1 | // RUN: %libomp-compile |
2 | // RUN: env OMP_SCHEDULE=guided %libomp-run |
3 | // RUN: env OMP_SCHEDULE=guided,1 %libomp-run 1 |
4 | // RUN: env OMP_SCHEDULE=guided,2 %libomp-run 2 |
5 | // RUN: env OMP_SCHEDULE=dynamic %libomp-run |
6 | // RUN: env OMP_SCHEDULE=dynamic,1 %libomp-run 1 |
7 | // RUN: env OMP_SCHEDULE=dynamic,2 %libomp-run 2 |
8 | // RUN: env OMP_SCHEDULE=auto %libomp-run |
9 | |
10 | // The test checks schedule(simd:runtime) |
11 | // in combination with OMP_SCHEDULE=guided[,chunk] |
12 | #include <stdio.h> |
13 | #include <stdlib.h> |
14 | #include <omp.h> |
15 | |
16 | #if defined(WIN32) || defined(_WIN32) |
17 | #include <windows.h> |
18 | #define delay() Sleep(1); |
19 | #define seten(a,b,c) _putenv_s((a),(b)) |
20 | #else |
21 | #include <unistd.h> |
22 | #define delay() usleep(10); |
23 | #define seten(a,b,c) setenv((a),(b),(c)) |
24 | #endif |
25 | |
26 | #define UBOUND 100 |
27 | #define SIMD_LEN 4 |
28 | int err = 0; |
29 | |
30 | // --------------------------------------------------------------------------- |
31 | // Various definitions copied from OpenMP RTL. |
32 | enum sched { |
33 | kmp_sch_static_balanced_chunked = 45, |
34 | kmp_sch_guided_simd = 46, |
35 | kmp_sch_runtime_simd = 47, |
36 | }; |
37 | typedef unsigned u32; |
38 | typedef long long i64; |
39 | typedef unsigned long long u64; |
40 | typedef struct { |
41 | int reserved_1; |
42 | int flags; |
43 | int reserved_2; |
44 | int reserved_3; |
45 | char *psource; |
46 | } id; |
47 | |
48 | #ifdef __cplusplus |
49 | extern "C" { |
50 | #endif |
51 | int __kmpc_global_thread_num(id*); |
52 | void __kmpc_barrier(id*, int gtid); |
53 | void __kmpc_dispatch_init_4(id*, int, enum sched, int, int, int, int); |
54 | void __kmpc_dispatch_init_8(id*, int, enum sched, i64, i64, i64, i64); |
55 | int __kmpc_dispatch_next_4(id*, int, void*, void*, void*, void*); |
56 | int __kmpc_dispatch_next_8(id*, int, void*, void*, void*, void*); |
57 | #ifdef __cplusplus |
58 | } // extern "C" |
59 | #endif |
60 | // End of definitions copied from OpenMP RTL. |
61 | // --------------------------------------------------------------------------- |
62 | static id loc = {.reserved_1: 0, .flags: 2, .reserved_2: 0, .reserved_3: 0, .psource: ";file;func;0;0;;" }; |
63 | |
64 | // --------------------------------------------------------------------------- |
65 | void |
66 | run_loop( |
67 | int loop_lb, // Loop lower bound. |
68 | int loop_ub, // Loop upper bound. |
69 | int loop_st, // Loop stride. |
70 | int lchunk |
71 | ) { |
72 | static int volatile loop_sync = 0; |
73 | int lb; // Chunk lower bound. |
74 | int ub; // Chunk upper bound. |
75 | int st; // Chunk stride. |
76 | int rc; |
77 | int nthreads = omp_get_num_threads(); |
78 | int tid = omp_get_thread_num(); |
79 | int gtid = __kmpc_global_thread_num(&loc); |
80 | int last; |
81 | int tc = (loop_ub - loop_lb) / loop_st + 1; |
82 | int ch; |
83 | int no_chunk = 0; |
84 | if (lchunk == 0) { |
85 | no_chunk = 1; |
86 | lchunk = 1; |
87 | } |
88 | ch = lchunk * SIMD_LEN; |
89 | #if _DEBUG > 1 |
90 | printf("run_loop gtid %d tid %d (lb=%d, ub=%d, st=%d, ch=%d)\n" , |
91 | gtid, tid, (int)loop_lb, (int)loop_ub, (int)loop_st, lchunk); |
92 | #endif |
93 | // Don't test degenerate cases that should have been discovered by codegen. |
94 | if (loop_st == 0) |
95 | return; |
96 | if (loop_st > 0 ? loop_lb > loop_ub : loop_lb < loop_ub) |
97 | return; |
98 | __kmpc_dispatch_init_4(&loc, gtid, kmp_sch_runtime_simd, |
99 | loop_lb, loop_ub, loop_st, SIMD_LEN); |
100 | { |
101 | // Let the master thread handle the chunks alone. |
102 | int chunk; // No of current chunk. |
103 | int last_ub; // Upper bound of the last processed chunk. |
104 | u64 cur; // Number of interations in current chunk. |
105 | u64 max; // Max allowed iterations for current chunk. |
106 | int undersized = 0; |
107 | last_ub = loop_ub; |
108 | chunk = 0; |
109 | max = (loop_ub - loop_lb) / loop_st + 1; |
110 | // The first chunk can consume all iterations. |
111 | while (__kmpc_dispatch_next_4(&loc, gtid, &last, &lb, &ub, &st)) { |
112 | ++ chunk; |
113 | #if _DEBUG |
114 | printf(format: "th %d: chunk=%d, lb=%d, ub=%d ch %d\n" , |
115 | tid, chunk, (int)lb, (int)ub, (int)(ub-lb+1)); |
116 | #endif |
117 | // Check if previous chunk (it is not the final chunk) is undersized. |
118 | if (undersized) |
119 | printf(format: "Error with chunk %d, th %d, err %d\n" , chunk, tid, ++err); |
120 | if (loop_st > 0) { |
121 | if (!(ub <= loop_ub)) |
122 | printf(format: "Error with ub %d, %d, ch %d, err %d\n" , |
123 | (int)ub, (int)loop_ub, chunk, ++err); |
124 | if (!(lb <= ub)) |
125 | printf(format: "Error with bounds %d, %d, %d, err %d\n" , |
126 | (int)lb, (int)ub, chunk, ++err); |
127 | } else { |
128 | if (!(ub >= loop_ub)) |
129 | printf(format: "Error with ub %d, %d, %d, err %d\n" , |
130 | (int)ub, (int)loop_ub, chunk, ++err); |
131 | if (!(lb >= ub)) |
132 | printf(format: "Error with bounds %d, %d, %d, err %d\n" , |
133 | (int)lb, (int)ub, chunk, ++err); |
134 | }; // if |
135 | // Stride should not change. |
136 | if (!(st == loop_st)) |
137 | printf(format: "Error with st %d, %d, ch %d, err %d\n" , |
138 | (int)st, (int)loop_st, chunk, ++err); |
139 | cur = ( ub - lb ) / loop_st + 1; |
140 | // Guided scheduling uses FP computations, so current chunk may |
141 | // be a bit bigger (+1) than allowed maximum. |
142 | if (!( cur <= max + 1)) |
143 | printf(format: "Error with iter %llu, %llu, err %d\n" , cur, max, ++err); |
144 | // Update maximum for the next chunk. |
145 | if (!last && cur % ch) |
146 | printf(format: "Error with chunk %d, %d, ch %d, tid %d, err %d\n" , |
147 | chunk, (int)cur, ch, tid, ++err); |
148 | if (last && !no_chunk && cur > ch && nthreads > 1) |
149 | printf(format: "Error: too big last chunk %d (%d), tid %d, err %d\n" , |
150 | (int)cur, ch, tid, ++err); |
151 | if (cur < max) |
152 | max = cur; |
153 | last_ub = ub; |
154 | undersized = (cur < ch); |
155 | #if _DEBUG > 1 |
156 | if (last) |
157 | printf("under%d cur %d, ch %d, tid %d, ub %d, lb %d, st %d =======\n" , |
158 | undersized,cur,ch,tid,ub,lb,loop_st); |
159 | #endif |
160 | } // while |
161 | // Must have the right last iteration index. |
162 | if (loop_st > 0) { |
163 | if (!(last_ub <= loop_ub)) |
164 | printf(format: "Error with last1 %d, %d, ch %d, err %d\n" , |
165 | (int)last_ub, (int)loop_ub, chunk, ++err); |
166 | if (last && !(last_ub + loop_st > loop_ub)) |
167 | printf(format: "Error with last2 %d, %d, %d, ch %d, err %d\n" , |
168 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk, ++err); |
169 | } else { |
170 | if (!(last_ub >= loop_ub)) |
171 | printf(format: "Error with last1 %d, %d, ch %d, err %d\n" , |
172 | (int)last_ub, (int)loop_ub, chunk, ++err); |
173 | if (last && !(last_ub + loop_st < loop_ub)) |
174 | printf(format: "Error with last2 %d, %d, %d, ch %d, err %d\n" , |
175 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk, ++err); |
176 | } // if |
177 | } |
178 | __kmpc_barrier(&loc, gtid); |
179 | } // run_loop |
180 | |
181 | int main(int argc, char *argv[]) |
182 | { |
183 | int chunk = 0; |
184 | if (argc > 1) { |
185 | // expect chunk size as a parameter |
186 | chunk = atoi(nptr: argv[1]); |
187 | } |
188 | #pragma omp parallel //num_threads(num_th) |
189 | run_loop(loop_lb: 0, UBOUND, loop_st: 1, lchunk: chunk); |
190 | if (err) { |
191 | printf(format: "failed, err = %d\n" , err); |
192 | return 1; |
193 | } else { |
194 | printf(format: "passed\n" ); |
195 | return 0; |
196 | } |
197 | } |
198 | |