1 @ check whether has main loop\n"
"blt 2f @ jump to tail\n"
/* main loop */
"1: @ main loop\n"
"vld1.8 {d24-d25}, [%[in]]! @ load input, q12\n"
"vld1.8 {d28-d29}, [%[w0]]! @ load weights q14\n"
/* mull int8 * int8 = int16*/
"vmull.s8 q1, d24, d28 @ mul add\n"
"vmlal.s8 q1, d25, d29 @ mul add\n"
"subs %[cnt] , #1 @ sub loop count \n"
/* pacc int16 + int32 = int32*/
"vpadal.s16 q0, q1 @ pair acc\n"
"bne 1b @ jump to main loop\n"
/* pair add to final result */
"2: @ end processing\n"
"vpadd.s32 d2, d0, d1 @ pair add, first step\n"
"vpadd.s32 d0, d2, d2 @ pair add, final step\n"
/* write output */
"vst1.32 {d0[0]}, [%[out]] @ save result\n"
: [in] "+r"(ptr_in), [w0] "+r"(ptr_w0), [cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc", "memory", "q0", "q1", "q12", "q13");
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 1, is_relu);
}
#endif // __aarch64__
return true;
}
#if defined(__aarch64__) && defined(WITH_ARM_DOTPROD)
template <typename dtype>
bool gemv_int8_sdot(const int8_t* A,
const int8_t* x,
dtype* y,
bool transA,
int M,
int N,
const float* scale,
bool is_bias,
const float* bias,
bool is_relu) {
if (transA) {
LOG(ERROR) << "ERROR: sgemv, transA is not supported now";
return false;
}
dtype* data_out = y;
const int8_t* data_in = x;
const int8_t* weights_ptr = A;
int cnt = N >> 4;
int tail = N & 15;
int size_m = (M >> 3) << 3;
#pragma omp parallel for
for (int j = 0; j < M - 7; j += 8) {
dtype* out_ptr = data_out + j;
const float* scale_ptr = scale + j;
auto bias_ptr = is_bias ? bias + j : nullptr;
int ptr_out[8] = {0, 0, 0, 0, 0, 0, 0, 0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * j);
const int8_t* ptr_w1 = ptr_w0 + N;
const int8_t* ptr_w2 = ptr_w1 + N;
const int8_t* ptr_w3 = ptr_w2 + N;
const int8_t* ptr_w4 = ptr_w3 + N;
const int8_t* ptr_w5 = ptr_w4 + N;
const int8_t* ptr_w6 = ptr_w5 + N;
const int8_t* ptr_w7 = ptr_w6 + N;
int cnt_loop = cnt;
if (cnt > 0) {
asm volatile(
"prfm pldl1keep, [%[in]] \n" /* preload din */
"prfm pldl1keep, [%[w0]] \n" /* preload w0 */
"prfm pldl1keep, [%[w1]] \n" /* preload w1 */
"prfm pldl1keep, [%[w2]] \n" /* preload w2 */
"prfm pldl1keep, [%[w3]] \n" /* preload w3 */
"prfm pldl1keep, [%[w4]] \n" /* preload w4 */
"prfm pldl1keep, [%[w5]] \n" /* preload w5 */
"prfm pldl1keep, [%[w6]] \n" /* preload w6 */
"prfm pldl1keep, [%[w7]] \n" /* preload w7 */
"movi v0.4s, #0 \n" /* set out0 to 0 */
"movi v1.4s, #0 \n" /* set out1 to 0 */
"movi v2.4s, #0 \n" /* set out2 to 0 */
"movi v3.4s, #0 \n" /* set out3 to 0 */
"movi v4.4s, #0 \n" /* set out4 to 0 */
"movi v5.4s, #0 \n" /* set out5 to 0 */
"movi v6.4s, #0 \n" /* set out6 to 0 */
"movi v7.4s, #0 \n" /* set out7 to 0 */
/* main loop */
"1: \n" /* main loop */
"ldr q8, [%[in]], #16 \n" /* load input, 16 int8 */
"ldr q9, [%[w0]], #16 \n" /* load w0, 16 int8 */
"ldr q10, [%[w1]], #16 \n" /* load w0, 16 int8 */
"ldr q11, [%[w2]], #16 \n" /* load w0, 16 int8 */
"ldr q12, [%[w3]], #16 \n" /* load w0, 16 int8 */
"ldr q13, [%[w4]], #16 \n" /* load w0, 16 int8 */
"ldr q14, [%[w5]], #16 \n" /* load w0, 16 int8 */
"ldr q15, [%[w6]], #16 \n" /* load w0, 16 int8 */
"ldr q16, [%[w7]], #16 \n" /* load w0, 16 int8 */
".word 0x4e899500 // sdot v0.4s, v8.16b, v9.16b \n" /* out0, out1,
out2, out3
*/
".word 0x4e8a9501 // sdot v1.4s, v8.16b, v10.16b \n" /* out4, out5,
out6, out7
*/
".word 0x4e8b9502 // sdot v2.4s, v8.16b, v11.16b \n" /* out0, out1,
out2, out3
*/
".word 0x4e8c9503 // sdot v3.4s, v8.16b, v12.16b \n" /* out4, out5,
out6, out7
*/
"subs %w[cnt], %w[cnt], #1 \n"
".word 0x4e8d9504 // sdot v4.4s, v8.16b, v13.16b \n" /* out0, out1,
out2, out3
*/
".word 0x4e8e9505 // sdot v5.4s, v8.16b, v14.16b \n" /* out4, out5,
out6, out7
*/
".word 0x4e8f9506 // sdot v6.4s, v8.16b, v15.16b \n" /* out0, out1,
out2, out3
*/
".word 0x4e909507 // sdot v7.4s, v8.16b, v16.16b \n" /* out4, out5,
out6, out7
*/
"bne 1b \n" /* jump to main loop */
/* pair add to final result */
"2: \n" /* reduce to scale */
"addp v10.4s , v0.4s , v1.4s \n" /* pair add to 4 int32*/
"addp v11.4s , v2.4s , v3.4s \n" /* pair add to 4 int32*/
"addp v12.4s , v4.4s , v5.4s \n" /* pair add to 4 int32*/
"addp v13.4s , v6.4s , v7.4s \n" /* pair add to 4 int32*/
"addp v0.4s , v10.4s , v11.4s \n" /* pair add to 4 int32*/
"addp v1.4s , v12.4s , v13.4s \n" /* pair add to 4 int32*/
/* write to output */
"stp q0, q1, [%[out]] \n" /* save result */
: [in] "+r"(ptr_in),
[w0] "+r"(ptr_w0),
[w1] "+r"(ptr_w1),
[w2] "+r"(ptr_w2),
[w3] "+r"(ptr_w3),
[w4] "+r"(ptr_w4),
[w5] "+r"(ptr_w5),
[w6] "+r"(ptr_w6),
[w7] "+r"(ptr_w7),
[cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18");
}
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
ptr_out[1] += ptr_in[i] * ptr_w1[i];
ptr_out[2] += ptr_in[i] * ptr_w2[i];
ptr_out[3] += ptr_in[i] * ptr_w3[i];
ptr_out[4] += ptr_in[i] * ptr_w4[i];
ptr_out[5] += ptr_in[i] * ptr_w5[i];
ptr_out[6] += ptr_in[i] * ptr_w6[i];
ptr_out[7] += ptr_in[i] * ptr_w7[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 8, is_relu);
}
//! deal with remains
#pragma omp parallel for
for (int j = size_m; j < M; j++) {
// int *ptr_out = data_out + j;
dtype* out_ptr = data_out + j;
const float* scale_ptr = scale + j;
int ptr_out[1] = {0};
const int8_t* ptr_in = data_in;
const int8_t* ptr_w0 = weights_ptr + (N * j);
int cnt_loop = cnt;
auto bias_ptr = is_bias ? bias + j : nullptr;
asm volatile(
"prfm pldl1keep, [%[in]] \n" /* preload din */
"prfm pldl1keep, [%[w0]] \n" /* preload w0 */
"cmp %w[cnt], #1 \n" /* check whether has main loop */
"movi v0.4s, #0 \n" /* set out0 to 0 */
/* check main loop */
"blt 2f \n" /* jump to tail */
/* main loop */
"1: \n" /* main loop */
"ldr q8, [%[in]], #16 \n" /* load input, 16 int8 */
"ldr q9, [%[w0]], #16 \n" /* load w0, 16 int8 */
"subs %w[cnt], %w[cnt], #1 \n" /* sub main loop count */
/* mul, lower 8 int8 * int8 = int16 */
".word 0x4e899500 // sdot v0.4s, v8.16b, v9.16b \n"
"bne 1b \n" /* jump to main loop */
/* pair add to final result */
"2: \n" /* reduce to scale */
"addp v1.4s, v0.4s, v0.4s \n" /* reduction to out0 */
"addp v2.4s, v1.4s, v1.4s \n" /* reduction to out0 */
/* write to output */
"str s2, [%[out]] \n" /* save result */
: [in] "+r"(ptr_in), [w0] "+r"(ptr_w0), [cnt] "+r"(cnt_loop)
: [out] "r"(ptr_out)
: "cc", "memory", "v0", "v1", "v2", "v8", "v9", "v18");
for (int i = 0; i < tail; ++i) {
ptr_out[0] += ptr_in[i] * ptr_w0[i];
}
write_gemv_out(ptr_out, out_ptr, scale_ptr, bias_ptr, 1, is_relu);
}
return true;
}
#endif // __aarch64__ && sdot
template <>
bool gemv_int8<float>(const int8_t* A,
const int8_t* x,
float* y,
bool transA,
int M,
int N,
const float* scale,
bool is_bias,
const float* bias,
bool is_relu,
const ARMContext* ctx) {
#if defined(__aarch64__) && defined(WITH_ARM_DOTPROD)
if (ctx->has_dot()) {
return gemv_int8_sdot<float>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
} else {
return gemv_int8_oth<float>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
}
#else
return gemv_int8_oth<float>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
#endif
}
template <>
bool gemv_int8<int8_t>(const int8_t* A,
const int8_t* x,
int8_t* y,
bool transA,
int M,
int N,
const float* scale,
bool is_bias,
const float* bias,
bool is_relu,
const ARMContext* ctx) {
#if defined(__aarch64__) && defined(WITH_ARM_DOTPROD)
if (ctx->has_dot()) {
return gemv_int8_sdot<int8_t>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
} else {
return gemv_int8_oth<int8_t>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
}
#else
return gemv_int8_oth<int8_t>(
A, x, y, transA, M, N, scale, is_bias, bias, is_relu);
#endif
}
} // namespace math
} // namespace arm
} // namespace lite
} // namespace paddle
