perf(restrict3): shrink X-pass ii sweep to required overlap window

- compute fi_min/fi_max from output i-range and derive ii_lo/ii_hi
 - replace full ii sweep (-1:extf(1)) with windowed sweep in Z/Y precompute passes
 - keep stencil math unchanged; add bounds sanity check for ii window

AMSS_NCKU_source/prolongrestrict_cell.f90

@@ -1955,11 +1955,11 @@
 
   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(-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
+  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
@@ -2072,26 +2072,26 @@
 
   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))
-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
+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))
@@ -2352,9 +2352,10 @@ end do
 
   real*8,dimension(3) :: CD,FD
 
-  real*8 :: tmp_xz_plane(-1:extf(1), 6)
-  real*8 :: tmp_x_line(-1: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
+  integer :: fi_min, fi_max, ii_lo, ii_hi
 
   if(wei.ne.3)then
      write(*,*)"prolongrestrict.f90::restrict3: this routine only surport 3 dimension"
@@ -2436,6 +2437,18 @@ end do
 
   call symmetry_bd(2,extf,funf,funff,SoA)
 
+  ! 仅计算 X 向最终写回所需的窗口：
+  ! func(i,j,k) 只访问 tmp_x_line(fi-2:fi+3)
+  fi_min = 2*(imino + lbc(1) - 1) - 1 - lbf(1) + 1
+  fi_max = 2*(imaxo + lbc(1) - 1) - 1 - lbf(1) + 1
+  ii_lo = fi_min - 2
+  ii_hi = fi_max + 3
+  if(ii_lo < -1 .or. ii_hi > extf(1))then
+      write(*,*)"restrict3: invalid ii window",ii_lo,ii_hi
+      write(*,*)"imino,imaxo,lbc(1),lbf(1),extf(1) = ",imino,imaxo,lbc(1),lbf(1),extf(1)
+      stop
+  endif
+
 !~~~~~~> restriction start...
 do k = kmino, kmaxo
     fk = 2*(k + lbc(3) - 1) - 1 - lbf(3) + 1
@@ -2446,7 +2459,7 @@ 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 = ii_lo, ii_hi
             ! 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)) + &
@@ -2471,7 +2484,7 @@ 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 = ii_lo, ii_hi
             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))