residual_calc.asm: templatize computing the mixed derivative

1 files changed, 39 insertions, 35 deletions

diff --git a/residual_calc.asm b/residual_calc.asm
index bfb6ce3..6424b2f 100644
--- a/residual_calc.asm
+++ b/residual_calc.asm
@@ -35,7 +35,6 @@ SECTION .rodata
 const8:  times 8 dq  8.0
 const16: times 8 dq 16.0
 const30: times 8 dq 30.0
-const64: times 8 dq 64.0
 
 SECTION .text
 
@@ -101,6 +100,43 @@ SECTION .text
     vfmadd231pd m0, m11, [diff_coeffs20q + offsetq]  ; res += d_xx u * diff_coeffs20
 %endmacro
 
+; calculate and add residual contributions from the second mixed derivative
+;
+; parameters:
+; %1: stencil
+;
+; register use (in addition to register allocation described above):
+; m6, m7: used for work (clobbered)
+%macro RES_ADD_DIFF_MIXED 1
+    movu  m6, [u_upq    + 1 * ELEM_SIZE + offsetq] ; m6 = u[y+1, x+1]
+    subpd m6, [u_upq    - 1 * ELEM_SIZE + offsetq] ;    - u[y+1, x-1]
+    subpd m6, [u_downq  + 1 * ELEM_SIZE + offsetq] ;    - u[y-1, x+1]
+    addpd m6, [u_downq  - 1 * ELEM_SIZE + offsetq] ;    + u[y-1, x-1]
+
+    %if %1 == 2
+        movu  m7, [u_up2q   - 1 * ELEM_SIZE + offsetq] ; m7 = u[y+2, x-1]
+        subpd m7, [u_up2q   + 1 * ELEM_SIZE + offsetq] ;    - u[y+2, x+1]
+        subpd m7, [u_upq    + 2 * ELEM_SIZE + offsetq] ;    - u[y+1, x+2]
+        addpd m7, [u_upq    - 2 * ELEM_SIZE + offsetq] ;    + u[y+1, x-2]
+        addpd m7, [u_downq  + 2 * ELEM_SIZE + offsetq] ;    + u[y-1, x+2]
+        subpd m7, [u_downq  - 2 * ELEM_SIZE + offsetq] ;    - u[y-1, x-2]
+        addpd m7, [u_down2q + 1 * ELEM_SIZE + offsetq] ;    + u[y-2, x+1]
+        subpd m7, [u_down2q - 1 * ELEM_SIZE + offsetq] ;    - u[y-2, x-1]
+
+        vfmadd123pd m6, m14, m7                        ; m6 = 8 m6 + m7
+
+        movu  m7, [u_up2q   + 2 * ELEM_SIZE + offsetq] ; m7 = u[y+2, x+2]
+        subpd m7, [u_up2q   - 2 * ELEM_SIZE + offsetq] ;    - u[y+2, x-2]
+        subpd m7, [u_down2q + 2 * ELEM_SIZE + offsetq] ;    - u[y-2, x+2]
+        addpd m7, [u_down2q - 2 * ELEM_SIZE + offsetq] ;    + u[y-2, x-2]
+
+        vfmadd123pd m6, m14, m7                        ; m6 = 8 m6 + m7
+    %endif
+
+    mulpd       m6, m3
+    vfmadd231pd m0, m6, [diff_coeffs11q + offsetq]     ; res += d_xy u * diff_coeffs11
+%endmacro
+
 INIT_YMM fma3
 cglobal residual_calc_line_s1, 7, 13, 12, linesize, dst, stride, u, rhs, diff_coeffs, fd_factors,\
                                           diff_coeffs00, diff_coeffs01, diff_coeffs10, diff_coeffs11, diff_coeffs02, u_up
@@ -156,15 +192,7 @@ cglobal residual_calc_line_s1, 7, 13, 12, linesize, dst, stride, u, rhs, diff_co
 
     RES_ADD_DIFF_SINGLEDIR 1, 0
     RES_ADD_DIFF_SINGLEDIR 1, 1
-
-    ; mixed d2xy
-    movu  m6, [u_upq   + offsetq + ELEM_SIZE]
-    subpd m6, [u_upq   + offsetq - ELEM_SIZE]
-    subpd m6, [u_downq + offsetq + ELEM_SIZE]
-    addpd m6, [u_downq + offsetq - ELEM_SIZE]
-
-    mulpd m6, m3
-    vfmadd231pd m0, m6, [diff_coeffs11q + offsetq]      ; res += d_xy u * diff_coeffs11
+    RES_ADD_DIFF_MIXED     1
 
     ; store the result
     movu [dstq + offsetq], m0
@@ -214,7 +242,6 @@ cglobal residual_calc_line_s2, 7, 15, 16, linesize, dst, stride, u, rhs, diff_co
     movu    m15, [const30]
     movu    m14, [const8]
     movu    m13, [const16]
-    movu    m12, [const64]
 
     ; setup pointers to the lines above and below
     lea u_upq,   [uq +     strideq]
@@ -237,30 +264,7 @@ cglobal residual_calc_line_s2, 7, 15, 16, linesize, dst, stride, u, rhs, diff_co
 
     RES_ADD_DIFF_SINGLEDIR 2, 0
     RES_ADD_DIFF_SINGLEDIR 2, 1
-
-    ; mixed d2xy
-    movu         m6,      [u_up2q + offsetq + 2 * ELEM_SIZE]   ; m6  =   u[y+2, x+2]
-    vfnmadd231pd m6, m14, [u_up2q + offsetq + 1 * ELEM_SIZE]   ; m6 -= 8 u[y+2, x+1]
-    vfmadd231pd  m6, m14, [u_up2q + offsetq - 1 * ELEM_SIZE]   ; m6 += 8 u[y+2, x-1]
-    subpd        m6,      [u_up2q + offsetq - 2 * ELEM_SIZE]   ; m6 -=   u[y+2, x-2]
-
-    vfnmadd231pd m6, m14, [u_upq  + offsetq + 2 * ELEM_SIZE]   ; m6 -=  8 u[y+1, x+2]
-    vfmadd231pd  m6, m12, [u_upq  + offsetq + 1 * ELEM_SIZE]   ; m6 += 64 u[y+1, x+1]
-    vfnmadd231pd m6, m12, [u_upq  + offsetq - 1 * ELEM_SIZE]   ; m6 -= 64 u[y+1, x-1]
-    vfmadd231pd  m6, m14, [u_upq  + offsetq - 2 * ELEM_SIZE]   ; m6 +=  8 u[y+1, x-2]
-
-    vfmadd231pd   m6, m14, [u_downq  + offsetq + 2 * ELEM_SIZE] ; m6 +=  8 u[y-1, x+2]
-    vfnmadd231pd  m6, m12, [u_downq  + offsetq + 1 * ELEM_SIZE] ; m6 -= 64 u[y-1, x+1]
-    vfmadd231pd   m6, m12, [u_downq  + offsetq - 1 * ELEM_SIZE] ; m6 += 64 u[y-1, x-1]
-    vfnmadd231pd  m6, m14, [u_downq  + offsetq - 2 * ELEM_SIZE] ; m6 -=  8 u[y-1, x-2]
-
-    subpd         m6,      [u_down2q + offsetq + 2 * ELEM_SIZE] ; m6 -=   u[y-2, x+2]
-    vfmadd231pd   m6, m14, [u_down2q + offsetq + 1 * ELEM_SIZE] ; m6 += 8 u[y-2, x+1]
-    vfnmadd231pd  m6, m14, [u_down2q + offsetq - 1 * ELEM_SIZE] ; m6 += 8 u[y-2, x-1]
-    addpd         m6,      [u_down2q + offsetq - 2 * ELEM_SIZE] ; m6 +=   u[y-2, x-2]
-
-    mulpd       m6, m3
-    vfmadd231pd m0, m6, [diff_coeffs11q + offsetq]             ; res += d_xy u * diff_coeffs11
+    RES_ADD_DIFF_MIXED     2
 
     ; store the result
     movu [dstq + offsetq], m0