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
|
/******************************************************************************
* Copyright (c) Intel Corporation - All rights reserved. *
* This file is part of the LIBXSMM library. *
* *
* For information on the license, see the LICENSE file. *
* Further information: https://github.com/hfp/libxsmm/ *
* SPDX-License-Identifier: BSD-3-Clause *
******************************************************************************/
/* Sasikanth Avancha, Dhiraj Kalamkar (Intel Corp.)
******************************************************************************/
#pragma once
#include <map>
#include <list>
#include <vector>
#include <algorithm>
#include <set>
#include <omp.h>
#include <sys/time.h>
#include <stdlib.h>
#include "proto/gxm.pb.h"
#include "Engine.fwd.hpp"
#include "MLNode.fwd.hpp"
#include "Config.hpp"
#include "Task.hpp"
#include "common.hpp"
#include "Solver.hpp"
#include "libxsmm.h"
#ifdef USE_MLSL
#include "mpi.h"
#endif
using namespace std;
using namespace gxm;
extern int iter;
#ifdef USE_MLSL
#include "mlsl.hpp"
//using namespace MLSL;
#endif
#define TRAIN 0
#define VAL 1
#define TEST 2
#define START_GUARD_BAND 64
#define END_GUARD_BAND 64
#define CANARY 0x7F
#define NDIFFS 10
struct dupChecker_ {
inline dupChecker_() : tmpSet() {}
inline bool operator()(Task *t) {
return tmpSet.insert(t).second;
}
private:
std::set<Task *> tmpSet;
};
class MLEngine
{
protected:
NTGParameter ntgparam_;
NodeParameter np_;
SolverParameter sparam_;
#ifdef USE_MLSL
MLSL::Distribution *data_parallelism;
MLSL::Session *session_;
#endif
vector<MLNode*> ntg_;
list<Task*> etg_[3]; // 0 - Training, 1 - Validation, 2 - testing
SolverParams *solverParams_;
SolverNode* solver_;
Tensor* tenScratch_;
TensorBuf* tenScratchBuf_;
struct TensorPair
{
string name;
Tensor* t;
};
typedef list<TensorPair> TensorList;
typedef TensorList::iterator Iter;
typedef map<string, Iter> Tmap;
Tmap inTensorMap_, outTensorMap_, weightTensorMap_, biasTensorMap_, statsTensorMap_;
TensorList defTList_, inTList_, outTList_, wTList_, biasTList_, statsTList_;
bool inferenceOnly_, load_from_checkpoint_;
string checkpoint_dir_, checkpoint_format_;
int num_epochs_, exec_mode_, current_epoch_, current_batch_;
int data_type_;
int num_machines_, num_machine_groups_, num_threads_;
int batch_size_, num_train_batches_, num_test_batches_, num_test_views_;
int global_node_id_;
float lr_, *wt_lr_mult_[NUM_NUMA_NODES], *wt_decay_mult_[NUM_NUMA_NODES];
float *bias_lr_mult_[NUM_NUMA_NODES], *bias_decay_mult_[NUM_NUMA_NODES];
float scf_=0;
void *input_buf_=NULL;
void *fact_buf_=NULL, *bact_buf_=NULL, *wbuf_=NULL;
void *weight_buf_[NUM_NUMA_NODES]={NULL}, *wdiff_buf_[NUM_NUMA_NODES]={NULL};
void *winc_buf_[NUM_NUMA_NODES]={NULL}, *lpweight_buf_[NUM_NUMA_NODES]={NULL};
void *lpwdiff_buf_[NUM_NUMA_NODES]={NULL};
#if 1
void *bias_buf_[NUM_NUMA_NODES]={NULL}, *bidiff_buf_[NUM_NUMA_NODES]={NULL};
void *biinc_buf_[NUM_NUMA_NODES]={NULL}, *stats_buf_[NUM_NUMA_NODES]={NULL};
#else
void *bias_buf_=NULL, *bidiff_buf_=NULL;
void *biinc_buf_=NULL, *stats_buf_=NULL;
#endif
int total_weights_, total_biases_, orig_total_weights_;
void *scratch[NUM_NUMA_NODES]={NULL};
vector<int> input_can_ptr;
vector<int> fact_can_ptr, bact_can_ptr;
vector<int> wt_can_ptr, wdiff_can_ptr, winc_can_ptr;
vector<int> bias_can_ptr, stats_can_ptr, bidiff_can_ptr, biinc_can_ptr;
#ifdef USE_MLSL
vector<MLSL::Operation*> wtgrad_comms_vec, bias_grad_comms_vec, combo_grad_comms_vec;
#endif
int ic, fac, bac, wtc, wdc, wic, bic, sic, bidc, biic;
void create_schedule(int);
void optimize_schedule(int);
void allocate_tensor_memory(Tensor*, int, void*);
void clear_history(TensorList);
int find_in_nodeTypeList(string);
void checkpoint(TensorList L, int);
void read_checkpoint_file(TensorBuf*, string, string);
void load_checkpoint(TensorList, int, string);
void canary_check(void*, vector<int>&, int);
void allocate_memory(string, TensorList, int, vector<int>&, int*, long long int*);
void* allocate_gradient_tensor(TensorList, int, int, long long int);
void insertSplitNodes(NTGParameter& p, NTGParameter* ps);
void convert_f32_bf16(float* in, libxsmm_bfloat16* out, int len, int numa_node);
void convert_f32_bf16(float** in, libxsmm_bfloat16** out, int len);
void convert_bf16_f32(libxsmm_bfloat16* in, float* out, int len);
void waitForComms(string);
public:
MLEngine() {}
virtual ~MLEngine() {}
void create(int mode, string ntgConfig, string solverConfig);
bool register_tensor(string name, int type, Tensor* t);
Tensor* get_tensor(string name, int type);
void execute_on_thread(int num_threads, MLNode* node, void (*fname)(int tid));
void set_global_strategy(MachineParameter* mparam);
void run(int mode);
SolverNode* getSolver() { return solver_; }
TensorBuf* getScratchBuffer() { return tenScratchBuf_; }
bool is_inference_only() { return inferenceOnly_; }
int get_num_threads() { return num_threads_; }
int get_num_machines() { return num_machines_; }
int get_num_machine_groups() { return num_machine_groups_; }
int get_num_epochs() { return num_epochs_;}
int get_current_epoch() { return current_epoch_; }
int get_current_batch() { return current_batch_; }
int get_execution_mode() { return exec_mode_; }
int get_global_node_id() { return global_node_id_; }
int get_num_train_batches() { return num_train_batches_; }
int get_num_test_batches() { return num_test_batches_; }
int get_num_test_views() {return num_test_views_; }
int get_batch_size() { return batch_size_; }
float get_scaling_factor() { return scf_; }
#ifdef USE_MLSL
vector<MLSL::Operation*>& get_wtgrad_comms_vec() { return wtgrad_comms_vec; }
vector<MLSL::Operation*>& get_bias_grad_comms_vec() { return bias_grad_comms_vec; }
vector<MLSL::Operation*>& get_combo_grad_comms_vec() { return combo_grad_comms_vec; }
#endif
void set_batch_size(int b) {batch_size_ = b; }
void set_num_train_batches(int ntrainb) {num_train_batches_ = ntrainb; }
void set_num_test_batches(int ntestb) {num_test_batches_ = ntestb; }
void set_num_test_views(int ntestv) {num_test_views_ = ntestv; }
void set_learning_rate(float lr) { lr_ = lr; }
void set_scaling_factor(float scf) { scf_ = scf; }
#ifdef USE_MLSL
MLSL::Distribution* get_distribution() { return data_parallelism; }
MLSL::Session *get_session() { return session_; }
#endif
};
|
