source: CIVL/examples/omp/DataRaceBench/micro-benchmarks/jacobi2d-tile-no.c@ af3b8e4

1.23 2.0 main test-branch
Last change on this file since af3b8e4 was 36a61f3, checked in by Ziqing Luo <ziqing@…>, 9 years ago

Commit DataRaceBench into CIVL examples

git-svn-id: ​svn://vsl.cis.udel.edu/civl/trunk@4225 fb995dde-84ed-4084-dfe6-e5aef3e2452c
  • Property mode set to 100644
File size: 13.5 KB
Line 
1/**
2 * jacobi-2d-imper.c: This file is part of the PolyBench/C 3.2 test suite.
3 * Jacobi with array copying, no reduction. with tiling and nested SIMD.
4 *
5 * Contact: Louis-Noel Pouchet <pouchet@cse.ohio-state.edu>
6 * Web address: http://polybench.sourceforge.net
7 */
8#include <stdio.h>
9#include <unistd.h>
10#include <string.h>
11#include <math.h>
12/* Include polybench common header. */
13#include <polybench.h>
14/* Include benchmark-specific header. */
15/* Default data type is double, default size is 20x1000. */
16#include "jacobi-2d-imper.h"
17/* Array initialization. */
18
19static void init_array(int n,double A[500 + 0][500 + 0],double B[500 + 0][500 + 0])
20{
21 int i;
22 int j;
23{
24 int c1;
25 int c2;
26 int c4;
27 int c3;
28 if (n >= 1) {
29#pragma omp parallel for private(c3, c4, c2)
30 for (c1 = 0; c1 <= (((n + -1) * 16 < 0?((16 < 0?-((-(n + -1) + 16 + 1) / 16) : -((-(n + -1) + 16 - 1) / 16))) : (n + -1) / 16)); c1++) {
31 for (c2 = 0; c2 <= (((n + -1) * 16 < 0?((16 < 0?-((-(n + -1) + 16 + 1) / 16) : -((-(n + -1) + 16 - 1) / 16))) : (n + -1) / 16)); c2++) {
32 for (c3 = 16 * c2; c3 <= ((16 * c2 + 15 < n + -1?16 * c2 + 15 : n + -1)); c3++) {
33#pragma ivdep
34#pragma vector always
35#pragma simd
36 for (c4 = 16 * c1; c4 <= ((16 * c1 + 15 < n + -1?16 * c1 + 15 : n + -1)); c4++) {
37 A[c4][c3] = (((double )c4) * (c3 + 2) + 2) / n;
38 B[c4][c3] = (((double )c4) * (c3 + 3) + 3) / n;
39 }
40 }
41 }
42 }
43 }
44 }
45}
46/* DCE code. Must scan the entire live-out data.
47 Can be used also to check the correctness of the output. */
48
49static void print_array(int n,double A[500 + 0][500 + 0])
50{
51 int i;
52 int j;
53 for (i = 0; i < n; i++)
54 for (j = 0; j < n; j++) {
55 fprintf(stderr,"%0.2lf ",A[i][j]);
56 if ((i * n + j) % 20 == 0)
57 fprintf(stderr,"\n");
58 }
59 fprintf(stderr,"\n");
60}
61/* Main computational kernel. The whole function will be timed,
62 including the call and return. */
63
64static void kernel_jacobi_2d_imper(int tsteps,int n,double A[500 + 0][500 + 0],double B[500 + 0][500 + 0])
65{
66 int t;
67 int i;
68 int j;
69
70#pragma scop
71{
72 int c0;
73 int c1;
74 int c3;
75 int c2;
76 int c4;
77 int c5;
78 if (n >= 3 && tsteps >= 1) {
79 for (c0 = 0; c0 <= (((n + 3 * tsteps + -4) * 16 < 0?((16 < 0?-((-(n + 3 * tsteps + -4) + 16 + 1) / 16) : -((-(n + 3 * tsteps + -4) + 16 - 1) / 16))) : (n + 3 * tsteps + -4) / 16)); c0++) {
80#pragma omp parallel for private(c5, c4, c2, c3)
81 for (c1 = (((2 * c0 * 3 < 0?-(-(2 * c0) / 3) : ((3 < 0?(-(2 * c0) + - 3 - 1) / - 3 : (2 * c0 + 3 - 1) / 3)))) > (((16 * c0 + -1 * tsteps + 1) * 16 < 0?-(-(16 * c0 + -1 * tsteps + 1) / 16) : ((16 < 0?(-(16 * c0 + -1 * tsteps + 1) + - 16 - 1) / - 16 : (16 * c0 + -1 * tsteps + 1 + 16 - 1) / 16))))?((2 * c0 * 3 < 0?-(-(2 * c0) / 3) : ((3 < 0?(-(2 * c0) + - 3 - 1) / - 3 : (2 * c0 + 3 - 1) / 3)))) : (((16 * c0 + -1 * tsteps + 1) * 16 < 0?-(-(16 * c0 + -1 * tsteps + 1) / 16) : ((16 < 0?(-(16 * c0 + -1 * tsteps + 1) + - 16 - 1) / - 16 : (16 * c0 + -1 * tsteps + 1 + 16 - 1) / 16))))); c1 <= (((((((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)) < (((32 * c0 + n + 29) * 48 < 0?((48 < 0?-((-(32 * c0 + n + 29) + 48 + 1) / 48) : -((-(32 * c0 + n + 29) + 48 - 1) / 48))) : (32 * c0 + n + 29) / 48))?(((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)) : (((32 * c0 + n + 29) * 48 < 0?((48 < 0?-((-(32 * c0 + n + 29) + 48 + 1) / 48) : -((-(32 * c0 + n + 29) + 48 - 1) / 48))) : (32 * c0 + n + 29) / 48)))) < c0?(((((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)) < (((32 * c0 + n + 29) * 48 < 0?((48 < 0?-((-(32 * c0 + n + 29) + 48 + 1) / 48) : -((-(32 * c0 + n + 29) + 48 - 1) / 48))) : (32 * c0 + n + 29) / 48))?(((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)) : (((32 * c0 + n + 29) * 48 < 0?((48 < 0?-((-(32 * c0 + n + 29) + 48 + 1) / 48) : -((-(32 * c0 + n + 29) + 48 - 1) / 48))) : (32 * c0 + n + 29) / 48)))) : c0)); c1++) {
82 for (c2 = ((((16 * c1 + -1 * n + -12) * 16 < 0?-(-(16 * c1 + -1 * n + -12) / 16) : ((16 < 0?(-(16 * c1 + -1 * n + -12) + - 16 - 1) / - 16 : (16 * c1 + -1 * n + -12 + 16 - 1) / 16)))) > 2 * c0 + -2 * c1?(((16 * c1 + -1 * n + -12) * 16 < 0?-(-(16 * c1 + -1 * n + -12) / 16) : ((16 < 0?(-(16 * c1 + -1 * n + -12) + - 16 - 1) / - 16 : (16 * c1 + -1 * n + -12 + 16 - 1) / 16)))) : 2 * c0 + -2 * c1); c2 <= (((((((16 * c1 + n + 12) * 16 < 0?((16 < 0?-((-(16 * c1 + n + 12) + 16 + 1) / 16) : -((-(16 * c1 + n + 12) + 16 - 1) / 16))) : (16 * c1 + n + 12) / 16)) < (((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16))?(((16 * c1 + n + 12) * 16 < 0?((16 < 0?-((-(16 * c1 + n + 12) + 16 + 1) / 16) : -((-(16 * c1 + n + 12) + 16 - 1) / 16))) : (16 * c1 + n + 12) / 16)) : (((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)))) < (((32 * c0 + -32 * c1 + n + 29) * 16 < 0?((16 < 0?-((-(32 * c0 + -32 * c1 + n + 29) + 16 + 1) / 16) : -((-(32 * c0 + -32 * c1 + n + 29) + 16 - 1) / 16))) : (32 * c0 + -32 * c1 + n + 29) / 16))?(((((16 * c1 + n + 12) * 16 < 0?((16 < 0?-((-(16 * c1 + n + 12) + 16 + 1) / 16) : -((-(16 * c1 + n + 12) + 16 - 1) / 16))) : (16 * c1 + n + 12) / 16)) < (((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16))?(((16 * c1 + n + 12) * 16 < 0?((16 < 0?-((-(16 * c1 + n + 12) + 16 + 1) / 16) : -((-(16 * c1 + n + 12) + 16 - 1) / 16))) : (16 * c1 + n + 12) / 16)) : (((n + 2 * tsteps + -3) * 16 < 0?((16 < 0?-((-(n + 2 * tsteps + -3) + 16 + 1) / 16) : -((-(n + 2 * tsteps + -3) + 16 - 1) / 16))) : (n + 2 * tsteps + -3) / 16)))) : (((32 * c0 + -32 * c1 + n + 29) * 16 < 0?((16 < 0?-((-(32 * c0 + -32 * c1 + n + 29) + 16 + 1) / 16) : -((-(32 * c0 + -32 * c1 + n + 29) + 16 - 1) / 16))) : (32 * c0 + -32 * c1 + n + 29) / 16)))); c2++) {
83 if (c0 <= (((32 * c1 + 16 * c2 + -1 * n + 1) * 32 < 0?((32 < 0?-((-(32 * c1 + 16 * c2 + -1 * n + 1) + 32 + 1) / 32) : -((-(32 * c1 + 16 * c2 + -1 * n + 1) + 32 - 1) / 32))) : (32 * c1 + 16 * c2 + -1 * n + 1) / 32)) && c1 <= c2 + -1) {
84 if ((n + 1) % 2 == 0) {
85 for (c4 = (16 * c1 > 16 * c2 + -1 * n + 3?16 * c1 : 16 * c2 + -1 * n + 3); c4 <= 16 * c1 + 15; c4++) {
86 A[-16 * c2 + c4 + n + -2][n + -2] = B[-16 * c2 + c4 + n + -2][n + -2];
87 }
88 }
89 }
90 if (c0 <= (((48 * c1 + -1 * n + 1) * 32 < 0?((32 < 0?-((-(48 * c1 + -1 * n + 1) + 32 + 1) / 32) : -((-(48 * c1 + -1 * n + 1) + 32 - 1) / 32))) : (48 * c1 + -1 * n + 1) / 32)) && c1 >= c2) {
91 if ((n + 1) % 2 == 0) {
92 for (c5 = (16 * c2 > 16 * c1 + -1 * n + 3?16 * c2 : 16 * c1 + -1 * n + 3); c5 <= ((16 * c1 < 16 * c2 + 15?16 * c1 : 16 * c2 + 15)); c5++) {
93 A[n + -2][-16 * c1 + c5 + n + -2] = B[n + -2][-16 * c1 + c5 + n + -2];
94 }
95 }
96 }
97 for (c3 = ((((((16 * c1 + -1 * n + 2) * 2 < 0?-(-(16 * c1 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c1 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c1 + -1 * n + 2 + 2 - 1) / 2)))) > (((16 * c2 + -1 * n + 2) * 2 < 0?-(-(16 * c2 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c2 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c2 + -1 * n + 2 + 2 - 1) / 2))))?(((16 * c1 + -1 * n + 2) * 2 < 0?-(-(16 * c1 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c1 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c1 + -1 * n + 2 + 2 - 1) / 2)))) : (((16 * c2 + -1 * n + 2) * 2 < 0?-(-(16 * c2 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c2 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c2 + -1 * n + 2 + 2 - 1) / 2)))))) > 16 * c0 + -16 * c1?(((((16 * c1 + -1 * n + 2) * 2 < 0?-(-(16 * c1 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c1 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c1 + -1 * n + 2 + 2 - 1) / 2)))) > (((16 * c2 + -1 * n + 2) * 2 < 0?-(-(16 * c2 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c2 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c2 + -1 * n + 2 + 2 - 1) / 2))))?(((16 * c1 + -1 * n + 2) * 2 < 0?-(-(16 * c1 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c1 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c1 + -1 * n + 2 + 2 - 1) / 2)))) : (((16 * c2 + -1 * n + 2) * 2 < 0?-(-(16 * c2 + -1 * n + 2) / 2) : ((2 < 0?(-(16 * c2 + -1 * n + 2) + - 2 - 1) / - 2 : (16 * c2 + -1 * n + 2 + 2 - 1) / 2)))))) : 16 * c0 + -16 * c1); c3 <= ((((((8 * c1 + 6 < 8 * c2 + 6?8 * c1 + 6 : 8 * c2 + 6)) < tsteps + -1?((8 * c1 + 6 < 8 * c2 + 6?8 * c1 + 6 : 8 * c2 + 6)) : tsteps + -1)) < 16 * c0 + -16 * c1 + 15?((((8 * c1 + 6 < 8 * c2 + 6?8 * c1 + 6 : 8 * c2 + 6)) < tsteps + -1?((8 * c1 + 6 < 8 * c2 + 6?8 * c1 + 6 : 8 * c2 + 6)) : tsteps + -1)) : 16 * c0 + -16 * c1 + 15)); c3++) {
98 if (c1 <= ((c3 * 8 < 0?((8 < 0?-((-c3 + 8 + 1) / 8) : -((-c3 + 8 - 1) / 8))) : c3 / 8))) {
99 for (c5 = (16 * c2 > 2 * c3 + 1?16 * c2 : 2 * c3 + 1); c5 <= ((16 * c2 + 15 < 2 * c3 + n + -2?16 * c2 + 15 : 2 * c3 + n + -2)); c5++) {
100 B[1][-2 * c3 + c5] = 0.2 * (A[1][-2 * c3 + c5] + A[1][-2 * c3 + c5 - 1] + A[1][1 + (-2 * c3 + c5)] + A[1 + 1][-2 * c3 + c5] + A[1 - 1][-2 * c3 + c5]);
101 }
102 }
103 for (c4 = (16 * c1 > 2 * c3 + 2?16 * c1 : 2 * c3 + 2); c4 <= ((16 * c1 + 15 < 2 * c3 + n + -2?16 * c1 + 15 : 2 * c3 + n + -2)); c4++) {
104 if (c2 <= ((c3 * 8 < 0?((8 < 0?-((-c3 + 8 + 1) / 8) : -((-c3 + 8 - 1) / 8))) : c3 / 8))) {
105 B[-2 * c3 + c4][1] = 0.2 * (A[-2 * c3 + c4][1] + A[-2 * c3 + c4][1 - 1] + A[-2 * c3 + c4][1 + 1] + A[1 + (-2 * c3 + c4)][1] + A[-2 * c3 + c4 - 1][1]);
106 }
107 for (c5 = (16 * c2 > 2 * c3 + 2?16 * c2 : 2 * c3 + 2); c5 <= ((16 * c2 + 15 < 2 * c3 + n + -2?16 * c2 + 15 : 2 * c3 + n + -2)); c5++) {
108 B[-2 * c3 + c4][-2 * c3 + c5] = 0.2 * (A[-2 * c3 + c4][-2 * c3 + c5] + A[-2 * c3 + c4][-2 * c3 + c5 - 1] + A[-2 * c3 + c4][1 + (-2 * c3 + c5)] + A[1 + (-2 * c3 + c4)][-2 * c3 + c5] + A[-2 * c3 + c4 - 1][-2 * c3 + c5]);
109 A[-2 * c3 + c4 + -1][-2 * c3 + c5 + -1] = B[-2 * c3 + c4 + -1][-2 * c3 + c5 + -1];
110 }
111 if (c2 >= (((2 * c3 + n + -16) * 16 < 0?-(-(2 * c3 + n + -16) / 16) : ((16 < 0?(-(2 * c3 + n + -16) + - 16 - 1) / - 16 : (2 * c3 + n + -16 + 16 - 1) / 16))))) {
112 A[-2 * c3 + c4 + -1][n + -2] = B[-2 * c3 + c4 + -1][n + -2];
113 }
114 }
115 if (c1 >= (((2 * c3 + n + -16) * 16 < 0?-(-(2 * c3 + n + -16) / 16) : ((16 < 0?(-(2 * c3 + n + -16) + - 16 - 1) / - 16 : (2 * c3 + n + -16 + 16 - 1) / 16))))) {
116 for (c5 = (16 * c2 > 2 * c3 + 2?16 * c2 : 2 * c3 + 2); c5 <= ((16 * c2 + 15 < 2 * c3 + n + -1?16 * c2 + 15 : 2 * c3 + n + -1)); c5++) {
117 A[n + -2][-2 * c3 + c5 + -1] = B[n + -2][-2 * c3 + c5 + -1];
118 }
119 }
120 }
121 if (c0 >= (((2 * c1 + c2 + -1) * 2 < 0?-(-(2 * c1 + c2 + -1) / 2) : ((2 < 0?(-(2 * c1 + c2 + -1) + - 2 - 1) / - 2 : (2 * c1 + c2 + -1 + 2 - 1) / 2)))) && c1 >= c2 + 1 && c2 <= (((tsteps + -8) * 8 < 0?((8 < 0?-((-(tsteps + -8) + 8 + 1) / 8) : -((-(tsteps + -8) + 8 - 1) / 8))) : (tsteps + -8) / 8))) {
122 for (c4 = 16 * c1; c4 <= ((16 * c1 + 15 < 16 * c2 + n + 12?16 * c1 + 15 : 16 * c2 + n + 12)); c4++) {
123 B[-16 * c2 + c4 + -14][1] = 0.2 * (A[-16 * c2 + c4 + -14][1] + A[-16 * c2 + c4 + -14][1 - 1] + A[-16 * c2 + c4 + -14][1 + 1] + A[1 + (-16 * c2 + c4 + -14)][1] + A[-16 * c2 + c4 + -14 - 1][1]);
124 }
125 }
126 if (c0 >= (((3 * c1 + -1) * 2 < 0?-(-(3 * c1 + -1) / 2) : ((2 < 0?(-(3 * c1 + -1) + - 2 - 1) / - 2 : (3 * c1 + -1 + 2 - 1) / 2)))) && c1 <= (((((tsteps + -8) * 8 < 0?((8 < 0?-((-(tsteps + -8) + 8 + 1) / 8) : -((-(tsteps + -8) + 8 - 1) / 8))) : (tsteps + -8) / 8)) < c2?(((tsteps + -8) * 8 < 0?((8 < 0?-((-(tsteps + -8) + 8 + 1) / 8) : -((-(tsteps + -8) + 8 - 1) / 8))) : (tsteps + -8) / 8)) : c2))) {
127 for (c5 = (16 * c2 > 16 * c1 + 15?16 * c2 : 16 * c1 + 15); c5 <= ((16 * c2 + 15 < 16 * c1 + n + 12?16 * c2 + 15 : 16 * c1 + n + 12)); c5++) {
128 B[1][-16 * c1 + c5 + -14] = 0.2 * (A[1][-16 * c1 + c5 + -14] + A[1][-16 * c1 + c5 + -14 - 1] + A[1][1 + (-16 * c1 + c5 + -14)] + A[1 + 1][-16 * c1 + c5 + -14] + A[1 - 1][-16 * c1 + c5 + -14]);
129 }
130 }
131 }
132 }
133 }
134 }
135 }
136
137#pragma endscop
138}
139
140int main(int argc,char **argv)
141{
142/* Retrieve problem size. */
143 int n = 500;
144 int tsteps = 10;
145/* Variable declaration/allocation. */
146 double (*A)[500 + 0][500 + 0];
147 A = ((double (*)[500 + 0][500 + 0])(polybench_alloc_data(((500 + 0) * (500 + 0)),(sizeof(double )))));
148 ;
149 double (*B)[500 + 0][500 + 0];
150 B = ((double (*)[500 + 0][500 + 0])(polybench_alloc_data(((500 + 0) * (500 + 0)),(sizeof(double )))));
151 ;
152/* Initialize array(s). */
153 init_array(n, *A, *B);
154/* Start timer. */
155 polybench_timer_start();
156 ;
157/* Run kernel. */
158 kernel_jacobi_2d_imper(tsteps,n, *A, *B);
159/* Stop and print timer. */
160 polybench_timer_stop();
161 ;
162 polybench_timer_print();
163 ;
164/* Prevent dead-code elimination. All live-out data must be printed
165 by the function call in argument. */
166 if (argc > 42 && !strcmp(argv[0],""))
167 print_array(n, *A);
168/* Be clean. */
169 free(((void *)A));
170 ;
171 free(((void *)B));
172 ;
173 return 0;
174}
Note: See TracBrowser for help on using the repository browser.