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
|
#include <cub/block/block_reduce.cuh>
#include "caffe2/core/common_gpu.h"
#include "caffe2/core/context_gpu.h"
#include "caffe2/sgd/adam_op.h"
#include "caffe2/utils/cub_namespace.cuh"
namespace caffe2 {
__global__ void AdamUpdate(
int N,
const float* g,
const float* m,
const float* v,
float* ng,
float* nm,
float* nv,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr) {
CUDA_1D_KERNEL_LOOP(i, N) {
float gi = g[i];
float mi = nm[i] = m[i] * beta1 + gi * (1 - beta1);
float vi = nv[i] = v[i] * beta2 + gi * gi * (1 - beta2);
ng[i] = lr[0] * correction * mi / (sqrtf(vi) + eps_hat);
}
}
template <>
void adam_update<CUDAContext>(
int N,
const float* g,
const float* m,
const float* v,
float* ng,
float* nm,
float* nv,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr,
CUDAContext* context) {
AdamUpdate<<<
CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N, g, m, v, ng, nm, nv, beta1, beta2, eps_hat, correction, lr);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
__global__ void AdamCompute(
int N,
const float* w,
const float* g,
const float* m,
const float* v,
float* nw,
float* nm,
float* nv,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr) {
CUDA_1D_KERNEL_LOOP(i, N) {
float gi = g[i];
float mi = nm[i] = m[i] * beta1 + gi * (1 - beta1);
float vi = nv[i] = v[i] * beta2 + gi * gi * (1 - beta2);
float ng = lr[0] * correction * mi / (sqrtf(vi) + eps_hat);
nw[i] = w[i] + ng;
}
}
template <>
void adam_compute<CUDAContext>(
int N,
const float* w,
const float* g,
const float* m,
const float* v,
float* nw,
float* nm,
float* nv,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr,
CUDAContext* context) {
AdamCompute<<<
CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N, w, g, m, v, nw, nm, nv, beta1, beta2, eps_hat, correction, lr);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
__global__ void AdamComputeOutputGrad(
int N,
const float* w,
const float* g,
const float* m,
const float* v,
float* nw,
float* nm,
float* nv,
float* ng,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr) {
CUDA_1D_KERNEL_LOOP(i, N) {
float gi = g[i];
float mi = nm[i] = m[i] * beta1 + gi * (1 - beta1);
float vi = nv[i] = v[i] * beta2 + gi * gi * (1 - beta2);
float ngi = ng[i] = correction * mi / (sqrtf(vi) + eps_hat);
nw[i] = w[i] + lr[0] * ngi;
}
}
template <>
void adam_compute_output_grad<CUDAContext>(
int N,
const float* w,
const float* g,
const float* m,
const float* v,
float* nw,
float* nm,
float* nv,
float* ng,
float beta1,
float beta2,
float eps_hat,
float correction,
const float* lr,
CUDAContext* context) {
AdamComputeOutputGrad<<<
CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(
N, w, g, m, v, nw, nm, nv, ng, beta1, beta2, eps_hat, correction, lr);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
template <typename SIndex>
__global__ void SparseAdamKernel(
const size_t N,
const size_t grad_slice_sz,
const float beta1,
const float beta2,
const float epsilon,
float* param,
float* mom1,
float* mom2,
const SIndex* indices,
const float* grad,
const float correction,
const float* lr,
const float iter) {
CUDA_1D_KERNEL_LOOP(i, N) {
const size_t gradIdx = i;
const SIndex index = indices[i / grad_slice_sz];
const size_t paramIdx = index * grad_slice_sz + (i % grad_slice_sz);
float m1n = mom1[paramIdx] =
mom1[paramIdx] * beta1 + grad[gradIdx] * (1.0f - beta1);
float m2n = mom2[paramIdx] =
mom2[paramIdx] * beta2 + grad[gradIdx] * grad[gradIdx] * (1.0f - beta2);
param[paramIdx] += lr[0] * correction * m1n / (sqrt(m2n) + epsilon);
}
}
template <typename SIndex>
__global__ void SparseAdamOutputGradKernel(
const size_t N,
const size_t grad_slice_sz,
const float beta1,
const float beta2,
const float epsilon,
float* param,
float* mom1,
float* mom2,
float* output_grad,
const SIndex* indices,
const float* grad,
const float correction,
const float* lr,
const float iter) {
CUDA_1D_KERNEL_LOOP(i, N) {
const size_t gradIdx = i;
const SIndex index = indices[i / grad_slice_sz];
const size_t paramIdx = index * grad_slice_sz + (i % grad_slice_sz);
float m1n = mom1[paramIdx] =
mom1[paramIdx] * beta1 + grad[gradIdx] * (1.0f - beta1);
float m2n = mom2[paramIdx] =
mom2[paramIdx] * beta2 + grad[gradIdx] * grad[gradIdx] * (1.0f - beta2);
float gradOut = output_grad[gradIdx] =
correction * m1n / (sqrt(m2n) + epsilon);
param[paramIdx] += lr[0] * gradOut;
}
}
template <typename SIndex>
__global__ void RowWiseSparseAdamKernel(
const int M,
const int N,
const float beta1,
const float beta2,
const float epsilon,
float* param,
float* mom1,
float* mom2,
const SIndex* indices,
const float* grad,
const float correction,
const float* lr) {
typedef cub::BlockReduce<float, CAFFE_CUDA_NUM_THREADS> BlockReduce;
__shared__ BlockReduce::TempStorage temp_storage;
int valid = min(N, CAFFE_CUDA_NUM_THREADS);
// in case gridDim is smaller than M
for (int i = blockIdx.x; i < M; i += gridDim.x) {
const SIndex index = indices[i];
float sum_squares = 0.0;
__shared__ float row_sum_squares_avg;
// in case N is bigger than block size which is 512 by default
for (int j = threadIdx.x; j < N; j += blockDim.x) {
const float x_ij = grad[i * N + j];
sum_squares += x_ij * x_ij;
}
float reduce_sum_squares =
BlockReduce(temp_storage).Sum(sum_squares, valid);
if (threadIdx.x == 0) {
row_sum_squares_avg = reduce_sum_squares / (float)N;
mom2[index] = mom2[index] * beta2 + row_sum_squares_avg * (1.0f - beta2);
}
__syncthreads();
// update param
float step = correction / (std::sqrt(mom2[index]) + epsilon);
for (int j = threadIdx.x; j < N; j += blockDim.x) {
mom1[index * N + j] =
mom1[index * N + j] * beta1 + grad[i * N + j] * (1.0f - beta1);
param[index * N + j] += lr[0] * mom1[index * N + j] * step;
}
}
}
template <typename SIndex>
__global__ void RowWiseSparseAdamOutputGradKernel(
const int M,
const int N,
const float beta1,
const float beta2,
const float epsilon,
float* param,
float* mom1,
float* mom2,
float* output_grad,
const SIndex* indices,
const float* grad,
const float correction,
const float* lr) {
typedef cub::BlockReduce<float, CAFFE_CUDA_NUM_THREADS> BlockReduce;
__shared__ BlockReduce::TempStorage temp_storage;
int valid = min(N, CAFFE_CUDA_NUM_THREADS);
// in case gridDim is smaller than M
for (int i = blockIdx.x; i < M; i += gridDim.x) {
const SIndex index = indices[i];
float sum_squares = 0.0;
__shared__ float row_sum_squares_avg;
// in case N is bigger than block size which is 512 by default
for (int j = threadIdx.x; j < N; j += blockDim.x) {
const float x_ij = grad[i * N + j];
sum_squares += x_ij * x_ij;
}
float reduce_sum_squares =
BlockReduce(temp_storage).Sum(sum_squares, valid);
if (threadIdx.x == 0) {
row_sum_squares_avg = reduce_sum_squares / (float)N;
mom2[index] = mom2[index] * beta2 + row_sum_squares_avg * (1.0f - beta2);
}
__syncthreads();
// update param
float step = correction / (std::sqrt(mom2[index]) + epsilon);
for (int j = threadIdx.x; j < N; j += blockDim.x) {
mom1[index * N + j] =
mom1[index * N + j] * beta1 + grad[i * N + j] * (1.0f - beta1);
output_grad[i * N + j] = mom1[index * N + j] * step;
param[index * N + j] += lr[0] * output_grad[i * N + j];
}
}
}
template <>
template <typename SIndex>
bool SparseAdamOp<float, CUDAContext>::DoRunWithType() {
Output(OUTPUT_PARAM)->ResizeLike(Input(PARAM));
Output(OUTPUT_MOMENT_1)->ResizeLike(Input(MOMENT_1));
Output(OUTPUT_MOMENT_2)->ResizeLike(Input(MOMENT_2));
auto N = Input(GRAD).size();
auto grad_slice_sz = Input(GRAD).size_from_dim(Input(INDICES).ndim());
const auto iter =
OperatorBase::Input<Tensor>(ITER, CPU).template data<int64_t>()[0];
const float correction = sqrtf(1.0f - std::pow(beta2_, iter + 1)) /
(1.0f - std::pow(beta1_, iter + 1));
if (OutputSize() == 3) {
SparseAdamKernel<SIndex>
<<<CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
N,
grad_slice_sz,
beta1_,
beta2_,
epsilon_,
Output(OUTPUT_PARAM)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_1)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_2)->template mutable_data<float>(),
Input(INDICES).template data<SIndex>(),
Input(GRAD).template data<float>(),
correction,
Input(LR).template data<float>(),
iter);
C10_CUDA_KERNEL_LAUNCH_CHECK();
} else {
Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD));
SparseAdamOutputGradKernel<SIndex>
<<<CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
N,
grad_slice_sz,
beta1_,
beta2_,
epsilon_,
Output(OUTPUT_PARAM)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_1)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_2)->template mutable_data<float>(),
Output(OUTPUT_GRAD)->template mutable_data<float>(),
Input(INDICES).template data<SIndex>(),
Input(GRAD).template data<float>(),
correction,
Input(LR).template data<float>(),
iter);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
return true;
}
template <>
template <typename SIndex>
bool RowWiseSparseAdamOp<float, CUDAContext>::DoRunWithType() {
Output(OUTPUT_PARAM)->ResizeLike(Input(PARAM));
Output(OUTPUT_MOMENT_1)->ResizeLike(Input(MOMENT_1));
Output(OUTPUT_MOMENT_2)->ResizeLike(Input(MOMENT_2));
auto N = Input(GRAD).size();
if (N == 0) {
// empty grad, nothing to do here, not even launching the kernel
return true;
}
const auto iter =
OperatorBase::Input<Tensor>(ITER, CPU).template data<int64_t>()[0];
const float correction = sqrtf(1.0f - std::pow(beta2_, iter + 1)) /
(1.0f - std::pow(beta1_, iter + 1));
// size of the 1st dimension of the input gradient
auto GRAD_M = Input(GRAD).dim32(0);
auto GRAD_N = N / GRAD_M;
if (OutputSize() == 3) {
RowWiseSparseAdamKernel<SIndex>
<<<std::min(GRAD_M, CAFFE_MAXIMUM_NUM_BLOCKS),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
GRAD_M,
GRAD_N,
beta1_,
beta2_,
epsilon_,
Output(OUTPUT_PARAM)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_1)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_2)->template mutable_data<float>(),
Input(INDICES).template data<SIndex>(),
Input(GRAD).template data<float>(),
correction,
Input(LR).template data<float>());
C10_CUDA_KERNEL_LAUNCH_CHECK();
} else {
Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD));
RowWiseSparseAdamOutputGradKernel<SIndex>
<<<std::min(GRAD_M, CAFFE_MAXIMUM_NUM_BLOCKS),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
GRAD_M,
GRAD_N,
beta1_,
beta2_,
epsilon_,
Output(OUTPUT_PARAM)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_1)->template mutable_data<float>(),
Output(OUTPUT_MOMENT_2)->template mutable_data<float>(),
Output(OUTPUT_GRAD)->template mutable_data<float>(),
Input(INDICES).template data<SIndex>(),
Input(GRAD).template data<float>(),
correction,
Input(LR).template data<float>());
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
return true;
}
REGISTER_CUDA_OPERATOR(Adam, AdamOp<float, CUDAContext>);
REGISTER_CUDA_OPERATOR(SparseAdam, SparseAdamOp<float, CUDAContext>);
REGISTER_CUDA_OPERATOR(
RowWiseSparseAdam,
RowWiseSparseAdamOp<float, CUDAContext>);
} // namespace caffe2
|
