prolong3: 减少Z-pass 冗余计算

AMSS_NCKU_source/prolongrestrict_cell.f90

@@ -1957,8 +1957,9 @@
   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
+  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
@@ -2070,8 +2071,34 @@
   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...
+#if 0
  do k = kmino, kmaxo
      pz = piz(k)
      kc = ciz(k)
@@ -2105,28 +2132,7 @@
 
 
 
-#if 0
-        ! 1. 【降维：Z 向】对当前 (j,k) 相关的 6 条 Y 偏置线进行 Z 向插值
-        ! 结果存入 tmp_yz(x_index, y_offset)
-        do jj = 1, 6
-           iy = jc - 3 + jj
-           do ii = 1, extc(1)
-              tmp_yz(ii, jj) = 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
-        end do
 
-        ! 2. 【降维：Y 向】将 Z 向结果合并，得到整条 X 轴线上的 Y-Z 融合值
-        do ii = 1, extc(1)
-           tmp_xyz_line(ii) = WC(1,py)*tmp_yz(ii, 1) + WC(2,py)*tmp_yz(ii, 2) + &
-                              WC(3,py)*tmp_yz(ii, 3) + WC(4,py)*tmp_yz(ii, 4) + &
-                              WC(5,py)*tmp_yz(ii, 5) + WC(6,py)*tmp_yz(ii, 6)
-        end do
-#endif
         ! 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
@@ -2144,7 +2150,7 @@
         end do
      end do
   end do
-
+#endif
   return
 
   end subroutine prolong3
@@ -2343,7 +2349,11 @@
   integer::imino,imaxo,jmino,jmaxo,kmino,kmaxo
 
   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
@@ -2425,6 +2435,56 @@
   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
@@ -2448,7 +2508,7 @@
     enddo
    enddo
   enddo
-  
+#endif
   return
 
   end subroutine restrict3