diff options
| -rw-r--r-- | residual_calc.asm | 8 |
1 files changed, 6 insertions, 2 deletions
diff --git a/residual_calc.asm b/residual_calc.asm index 95eb226..c51ba5e 100644 --- a/residual_calc.asm +++ b/residual_calc.asm @@ -64,11 +64,15 @@ SECTION .text %define up2q uq + ELEM_SIZE * 2 %define um1q uq - ELEM_SIZE %define um2q uq - ELEM_SIZE * 2 + %define coeffs1q diff_coeffs10q + %define coeffs2q diff_coeffs20q %else %define up1q u_upq %define up2q u_up2q %define um1q u_downq %define um2q u_down2q + %define coeffs1q diff_coeffs01q + %define coeffs2q diff_coeffs02q %endif ; load the function values @@ -87,7 +91,7 @@ SECTION .text addpd m11, m10 ; m11 += u[x-2] %endif mulpd m11, m2 - vfmadd231pd m0, m11, [diff_coeffs10q + offsetq] ; res += d_x u * diff_coeffs10 + vfmadd231pd m0, m11, [coeffs1q + offsetq] ; res += d_x u * diff_coeffs10 ; second derivative addpd m11, m7, m9 ; m11 = u[x+1] + u[x-1] @@ -99,7 +103,7 @@ SECTION .text %endif subpd m11, m6 ; m11 -= fd0 u[x] mulpd m11, m5 - vfmadd231pd m0, m11, [diff_coeffs20q + offsetq] ; res += d_xx u * diff_coeffs20 + vfmadd231pd m0, m11, [coeffs2q + offsetq] ; res += d_xx u * diff_coeffs20 %endmacro ; calculate and add residual contributions from the second mixed derivative |