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
|
#if GOOGLE_CUDA
#include "tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.h"
#include <algorithm>
#include <vector>
#include "tensorflow/stream_executor/stream_executor.h"
#include <gtest/gtest.h>
#include "tensorflow/core/common_runtime/gpu/gpu_init.h"
#include "tensorflow/core/lib/gtl/inlined_vector.h"
#include "tensorflow/core/lib/random/simple_philox.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/port.h"
namespace gpu = ::perftools::gputools;
namespace tensorflow {
namespace {
TEST(GPUBFCAllocatorTest, NoDups) {
GPUBFCAllocator a(0, 1 << 30);
// Allocate a lot of raw pointers
std::vector<void*> ptrs;
for (int s = 1; s < 1024; s++) {
void* raw = a.AllocateRaw(1, s);
ptrs.push_back(raw);
}
std::sort(ptrs.begin(), ptrs.end());
// Make sure none of them are equal, and that none of them overlap.
for (int i = 0; i < ptrs.size(); i++) {
if (i > 0) {
ASSERT_NE(ptrs[i], ptrs[i - 1]); // No dups
size_t req_size = a.RequestedSize(ptrs[i - 1]);
ASSERT_GT(req_size, 0);
ASSERT_GE(static_cast<char*>(ptrs[i]) - static_cast<char*>(ptrs[i - 1]),
req_size);
}
}
for (int i = 0; i < ptrs.size(); i++) {
a.DeallocateRaw(ptrs[i]);
}
}
TEST(GPUBFCAllocatorTest, AllocationsAndDeallocations) {
GPUBFCAllocator a(0, 1 << 30);
// Allocate 256 raw pointers of sizes between 100 bytes and about
// a meg
random::PhiloxRandom philox(123, 17);
random::SimplePhilox rand(&philox);
std::vector<void*> initial_ptrs;
for (int s = 1; s < 256; s++) {
size_t size = std::min<size_t>(
std::max<size_t>(rand.Rand32() % 1048576, 100), 1048576);
void* raw = a.AllocateRaw(1, size);
initial_ptrs.push_back(raw);
}
// Deallocate half of the memory, and keep track of the others.
std::vector<void*> existing_ptrs;
for (int i = 0; i < initial_ptrs.size(); i++) {
if (i % 2 == 1) {
a.DeallocateRaw(initial_ptrs[i]);
} else {
existing_ptrs.push_back(initial_ptrs[i]);
}
}
// Allocate a lot of raw pointers
for (int s = 1; s < 256; s++) {
size_t size = std::min<size_t>(
std::max<size_t>(rand.Rand32() % 1048576, 100), 1048576);
void* raw = a.AllocateRaw(1, size);
existing_ptrs.push_back(raw);
}
std::sort(existing_ptrs.begin(), existing_ptrs.end());
// Make sure none of them are equal
for (int i = 0; i < existing_ptrs.size(); i++) {
if (i > 0) {
CHECK_NE(existing_ptrs[i], existing_ptrs[i - 1]); // No dups
size_t req_size = a.RequestedSize(existing_ptrs[i - 1]);
ASSERT_GT(req_size, 0);
// Check that they don't overlap.
ASSERT_GE(static_cast<char*>(existing_ptrs[i]) -
static_cast<char*>(existing_ptrs[i - 1]),
req_size);
}
}
for (int i = 0; i < existing_ptrs.size(); i++) {
a.DeallocateRaw(existing_ptrs[i]);
}
}
TEST(GPUBFCAllocatorTest, ExerciseCoalescing) {
GPUBFCAllocator a(0, 1 << 30);
float* first_ptr = a.Allocate<float>(1024);
a.Deallocate(first_ptr);
for (int i = 0; i < 1024; ++i) {
// Allocate several buffers of different sizes, and then clean them
// all up. We should be able to repeat this endlessly without
// causing fragmentation and growth.
float* t1 = a.Allocate<float>(1024);
int64* t2 = a.Allocate<int64>(1048576);
double* t3 = a.Allocate<double>(2048);
float* t4 = a.Allocate<float>(10485760);
a.Deallocate(t1);
a.Deallocate(t2);
a.Deallocate(t3);
a.Deallocate(t4);
}
// At the end, we should have coalesced all memory into one region
// starting at the beginning, so validate that allocating a pointer
// starts from this region.
float* first_ptr_after = a.Allocate<float>(1024);
EXPECT_EQ(first_ptr, first_ptr_after);
a.Deallocate(first_ptr_after);
}
TEST(GPUBFCAllocatorTest, AllocateZeroBufSize) {
GPUBFCAllocator a(0, 1 << 30);
float* ptr = a.Allocate<float>(0);
EXPECT_EQ(nullptr, ptr);
}
TEST(GPUBFCAllocatorTest, TracksSizes) {
GPUBFCAllocator a(0, 1 << 30);
EXPECT_EQ(true, a.TracksAllocationSizes());
}
TEST(GPUBFCAllocatorTest, AllocatedVsRequested) {
GPUBFCAllocator a(0, 1 << 30);
float* t1 = a.Allocate<float>(1);
EXPECT_EQ(4, a.RequestedSize(t1));
EXPECT_EQ(256, a.AllocatedSize(t1));
a.Deallocate(t1);
}
TEST(GPUBFCAllocatorTest, TestCustomMemoryLimit) {
// Configure a 1MiB byte limit
GPUBFCAllocator a(0, 1 << 20);
float* first_ptr = a.Allocate<float>(1 << 6);
float* second_ptr = a.Allocate<float>(1 << 20);
EXPECT_NE(nullptr, first_ptr);
EXPECT_EQ(nullptr, second_ptr);
a.Deallocate(first_ptr);
}
} // namespace
} // namespace tensorflow
#endif // GOOGLE_CUDA
|
