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

Optimize fdderivs: skip redundant 2nd-order work in 4th-order overlap

Browse Source
This commit is contained in:
CGH0S7
2026-03-02 03:21:21 +08:00
committed by ianchb
parent f561522d89
commit 90620c2aec
1 changed files with 206 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

348
AMSS_NCKU_source/fdderivs_c.C
View File

@@ -71,149 +71,99 @@ void fdderivs(const int ex[3],
const double Fdxdz = F1o144 / (dX * dZ);
const double Fdydz = F1o144 / (dY * dZ);
/* 输出清零：fxx,fyy,fzz,fxy,fxz,fyz = 0 */
const size_t all = (size_t)ex1 * (size_t)ex2 * (size_t)ex3;
for (size_t p = 0; p < all; ++p) {
fxx[p] = ZEO; fyy[p] = ZEO; fzz[p] = ZEO;
fxy[p] = ZEO; fxz[p] = ZEO; fyz[p] = ZEO;
/* 只清零不被主循环覆盖的边界面 */
{
/* 高边界：k0=ex3-1 */
for (int j0 = 0; j0 < ex2; ++j0)
for (int i0 = 0; i0 < ex1; ++i0) {
const size_t p = idx_ex(i0, j0, ex3 - 1, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
/* 高边界：j0=ex2-1 */
for (int k0 = 0; k0 < ex3 - 1; ++k0)
for (int i0 = 0; i0 < ex1; ++i0) {
const size_t p = idx_ex(i0, ex2 - 1, k0, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
/* 高边界：i0=ex1-1 */
for (int k0 = 0; k0 < ex3 - 1; ++k0)
for (int j0 = 0; j0 < ex2 - 1; ++j0) {
const size_t p = idx_ex(ex1 - 1, j0, k0, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
/* 低边界：当二阶模板也不可用时，对应 i0/j0/k0=0 面 */
if (kminF == 1) {
for (int j0 = 0; j0 < ex2; ++j0)
for (int i0 = 0; i0 < ex1; ++i0) {
const size_t p = idx_ex(i0, j0, 0, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
}
if (jminF == 1) {
for (int k0 = 0; k0 < ex3; ++k0)
for (int i0 = 0; i0 < ex1; ++i0) {
const size_t p = idx_ex(i0, 0, k0, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
}
if (iminF == 1) {
for (int k0 = 0; k0 < ex3; ++k0)
for (int j0 = 0; j0 < ex2; ++j0) {
const size_t p = idx_ex(0, j0, k0, ex);
fxx[p]=ZEO; fyy[p]=ZEO; fzz[p]=ZEO;
fxy[p]=ZEO; fxz[p]=ZEO; fyz[p]=ZEO;
}
}
}
/*
* Fortran:
* do k=1,ex3-1
* do j=1,ex2-1
* do i=1,ex1-1
* 两段式：
* 1) 二阶可用区域先计算二阶模板
* 2) 高阶可用区域再覆盖四阶模板
*/
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;
/* 高阶分支：i±2,j±2,k±2 都在范围内 */
if ((iF + 2) <= imaxF && (iF - 2) >= iminF &&
(jF + 2) <= jmaxF && (jF - 2) >= jminF &&
(kF + 2) <= kmaxF && (kF - 2) >= kminF)
{
fxx[p] = Fdxdx * (
-fh[idx_fh_F_ord2(iF - 2, jF, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF + 2, jF, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF + 1, jF, kF, ex)]
);
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;
fyy[p] = Fdydy * (
-fh[idx_fh_F_ord2(iF, jF - 2, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF - 1, kF, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF, jF + 2, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF + 1, kF, ex)]
);
/*
* Strategy A:
* Avoid redundant work in overlap of 2nd/4th-order regions.
* Only compute 2nd-order on shell points that are NOT overwritten by
* the 4th-order pass.
*/
const int has4 = (i4_lo <= i4_hi && j4_lo <= j4_hi && k4_lo <= k4_hi);
fzz[p] = Fdzdz * (
-fh[idx_fh_F_ord2(iF, jF, kF - 2, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF, kF - 1, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF, jF, kF + 2, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF, kF + 1, ex)]
);
/* fxy 高阶：完全照搬 Fortran 的括号结构 */
{
const double t_jm2 =
( fh[idx_fh_F_ord2(iF - 2, jF - 2, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF - 2, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF - 2, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF - 2, kF, ex)] );
const double t_jm1 =
( fh[idx_fh_F_ord2(iF - 2, jF - 1, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF - 1, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF - 1, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF - 1, kF, ex)] );
const double t_jp1 =
( fh[idx_fh_F_ord2(iF - 2, jF + 1, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF + 1, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF + 1, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF + 1, kF, ex)] );
const double t_jp2 =
( fh[idx_fh_F_ord2(iF - 2, jF + 2, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF + 2, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF + 2, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF + 2, kF, ex)] );
fxy[p] = Fdxdy * ( t_jm2 - F8 * t_jm1 + F8 * t_jp1 - t_jp2 );
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) {
if (has4 &&
i0 >= i4_lo && i0 <= i4_hi &&
j0 >= j4_lo && j0 <= j4_hi &&
k0 >= k4_lo && k0 <= k4_hi) {
continue;
}
const int iF = i0 + 1;
const size_t p = idx_ex(i0, j0, k0, ex);
/* fxz 高阶 */
{
const double t_km2 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF - 2, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF - 2, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF - 2, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF - 2, ex)] );
const double t_km1 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF - 1, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF - 1, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF - 1, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF - 1, ex)] );
const double t_kp1 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF + 1, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF + 1, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF + 1, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF + 1, ex)] );
const double t_kp2 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF + 2, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF + 2, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF + 2, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF + 2, ex)] );
fxz[p] = Fdxdz * ( t_km2 - F8 * t_km1 + F8 * t_kp1 - t_kp2 );
}
/* fyz 高阶 */
{
const double t_km2 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF - 2, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF - 2, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF - 2, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF - 2, ex)] );
const double t_km1 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF - 1, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF - 1, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF - 1, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF - 1, ex)] );
const double t_kp1 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF + 1, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF + 1, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF + 1, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF + 1, ex)] );
const double t_kp2 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF + 2, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF + 2, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF + 2, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF + 2, ex)] );
fyz[p] = Fdydz * ( t_km2 - F8 * t_km1 + F8 * t_kp1 - t_kp2 );
}
}
/* 二阶分支：i±1,j±1,k±1 在范围内 */
else if ((iF + 1) <= imaxF && (iF - 1) >= iminF &&
(jF + 1) <= jmaxF && (jF - 1) >= jminF &&
(kF + 1) <= kmaxF && (kF - 1) >= kminF)
{
fxx[p] = Sdxdx * (
fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] -
TWO * fh[idx_fh_F_ord2(iF, jF, kF, ex)] +
@@ -252,17 +202,131 @@ void fdderivs(const int ex[3],
fh[idx_fh_F_ord2(iF, jF - 1, kF + 1, ex)] +
fh[idx_fh_F_ord2(iF, jF + 1, kF + 1, ex)]
);
}else{
fxx[p] = 0.0;
fyy[p] = 0.0;
fzz[p] = 0.0;
fxy[p] = 0.0;
fxz[p] = 0.0;
fyz[p] = 0.0;
}
}
}
}
if (has4) {
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);
fxx[p] = Fdxdx * (
-fh[idx_fh_F_ord2(iF - 2, jF, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF - 1, jF, kF, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF + 2, jF, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF + 1, jF, kF, ex)]
);
fyy[p] = Fdydy * (
-fh[idx_fh_F_ord2(iF, jF - 2, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF - 1, kF, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF, jF + 2, kF, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF + 1, kF, ex)]
);
fzz[p] = Fdzdz * (
-fh[idx_fh_F_ord2(iF, jF, kF - 2, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF, kF - 1, ex)] -
F30 * fh[idx_fh_F_ord2(iF, jF, kF, ex)] -
fh[idx_fh_F_ord2(iF, jF, kF + 2, ex)] +
F16 * fh[idx_fh_F_ord2(iF, jF, kF + 1, ex)]
);
{
const double t_jm2 =
( fh[idx_fh_F_ord2(iF - 2, jF - 2, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF - 2, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF - 2, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF - 2, kF, ex)] );
const double t_jm1 =
( fh[idx_fh_F_ord2(iF - 2, jF - 1, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF - 1, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF - 1, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF - 1, kF, ex)] );
const double t_jp1 =
( fh[idx_fh_F_ord2(iF - 2, jF + 1, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF + 1, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF + 1, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF + 1, kF, ex)] );
const double t_jp2 =
( fh[idx_fh_F_ord2(iF - 2, jF + 2, kF, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF + 2, kF, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF + 2, kF, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF + 2, kF, ex)] );
fxy[p] = Fdxdy * ( t_jm2 - F8 * t_jm1 + F8 * t_jp1 - t_jp2 );
}
{
const double t_km2 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF - 2, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF - 2, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF - 2, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF - 2, ex)] );
const double t_km1 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF - 1, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF - 1, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF - 1, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF - 1, ex)] );
const double t_kp1 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF + 1, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF + 1, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF + 1, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF + 1, ex)] );
const double t_kp2 =
( fh[idx_fh_F_ord2(iF - 2, jF, kF + 2, ex)]
-F8*fh[idx_fh_F_ord2(iF - 1, jF, kF + 2, ex)]
+F8*fh[idx_fh_F_ord2(iF + 1, jF, kF + 2, ex)]
- fh[idx_fh_F_ord2(iF + 2, jF, kF + 2, ex)] );
fxz[p] = Fdxdz * ( t_km2 - F8 * t_km1 + F8 * t_kp1 - t_kp2 );
}
{
const double t_km2 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF - 2, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF - 2, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF - 2, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF - 2, ex)] );
const double t_km1 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF - 1, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF - 1, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF - 1, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF - 1, ex)] );
const double t_kp1 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF + 1, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF + 1, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF + 1, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF + 1, ex)] );
const double t_kp2 =
( fh[idx_fh_F_ord2(iF, jF - 2, kF + 2, ex)]
-F8*fh[idx_fh_F_ord2(iF, jF - 1, kF + 2, ex)]
+F8*fh[idx_fh_F_ord2(iF, jF + 1, kF + 2, ex)]
- fh[idx_fh_F_ord2(iF, jF + 2, kF + 2, ex)] );
fyz[p] = Fdydz * ( t_km2 - F8 * t_km1 + F8 * t_kp1 - t_kp2 );
}
}
}
}
}
// free(fh);
}
}