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
|
// Copyright (c) Facebook, Inc. and its affiliates.
//
// This source code is licensed under the MIT license found in the
// LICENSE file in the root directory of this source tree.
#include <ops.cuh>
#include <kernels.cuh>
#include <cub/device/device_scan.cuh>
#include <limits>
#include <BinSearch.h>
#include <common.h>
using namespace BinSearch;
using std::cout;
using std::endl;
void histogramScatterAdd2D(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n)
{
int threads = 512;
int blocks = n/threads;
blocks = n % threads == 0 ? blocks : blocks + 1;
kHistogramScatterAdd2D<<<blocks, 512>>>(histogram, index1, index2, src, maxidx1, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
template <typename T> void estimateQuantiles(T *A, float *code, float offset, int n)
{
int blocks = n/4096;
blocks = n % 4096 == 0 ? blocks : blocks + 1;
CUDA_CHECK_RETURN(cudaMemset(code, 0, 256*sizeof(float)));
kEstimateQuantiles<T><<<blocks, 512>>>(A, code, offset, std::numeric_limits<T>::max(), n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
void quantize(float *code, float *A, unsigned char *out, int n)
{
int blocks = n/1024;
blocks = n % 1024 == 0 ? blocks : blocks + 1;
kQuantize<<<blocks, 1024>>>(code, A, out, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n)
{
int blocks = n/4096;
blocks = n % 4096 == 0 ? blocks : blocks + 1;
kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n)
{
int blocks = n/blocksize;
blocks = n % blocksize == 0 ? blocks : blocks + 1;
if(blocksize == 4096)
kDequantizeBlockwise<T, 4096, 1024, 4><<<blocks, 4096/4>>>(code, A, absmax, out, n);
else if(blocksize == 2048)
kDequantizeBlockwise<T, 2048, 512, 4><<<blocks, 2048/4>>>(code, A, absmax, out, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
const float beta1, const float beta2, const float eps, const float weight_decay,
const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n)
{
int blocks = n/4096;
blocks = n % 4096 == 0 ? blocks : blocks + 1;
switch(OPTIMIZER)
{
case ADAM:
if(max_unorm > 0.0f)
{
CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
kPreconditionOptimizer32bit2State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
case MOMENTUM:
case RMSPROP:
case ADAGRAD:
if(max_unorm > 0.0f)
{
CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
}
}
template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
unsigned char* state1, unsigned char* state2,
float *unorm, float max_unorm, float param_norm,
float beta1, float beta2,
float eps, int step, float lr,
float* quantiles1, float* quantiles2,
float* max1, float* max2, float* new_max1, float* new_max2,
float weight_decay,
const float gnorm_scale, int n)
{
int blocks = n/4096;
blocks = n % 4096 == 0 ? blocks : blocks + 1;
if(max_unorm > 0.0f){ CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); }
switch(OPTIMIZER)
{
case ADAM:
CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1*sizeof(float)));
kPreconditionOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
kOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
case MOMENTUM:
case RMSPROP:
case ADAGRAD:
CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
kOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
default:
break;
}
}
#define BLOCKSIZE_2STATE 2048
#define NUM_2STATE 8
#define BLOCKSIZE_1STATE 2048
#define NUM_1STATE 8
template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g,
unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,
float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)
{
int blocks = 0;
switch(OPTIMIZER)
{
case ADAM:
blocks = n/BLOCKSIZE_2STATE;
blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
kOptimizerStatic8bit2StateBlockwise<T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE><<<blocks, BLOCKSIZE_2STATE/NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
case MOMENTUM:
case RMSPROP:
case ADAGRAD:
blocks = n/BLOCKSIZE_1STATE;
blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
break;
}
}
template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n)
{
int blocks = n/2048;
blocks = n % 2048 == 0 ? blocks : blocks + 1;
CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1*sizeof(float)));
kPercentileClipping<T, 2048, 4><<<blocks, 512>>>(g, gnorm_vec, step, n);
CUDA_CHECK_RETURN(cudaPeekAtLastError());
}
//==============================================================
// TEMPLATE DEFINITIONS
//==============================================================
template void estimateQuantiles(half *A, float *code, float offset, int n);
template void estimateQuantiles(float *A, float *code, float offset, int n);
template void quantizeBlockwise<half, 0>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
template void quantizeBlockwise<float, 0>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
template void quantizeBlockwise<half, 1>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
template void quantizeBlockwise<float, 1>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
template void dequantizeBlockwise<half>(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n);
template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n);
#define MAKE_optimizer32bit(name, gtype) \
template void optimizer32bit<gtype, name>(gtype* g, gtype* p, \
float* state1, float* state2, float* unorm, float max_unorm, float param_norm, \
const float beta1, const float beta2, const float eps, const float weight_decay, \
const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
MAKE_optimizer32bit(ADAM, half)
MAKE_optimizer32bit(ADAM, float)
MAKE_optimizer32bit(MOMENTUM, half)
MAKE_optimizer32bit(MOMENTUM, float)
MAKE_optimizer32bit(RMSPROP, half)
MAKE_optimizer32bit(RMSPROP, float)
MAKE_optimizer32bit(ADAGRAD, half)
MAKE_optimizer32bit(ADAGRAD, float)
#define MAKE_optimizerStatic8bit(name, gtype) \
template void optimizerStatic8bit<gtype, name>(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
float *unorm, float max_unorm, float param_norm, \
float beta1, float beta2, \
float eps, int step, float lr, \
float* quantiles1, float* quantiles2, \
float* max1, float* max2, float* new_max1, float* new_max2, \
float weight_decay, \
const float gnorm_scale, int n); \
MAKE_optimizerStatic8bit(ADAM, half)
MAKE_optimizerStatic8bit(ADAM, float)
MAKE_optimizerStatic8bit(MOMENTUM, half)
MAKE_optimizerStatic8bit(MOMENTUM, float)
MAKE_optimizerStatic8bit(RMSPROP, half)
MAKE_optimizerStatic8bit(RMSPROP, float)
#define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
template void optimizerStatic8bitBlockwise<gtype, optim_name>(gtype* p, gtype* g, \
unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, \
float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n); \
MAKE_optimizerStatic8bitBlockwise(half, ADAM);
MAKE_optimizerStatic8bitBlockwise(float, ADAM);
MAKE_optimizerStatic8bitBlockwise(half, MOMENTUM);
MAKE_optimizerStatic8bitBlockwise(float, MOMENTUM);
MAKE_optimizerStatic8bitBlockwise(half, RMSPROP);
MAKE_optimizerStatic8bitBlockwise(float, RMSPROP);
MAKE_optimizerStatic8bitBlockwise(half, ADAGRAD);
MAKE_optimizerStatic8bitBlockwise(float, ADAGRAD);
template void percentileClipping(float * g, float *gnorm_vec, int step, const int n);
template void percentileClipping(half * g, float *gnorm_vec, int step, const int n);
|
