arm: vp9itxfm16: Fix vertical alignment

Signed-off-by: Martin Storsjö <martin@martin.st>
pull/244/merge
Martin Storsjö 8 years ago
parent b46d37e93a
commit c1619318e5
  1. 20
      libavcodec/arm/vp9itxfm_16bpp_neon.S

@ -1395,25 +1395,25 @@ function idct32_1d_2x32_pass2_neon
vld1.32 {d4}, [r2,:64], r12
vld1.32 {d5}, [r2,:64], r12
.if \neg == 0
vadd.s32 d4, d4, d\a
vadd.s32 d4, d4, d\a
vld1.32 {d6}, [r2,:64], r12
vadd.s32 d5, d5, d\b
vadd.s32 d5, d5, d\b
vld1.32 {d7}, [r2,:64], r12
vadd.s32 d6, d6, d\c
vadd.s32 d7, d7, d\d
vadd.s32 d6, d6, d\c
vadd.s32 d7, d7, d\d
.else
vsub.s32 d4, d4, d\a
vsub.s32 d4, d4, d\a
vld1.32 {d6}, [r2,:64], r12
vsub.s32 d5, d5, d\b
vsub.s32 d5, d5, d\b
vld1.32 {d7}, [r2,:64], r12
vsub.s32 d6, d6, d\c
vsub.s32 d7, d7, d\d
vsub.s32 d6, d6, d\c
vsub.s32 d7, d7, d\d
.endif
vld1.32 {d2[]}, [r0,:32], r1
vld1.32 {d2[1]}, [r0,:32], r1
vrshr.s32 q2, q2, #6
vrshr.s32 q2, q2, #6
vld1.32 {d3[]}, [r0,:32], r1
vrshr.s32 q3, q3, #6
vrshr.s32 q3, q3, #6
vld1.32 {d3[1]}, [r0,:32], r1
sub r0, r0, r1, lsl #2
vaddw.u16 q2, q2, d2

Loading…
Cancel
Save