-
Notifications
You must be signed in to change notification settings - Fork 2
/
saga.hpp
617 lines (455 loc) · 17.4 KB
/
saga.hpp
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
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
// -*-c++-*-
/*
* Copyright (c) 2019, Andreas Smas
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once
#include <memory>
#include <vector>
#include <unordered_map>
#include <functional>
#include <variant>
#include <unordered_set>
#include <optional>
#include <string>
#include <stdexcept>
#include <assert.h>
namespace saga {
class Context;
//------------------------------------------------------------------------
//------------------------------------------------------------------------
enum class DimParam {
BATCH_SIZE,
UNCHANGED,
REDUCE,
};
struct Dim : public std::variant<int64_t, DimParam> {
using std::variant<int64_t, DimParam>::variant;
std::string to_string() const;
operator int() const
{
if(auto v = std::get_if<int64_t>(&*this)) {
return *v;
} else {
throw std::runtime_error(
"Unable to convert parameterized dim to scalar");
}
}
Dim &operator++();
Dim &operator+=(const Dim &);
};
class Dims : public std::vector<Dim> {
using std::vector<Dim>::vector;
public:
Dims(const std::vector<Dim> &v) : std::vector<Dim>(v) {}
Dims(const std::vector<int> &v);
Dims batch(int64_t v) const;
std::vector<int64_t> i64() const;
std::vector<int32_t> i32() const;
size_t elements(size_t from_rank = 0) const;
std::string to_string() const;
bool similar(const Dims &) const;
Dims transform(std::function<Dim(const DimParam &dp, size_t i)> fn) const;
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
struct Attribute
: public std::variant<float, int, std::vector<int>, bool, Dims> {
using std::variant<float, int, std::vector<int>, bool, Dims>::variant;
std::string to_string() const;
};
class Attributes : public std::unordered_map<std::string, Attribute> {
public:
using std::unordered_map<std::string, Attribute>::unordered_map;
template <typename T>
T get(const std::string &n, T def) const
{
auto it = find(n);
if(it == end())
return def;
auto p = std::get_if<T>(&it->second);
if(p == NULL)
return def;
return *p;
}
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
enum class TensorLayout {
Auto,
NHWC,
NCHW,
};
class Tensors;
class TensorAccess {
protected:
TensorAccess(){};
public:
virtual ~TensorAccess(){};
virtual Dims strides() = 0;
virtual void *data() = 0;
virtual void copyBytesFrom(const Dims &element, const void *data,
size_t size) = 0;
virtual void *getAddr(const Dims &element) { return nullptr; }
virtual double get(const Dims &element) = 0;
virtual void set(const Dims &element, double value) = 0;
TensorAccess(TensorAccess const &) = delete;
TensorAccess &operator=(TensorAccess const &) = delete;
};
class Tensor {
public:
Tensor &operator=(Tensor const &) = delete;
Tensor(Tensor const &) = delete;
struct Stats {
double min;
double max;
double mean;
double stddev;
};
enum class DataType {
U8,
HALF,
FLOAT,
INT64,
I32,
I16,
};
static size_t DataTypeSize(DataType dt);
static const char *DataTypeStr(DataType dt);
virtual ~Tensor(){};
virtual std::string info() const;
// Make this const?
virtual std::unique_ptr<TensorAccess> access() { return nullptr; }
virtual std::shared_ptr<Tensor> slice(const Dims &offset, const Dims &size);
virtual std::shared_ptr<Tensor> grad(bool create = true);
virtual std::shared_ptr<Tensor> value() const;
virtual void copyFrom(Tensor &t);
double sse(Tensor &t);
static std::shared_ptr<Tensor> loadProtoBuf(const char *path);
static std::shared_ptr<Tensor> load(
const char *path, const std::optional<const std::string> &name);
bool save(const char *path);
static std::shared_ptr<Tensor> find(
Tensor::DataType data_type, const Dims &size, double init_mean,
double init_stddev, Tensors &named_tensors,
const std::optional<const std::string> &name = std::nullopt);
static std::shared_ptr<Tensor> make(Tensor::DataType data_type,
const Dims &size, double init_mean,
double init_stddev);
// Info / Debug / etc
std::shared_ptr<Tensor> toRGB(
std::optional<std::pair<float, float>> range = std::nullopt);
void print(const char *prefix, int elements_per_rank = 0);
void print_anomaly(const char *prefix);
void printRGB(const char *prefix,
std::optional<std::pair<float, float>> range = std::nullopt);
virtual Stats stats();
void printStats(const char *prefix);
std::string statsString(void);
std::optional<const std::string> namePostfix(
const std::string &postfix) const;
const std::optional<const std::string> m_name;
const DataType m_data_type;
const Dims m_dims;
protected:
Tensor(DataType data_type, const Dims &size,
const std::optional<const std::string> &name = std::nullopt);
};
std::shared_ptr<Tensor> makeCPUTensor(
Tensor::DataType data_type, const Dims &size,
const std::optional<const std::string> &name = std::nullopt);
std::shared_ptr<Tensor> makeTensor(
Tensor::DataType data_type, const Dims &size,
const std::optional<const std::string> &name = std::nullopt);
class Tensors
: public std::unordered_map<std::string, std::shared_ptr<Tensor>> {
public:
using std::unordered_map<std::string,
std::shared_ptr<Tensor>>::unordered_map;
bool has(const std::string &n) const;
const std::shared_ptr<Tensor> operator[](const std::string &n) const;
std::shared_ptr<Tensor> &operator[](const std::string &n);
std::vector<std::shared_ptr<Tensor>> getv(const std::string &n) const
{
std::vector<std::shared_ptr<Tensor>> v;
for(int i = 0;; i++) {
auto it = find(n + std::to_string(i));
if(it == end())
break;
v.push_back(it->second);
}
return v;
}
void loadTensors(const char *path);
bool saveTensors(const char *path, Context *ctx);
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
typedef std::function<size_t(long batch, int n, uint8_t *data, size_t capacity)>
Loader;
class Node {
public:
Node(const std::string &type,
const std::optional<const std::string> &name = std::nullopt)
: m_type(type), m_name(name)
{
}
Node(const Node &n)
: m_type(n.m_type)
, m_name(n.m_name)
, m_inputs(n.m_inputs)
, m_attributes(n.m_attributes)
, m_outputs(n.m_outputs)
{
}
const std::string m_type;
const std::optional<const std::string> m_name;
Tensors m_inputs;
Attributes m_attributes;
Tensors m_outputs;
Loader m_loader;
std::shared_ptr<Tensor> inferTensor_y(
const std::optional<const std::string> &name = std::nullopt);
void print() const;
static std::vector<std::shared_ptr<Node>> make(
const std::string &type, const Tensors &inputs,
const Attributes &attributes, Tensors &named_tensors,
const std::optional<const std::string> &name = std::nullopt);
static std::vector<std::shared_ptr<Node>> make(
const std::string &type, Loader loader, const Attributes &attributes);
std::shared_ptr<Tensor> y() const;
};
class Nodes : public std::vector<std::shared_ptr<Node>> {
public:
using std::vector<std::shared_ptr<Node>>::vector;
iterator findSingleDownStreamNode(std::shared_ptr<Tensor> t,
const std::string &type);
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
typedef std::unordered_map<
std::shared_ptr<Tensor>,
std::vector<std::pair<std::string, std::shared_ptr<Node>>>>
TensorMapping;
class Graph {
public:
Graph() : m_named_tensors(std::make_shared<Tensors>()) {}
Graph(std::shared_ptr<Tensors> named_tensors)
: m_named_tensors(named_tensors)
{
}
Nodes m_nodes;
std::unordered_set<std::shared_ptr<Tensor>> m_inputs;
std::unordered_set<std::shared_ptr<Tensor>> m_outputs;
const std::shared_ptr<Tensors> m_named_tensors;
std::shared_ptr<Node> addNode(
const std::string &type, const std::shared_ptr<Tensor> &t,
const Attributes &attributes = {},
const std::optional<const std::string> &name = std::nullopt);
std::shared_ptr<Node> addNode(
const std::string &type, const std::shared_ptr<Node> &n,
const Attributes &attributes = {},
const std::optional<const std::string> &name = std::nullopt);
std::shared_ptr<Node> addNode(
const std::string &type, const Tensors &inputs,
const Attributes &attributes = {},
const std::optional<const std::string> &name = std::nullopt);
std::shared_ptr<Node> addNode(const std::string &type, Loader loader,
const Attributes &attributes);
static std::shared_ptr<Graph> load(const char *path);
void loadTensors(const char *path);
bool saveTensors(const char *path, Context *p);
void print() const;
void statsTensors(Context *p);
std::pair<TensorMapping, TensorMapping> tensorMappings() const;
std::unordered_set<std::shared_ptr<Tensor>> inputTensors() const;
std::unordered_set<std::shared_ptr<Tensor>> outputTensors() const;
// Build helpers
std::shared_ptr<Node> addJpegDecoder(int width, int height,
Tensor::DataType output_data_type,
Loader loader);
std::shared_ptr<Node> addConvert(std::shared_ptr<Tensor> input,
Tensor::DataType dt, float scale);
std::shared_ptr<Node> addSpatialTransform(
std::shared_ptr<Tensor> input, std::shared_ptr<Tensor> theta,
int output_width = -1, int output_height = -1,
bool also_during_inference = false);
std::shared_ptr<Node> addResNetBottleNeck(std::shared_ptr<Tensor> input,
int squeeze_activations,
int output_activations,
bool downsample,
const std::string &name);
std::shared_ptr<Node> addResNetBottleNeck_transposed(
std::shared_ptr<Tensor> input, int squeeze_activations,
int output_activations, bool upsample, const std::string &name);
std::shared_ptr<Node> addResNet(std::shared_ptr<Tensor> input,
int output_activations, bool downsample,
const std::string &name);
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
struct UI {
virtual ~UI() {}
enum class Align {
LEFT,
CENTER,
RIGHT,
};
enum Page { SYS = 0, DATA = 100, CTX = 200, USER = 300 };
virtual void updateCell(Page page, size_t row, size_t column, Align a,
const char *fmt, ...)
__attribute__((format(printf, 6, 7))) = 0;
virtual void refresh() = 0;
};
std::shared_ptr<UI> make_tui();
std::shared_ptr<UI> make_nui();
//------------------------------------------------------------------------
//------------------------------------------------------------------------
enum class ProgramType {
INFERENCE = 0x1,
TRAINING = 0x2,
};
inline constexpr ProgramType
operator|(ProgramType a, ProgramType b)
{
return static_cast<ProgramType>(static_cast<int>(a) | static_cast<int>(b));
}
inline constexpr ProgramType
operator&(ProgramType a, ProgramType b)
{
return static_cast<ProgramType>(static_cast<int>(a) & static_cast<int>(b));
}
inline constexpr bool
operator!(ProgramType a)
{
return static_cast<bool>(!static_cast<int>(a));
}
enum class Phase { PRE = 0x1, POST = 0x2 };
inline constexpr Phase
operator|(Phase a, Phase b)
{
return static_cast<Phase>(static_cast<int>(a) | static_cast<int>(b));
}
inline constexpr Phase
operator&(Phase a, Phase b)
{
return static_cast<Phase>(static_cast<int>(a) & static_cast<int>(b));
}
inline constexpr bool
operator!(Phase a)
{
return static_cast<bool>(!static_cast<int>(a));
}
typedef std::unordered_map<std::shared_ptr<Tensor>, Phase> BatchedTensors;
class Program;
typedef std::function<void(
long batch, const Program &p,
std::unordered_map<std::shared_ptr<Tensor>, TensorAccess *>)>
TensorBatchCallback;
typedef std::function<bool(void)> StopCheck;
struct ProgramSource {
const Graph &graph;
BatchedTensors batched_tensors;
int batch_size{1};
};
struct ProgramConfig {
TensorBatchCallback pre_ops;
TensorBatchCallback post_ops;
float learning_rate{1e-4};
// L2 regularization term
float l2_lambda{0};
float bn_expavg{0.1};
TensorLayout tensor_layout{TensorLayout::Auto};
// Scan tensors for NAN values
// Caution: Makes everything slower.
// Only use for debug
bool anomaly_detect{false};
// Useful for debug
bool disable_op_fusing{false};
};
enum class ExecResult { OK, STOPPED };
class Program {
public:
virtual ~Program() {}
virtual void finalize() = 0;
virtual ExecResult run(long batches = 1, StopCheck stop_check = nullptr,
long batch_offset = 0) = 0;
virtual void dump(FILE *output, bool detailed = false) = 0;
virtual void debug(bool on) = 0;
virtual bool dumpGraph(const char *path) { return false; }
virtual ProgramType getType() const = 0;
virtual int getUiRow() const = 0;
virtual double getMPS() const = 0;
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
class Context {
public:
virtual ~Context() {}
virtual std::shared_ptr<Program> createMultiProgram(
const std::vector<ProgramSource> &sources, ProgramType pt,
const ProgramConfig &pc) = 0;
std::shared_ptr<Program> createProgram(const ProgramSource &ps,
ProgramType pt,
const ProgramConfig &pc)
{
return createMultiProgram({ps}, pt, pc);
}
virtual std::shared_ptr<Tensor> resolveTensor(
std::shared_ptr<Tensor> t) = 0;
virtual int getId() const = 0;
virtual UI::Page getUiPage() const = 0;
virtual void reset() = 0;
virtual void bindToHostThread() = 0;
};
//------------------------------------------------------------------------
//------------------------------------------------------------------------
class Engine {
public:
virtual ~Engine() {}
virtual std::vector<std::shared_ptr<Context>> createContexts(
bool multi = false) = 0;
};
std::shared_ptr<Engine> createEngine(const std::shared_ptr<UI> &ui);
std::shared_ptr<Context> createContext(const std::shared_ptr<UI> &ui = nullptr);
//------------------------------------------------------------------------
//------------------------------------------------------------------------
class Publisher {
public:
virtual ~Publisher(){};
virtual void publish(const char *id, Tensor &t, const Dims &offset,
TensorAccess *ta) = 0;
virtual void sync() = 0;
};
std::shared_ptr<Publisher> make_tcp_publisher(int bind_port);
//------------------------------------------------------------------------
//------------------------------------------------------------------------
int64_t now();
std::string fmt(const char *fmt, ...) __attribute__((format(printf, 1, 2)));
struct Barrier {
virtual ~Barrier(){};
virtual void wait() = 0;
static std::shared_ptr<Barrier> make(size_t count);
};
}; // namespace saga