diff options
Diffstat (limited to 'residual_calc.asm')
-rw-r--r-- | residual_calc.asm | 22 |
1 files changed, 11 insertions, 11 deletions
diff --git a/residual_calc.asm b/residual_calc.asm index 5eea31c..9cd530d 100644 --- a/residual_calc.asm +++ b/residual_calc.asm @@ -166,7 +166,7 @@ SECTION .text mov diff_coeffs20q, [diff_coeffsq + OFF_DIFF_COEFF_20] ; setup the data pointers and the loop counter - shl strideq, 3 + shl u_strideq, 3 shl linesizeq, 3 add dstq, linesizeq add uq, linesizeq @@ -182,14 +182,14 @@ SECTION .text %define offsetq linesizeq ; setup pointers to the line above and below - lea u_upq, [uq + strideq] + lea u_upq, [uq + u_strideq] mov u_downq, uq - sub u_downq, strideq + sub u_downq, u_strideq %if stencil == 2 - lea u_up2q, [uq + 2 * strideq] - neg strideq - add strideq, u_downq - %define u_down2q strideq ; reuse the stride register for the u[y-2] line + lea u_up2q, [uq + 2 * u_strideq] + neg u_strideq + add u_strideq, u_downq + %define u_down2q u_strideq ; reuse the stride register for the u[y-2] line movu m15, [const30] movu m14, [const8] @@ -267,18 +267,18 @@ SECTION .text %endmacro INIT_YMM fma3 -cglobal residual_calc_line_s1, 7, 14, 14, linesize, dst, res_max, stride, u, rhs, diff_coeffs,\ +cglobal residual_calc_line_s1, 7, 14, 14, linesize, dst, res_max, u_stride, u, rhs, diff_coeffs,\ diff_coeffs00, diff_coeffs01, diff_coeffs10, diff_coeffs11, diff_coeffs02, u_down, u_up RESIDUAL_CALC 1, 0 -cglobal residual_add_line_s1, 7, 14, 14, linesize, dst, res_max, stride, u, rhs, diff_coeffs,\ +cglobal residual_add_line_s1, 7, 14, 14, linesize, dst, res_max, u_stride, u, rhs, diff_coeffs,\ diff_coeffs00, diff_coeffs01, diff_coeffs10, diff_coeffs11, diff_coeffs02, u_down, u_up RESIDUAL_CALC 1, 1 INIT_YMM fma3 -cglobal residual_calc_line_s2, 7, 15, 16, linesize, dst, res_max, stride, u, rhs, diff_coeffs,\ +cglobal residual_calc_line_s2, 7, 15, 16, linesize, dst, res_max, u_stride, u, rhs, diff_coeffs,\ diff_coeffs00, diff_coeffs01, diff_coeffs10, diff_coeffs11, diff_coeffs02, u_down, u_up, u_up2 RESIDUAL_CALC 2, 0 -cglobal residual_add_line_s2, 7, 15, 16, linesize, dst, res_max, stride, u, rhs, diff_coeffs,\ +cglobal residual_add_line_s2, 7, 15, 16, linesize, dst, res_max, u_stride, u, rhs, diff_coeffs,\ diff_coeffs00, diff_coeffs01, diff_coeffs10, diff_coeffs11, diff_coeffs02, u_down, u_up, u_up2 RESIDUAL_CALC 2, 1 |