Compare commits
4 Commits
main
...
yx-prolong
| Author | SHA1 | Date | |
|---|---|---|---|
| 12e1f63d50 | |||
| 47f91ff46f | |||
|
672b7ebee2 | ||
|
|
63bf180159 |
@@ -1955,7 +1955,12 @@
|
||||
integer::maxcx,maxcy,maxcz
|
||||
|
||||
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
|
||||
if(wei.ne.3)then
|
||||
write(*,*)"prolongrestrict.f90::prolong3: this routine only surport 3 dimension"
|
||||
write(*,*)"dim = ",wei
|
||||
@@ -2067,121 +2072,86 @@
|
||||
endif
|
||||
|
||||
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
|
||||
! --- 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))
|
||||
end do
|
||||
end do
|
||||
end do
|
||||
|
||||
!~~~~~~> prolongation start...
|
||||
do k = kmino,kmaxo
|
||||
do j = jmino,jmaxo
|
||||
do i = imino,imaxo
|
||||
cxI(1) = cix(i)
|
||||
cxI(2) = ciy(j)
|
||||
cxI(3) = ciz(k)
|
||||
px = pix(i)
|
||||
py = piy(j)
|
||||
pz = piz(k)
|
||||
#if 0
|
||||
if(ii/2*2==ii)then
|
||||
if(jj/2*2==jj)then
|
||||
if(kk/2*2==kk)then
|
||||
tmp2= C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C1*tmp2(:,1)+C2*tmp2(:,2)+C3*tmp2(:,3)+C4*tmp2(:,4)+C5*tmp2(:,5)+C6*tmp2(:,6)
|
||||
funf(i,j,k)= C1*tmp1(1)+C2*tmp1(2)+C3*tmp1(3)+C4*tmp1(4)+C5*tmp1(5)+C6*tmp1(6)
|
||||
else
|
||||
tmp2= C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C1*tmp2(:,1)+C2*tmp2(:,2)+C3*tmp2(:,3)+C4*tmp2(:,4)+C5*tmp2(:,5)+C6*tmp2(:,6)
|
||||
funf(i,j,k)= C1*tmp1(1)+C2*tmp1(2)+C3*tmp1(3)+C4*tmp1(4)+C5*tmp1(5)+C6*tmp1(6)
|
||||
endif
|
||||
else
|
||||
if(kk/2*2==kk)then
|
||||
tmp2= C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C6*tmp2(:,1)+C5*tmp2(:,2)+C4*tmp2(:,3)+C3*tmp2(:,4)+C2*tmp2(:,5)+C1*tmp2(:,6)
|
||||
funf(i,j,k)= C1*tmp1(1)+C2*tmp1(2)+C3*tmp1(3)+C4*tmp1(4)+C5*tmp1(5)+C6*tmp1(6)
|
||||
else
|
||||
tmp2= C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C6*tmp2(:,1)+C5*tmp2(:,2)+C4*tmp2(:,3)+C3*tmp2(:,4)+C2*tmp2(:,5)+C1*tmp2(:,6)
|
||||
funf(i,j,k)= C1*tmp1(1)+C2*tmp1(2)+C3*tmp1(3)+C4*tmp1(4)+C5*tmp1(5)+C6*tmp1(6)
|
||||
endif
|
||||
endif
|
||||
else
|
||||
if(jj/2*2==jj)then
|
||||
if(kk/2*2==kk)then
|
||||
tmp2= C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C1*tmp2(:,1)+C2*tmp2(:,2)+C3*tmp2(:,3)+C4*tmp2(:,4)+C5*tmp2(:,5)+C6*tmp2(:,6)
|
||||
funf(i,j,k)= C6*tmp1(1)+C5*tmp1(2)+C4*tmp1(3)+C3*tmp1(4)+C2*tmp1(5)+C1*tmp1(6)
|
||||
else
|
||||
tmp2= C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C1*tmp2(:,1)+C2*tmp2(:,2)+C3*tmp2(:,3)+C4*tmp2(:,4)+C5*tmp2(:,5)+C6*tmp2(:,6)
|
||||
funf(i,j,k)= C6*tmp1(1)+C5*tmp1(2)+C4*tmp1(3)+C3*tmp1(4)+C2*tmp1(5)+C1*tmp1(6)
|
||||
endif
|
||||
else
|
||||
if(kk/2*2==kk)then
|
||||
tmp2= C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C6*tmp2(:,1)+C5*tmp2(:,2)+C4*tmp2(:,3)+C3*tmp2(:,4)+C2*tmp2(:,5)+C1*tmp2(:,6)
|
||||
funf(i,j,k)= C6*tmp1(1)+C5*tmp1(2)+C4*tmp1(3)+C3*tmp1(4)+C2*tmp1(5)+C1*tmp1(6)
|
||||
else
|
||||
tmp2= C6*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
C5*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
C4*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
C3*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
C2*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
C1*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
tmp1= C6*tmp2(:,1)+C5*tmp2(:,2)+C4*tmp2(:,3)+C3*tmp2(:,4)+C2*tmp2(:,5)+C1*tmp2(:,6)
|
||||
funf(i,j,k)= C6*tmp1(1)+C5*tmp1(2)+C4*tmp1(3)+C3*tmp1(4)+C2*tmp1(5)+C1*tmp1(6)
|
||||
endif
|
||||
endif
|
||||
endif
|
||||
#else
|
||||
tmp2= WC(1,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-2)+&
|
||||
WC(2,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)-1)+&
|
||||
WC(3,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3) )+&
|
||||
WC(4,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+1)+&
|
||||
WC(5,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+2)+&
|
||||
WC(6,pz)*funcc(cxI(1)-2:cxI(1)+3,cxI(2)-2:cxI(2)+3,cxI(3)+3)
|
||||
do k = kmino, kmaxo
|
||||
pz = piz(k)
|
||||
kc = ciz(k)
|
||||
|
||||
tmp1= WC(1,py)*tmp2(:,1)+WC(2,py)*tmp2(:,2)+WC(3,py)*tmp2(:,3)+&
|
||||
WC(4,py)*tmp2(:,4)+WC(5,py)*tmp2(:,5)+WC(6,py)*tmp2(:,6)
|
||||
do j = jmino, jmaxo
|
||||
py = piy(j)
|
||||
jc = ciy(j)
|
||||
|
||||
funf(i,j,k)= WC(1,px)*tmp1(1)+WC(2,px)*tmp1(2)+WC(3,px)*tmp1(3)+&
|
||||
WC(4,px)*tmp1(4)+WC(5,px)*tmp1(5)+WC(6,px)*tmp1(6)
|
||||
! --- 步骤 1 & 2 融合：分段处理 X 轴，提升 Cache 命中率 ---
|
||||
! 我们将 ii 循环逻辑重组，减少对 funcc 的跨行重复访问
|
||||
do ii = 1, extc(1)
|
||||
! 1. 先做 Z 方向的 6 条线插值（针对当前的 ii 和当前的 6 个 iy）
|
||||
! 我们直接在这里把 Y 方向的加权也做了，省去 tmp_yz 数组
|
||||
! 这样 funcc 的数据读进来后立即完成所有维度的贡献，不再写回内存
|
||||
|
||||
res_line = 0.0d0
|
||||
do jj = 1, 6
|
||||
iy = jc - 3 + jj
|
||||
! 这一行代码是核心：一次性完成 Z 插值并加上 Y 的权重
|
||||
! 编译器会把 WC(jj, py) 存在寄存器里
|
||||
res_line = res_line + WC(jj, py) * ( &
|
||||
WC(1, pz) * funcc(ii, iy, kc-2) + &
|
||||
WC(2, pz) * funcc(ii, iy, kc-1) + &
|
||||
WC(3, pz) * funcc(ii, iy, kc ) + &
|
||||
WC(4, pz) * funcc(ii, iy, kc+1) + &
|
||||
WC(5, pz) * funcc(ii, iy, kc+2) + &
|
||||
WC(6, pz) * funcc(ii, iy, kc+3) )
|
||||
end do
|
||||
tmp_xyz_line(ii) = res_line
|
||||
end do
|
||||
|
||||
|
||||
|
||||
|
||||
! 3. 【降维：X 向】最后在最内层只处理 X 方向的 6 点加权
|
||||
! 此时每个点的计算量从原来的 200+ 次乘法降到了仅 6 次
|
||||
do i = imino, imaxo
|
||||
px = pix(i)
|
||||
ic = cix(i)
|
||||
|
||||
! 直接从预计算好的 line 中读取连续的 6 个点
|
||||
! ic-2 到 ic+3 对应原始 6 点算子
|
||||
funf(i,j,k) = WC(1,px)*tmp_xyz_line(ic-2) + &
|
||||
WC(2,px)*tmp_xyz_line(ic-1) + &
|
||||
WC(3,px)*tmp_xyz_line(ic ) + &
|
||||
WC(4,px)*tmp_xyz_line(ic+1) + &
|
||||
WC(5,px)*tmp_xyz_line(ic+2) + &
|
||||
WC(6,px)*tmp_xyz_line(ic+3)
|
||||
end do
|
||||
end do
|
||||
end do
|
||||
#endif
|
||||
enddo
|
||||
enddo
|
||||
enddo
|
||||
|
||||
return
|
||||
|
||||
end subroutine prolong3
|
||||
@@ -2381,6 +2351,10 @@
|
||||
|
||||
real*8,dimension(3) :: CD,FD
|
||||
|
||||
real*8 :: tmp_xz_plane(extf(1), 6)
|
||||
real*8 :: tmp_x_line(extf(1))
|
||||
integer :: fi, fj, fk, ii, jj, kk
|
||||
|
||||
if(wei.ne.3)then
|
||||
write(*,*)"prolongrestrict.f90::restrict3: this routine only surport 3 dimension"
|
||||
write(*,*)"dim = ",wei
|
||||
@@ -2462,6 +2436,56 @@
|
||||
call symmetry_bd(2,extf,funf,funff,SoA)
|
||||
|
||||
!~~~~~~> restriction start...
|
||||
do k = kmino, kmaxo
|
||||
fk = 2*(k + lbc(3) - 1) - 1 - lbf(3) + 1
|
||||
|
||||
do j = jmino, jmaxo
|
||||
fj = 2*(j + lbc(2) - 1) - 1 - lbf(2) + 1
|
||||
|
||||
! 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)
|
||||
! 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)) + &
|
||||
C2*(funff(ii,fj-2,fk-1)+funff(ii,fj-2,fk+2)) + &
|
||||
C3*(funff(ii,fj-2,fk )+funff(ii,fj-2,fk+1))
|
||||
tmp_xz_plane(ii, 2) = C1*(funff(ii,fj-1,fk-2)+funff(ii,fj-1,fk+3)) + &
|
||||
C2*(funff(ii,fj-1,fk-1)+funff(ii,fj-1,fk+2)) + &
|
||||
C3*(funff(ii,fj-1,fk )+funff(ii,fj-1,fk+1))
|
||||
tmp_xz_plane(ii, 3) = C1*(funff(ii,fj ,fk-2)+funff(ii,fj ,fk+3)) + &
|
||||
C2*(funff(ii,fj ,fk-1)+funff(ii,fj ,fk+2)) + &
|
||||
C3*(funff(ii,fj ,fk )+funff(ii,fj ,fk+1))
|
||||
tmp_xz_plane(ii, 4) = C1*(funff(ii,fj+1,fk-2)+funff(ii,fj+1,fk+3)) + &
|
||||
C2*(funff(ii,fj+1,fk-1)+funff(ii,fj+1,fk+2)) + &
|
||||
C3*(funff(ii,fj+1,fk )+funff(ii,fj+1,fk+1))
|
||||
tmp_xz_plane(ii, 5) = C1*(funff(ii,fj+2,fk-2)+funff(ii,fj+2,fk+3)) + &
|
||||
C2*(funff(ii,fj+2,fk-1)+funff(ii,fj+2,fk+2)) + &
|
||||
C3*(funff(ii,fj+2,fk )+funff(ii,fj+2,fk+1))
|
||||
tmp_xz_plane(ii, 6) = C1*(funff(ii,fj+3,fk-2)+funff(ii,fj+3,fk+3)) + &
|
||||
C2*(funff(ii,fj+3,fk-1)+funff(ii,fj+3,fk+2)) + &
|
||||
C3*(funff(ii,fj+3,fk )+funff(ii,fj+3,fk+1))
|
||||
end do
|
||||
|
||||
! 优化点 2: 同样向量化 Y 方向压缩
|
||||
!DIR$ VECTOR ALWAYS
|
||||
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))
|
||||
end do
|
||||
|
||||
! 优化点 3: 最终写入，利用已经缓存在 tmp_x_line 的数据
|
||||
do i = imino, imaxo
|
||||
fi = 2*(i + lbc(1) - 1) - 1 - lbf(1) + 1
|
||||
func(i, j, k) = C1*(tmp_x_line(fi-2) + tmp_x_line(fi+3)) + &
|
||||
C2*(tmp_x_line(fi-1) + tmp_x_line(fi+2)) + &
|
||||
C3*(tmp_x_line(fi ) + tmp_x_line(fi+1))
|
||||
end do
|
||||
end do
|
||||
end do
|
||||
#if 0
|
||||
do k = kmino,kmaxo
|
||||
do j = jmino,jmaxo
|
||||
do i = imino,imaxo
|
||||
@@ -2485,7 +2509,7 @@
|
||||
enddo
|
||||
enddo
|
||||
enddo
|
||||
|
||||
#endif
|
||||
return
|
||||
|
||||
end subroutine restrict3
|
||||
|
||||
Reference in New Issue
Block a user