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

cjy-dystopia #3

Merged
gh0s7 merged 13 commits from cjy-dystopia into main 2026-03-02 21:36:26 +08:00
Conversation 0 Commits 13 Files Changed 8 +29 -27
1 changed files with 29 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 466b084a58 - Show all commits

56
AMSS_NCKU_source/prolongrestrict_cell.f90
View File

@@ -1956,11 +1956,11 @@
real*8,dimension(3) :: CD,FD
real*8 :: tmp_yz(extc(1), 6) ! 存储整条 X 线上 6 个 Y 轴偏置的 Z 向插值结果
real*8 :: tmp_xyz_line(extc(1)) ! 存储整条 X 线上完成 Y 向融合后的结果
real*8 :: v1, v2, v3, v4, v5, v6
integer :: ic, jc, kc, ix_offset,ix,iy,iz,jc_min,jc_max
real*8 :: res_line
real*8 :: tmp_z_slab(extc(1), extc(2)) ! 分配在 k 循环外
real*8 :: tmp_xyz_line(-2:extc(1)) ! 包含 X 向 6 点模板访问所需下界
real*8 :: v1, v2, v3, v4, v5, v6
integer :: ic, jc, kc, ix_offset,ix,iy,iz,jc_min,jc_max,ic_min,ic_max
real*8 :: res_line
real*8 :: tmp_z_slab(-2:extc(1), -2:extc(2)) ! 包含 Y/X 向模板访问所需下界
if(wei.ne.3)then
write(*,*)"prolongrestrict.f90::prolong3: this routine only surport 3 dimension"
write(*,*)"dim = ",wei
@@ -2073,24 +2073,26 @@
call symmetry_bd(3,extc,func,funcc,SoA)
! 对每个 k（pz, kc 固定）预计算 Z 向插值的 2D 切片
jc_min = minval(ciy(jmino:jmaxo))
jc_max = maxval(ciy(jmino:jmaxo))
do k = kmino, kmaxo
pz = piz(k); kc = ciz(k)
! --- Pass 1: Z 方向，只算一次 ---
do iy = jc_min-3, jc_max+3 ! 仅需的 iy 范围
do ii = imini-3, imaxi+3 ! 仅需的 ii 范围
tmp_z_slab(ii, iy) = sum(WC(:,pz) * funcc(ii, iy, kc-2:kc+3))
end do
end do
do j = jmino, jmaxo
py = piy(j); jc = ciy(j)
! --- Pass 2: Y 方向 ---
do ii = imini-3, imaxi+3
tmp_xyz_line(ii) = sum(WC(:,py) * tmp_z_slab(ii, jc-2:jc+3))
end do
jc_min = minval(ciy(jmino:jmaxo))
jc_max = maxval(ciy(jmino:jmaxo))
ic_min = minval(cix(imino:imaxo))
ic_max = maxval(cix(imino:imaxo))
do k = kmino, kmaxo
pz = piz(k); kc = ciz(k)
! --- Pass 1: Z 方向，只算一次 ---
do iy = jc_min-2, jc_max+3 ! 仅需的 iy 范围（对应 jc-2:jc+3）
do ii = ic_min-2, ic_max+3 ! 仅需的 ii 范围（对应 cix-2:cix+3）
tmp_z_slab(ii, iy) = sum(WC(:,pz) * funcc(ii, iy, kc-2:kc+3))
end do
end do
do j = jmino, jmaxo
py = piy(j); jc = ciy(j)
! --- Pass 2: Y 方向 ---
do ii = ic_min-2, ic_max+3
tmp_xyz_line(ii) = sum(WC(:,py) * tmp_z_slab(ii, jc-2:jc+3))
end do
! --- Pass 3: X 方向 ---
do i = imino, imaxo
funf(i,j,k) = sum(WC(:,pix(i)) * tmp_xyz_line(cix(i)-2:cix(i)+3))
@@ -2351,8 +2353,8 @@ end do
real*8,dimension(3) :: CD,FD
real*8 :: tmp_xz_plane(extf(1), 6)
real*8 :: tmp_x_line(extf(1))
real*8 :: tmp_xz_plane(-1:extf(1), 6)
real*8 :: tmp_x_line(-1:extf(1))
integer :: fi, fj, fk, ii, jj, kk
if(wei.ne.3)then
@@ -2445,7 +2447,7 @@ do k = kmino, kmaxo
! 优化点 1: 显式展开 Z 方向计算，减少循环开销
! 确保 ii 循环是最内层且连续访问
!DIR$ VECTOR ALWAYS
do ii = 1, extf(1)
do ii = -1, extf(1)
! 预计算当前 j 对应的 6 行在 Z 方向的压缩结果
! 这里直接硬编码 jj 的偏移，彻底消除一层循环
tmp_xz_plane(ii, 1) = C1*(funff(ii,fj-2,fk-2)+funff(ii,fj-2,fk+3)) + &
@@ -2470,7 +2472,7 @@ do k = kmino, kmaxo
! 优化点 2: 同样向量化 Y 方向压缩
!DIR$ VECTOR ALWAYS
do ii = 1, extf(1)
do ii = -1, extf(1)
tmp_x_line(ii) = C1*(tmp_xz_plane(ii, 1) + tmp_xz_plane(ii, 6)) + &
C2*(tmp_xz_plane(ii, 2) + tmp_xz_plane(ii, 5)) + &
C3*(tmp_xz_plane(ii, 3) + tmp_xz_plane(ii, 4))