1 | // RUN: %libomptarget-compile-generic && env LIBOMPTARGET_DEBUG=1 %libomptarget-run-generic 2>&1 | %fcheck-generic -allow-empty -check-prefix=DEBUG |
2 | // REQUIRES: libomptarget-debug |
3 | |
4 | #include <cassert> |
5 | #include <cstdio> |
6 | #include <cstdlib> |
7 | |
8 | // Data structure definitions copied from OpenMP RTL. |
9 | struct __tgt_target_non_contig { |
10 | int64_t offset; |
11 | int64_t width; |
12 | int64_t stride; |
13 | }; |
14 | |
15 | enum tgt_map_type { OMP_TGT_MAPTYPE_NON_CONTIG = 0x100000000000 }; |
16 | |
17 | // OpenMP RTL interfaces |
18 | #ifdef __cplusplus |
19 | extern "C" { |
20 | #endif |
21 | void __tgt_target_data_update(int64_t device_id, int32_t arg_num, |
22 | void **args_base, void **args, int64_t *arg_sizes, |
23 | int64_t *arg_types); |
24 | #ifdef __cplusplus |
25 | } |
26 | #endif |
27 | |
28 | int main() { |
29 | // case 1 |
30 | // int arr[3][4][5][6]; |
31 | // #pragma omp target update to(arr[0:2][1:3][1:2][:]) |
32 | // set up descriptor |
33 | __tgt_target_non_contig non_contig[5] = { |
34 | {.offset: 0, .width: 2, .stride: 480}, {.offset: 1, .width: 3, .stride: 120}, {.offset: 1, .width: 2, .stride: 24}, {.offset: 0, .width: 6, .stride: 4}, {.offset: 0, .width: 1, .stride: 4}}; |
35 | int64_t size = 4, type = OMP_TGT_MAPTYPE_NON_CONTIG; |
36 | |
37 | void *base; |
38 | void *begin = &non_contig; |
39 | int64_t *sizes = &size; |
40 | int64_t *types = &type; |
41 | |
42 | // The below diagram is the visualization of the non-contiguous transfer after |
43 | // optimization. Note that each element represent the innermost dimension |
44 | // (unit size = 24) since the stride * count of last dimension is equal to the |
45 | // stride of second last dimension. |
46 | // |
47 | // OOOOO OOOOO OOOOO |
48 | // OXXOO OXXOO OOOOO |
49 | // OXXOO OXXOO OOOOO |
50 | // OXXOO OXXOO OOOOO |
51 | __tgt_target_data_update(/*device_id*/ -1, /*arg_num*/ 1, args_base: &base, args: &begin, |
52 | arg_sizes: sizes, arg_types: types); |
53 | // DEBUG: offset 144 |
54 | // DEBUG: offset 264 |
55 | // DEBUG: offset 384 |
56 | // DEBUG: offset 624 |
57 | // DEBUG: offset 744 |
58 | // DEBUG: offset 864 |
59 | |
60 | // case 2 |
61 | // double darr[3][4][5]; |
62 | // #pragma omp target update to(darr[0:2:2][2:2][:2:2]) |
63 | // set up descriptor |
64 | __tgt_target_non_contig non_contig_2[4] = { |
65 | {.offset: 0, .width: 2, .stride: 320}, {.offset: 2, .width: 2, .stride: 40}, {.offset: 0, .width: 2, .stride: 16}, {.offset: 0, .width: 1, .stride: 8}}; |
66 | int64_t size_2 = 4, type_2 = OMP_TGT_MAPTYPE_NON_CONTIG; |
67 | |
68 | void *base_2; |
69 | void *begin_2 = &non_contig_2; |
70 | int64_t *sizes_2 = &size_2; |
71 | int64_t *types_2 = &type_2; |
72 | |
73 | // The below diagram is the visualization of the non-contiguous transfer after |
74 | // optimization. Note that each element represent the innermost dimension |
75 | // (unit size = 24) since the stride * count of last dimension is equal to the |
76 | // stride of second last dimension. |
77 | // |
78 | // OOOOO OOOOO OOOOO |
79 | // OOOOO OOOOO OOOOO |
80 | // XOXOO OOOOO XOXOO |
81 | // XOXOO OOOOO XOXOO |
82 | __tgt_target_data_update(/*device_id*/ -1, /*arg_num*/ 1, args_base: &base_2, args: &begin_2, |
83 | arg_sizes: sizes_2, arg_types: types_2); |
84 | // DEBUG: offset 80 |
85 | // DEBUG: offset 96 |
86 | // DEBUG: offset 120 |
87 | // DEBUG: offset 136 |
88 | // DEBUG: offset 400 |
89 | // DEBUG: offset 416 |
90 | // DEBUG: offset 440 |
91 | // DEBUG: offset 456 |
92 | return 0; |
93 | } |
94 | |