forked from kaldi-asr/kaldi
-
Notifications
You must be signed in to change notification settings - Fork 0
/
cu-math.cc
241 lines (194 loc) · 7.45 KB
/
cu-math.cc
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
// cudamatrix/cu-math.cc
// Copyright 2009-2012 Karel Vesely
// Johns Hopkins University (author: Daniel Povey)
// See ../../COPYING for clarification regarding multiple authors
//
// 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
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "base/timer.h"
#include "cudamatrix/cu-common.h"
#include "cudamatrix/cu-matrix.h"
#include "cudamatrix/cu-device.h"
#include "cudamatrix/cu-kernels.h"
namespace kaldi {
namespace cu {
/*
* templated functions wrapping the ANSI-C CUDA kernel functions
*/
template<typename Real>
void RegularizeL1(CuMatrixBase<Real> *weight, CuMatrixBase<Real> *grad, Real l1, Real lr) {
KALDI_ASSERT(SameDim(*weight, *grad));
#if HAVE_CUDA == 1
if (CuDevice::Instantiate().Enabled()) {
Timer tim;
dim3 dimBlock(CU2DBLOCK, CU2DBLOCK);
dim3 dimGrid(n_blocks(weight->NumCols(), CU2DBLOCK), n_blocks(weight->NumRows(), CU2DBLOCK));
cuda_regularize_l1(dimGrid, dimBlock, weight->Data(), grad->Data(), l1, lr,
weight->Dim(), grad->Stride());
CU_SAFE_CALL(cudaGetLastError());
CuDevice::Instantiate().AccuProfile(__func__, tim.Elapsed());
} else
#endif
{
MatrixBase<Real> &weight2 = weight->Mat();
MatrixBase<Real> &grad2 = grad->Mat();
for(MatrixIndexT r=0; r<weight2.NumRows(); r++) {
for(MatrixIndexT c=0; c<weight2.NumCols(); c++) {
if(weight2(r,c)==0.0) continue; // skip L1 if zero weightght!
Real l1_signed = l1;
if (weight2(r, c) < 0.0)
l1_signed = -l1;
Real before = weight2(r, c);
Real after = weight2(r, c) - lr*grad2(r, c) - l1_signed;
if ((after > 0.0) ^ (before > 0.0)) {
weight2(r, c) = 0.0;
grad2(r, c) = 0.0;
} else {
weight2(r, c) -= l1_signed;
}
}
}
}
}
template<typename Real>
void Randomize(const CuMatrixBase<Real> &src,
const CuArray<int32> ©_from_idx,
CuMatrixBase<Real> *tgt) {
KALDI_ASSERT(src.NumCols() == tgt->NumCols());
KALDI_ASSERT(src.NumRows() == tgt->NumRows());
KALDI_ASSERT(copy_from_idx.Dim() <= tgt->NumRows());
#if HAVE_CUDA == 1
if (CuDevice::Instantiate().Enabled()) {
Timer tim;
/*
Note: default 16x16 block-size limits the --cachesize to matrix size 16*65535 x 16*65535
dim3 dimBlock(CU2DBLOCK, CU2DBLOCK);
dim3 dimGrid(n_blocks(tgt->NumCols(), CU2DBLOCK), n_blocks(copy_from_idx.Dim(), CU2DBLOCK));
*/
/*
* Let's use blocksize 4 x 128 (512 threads/block)
* and extend the randomizable matrices to: col 4*65535, row 128*65535
* (ie. max-cols:262140 (dim), max-rows:8388480 (datapoints))
*/
dim3 dimBlock(4, 128);
dim3 dimGrid(n_blocks(tgt->NumCols(), 4), n_blocks(copy_from_idx.Dim(), 128));
/*
*/
MatrixDim dimsrc = src.Dim(); dimsrc.rows=copy_from_idx.Dim();
MatrixDim dimtgt = tgt->Dim(); dimtgt.rows=copy_from_idx.Dim();
cuda_randomize(dimGrid, dimBlock, tgt->Data(), src.Data(),
copy_from_idx.Data(), dimtgt, dimsrc);
CU_SAFE_CALL(cudaGetLastError());
CuDevice::Instantiate().AccuProfile(__func__, tim.Elapsed());
} else
#endif
{
// randomize in CPU
const MatrixBase<Real> &srcmat = src.Mat();
const int32 *copy_from_idxvec = copy_from_idx.Data();
MatrixBase<Real> &tgtmat = tgt->Mat();
for(int32 i=0; i<copy_from_idx.Dim(); i++) {
tgtmat.Row(i).CopyFromVec(srcmat.Row(copy_from_idxvec[i]));
}
}
}
template<typename Real>
void Splice(const CuMatrixBase<Real> &src, const CuArray<int32> &frame_offsets,
CuMatrixBase<Real> *tgt) {
KALDI_ASSERT(src.NumCols()*frame_offsets.Dim() == tgt->NumCols());
KALDI_ASSERT(src.NumRows() == tgt->NumRows());
#if HAVE_CUDA == 1
if (CuDevice::Instantiate().Enabled()) {
Timer tim;
dim3 dimBlock(CU2DBLOCK, CU2DBLOCK);
dim3 dimGrid(n_blocks(tgt->NumCols(), CU2DBLOCK), n_blocks(tgt->NumRows(), CU2DBLOCK));
cuda_splice(dimGrid, dimBlock, tgt->Data(), src.Data(),
frame_offsets.Data(), tgt->Dim(), src.Dim());
CU_SAFE_CALL(cudaGetLastError());
CuDevice::Instantiate().AccuProfile(__func__, tim.Elapsed());
} else
#endif
{
// expand in CPU
const MatrixBase<Real> &srcmat = src.Mat();
const int32 *frame_offsetvec = frame_offsets.Data();
int32 dim = frame_offsets.Dim();
MatrixBase<Real> &tgtmat = tgt->Mat();
//
for(int32 r=0; r < tgtmat.NumRows(); r++) {
for(int32 off=0; off < dim; off++) {
int32 r_off = r + frame_offsetvec[off];
if(r_off < 0) r_off = 0;
if(r_off >= srcmat.NumRows()) r_off = srcmat.NumRows()-1;
memcpy(tgtmat.RowData(r)+off*srcmat.NumCols(),srcmat.RowData(r_off),sizeof(Real)*srcmat.NumCols());
}
}
}
}
template<typename Real>
void Copy(const CuMatrixBase<Real> &src, const CuArray<int32> ©_from_indices,
CuMatrixBase<Real> *tgt) {
KALDI_ASSERT(copy_from_indices.Dim() == tgt->NumCols());
KALDI_ASSERT(src.NumRows() == tgt->NumRows());
#if HAVE_CUDA == 1
if (CuDevice::Instantiate().Enabled()) {
Timer tim;
dim3 dimBlock(CU2DBLOCK, CU2DBLOCK);
dim3 dimGrid(n_blocks(tgt->NumCols(), CU2DBLOCK), n_blocks(tgt->NumRows(), CU2DBLOCK));
cuda_copy(dimGrid, dimBlock, tgt->Data(), src.Data(),
copy_from_indices.Data(), tgt->Dim(), src.Dim());
CU_SAFE_CALL(cudaGetLastError());
CuDevice::Instantiate().AccuProfile(__func__, tim.Elapsed());
} else
#endif
{
// expand in CPU
const MatrixBase<Real> &srcmat = src.Mat();
const int32 *copy_from_indicesvec = copy_from_indices.Data();
int32 dim = copy_from_indices.Dim();
MatrixBase<Real> &tgtmat = tgt->Mat();
//
for(int32 r = 0; r < tgtmat.NumRows(); r++) {
for(int32 c = 0; c < dim; c++) {
tgtmat(r,c) = srcmat(r,copy_from_indicesvec[c]);
}
}
}
}
// instantiate the templates.
template
void RegularizeL1(CuMatrixBase<float> *weight, CuMatrixBase<float> *grad, float l1, float lr);
template
void RegularizeL1(CuMatrixBase<double> *weight, CuMatrixBase<double> *grad, double l1, double lr);
template
void Splice(const CuMatrixBase<float> &src, const CuArray<int32> &frame_offsets,
CuMatrixBase<float> *tgt);
template
void Splice(const CuMatrixBase<double> &src, const CuArray<int32> &frame_offsets,
CuMatrixBase<double> *tgt);
template
void Copy(const CuMatrixBase<float> &src, const CuArray<int32> ©_from_indices,
CuMatrixBase<float> *tgt);
template
void Copy(const CuMatrixBase<double> &src, const CuArray<int32> ©_from_indices,
CuMatrixBase<double> *tgt);
template
void Randomize(const CuMatrixBase<float> &src,
const CuArray<int32> ©_from_idx,
CuMatrixBase<float> *tgt);
template
void Randomize(const CuMatrixBase<double> &src,
const CuArray<int32> ©_from_idx,
CuMatrixBase<double> *tgt);
} //namespace cu
} //namespace kaldi