如何安全地平行for
我正在KSVD软件包中开发一个原本的串行代码来支持OpenMP。 原始代码在MATLAB中用作im2col并从图像中提取补丁，如下所示：
/* n stands for the size of an image, sz stands for the patch size to extract */
int blocknum = 0;
for (k=0; k<=n[2]-sz[2]; k+=1) {
for (j=0; j<=n[1]-sz[1]; j+=1) {
for (i=0; i<=n[0]-sz[0]; i+=1) {
/* copy single block */
for (m=0; m<sz[2]; m++) {
for (l=0; l<sz[1]; l++) {
memcpy(b + blocknum*sz[0]*sz[1]*sz[2] + m*sz[0]*sz[1] + l*sz[0], x+(k+m)*n[0]*n[1]+(j+l)*n[0]+i, sz[0]*sizeof(double));
}
}
blocknum ++;
}
}
}
虽然，我想通过用索引变量blockid替换增量blocknum来使其平行。
/* n stands for the size of an image, sz stands for the patch size to extract */
int blockid3, blockid2, blockid;
for (k=0; k<=n[2]-sz[2]; k+=1) {
blockid3 = k * (n[1]-sz[1]+1) * (n[0]-sz[0]+1);
#pragma omp parallel for
for (j=0; j<=n[1]-sz[1]; j+=1) {
blockid2 = j * (n[0]-sz[0]+1);
for (i=0; i<=n[0]-sz[0]; i+=1) {
blockid = i + blockid2 + blockid3;
/* copy single block */
for (m=0; m<sz[2]; m++) {
for (l=0; l<sz[1]; l++) {
memcpy(b + blockid*sz[0]*sz[1]*sz[2] + m*sz[0]*sz[1] + l*sz[0], x+(k+m)*n[0]*n[1]+(j+l)*n[0]+i, sz[0]*sizeof(double));
}
}
}
}
}
然后运行导致我致命的分段错误。 我不知道为什么（根据堆栈跟踪，它似乎与安全线程有关）。 因为我认为并行线程不应该访问同一个地址一次。 我应该设置一些变量的属性，即静态或共享或私有？ 这是堆栈跟踪：
Stack Trace (from fault):
[ 0] 0x00007f9bcaa695de /usr/local/MATLAB/R2011b/bin/glnxa64/libmwfl.so+00210398 _ZN2fl4diag15stacktrace_base7capt
ureERKNS0_14thread_contextEm+000158
[ 1] 0x00007f9bcaa6b62d /usr/local/MATLAB/R2011b/bin/glnxa64/libmwfl.so+00218669
[ 2] 0x00007f9bcaa6b8f5 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwfl.so+00219381 _ZN2fl4diag13terminate_logEPKcRKN
S0_14thread_contextEb+000165
[ 3] 0x00007f9bc9a714f5 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00447733 _ZN2fl4diag13terminate_logEPKcPK8
ucontextb+000085
[ 4] 0x00007f9bc9a6e5b4 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00435636
[ 5] 0x00007f9bc9a6f333 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00439091
[ 6] 0x00007f9bc9a6f4c7 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00439495
[ 7] 0x00007f9bc9a7085f /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00444511
[ 8] 0x00007f9bc9a70a15 /usr/local/MATLAB/R2011b/bin/glnxa64/libmwmcr.so+00444949
[ 9] 0x00007f9bc89f0cb0 /lib/x86_64-linux-gnu/libpthread.so.0+00064688
[ 10] 0x00007f9bc876cb8e /lib/x86_64-linux-gnu/libc.so.6+01346446
[ 11] 0x00007f9b88238bb8 /home/peiyun/schmax3.0/test_im2col/mex_im2colstep.mexa64+00003000
[ 12] 0x00007f9bcb004eea /usr/lib/gcc/x86_64-linux-gnu/4.6.3//libgomp.so+00032490
[ 13] 0x00007f9bc89e8e9a /lib/x86_64-linux-gnu/libpthread.so.0+00032410
[ 14] 0x00007f9bc87164bd /lib/x86_64-linux-gnu/libc.so.6+00992445 clone+000109
顺便说一下，如果他们正在写入不同的地址，那么在omp for-loop中是否有任何关于memcpy的竞争条件？
您的代码中有多个数据竞赛，即：
/* n stands for the size of an image, sz stands for the patch size to extract */
int blockid3, blockid2, blockid;
for (k=0; k<=n[2]-sz[2]; k+=1) {
blockid3 = k * (n[1]-sz[1]+1) * (n[0]-sz[0]+1);
#pragma omp parallel for
for (j=0; j<=n[1]-sz[1]; j+=1) {
blockid2 = j * (n[0]-sz[0]+1); // <--- here
for (i=0; i<=n[0]-sz[0]; i+=1) { // <--- here
blockid = i + blockid2 + blockid3; // <--- here
/* copy single block */
for (m=0; m<sz[2]; m++) { // <--- here
for (l=0; l<sz[1]; l++) { // <--- and here
memcpy(b + blockid*sz[0]*sz[1]*sz[2] + m*sz[0]*sz[1] + l*sz[0], x+(k+m)*n[0]*n[1]+(j+l)*n[0]+i, sz[0]*sizeof(double));
}
}
}
}
}
按照OpenMP blockid2的规则， i ， blockid ， m和l都是隐式共享的，这不是你想要的。 你应该让它们private ，或者更好地在并行区域内声明它们，从而使它们隐含私有：
#pragma omp parallel for private(i,m,l,blockid,blockid2)
...
要么
int blockid3;
for (k=0; k<=n[2]-sz[2]; k+=1) {
blockid3 = k * (n[1]-sz[1]+1) * (n[0]-sz[0]+1);
#pragma omp parallel for
for (j=0; j<=n[1]-sz[1]; j+=1) {
int blockid2 = j * (n[0]-sz[0]+1);
for (int i=0; i<=n[0]-sz[0]; i+=1) {
int blockid = i + blockid2 + blockid3;
/* copy single block */
for (int m=0; m<sz[2]; m++) {
for (int l=0; l<sz[1]; l++) {
memcpy(b + blockid*sz[0]*sz[1]*sz[2] + m*sz[0]*sz[1] + l*sz[0], x+(k+m)*n[0]*n[1]+(j+l)*n[0]+i, sz[0]*sizeof(double));
}
}
}
}
}
后者需要符合C99的编译器（因为声明了循环变量的方式）。 您的GCC 4.6.3需要-std=c99选项才能启用C99合规性。 如果没有这样的编译器可用（是否还有通常使用的非C99编译器？），则应该添加private(i,l,m)子句。 您可能还想将并行化移至最外层循环，以便将OpenMP开销降至最低。
上一篇: How to safely parallel the for
下一篇: C++ OpenMP writing to specific element of a shared array/vector