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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
|
#include "Halide.h"
using namespace Halide;
using namespace Halide::Internal;
class Checker : public IRMutator {
Stmt visit(const Atomic *op) override {
count_atomics++;
if (!op->mutex_name.empty()) {
count_atomics_with_mutexes++;
}
return IRMutator::visit(op);
}
public:
int count_atomics = 0;
int count_atomics_with_mutexes = 0;
};
int main(int argc, char **argv) {
if (get_jit_target_from_environment().arch == Target::WebAssembly) {
printf("[SKIP] Skipping test for WebAssembly as it does not support atomics yet.\n");
return 0;
}
{
Func f, g;
Var x, y;
f(x, y) = {x, y};
f(x, y) = {f(x, y)[0] + x,
f(x, y)[1] + y};
// The summation is independent in the two tuple components,
// so it can just be two atomic add instructions. No CAS loop
// required.
f.compute_root().update().parallel(y).atomic();
g(x, y) = f(x, y)[0] + f(x, y)[1];
Checker checker;
g.add_custom_lowering_pass(&checker, []() {});
Buffer<int> out = g.realize({128, 128});
for (int y = 0; y < 128; y++) {
for (int x = 0; x < 128; x++) {
int correct = 2 * x + 2 * y;
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
return 1;
}
}
}
if (checker.count_atomics != 2 || checker.count_atomics_with_mutexes != 0) {
printf("Expected two atomic nodes, neither of them with mutexes\n");
return 1;
}
}
{
Func f, g;
Var x, y;
f(x, y) = {x, y};
f(x, y) = {f(x, y)[1] + x,
f(x, y)[0] + y};
// The summation is coupled across the two tuple components
// and there are two stores, so we need a mutex.
f.compute_root().update().parallel(y).atomic();
g(x, y) = f(x, y)[0] + f(x, y)[1];
Checker checker;
g.add_custom_lowering_pass(&checker, []() {});
Buffer<int> out = g.realize({128, 128});
for (int y = 0; y < 128; y++) {
for (int x = 0; x < 128; x++) {
int correct = 2 * x + 2 * y;
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
return 1;
}
}
}
if (checker.count_atomics != 1 || checker.count_atomics_with_mutexes != 1) {
printf("Expected one atomic node, with mutex\n");
return 1;
}
}
{
Func f, g;
Var x, y;
f(x, y) = {x, y, 0};
f(x, y) = {f(x, y)[1] + x,
f(x, y)[0] + y,
f(x, y)[2] + 1};
// The summation is coupled across the first two tuple
// components and there are two stores, so we need a mutex
// there. The last store could in principle be a separate atomic
// add, but we instead just pack it into the critical section.
f.compute_root().update().parallel(y).atomic();
g(x, y) = f(x, y)[0] + f(x, y)[1] + f(x, y)[2];
Checker checker;
g.add_custom_lowering_pass(&checker, []() {});
Buffer<int> out = g.realize({128, 128});
for (int y = 0; y < 128; y++) {
for (int x = 0; x < 128; x++) {
int correct = 2 * x + 2 * y + 1;
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
return 1;
}
}
}
if (checker.count_atomics != 1 || checker.count_atomics_with_mutexes != 1) {
printf("Expected one atomic nodes, with mutex\n");
return 1;
}
}
{
Func f, g;
Var x, y;
f(x, y) = {x, y, x, y};
f(x, y) = {f(x, y)[1] + x,
f(x, y)[0] + y,
f(x, y)[3] + x,
f(x, y)[2] + y};
// The summation is coupled across the first two tuple
// components and the last two components, but they're
// independent so they *could* get two critical sections, but
// it would be on the same mutex, so we just pack them all
// into one critical section.
f.compute_root().update().parallel(y).atomic();
g(x, y) = f(x, y)[0] + f(x, y)[1] + f(x, y)[2] + f(x, y)[3];
Checker checker;
g.add_custom_lowering_pass(&checker, []() {});
Buffer<int> out = g.realize({128, 128});
for (int y = 0; y < 128; y++) {
for (int x = 0; x < 128; x++) {
int correct = 4 * x + 4 * y;
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
return 1;
}
}
}
if (checker.count_atomics != 1 || checker.count_atomics_with_mutexes != 1) {
printf("Expected one atomic node, with mutex\n");
return 1;
}
}
{
Func f, g;
Var x, y;
f(x, y) = {x, y};
RDom r(0, 65);
// Update even rows
f(x, r * 2) = {f(x, r * 2 + 1)[1] + x,
f(x, r * 2 - 1)[0] + r * 2};
// Update odd rows using even rows
f(x, r * 2 + 1) = {f(x, r * 2)[1] + x,
f(x, r * 2 + 2)[0] + r * 2 + 1};
// The tuple components have cross-talk, but the loads
// couldn't possibly alias with the stores because of the
// even/odd split. We can just use four atomic adds safely.
f.compute_root();
f.update(0)
.parallel(r)
.atomic();
f.update(1)
.parallel(r)
.atomic();
g(x, y) = f(x, y)[0] + f(x, y)[1];
Checker checker;
g.add_custom_lowering_pass(&checker, []() {});
Buffer<int> out = g.realize({128, 128});
for (int y = 0; y < 128; y++) {
for (int x = 0; x < 128; x++) {
int correct = 2 * x + 2 * y + 1;
if (y & 1) {
// The odd rows happen after the even rows, so
// they get another dose of x + y.
correct += x + y;
}
if (out(x, y) != correct) {
printf("out(%d, %d) = %d instead of %d\n",
x, y, out(x, y), correct);
// return 1;
}
}
}
if (checker.count_atomics != 4 || checker.count_atomics_with_mutexes != 0) {
printf("Expected four atomic nodes, with no mutexes\n");
return 1;
}
}
printf("Success!\n");
return 0;
}
|
