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
|
// TODO(vrv): Switch this to an open-sourced version of Arena.
#ifndef TENSORFLOW_LIB_CORE_ARENA_H_
#define TENSORFLOW_LIB_CORE_ARENA_H_
#include <assert.h>
#include <vector>
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/port.h"
namespace tensorflow {
namespace core {
// This class is "thread-compatible": different threads can access the
// arena at the same time without locking, as long as they use only
// const methods.
class Arena {
public:
// Allocates a thread-compatible arena with the specified block size.
explicit Arena(const size_t block_size);
~Arena();
char* Alloc(const size_t size) {
return reinterpret_cast<char*>(GetMemory(size, 1));
}
void Reset();
// This should be the worst-case alignment for any type. This is
// good for IA-32, SPARC version 7 (the last one I know), and
// supposedly Alpha. i386 would be more time-efficient with a
// default alignment of 8, but ::operator new() uses alignment of 4,
// and an assertion will fail below after the call to MakeNewBlock()
// if you try to use a larger alignment.
#ifdef __i386__
static const int kDefaultAlignment = 4;
#else
static const int kDefaultAlignment = 8;
#endif
protected:
bool SatisfyAlignment(const size_t alignment);
void MakeNewBlock(const uint32 alignment);
void* GetMemoryFallback(const size_t size, const int align);
void* GetMemory(const size_t size, const int align) {
assert(remaining_ <= block_size_); // an invariant
if (size > 0 && size < remaining_ && align == 1) { // common case
void* result = freestart_;
freestart_ += size;
remaining_ -= size;
return result;
}
return GetMemoryFallback(size, align);
}
size_t remaining_;
private:
struct AllocatedBlock {
char* mem;
size_t size;
};
// Allocate new new block of at least block_size, with the specified
// alignment.
// The returned AllocatedBlock* is valid until the next call to AllocNewBlock
// or Reset (i.e. anything that might affect overflow_blocks_).
AllocatedBlock* AllocNewBlock(const size_t block_size,
const uint32 alignment);
const size_t block_size_;
char* freestart_; // beginning of the free space in most recent block
char* freestart_when_empty_; // beginning of the free space when we're empty
// STL vector isn't as efficient as it could be, so we use an array at first
size_t blocks_alloced_; // how many of the first_blocks_ have been alloced
AllocatedBlock first_blocks_[16]; // the length of this array is arbitrary
// if the first_blocks_ aren't enough, expand into overflow_blocks_.
std::vector<AllocatedBlock>* overflow_blocks_;
void FreeBlocks(); // Frees all except first block
TF_DISALLOW_COPY_AND_ASSIGN(Arena);
};
} // namespace core
} // namespace tensorflow
#endif // TENSORFLOW_LIB_CORE_ARENA_H_
|
