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
|
#include "Halide.h"
#include <algorithm>
#include "halide_benchmark.h"
// This test makes sure GPU runtimes can handle many different small
// kernels and can handle releasing a device context and making a new
// one and still have many kernels work. This is needed due to kernel
// compilation caching mechanisms in the GPU runtimes.
using namespace Halide;
constexpr size_t kNumKernels = 70;
int main(int argc, char **argv) {
Var x, y, xi, yi;
Func adders[kNumKernels];
ImageParam input(Int(32), 2);
Target target = get_jit_target_from_environment();
int i = 1;
for (Func &f : adders) {
f(x, y) = input(x, y) + i;
if (target.has_gpu_feature()) {
f.compute_root().gpu_tile(x, y, xi, yi, 16, 16);
} else {
f.compute_root().vectorize(x, target.natural_vector_size<int32_t>());
}
i += 1;
}
auto start = Halide::Tools::benchmark_now();
Buffer<int32_t> buf_a_store(32, 32);
Buffer<int32_t> buf_b_store(32, 32);
Buffer<int32_t> *buf_in = &buf_a_store;
Buffer<int32_t> *buf_out = &buf_b_store;
buf_in->fill(0);
for (Func &f : adders) {
input.set(*buf_in);
f.realize(*buf_out);
std::swap(buf_in, buf_out);
}
buf_in->copy_to_host();
auto end = Halide::Tools::benchmark_now();
double initial_runtime = Halide::Tools::benchmark_duration_seconds(start, end);
buf_in->for_each_value([](int32_t x) { assert(x == (kNumKernels * (kNumKernels + 1)) / 2); });
start = Halide::Tools::benchmark_now();
buf_in->fill(0);
for (Func &f : adders) {
input.set(*buf_in);
f.realize(*buf_out);
std::swap(buf_in, buf_out);
}
buf_in->copy_to_host();
end = Halide::Tools::benchmark_now();
double precompiled_runtime = Halide::Tools::benchmark_duration_seconds(start, end);
buf_in->for_each_value([](int32_t x) { assert(x == (kNumKernels * (kNumKernels + 1)) / 2); });
buf_a_store.device_free();
buf_b_store.device_free();
const halide_device_interface_t *device = get_device_interface_for_device_api(DeviceAPI::Default_GPU, target);
if (device != nullptr) {
device->device_release(nullptr, device);
}
start = Halide::Tools::benchmark_now();
buf_in->fill(0);
for (Func &f : adders) {
input.set(*buf_in);
f.realize(*buf_out);
std::swap(buf_in, buf_out);
}
buf_in->copy_to_host();
end = Halide::Tools::benchmark_now();
double second_runtime = Halide::Tools::benchmark_duration_seconds(start, end);
buf_in->for_each_value([](int32_t x) { assert(x == (kNumKernels * (kNumKernels + 1)) / 2); });
printf("Initial runtime %f, precompiled runtime %f, second runtime %f.\n", initial_runtime, precompiled_runtime, second_runtime);
printf("Success!\n");
return 0;
}
|
