Debian Codegraph @

libgomp/testsuite/libgomp.c-c++-common/interop-2.c

raw
1 /* { dg-do run } */
2 /* { dg-additional-options "-lm" } */
3
4 /* Note: At the time this program was written, Nvptx was not asynchronous
5 enough to trigger the issue (with a 'nowait' added); however, one
6 AMD GPUs, it triggered. */
7
8 /* Test whether nowait / dependency is handled correctly.
9 Motivated by OpenMP_VV's 5.1/interop/test_interop_target.c
10
11 The code actually only creates a streaming object without actually using it,
12 except for dependency tracking.
13
14 Note that there is a difference between having a steaming (targetsync) object
15 and not (= omp_interop_none); at least if one assumes that omp_interop_none
16 does not include 'targetsync' as (effective) interop type - in that case,
17 'nowait' has no effect and the 'depend' is active as included task, otherwise
18 the code continues with the depend being active only for the about to be
19 destroyed or used thread.
20
21 The OpenMP spec states (here 6.0):
22 "If the interop-type set includes 'targetsync', an empty mergeable task is
23 generated. If the 'nowait' clause is not present on the construct then
24 the task is also an included task. If the interop-type set does not
25 include 'targetsync', the 'nowait' clause has no effect. Any depend
26 clauses that are present on the construct apply to the generated task. */
27
28 #include <omp.h>
29
30 void
31 test_async (const int dev)
32 {
33 constexpr int N = 2048;
34 constexpr int ulp = 4;
35 constexpr double M_PI = 2.0 * __builtin_acos (0.0);
36 omp_interop_t obj1, obj2;
37 double A[N] = { };
38 int B[N] = { };
39
40 /* Create interop object. */
41 #pragma omp interop device(dev) init(targetsync : obj1, obj2)
42
43 if (dev == omp_initial_device || dev == omp_get_num_devices ())
44 {
45 if (obj1 != omp_interop_none || obj2 != omp_interop_none)
46 __builtin_abort ();
47 }
48 else
49 {
50 if (obj1 == omp_interop_none || obj2 == omp_interop_none)
51 __builtin_abort ();
52 }
53
54 /* DOUBLE */
55
56 /* Now in the background update it, slowly enough that the
57 code afterwards is reached while still running asynchronously.
58 As OpenMP_VV's Issue #863 shows, the overhead is high enough to
59 fail even when only doing an atomic integer increment. */
60
61 #pragma omp target device(dev) map(A) depend(out: A[:N]) nowait
62 for (int i = 0; i < N; i++)
63 #pragma omp atomic update
64 A[i] += __builtin_sin (2*i*M_PI/N);
65
66 /* DESTROY take care of the dependeny such that ... */
67
68 if (obj1 == omp_interop_none)
69 {
70 // Same as below as 'nowait' is ignored.
71 #pragma omp interop destroy(obj1) depend(in: A[:N]) nowait
72 }
73 else
74 {
75 #pragma omp interop destroy(obj1) depend(in: A[:N])
76 }
77
78 /* ... this code is only executed once the dependency as been fulfilled. */
79
80 /* Check the value - part I: quick, avoid A[0] == sin(0) = 0. */
81 for (int i = 1; i < N; i++)
82 if (A[i] == 0.0)
83 __builtin_abort ();
84
85 /* Check the value - part II: throughly */
86 for (int i = 0; i < N; i++)
87 {
88 double x = A[i];
89 double y = __builtin_sin (2*i*M_PI/N);
90 if (__builtin_fabs (x - y) > ulp * __builtin_fabs (x+y) * __DBL_EPSILON__)
91 __builtin_abort ();
92 }
93
94 /* Integer */
95
96 #pragma omp target device(dev) map(B) depend(out: B[:N]) nowait
97 for (int i = 0; i < N; i++)
98 #pragma omp atomic update
99 B[i] += 42;
100
101 /* Same - but using USE. */
102 if (obj2 == omp_interop_none)
103 {
104 // Same as below as 'nowait' is ignored.
105 #pragma omp interop use(obj2) depend(in: B[:N]) nowait
106 }
107 else
108 {
109 #pragma omp interop use(obj2) depend(in: B[:N])
110 }
111
112 for (int i = 0; i < N; i++)
113 if (B[i] != 42)
114 __builtin_abort ();
115
116 #pragma omp interop destroy(obj2)
117 }
118
119 int
120 main ()
121 {
122 int ndev = omp_get_num_devices ();
123
124 for (int dev = 0; dev <= ndev; dev++)
125 test_async (dev);
126 test_async (omp_initial_device);
127
128 return 0;
129 }