-
Notifications
You must be signed in to change notification settings - Fork 3
/
xnet_function.h
489 lines (447 loc) · 16.6 KB
/
xnet_function.h
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
#ifndef BIN_MATH_FUNCTION
#define BIN_MATH_FUNCTION
#include "xnet_common.h"
#include<cblas.h>
#include<math.h>
#include<stdint.h>
#include<x86intrin.h>
#include<immintrin.h>
#include<vector>
using std::vector;
#ifdef DEBUG_XNOR
#include<iostream>
using namespace std;
#endif
namespace xnet{
#define BIT_SET(var, pos, val) var |= (val<<pos)
template<typename Dtype>
Dtype xnet_cpu_asum(const int n , const Dtype* x);
template<>
float xnet_cpu_asum<float>(const int n, const float* x){
return cblas_sasum(n, x, 1);
}
template<>
double xnet_cpu_asum<double>(const int n, const double* x){
return cblas_dasum(n, x, 1);
}
template<typename Dtype>
inline int8_t xnet_sign(Dtype val){
return (Dtype(0) < val) - ( val < Dtype(0));
}
//
//@code from https://github.com/WojciechMula/sse-popcount
uint8_t lookup8bit[256] = {
/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
/* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
/* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
/* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4,
/* 10 */ 1, /* 11 */ 2, /* 12 */ 2, /* 13 */ 3,
/* 14 */ 2, /* 15 */ 3, /* 16 */ 3, /* 17 */ 4,
/* 18 */ 2, /* 19 */ 3, /* 1a */ 3, /* 1b */ 4,
/* 1c */ 3, /* 1d */ 4, /* 1e */ 4, /* 1f */ 5,
/* 20 */ 1, /* 21 */ 2, /* 22 */ 2, /* 23 */ 3,
/* 24 */ 2, /* 25 */ 3, /* 26 */ 3, /* 27 */ 4,
/* 28 */ 2, /* 29 */ 3, /* 2a */ 3, /* 2b */ 4,
/* 2c */ 3, /* 2d */ 4, /* 2e */ 4, /* 2f */ 5,
/* 30 */ 2, /* 31 */ 3, /* 32 */ 3, /* 33 */ 4,
/* 34 */ 3, /* 35 */ 4, /* 36 */ 4, /* 37 */ 5,
/* 38 */ 3, /* 39 */ 4, /* 3a */ 4, /* 3b */ 5,
/* 3c */ 4, /* 3d */ 5, /* 3e */ 5, /* 3f */ 6,
/* 40 */ 1, /* 41 */ 2, /* 42 */ 2, /* 43 */ 3,
/* 44 */ 2, /* 45 */ 3, /* 46 */ 3, /* 47 */ 4,
/* 48 */ 2, /* 49 */ 3, /* 4a */ 3, /* 4b */ 4,
/* 4c */ 3, /* 4d */ 4, /* 4e */ 4, /* 4f */ 5,
/* 50 */ 2, /* 51 */ 3, /* 52 */ 3, /* 53 */ 4,
/* 54 */ 3, /* 55 */ 4, /* 56 */ 4, /* 57 */ 5,
/* 58 */ 3, /* 59 */ 4, /* 5a */ 4, /* 5b */ 5,
/* 5c */ 4, /* 5d */ 5, /* 5e */ 5, /* 5f */ 6,
/* 60 */ 2, /* 61 */ 3, /* 62 */ 3, /* 63 */ 4,
/* 64 */ 3, /* 65 */ 4, /* 66 */ 4, /* 67 */ 5,
/* 68 */ 3, /* 69 */ 4, /* 6a */ 4, /* 6b */ 5,
/* 6c */ 4, /* 6d */ 5, /* 6e */ 5, /* 6f */ 6,
/* 70 */ 3, /* 71 */ 4, /* 72 */ 4, /* 73 */ 5,
/* 74 */ 4, /* 75 */ 5, /* 76 */ 5, /* 77 */ 6,
/* 78 */ 4, /* 79 */ 5, /* 7a */ 5, /* 7b */ 6,
/* 7c */ 5, /* 7d */ 6, /* 7e */ 6, /* 7f */ 7,
/* 80 */ 1, /* 81 */ 2, /* 82 */ 2, /* 83 */ 3,
/* 84 */ 2, /* 85 */ 3, /* 86 */ 3, /* 87 */ 4,
/* 88 */ 2, /* 89 */ 3, /* 8a */ 3, /* 8b */ 4,
/* 8c */ 3, /* 8d */ 4, /* 8e */ 4, /* 8f */ 5,
/* 90 */ 2, /* 91 */ 3, /* 92 */ 3, /* 93 */ 4,
/* 94 */ 3, /* 95 */ 4, /* 96 */ 4, /* 97 */ 5,
/* 98 */ 3, /* 99 */ 4, /* 9a */ 4, /* 9b */ 5,
/* 9c */ 4, /* 9d */ 5, /* 9e */ 5, /* 9f */ 6,
/* a0 */ 2, /* a1 */ 3, /* a2 */ 3, /* a3 */ 4,
/* a4 */ 3, /* a5 */ 4, /* a6 */ 4, /* a7 */ 5,
/* a8 */ 3, /* a9 */ 4, /* aa */ 4, /* ab */ 5,
/* ac */ 4, /* ad */ 5, /* ae */ 5, /* af */ 6,
/* b0 */ 3, /* b1 */ 4, /* b2 */ 4, /* b3 */ 5,
/* b4 */ 4, /* b5 */ 5, /* b6 */ 5, /* b7 */ 6,
/* b8 */ 4, /* b9 */ 5, /* ba */ 5, /* bb */ 6,
/* bc */ 5, /* bd */ 6, /* be */ 6, /* bf */ 7,
/* c0 */ 2, /* c1 */ 3, /* c2 */ 3, /* c3 */ 4,
/* c4 */ 3, /* c5 */ 4, /* c6 */ 4, /* c7 */ 5,
/* c8 */ 3, /* c9 */ 4, /* ca */ 4, /* cb */ 5,
/* cc */ 4, /* cd */ 5, /* ce */ 5, /* cf */ 6,
/* d0 */ 3, /* d1 */ 4, /* d2 */ 4, /* d3 */ 5,
/* d4 */ 4, /* d5 */ 5, /* d6 */ 5, /* d7 */ 6,
/* d8 */ 4, /* d9 */ 5, /* da */ 5, /* db */ 6,
/* dc */ 5, /* dd */ 6, /* de */ 6, /* df */ 7,
/* e0 */ 3, /* e1 */ 4, /* e2 */ 4, /* e3 */ 5,
/* e4 */ 4, /* e5 */ 5, /* e6 */ 5, /* e7 */ 6,
/* e8 */ 4, /* e9 */ 5, /* ea */ 5, /* eb */ 6,
/* ec */ 5, /* ed */ 6, /* ee */ 6, /* ef */ 7,
/* f0 */ 4, /* f1 */ 5, /* f2 */ 5, /* f3 */ 6,
/* f4 */ 5, /* f5 */ 6, /* f6 */ 6, /* f7 */ 7,
/* f8 */ 5, /* f9 */ 6, /* fa */ 6, /* fb */ 7,
/* fc */ 6, /* fd */ 7, /* fe */ 7, /* ff */ 8
};
uint64_t popcnt_lookup_8bit(const uint64_t* x){
uint64_t result = 0;
const uint8_t* data = reinterpret_cast<const uint8_t*>(x);
for(int i = 0; i < 8; i++)
result += lookup8bit[data[i]];
return result;
}
#ifdef HAVE_NOTHING
#define POPCNT(val) popcnt_lookup_8bit(val)
#endif
#ifdef HAVE_BUILTIN
#define POPCNT(val) __builtin_popcountll(val)
#endif
#ifdef HAVE_SSE_INSTRUCTIONS
#define POPCNT(val) _popcnt64(val)
#endif
#ifdef HAVE_NEON_INSTRUCTIONS
#define POPCNT(val)
#endif
/*
@brief Binarize convolution weights.
This function calculates the scale factor $\alpha$
and times sign(w) to estimate the real value weights.
@input:
- real_weights. BinBlob store both real value weights and binary value weights.
When accessing, we use filter_num and kernel_count to get the right offset.
- alpha. scale factor.
*/
template <typename Dtype>
void binarizeWeights(const Dtype* real_weights, BinBlob<Dtype>& binary_weights,
vector<Dtype>& alpha){
const int* shape = binary_weights.shape();
int filter_num = shape[0];
int kernel_count = shape[1] * shape[2] * shape[3];
const Dtype* r_data = real_weights;
//caculate alpha
alpha.resize(filter_num);
for(int filter_idx = 0; filter_idx < filter_num; filter_idx++)
alpha[filter_idx] = xnet_cpu_asum<Dtype>(kernel_count, r_data+filter_idx*
kernel_count) / kernel_count;
//sign
int filter_offset = ceil((float)kernel_count/BIN_SIZE);
binary_weights.allocateBC(filter_num * filter_offset, true);
BinaryCode* b_data = binary_weights.mutable_b_data();
//The stored style is NCHW
for (uint64_t filter_idx = 0; filter_idx < filter_num; filter_idx++){
int out_idx = filter_idx*filter_offset;
for (int weight_idx = 0 ; weight_idx < kernel_count; weight_idx++){
uint64_t sign = (r_data[filter_idx*kernel_count + weight_idx] > 0);
BIT_SET(b_data[out_idx+weight_idx/BIN_SIZE], (int)(weight_idx%BIN_SIZE),
sign);
}
}
}
template <typename Dtype>
void binarizeWeights_omp(const Dtype* real_weights, BinBlob<Dtype>& binary_weights,
vector<Dtype>& alpha){
const int* shape = binary_weights.shape();
int filter_num = shape[0];
int kernel_count = shape[1] * shape[2] * shape[3];
const Dtype* r_data = real_weights;
//caculate alpha
alpha.resize(filter_num);
#pragma omp parallel for
for(int filter_idx = 0; filter_idx < filter_num; filter_idx++)
alpha[filter_idx] = xnet_cpu_asum<Dtype>(kernel_count, r_data+filter_idx*
kernel_count) / kernel_count;
//sign
int filter_offset = ceil((float)kernel_count/BIN_SIZE);
binary_weights.allocateBC(filter_num * filter_offset, true);
BinaryCode* b_data = binary_weights.mutable_b_data();
//The stored style is NCHW
#pragma omp parallel for
//binary matrix
for (int filter_idx = 0; filter_idx < filter_num; filter_idx++){
int out_idx = filter_idx*filter_offset;
for (int weight_idx = 0 ; weight_idx < kernel_count; weight_idx++){
uint64_t sign = (r_data[filter_idx*kernel_count + weight_idx] > 0);
BIT_SET(b_data[out_idx+weight_idx/BIN_SIZE], weight_idx%BIN_SIZE, sign);
}
}
}
/*
@brief code from caffe.The casting allows to use one condition instead of two.
*/
inline bool is_a_ge_zero_and_a_lt_b(int a, int b){
return static_cast<unsigned>(a) < static_cast<unsigned>(b);
}
/*
@brief Combine the binarization and im2col. It's easier way than separating
binarization and im2col.
The binary col data is stored in input BinBlob as colum major array(
regared as the transpose of row major array).
finally we caculate the convolution results using specific bit operation
gemm.
@param
*/
template <typename Dtype>
void binarizeIm2Col(const Dtype* input_data, BinBlob<Dtype>& col_buf,
const int channels, const int height, const int width, const int kernel_h,
const int kernel_w, const int pad_h,const int pad_w, const int stride_h,
const int stride_w, const int dilation_h,const int dilation_w)
{
const int output_h = (height + 2*pad_h - (dilation_h * (kernel_h - 1)+1))/stride_h + 1;
const int output_w = (width + 2*pad_w - (dilation_w*(kernel_h - 1)+1))/stride_w + 1;
const int channel_size = height * width;
const int kernel_size = kernel_h * kernel_w;
const uint64_t output_spatial_size = output_h*output_w;
const uint64_t output_cnt = output_spatial_size*
ceil((float)channels*kernel_size/BIN_SIZE);
col_buf.Reshape(1, 1, channels*kernel_size, output_spatial_size);
col_buf.allocateBC(output_cnt, true);
BinaryCode* col_data = col_buf.mutable_b_data();
uint64_t position = 0;
for(int channel = -1;++channel< channels;input_data += channel_size){
for(int kernel_row = 0; kernel_row < kernel_h; kernel_row++){
for(int kernel_col = 0; kernel_col < kernel_w; kernel_col++){
int input_row = -pad_h + kernel_row * dilation_h;
uint64_t bv_idx = position / BIN_SIZE;
uint64_t bv_offset = position % BIN_SIZE;
position++;
uint64_t output_offset = 0;
for(int output_row = 0; output_row < output_h; output_row++){
if(!is_a_ge_zero_and_a_lt_b(input_row, height)){
output_offset += output_w;
}else{
int input_col = -pad_w + kernel_col*dilation_w;
for(int output_col = 0;output_col< output_w; output_col++){
if (is_a_ge_zero_and_a_lt_b(input_col, width)){
uint64_t sign = (input_data[input_row*width + input_col]>0);
BIT_SET(col_data[bv_idx*output_spatial_size+output_offset], bv_offset,
sign);
}else{
}
output_offset++;
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
template <typename Dtype>
void binarizeIm2Col_omp(const Dtype* input_data, BinBlob<Dtype>& col_buf,
const int channels, const int height,const int width, const int kernel_h,
const int kernel_w, const int pad_h, const int pad_w, const int stride_h,
const int stride_w, const int dilation_h, const int dilation_w)
{
const int output_h = (height + 2*pad_h - (dilation_h * (kernel_h - 1)+1))/stride_h + 1;
const int output_w = (width + 2*pad_w - (dilation_w*(kernel_h - 1)+1))/stride_w + 1;
const int channel_size = height * width;
const int kernel_size = kernel_h * kernel_w;
const int output_spatial_size = output_h*output_w;
//initialize bin_data:
const uint64_t output_cnt = output_h*output_w*channels*kernel_size;
col_buf.Reshape(1, 1, channels*kernel_size, output_spatial_size);
col_buf.allocateBC(output_cnt, true);
BinaryCode* col_data = col_buf.mutable_b_data();
//caffe im2col
for(int channel = -1;++channel < channels;input_data += channel_size){
#pragma omp parallel for
for(int kernel_row = 0; kernel_row < kernel_h; kernel_row++){
#pragma omp parallel for
for(int kernel_col = 0; kernel_col < kernel_w; kernel_col++){
int input_row = -pad_h + kernel_row * dilation_h;
uint64_t rv_idx = channel*kernel_size + kernel_row*kernel_w
+ kernel_col;
uint64_t bv_idx = rv_idx / BIN_SIZE;
uint64_t bv_offset = rv_idx % BIN_SIZE;
uint64_t output_offset = 0;
for(int output_row = 0; output_row < output_h; output_row++){
if(!is_a_ge_zero_and_a_lt_b(input_row, height)){
output_offset += output_w;
}else{
int input_col = -pad_w + kernel_col*dilation_w;
for(int output_col = 0;output_col< output_w; output_col++){
//in col data:
// 1) row:
// channel*kernel_size + kernel_row*kernel_w + kernel_col
// 2) col:
// output_row*output_w+output_col
if (is_a_ge_zero_and_a_lt_b(input_col, width)){
uint64_t sign = (input_data[input_row*width + input_col]>0);
BIT_SET(col_data[bv_idx*output_spatial_size+output_offset],
bv_offset, sign);
}else{
}
output_offset++;
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
/*
* @brief: GEMM using xnor operation.
* In convolution, M means the number of filters and K means (ck^2)/BIN_SIZE.
* N means the output_spatial_size. A is weights pointer. B is the input
* column pointer. C is the output pointer.
*
* @TODO inner_product
* @param:
* int M, K, N;
* int lda = K , ldb = N, ldc = N ;
* int size, the actual size of a vector;
*
*/
template <typename Dtype>
void xorGEMM_baseline(int M,int K,int N,
const BinaryCode* A, int lda ,
const BinaryCode* B, int ldb,
Dtype* C, int ldc,
int size, vector<Dtype>& alphas){
int m,k,n;
for (m = 0; m < M; m++){
for(k = 0; k < K; k++){
register BinaryCode temp_A= A[m*lda+k];
for(n = 0; n < N; n++){
C[m*ldc+n] += POPCNT(temp_A^B[k*ldb+n]);
}
}
}
for(m = 0; m < M; m++)
for(n = 0; n < N; n++){
C[m*ldc+n] = size - 2*C[m*ldc+n];
C[m*ldc+n] *= alphas[m];
}
}
template <typename Dtype>
void xorGEMM_omp_baseline(int M,int K,int N,
const BinaryCode* A, int lda,
const BinaryCode* B, int ldb,
Dtype* C, int ldc,
int size, vector<Dtype>& alphas){
int m,k,n;
#pragma omp parallel for
for (m = 0; m < M; m++){
#pragma omp parallel for
for(k = 0; k <K; k++){
BinaryCode temp_A= A[m*lda+k];
#pragma omp parallel for
for(n = 0; n < N; n++){
//C[m*ldc+n] += __builtin_popcountll(temp_A^B[k*ldb+n]);
C[m*ldc+n] += POPCNT(temp_A^B[k*ldb+n]);
}
}
}
#pragma omp parallel for
for(m = 0; m < M; m++)
for(n = 0; n < N; n++){
C[m*ldc+n] = size - 2*C[m*ldc+n];
C[m*ldc+n] *= alphas[m];
}
}
/*Code From https://github.com/hpi-xnor/BMXNet
* unroll the for loop to balance the computation load and the threads
* scheduling load.
*/
#define UNROLLN 6
template <typename Dtype>
void xorGEMM_omp_unrolled(int M,int K,int N,
const BinaryCode* A, int lda,
const BinaryCode* B, int ldb,
Dtype* C, int ldc,
int size, vector<Dtype>& alphas){
int m,k,n;
#pragma omp parallel for
for (m = 0; m < M; m++){
#pragma omp parallel for
for(k = 0; k <(K/UNROLLN)*UNROLLN; k++){
BinaryCode A_PART[UNROLLN];
A_PART[0] = A[m*lda + k];
A_PART[1] = A[m*lda + k + 1];
A_PART[2] = A[m*lda + k + 2];
A_PART[3] = A[m*lda + k + 3];
A_PART[4] = A[m*lda + k + 4];
A_PART[5] = A[m*lda + k + 5];
#pragma omp parallel for
for(n = 0; n < N; n++){
int popc[UNROLLN];
popc[0] = POPCNT(A_PART[0]^B[k*ldb+n]);
popc[1] = POPCNT(A_PART[1]^B[(k+1)*ldb+n]);
popc[2] = POPCNT(A_PART[2]^B[(k+2)*ldb+n]);
popc[3] = POPCNT(A_PART[3]^B[(k+3)*ldb+n]);
popc[4] = POPCNT(A_PART[4]^B[(k+4)*ldb+n]);
popc[5] = POPCNT(A_PART[1]^B[(k+5)*ldb+n]);
//C[m*ldc+n] += __builtin_popcountll(temp_A^B[k*ldb+n]);
C[m*ldc+n] += popc[0]+popc[1]+popc[2]+popc[3]+popc[4]+popc[5];
}
}
#pragma omp parallel for
for(k=(K/UNROLLN)*UNROLLN;k<K;k++){
BinaryCode temp_A= A[m*lda+k];
#pragma omp parallel for
for(n = 0; n < N; n++){
C[m*ldc+n] += POPCNT(temp_A^B[k*ldb+n]);
}
}
}
#pragma omp parallel for
for(m = 0; m < M; m++)
#pragma omp parallel for
for(n = 0; n < N; n++){
C[m*ldc+n] = size - 2*C[m*ldc+n];
C[m*ldc+n] *= alphas[m];
}
}
/*
* @brief: GEMM using xnor operation.
* In convolution, M means the number of filters and K means (ck^2)/BIN_SIZE.
* N means the output_spatial_size. A is weights pointer. B is the input
* column pointer. C is the output pointer.
*
* @TODO inner_product
* @param:
* int M, K, N;
* int lda = K , ldb = N, ldc = N ;
* int size, the actual size of a vector;
*
*/
template <typename Dtype>
void xnorGEMM_baseline(int M,int K,int N,
const BinaryCode* A, int lda ,
const BinaryCode* B, int ldb,
Dtype* C, int ldc,
vector<Dtype>& alphas){
int m,k,n;
for (m = 0; m < M; m++){
for(k = 0; k < K; k++){
register BinaryCode temp_A= A[m*lda+k];
for(n = 0; n < N; n++){
C[m*ldc+n] += POPCNT(~(temp_A^B[k*ldb+n]));
}
}
}
for(m = 0; m < M; m++){
register Dtype a = alphas[m];
for(n = 0; n < N; n++){
C[m*ldc+n] *= a;
}
}
}
}
#endif