64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

derivs_c: split low/high stencil regions to reduce branch overhead

Browse Source
This commit is contained in:
CGH0S7
2026-02-28 15:42:31 +08:00
parent bc457d981e
commit 6024708a48
2 changed files with 236 additions and 188 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
AMSS_NCKU_source/fderivs_c.C
View File

@@ -81,26 +81,63 @@ void fderivs(const int ex[3],
}
/*
* Fortran loops:
* do k=1,ex3-1
* do j=1,ex2-1
* do i=1,ex1-1
* 两段式：
* 1) 先在二阶可用区域计算二阶模板
* 2) 再在高阶可用区域覆盖为四阶模板
*
* C: k0=0..ex3-2, j0=0..ex2-2, i0=0..ex1-2
* 与原 if/elseif 逻辑等价，但减少逐点分支判断。
*/
for (int k0 = 0; k0 <= ex3 - 2; ++k0) {
const int kF = k0 + 1;
for (int j0 = 0; j0 <= ex2 - 2; ++j0) {
const int jF = j0 + 1;
for (int i0 = 0; i0 <= ex1 - 2; ++i0) {
const int iF = i0 + 1;
const size_t p = idx_ex(i0, j0, k0, ex);
const int i2_lo = (iminF > 0) ? iminF : 0;
const int j2_lo = (jminF > 0) ? jminF : 0;
const int k2_lo = (kminF > 0) ? kminF : 0;
const int i2_hi = ex1 - 2;
const int j2_hi = ex2 - 2;
const int k2_hi = ex3 - 2;
const int i4_lo = (iminF + 1 > 0) ? (iminF + 1) : 0;
const int j4_lo = (jminF + 1 > 0) ? (jminF + 1) : 0;
const int k4_lo = (kminF + 1 > 0) ? (kminF + 1) : 0;
const int i4_hi = ex1 - 3;
const int j4_hi = ex2 - 3;
const int k4_hi = ex3 - 3;
if (i2_lo <= i2_hi && j2_lo <= j2_hi && k2_lo <= k2_hi) {
for (int k0 = k2_lo; k0 <= k2_hi; ++k0) {
const int kF = k0 + 1;
for (int j0 = j2_lo; j0 <= j2_hi; ++j0) {
const int jF = j0 + 1;
for (int i0 = i2_lo; i0 <= i2_hi; ++i0) {
const int iF = i0 + 1;
const size_t p = idx_ex(i0, j0, k0, ex);
fx[p] = d2dx * (
-fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] +
fh[idx_fh_F_ord2(iF + 1, jF, kF, ex)]
);
fy[p] = d2dy * (
-fh[idx_fh_F_ord2(iF, jF - 1, kF, ex)] +
fh[idx_fh_F_ord2(iF, jF + 1, kF, ex)]
);
fz[p] = d2dz * (
-fh[idx_fh_F_ord2(iF, jF, kF - 1, ex)] +
fh[idx_fh_F_ord2(iF, jF, kF + 1, ex)]
);
}
}
}
}
if (i4_lo <= i4_hi && j4_lo <= j4_hi && k4_lo <= k4_hi) {
for (int k0 = k4_lo; k0 <= k4_hi; ++k0) {
const int kF = k0 + 1;
for (int j0 = j4_lo; j0 <= j4_hi; ++j0) {
const int jF = j0 + 1;
for (int i0 = i4_lo; i0 <= i4_hi; ++i0) {
const int iF = i0 + 1;
const size_t p = idx_ex(i0, j0, k0, ex);
// if(i+2 <= imax .and. i-2 >= imin ... ) (全是 Fortran 索引)
if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
(jF + 2) <= jmaxF && (jF - 2) >= jminF &&
(kF + 2) <= kmaxF && (kF - 2) >= kminF)
{
fx[p] = d12dx * (
fh[idx_fh_F_ord2(iF - 2, jF, kF, ex)] -
EIT * fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] +
@@ -122,29 +159,9 @@ void fderivs(const int ex[3],
fh[idx_fh_F_ord2(iF, jF, kF + 2, ex)]
);
}
// elseif(i+1 <= imax .and. i-1 >= imin ...)
else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
(jF + 1) <= jmaxF && (jF - 1) >= jminF &&
(kF + 1) <= kmaxF && (kF - 1) >= kminF)
{
fx[p] = d2dx * (
-fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] +
fh[idx_fh_F_ord2(iF + 1, jF, kF, ex)]
);
fy[p] = d2dy * (
-fh[idx_fh_F_ord2(iF, jF - 1, kF, ex)] +
fh[idx_fh_F_ord2(iF, jF + 1, kF, ex)]
);
fz[p] = d2dz * (
-fh[idx_fh_F_ord2(iF, jF, kF - 1, ex)] +
fh[idx_fh_F_ord2(iF, jF, kF + 1, ex)]
);
}
}
}
}
// free(fh);
}
}