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
|
#include "Halide.h"
#include "halide_benchmark.h"
#include <cstdio>
#include <functional>
#include <thread>
/** \file Test to demonstrate using JIT across multiple threads with
* varying parameters passed to realizations. Performance is tested
* by comparing a technique that recompiles vs one that should not.
*/
using namespace Halide;
using namespace Halide::Tools;
struct test_func {
Param<int32_t> p;
ImageParam in{Int(32), 1};
Func f;
Var x;
test_func() {
Expr big = 0;
for (int i = 0; i < 75; i++) {
big += p;
}
Func inner;
inner(x) = x * in(clamp(x, 0, 9)) + big;
f(x) = inner(x - 1) + inner(x) + inner(x + 1);
inner.compute_at(f, x);
}
};
// The Halide compiler is currently not guaranteed to be thread safe.
std::mutex compiler_mutex;
Buffer<int32_t> bufs[16];
void separate_func_per_thread_executor(int index) {
test_func test;
{
std::lock_guard<std::mutex> lock(compiler_mutex);
test.f.compile_jit();
}
test.p.set(index);
test.in.set(bufs[index]);
for (int i = 0; i < 10; i++) {
Buffer<int32_t> result = test.f.realize({10});
for (int j = 0; j < 10; j++) {
int64_t left = ((j - 1) * (int64_t)bufs[index](std::min(std::max(0, j - 1), 9)) + index * 75);
int64_t middle = (j * (int64_t)bufs[index](std::min(std::max(0, j), 9)) + index * 75);
int64_t right = ((j + 1) * (int64_t)bufs[index](std::min(std::max(0, j + 1), 9)) + index * 75);
assert(result(j) == (int32_t)(left + middle + right));
}
}
}
void separate_func_per_thread() {
std::thread threads[16];
for (auto &thread : threads) {
thread = std::thread(separate_func_per_thread_executor,
(int)(&thread - threads));
}
for (auto &thread : threads) {
thread.join();
}
}
void same_func_per_thread_executor(int index, test_func &test) {
for (int i = 0; i < 10; i++) {
Buffer<int32_t> result = test.f.realize({10}, get_jit_target_from_environment(),
{{test.p, index},
{test.in, bufs[index]}});
for (int j = 0; j < 10; j++) {
int64_t left = ((j - 1) * (int64_t)bufs[index](std::min(std::max(0, j - 1), 9)) + index * 75);
int64_t middle = (j * (int64_t)bufs[index](std::min(std::max(0, j), 9)) + index * 75);
int64_t right = ((j + 1) * (int64_t)bufs[index](std::min(std::max(0, j + 1), 9)) + index * 75);
assert(result(j) == (int32_t)(left + middle + right));
}
}
}
void same_func_per_thread() {
std::thread threads[16];
test_func test;
// In this program, only one thread can call into the compiler
// at this point. The mutex guard is still included both to show that in
// general Halide compilation is not thread ssafe and also to keep
// the performance comparison slightly more equal by including
// (minimal) mutex cost on both paths.
{
std::lock_guard<std::mutex> lock(compiler_mutex);
test.f.compile_jit();
}
for (auto &thread : threads) {
thread = std::thread(same_func_per_thread_executor,
(int)(&thread - threads), std::ref(test));
}
for (auto &thread : threads) {
thread.join();
}
}
int main(int argc, char **argv) {
Target target = get_jit_target_from_environment();
if (target.arch == Target::WebAssembly) {
printf("[SKIP] Performance tests are meaningless and/or misleading under WebAssembly interpreter.\n");
return 0;
}
for (auto &buf : bufs) {
buf = Buffer<int32_t>(10);
for (int i = 0; i < 10; i++) {
buf(i) = std::rand();
}
}
double separate_time = benchmark(separate_func_per_thread);
printf("Separate compilations time: %fs.\n", separate_time);
double same_time = benchmark(same_func_per_thread);
printf("One compilation time: %fs.\n", same_time);
assert(same_time < separate_time);
printf("Success!\n");
return 0;
}
|
