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
|
#include "Halide.h"
#include "halide_benchmark.h"
using namespace Halide;
int fib(int N, int a, int b) {
while (N > 2) {
a += b;
std::swap(a, b);
N--;
}
return b;
}
int main(int argc, char **argv) {
Target target = get_jit_target_from_environment();
if (!target.has_gpu_feature()) {
printf("[SKIP] No GPU target enabled.\n");
return 0;
}
if (target.has_feature(Target::D3D12Compute)) {
// https://github.com/halide/Halide/issues/5000
printf("[SKIP] Allocation cache not yet implemented for D3D12Compute.\n");
return 0;
}
const int N = 30;
Var x, y, xi, yi;
// Fixed size, overlapping lifetimes, looped 300 times. Should have 3 allocations live and OOM if there's a leak.
Func f1[N];
f1[0](x, y) = 1.0f;
f1[0].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
f1[1](x, y) = 2.0f;
f1[1].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
for (int i = 2; i < N; i++) {
f1[i](x, y) = f1[i - 1](x, y) + f1[i - 2](x, y);
f1[i].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
}
// Decreasing size, overlapping lifetimes, looped 300 times. Should OOM on leak.
Func f2[N];
f2[0](x, y) = 3.0f;
f2[0].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
f2[1](x, y) = 4.0f;
f2[1].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
for (int i = 2; i < N; i++) {
f2[i](x, y) = f2[i - 1](x + 1, y) + f2[i - 2](x, y);
f2[i].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
}
Func f3[N];
f3[0](x, y) = 5.0f;
f3[0].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
f3[1](x, y) = 6.0f;
f3[1].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
for (int i = 2; i < N; i++) {
f3[i](x, y) = f3[i - 1](x, clamp(y, 0, i)) + f3[i - 2](x, clamp(y, 0, i));
f3[i].compute_root().gpu_tile(x, y, xi, yi, 8, 8);
}
float correct1 = fib(N, 1, 2), correct2 = fib(N, 3, 4), correct3 = fib(N, 5, 6);
auto test1 = [&](bool use_cache, bool validate = true) {
Halide::Internal::JITSharedRuntime::reuse_device_allocations(use_cache);
for (int i = 0; i < 300; i++) {
Buffer<float> result = f1[N - 1].realize({128, 128});
if (validate) {
result.copy_to_host();
result.for_each_value([=](float f) {
if (f != correct1) {
printf("result is %f instead of %f\n", f, correct1);
abort();
}
});
} else {
result.device_sync();
}
}
// We don't want the cache to persist across these tests
Halide::Internal::JITSharedRuntime::reuse_device_allocations(false);
};
auto test2 = [&](bool use_cache, bool validate = true) {
Halide::Internal::JITSharedRuntime::reuse_device_allocations(use_cache);
for (int i = 0; i < 300; i++) {
Buffer<float> result = f2[N - 1].realize({128, 128});
if (validate) {
result.copy_to_host();
result.for_each_value([=](float f) {
if (f != correct2) {
printf("result is %f instead of %f\n", f, correct2);
abort();
}
});
} else {
result.device_sync();
}
}
// We don't want the cache to persist across these tests
Halide::Internal::JITSharedRuntime::reuse_device_allocations(false);
};
auto test3 = [&](bool use_cache, bool validate = true) {
Halide::Internal::JITSharedRuntime::reuse_device_allocations(use_cache);
// Increasing size, overlapping lifetimes, looped 300 times. Should OOM on leak.
for (int i = 0; i < 300; i++) {
Buffer<float> result = f3[N - 1].realize({128, 128});
if (validate) {
result.copy_to_host();
result.for_each_value([=](float f) {
if (f != correct3) {
printf("result is %f instead of %f\n", f, correct3);
abort();
}
});
} else {
result.device_sync();
}
}
// We don't want the cache to persist across these tests
Halide::Internal::JITSharedRuntime::reuse_device_allocations(false);
};
// First run them serially (compilation of a Func isn't thread-safe).
//test1(true);
//test2(true);
//test3(true);
//return 0;
// Now run all at the same time to check for concurrency issues.
// FIXME: Skipping OpenGLCompute, which has concurrency
// issues. Probably due to using the GL context on the wrong
// thread.
if (!target.has_feature(Target::OpenGLCompute)) {
Halide::Internal::ThreadPool<void> pool(1);
std::vector<std::future<void>> futures;
futures.emplace_back(pool.async(test1, true));
futures.emplace_back(pool.async(test1, true));
futures.emplace_back(pool.async(test2, true));
futures.emplace_back(pool.async(test2, true));
futures.emplace_back(pool.async(test3, true));
futures.emplace_back(pool.async(test3, true));
for (auto &f : futures) {
f.get();
}
}
// Now benchmark with and without, (just informational, as this isn't a performance test)
double t1 = Tools::benchmark([&]() {
test1(true, false);
test2(true, false);
test3(true, false);
});
double t2 = Tools::benchmark([&]() {
test1(false, false);
test2(false, false);
test3(false, false);
});
printf("Runtime with cache: %f\n"
"Without cache: %f\n",
t1, t2);
printf("Success!\n");
return 0;
}
|
