diff options
Diffstat (limited to 'Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc')
-rw-r--r-- | Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc | 105 |
1 files changed, 105 insertions, 0 deletions
diff --git a/Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc b/Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc index 3d1ce9acf..7f5875630 100644 --- a/Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc +++ b/Carpet/CarpetLib/src/restrict_3d_cc_o3_rf2.cc @@ -139,6 +139,109 @@ namespace CarpetLib { + if (not use_loopcontrol_in_operators) { + + // Loop over coarse region +#pragma omp parallel for collapse(3) + for (int k=0; k<regkext; ++k) { + for (int j=0; j<regjext; ++j) { + for (int i=0; i<regiext; ++i) { + +#ifdef CARPET_DEBUG + if(not (2 * k + 2 + srckoff < srckext and + 2 * j + 2 + srcjoff < srcjext and + 2 * i + 2 + srcioff < srciext and + 2 * k - 1 + srckoff >= 0 and + 2 * j - 1 + srcjoff >= 0 and + 2 * i - 1 + srcioff >= 0)) + { + cout << "restrict_3d_cc_o3_rf2.cc\n"; + cout << "regext " << regext << "\n"; + cout << "srcext " << srcext << "\n"; + cout << "srcbbox=" << srcbbox << "\n"; + cout << "dstbbox=" << dstbbox << "\n"; + cout << "regbbox=" << regbbox << "\n"; + cout << "i,j,k=" << i << " " << j << " " << k << "\n"; + assert(2 * k + 2 + srckoff < srckext); + assert(2 * j + 2 + srcjoff < srcjext); + assert(2 * i + 2 + srcioff < srciext); + assert(2 * k - 1 + srckoff >= 0); + assert(2 * j - 1 + srcjoff >= 0); + assert(2 * i - 1 + srcioff >= 0); + } +#endif + dst [DSTIND3(i, j, k)] = + - f1 * src [SRCIND3(2*i-1, 2*j-1, 2*k-1)] + + f2 * src [SRCIND3(2*i , 2*j-1, 2*k-1)] + + f2 * src [SRCIND3(2*i+1, 2*j-1, 2*k-1)] + - f1 * src [SRCIND3(2*i+2, 2*j-1, 2*k-1)] + + f2 * src [SRCIND3(2*i-1, 2*j , 2*k-1)] + - f3 * src [SRCIND3(2*i , 2*j , 2*k-1)] + - f3 * src [SRCIND3(2*i+1, 2*j , 2*k-1)] + + f2 * src [SRCIND3(2*i+2, 2*j , 2*k-1)] + + f2 * src [SRCIND3(2*i-1, 2*j+1, 2*k-1)] + - f3 * src [SRCIND3(2*i , 2*j+1, 2*k-1)] + - f3 * src [SRCIND3(2*i+1, 2*j+1, 2*k-1)] + + f2 * src [SRCIND3(2*i+2, 2*j+1, 2*k-1)] + - f1 * src [SRCIND3(2*i-1, 2*j+2, 2*k-1)] + + f2 * src [SRCIND3(2*i , 2*j+2, 2*k-1)] + + f2 * src [SRCIND3(2*i+1, 2*j+2, 2*k-1)] + - f1 * src [SRCIND3(2*i+2, 2*j+2, 2*k-1)] + + f2 * src [SRCIND3(2*i-1, 2*j-1, 2*k )] + - f3 * src [SRCIND3(2*i , 2*j-1, 2*k )] + - f3 * src [SRCIND3(2*i+1, 2*j-1, 2*k )] + + f2 * src [SRCIND3(2*i+2, 2*j-1, 2*k )] + - f3 * src [SRCIND3(2*i-1, 2*j , 2*k )] + + f4 * src [SRCIND3(2*i , 2*j , 2*k )] + + f4 * src [SRCIND3(2*i+1, 2*j , 2*k )] + - f3 * src [SRCIND3(2*i+2, 2*j , 2*k )] + - f3 * src [SRCIND3(2*i-1, 2*j+1, 2*k )] + + f4 * src [SRCIND3(2*i , 2*j+1, 2*k )] + + f4 * src [SRCIND3(2*i+1, 2*j+1, 2*k )] + - f3 * src [SRCIND3(2*i+2, 2*j+1, 2*k )] + + f2 * src [SRCIND3(2*i-1, 2*j+2, 2*k )] + - f3 * src [SRCIND3(2*i , 2*j+2, 2*k )] + - f3 * src [SRCIND3(2*i+1, 2*j+2, 2*k )] + + f2 * src [SRCIND3(2*i+2, 2*j+2, 2*k )] + + f2 * src [SRCIND3(2*i-1, 2*j-1, 2*k+1)] + - f3 * src [SRCIND3(2*i , 2*j-1, 2*k+1)] + - f3 * src [SRCIND3(2*i+1, 2*j-1, 2*k+1)] + + f2 * src [SRCIND3(2*i+2, 2*j-1, 2*k+1)] + - f3 * src [SRCIND3(2*i-1, 2*j , 2*k+1)] + + f4 * src [SRCIND3(2*i , 2*j , 2*k+1)] + + f4 * src [SRCIND3(2*i+1, 2*j , 2*k+1)] + - f3 * src [SRCIND3(2*i+2, 2*j , 2*k+1)] + - f3 * src [SRCIND3(2*i-1, 2*j+1, 2*k+1)] + + f4 * src [SRCIND3(2*i , 2*j+1, 2*k+1)] + + f4 * src [SRCIND3(2*i+1, 2*j+1, 2*k+1)] + - f3 * src [SRCIND3(2*i+2, 2*j+1, 2*k+1)] + + f2 * src [SRCIND3(2*i-1, 2*j+2, 2*k+1)] + - f3 * src [SRCIND3(2*i , 2*j+2, 2*k+1)] + - f3 * src [SRCIND3(2*i+1, 2*j+2, 2*k+1)] + + f2 * src [SRCIND3(2*i+2, 2*j+2, 2*k+1)] + - f1 * src [SRCIND3(2*i-1, 2*j-1, 2*k+2)] + + f2 * src [SRCIND3(2*i , 2*j-1, 2*k+2)] + + f2 * src [SRCIND3(2*i+1, 2*j-1, 2*k+2)] + - f1 * src [SRCIND3(2*i+2, 2*j-1, 2*k+2)] + + f2 * src [SRCIND3(2*i-1, 2*j , 2*k+2)] + - f3 * src [SRCIND3(2*i , 2*j , 2*k+2)] + - f3 * src [SRCIND3(2*i+1, 2*j , 2*k+2)] + + f2 * src [SRCIND3(2*i+2, 2*j , 2*k+2)] + + f2 * src [SRCIND3(2*i-1, 2*j+1, 2*k+2)] + - f3 * src [SRCIND3(2*i , 2*j+1, 2*k+2)] + - f3 * src [SRCIND3(2*i+1, 2*j+1, 2*k+2)] + + f2 * src [SRCIND3(2*i+2, 2*j+1, 2*k+2)] + - f1 * src [SRCIND3(2*i-1, 2*j+2, 2*k+2)] + + f2 * src [SRCIND3(2*i , 2*j+2, 2*k+2)] + + f2 * src [SRCIND3(2*i+1, 2*j+2, 2*k+2)] + - f1 * src [SRCIND3(2*i+2, 2*j+2, 2*k+2)]; + + } + } + } + + } else { + // Loop over coarse region #pragma omp parallel CCTK_LOOP3(restrict_3d_cc_o3_rf2, @@ -237,6 +340,8 @@ namespace CarpetLib { } CCTK_ENDLOOP3(restrict_3d_cc_o3_rf2); + } + } |