On 2016-11-23 15:00:49 +0200, Martin Storsjö wrote:
> ---
> libavcodec/aarch64/vp9itxfm_neon.S | 2 --
> 1 file changed, 2 deletions(-)
>
> diff --git a/libavcodec/aarch64/vp9itxfm_neon.S
> b/libavcodec/aarch64/vp9itxfm_neon.S
> index 2dc6b75..21352f0 100644
> --- a/libavcodec/aarch64/vp9itxfm_neon.S
> +++ b/libavcodec/aarch64/vp9itxfm_neon.S
> @@ -1001,7 +1001,6 @@ function idct32_1d_8x32_pass2_neon
>
> idct16
>
> - mov x9, #128
> .irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
> st1 {v\i\().8h}, [x2], x9
> .endr
> @@ -1018,7 +1017,6 @@ function idct32_1d_8x32_pass2_neon
>
> idct32_odd
>
> - mov x9, #128
> .macro load_acc_store a, b, c, d, neg=0
> ld1 {v4.8h}, [x2], x9
> ld1 {v5.8h}, [x2], x9
patch ok although there still many redundant mov x9 left. It's enough to
have it once in ff_vp9_idct_idct_32x32_add_neon (and a neg x7, x9) and
use x7 instead of x9 in the neg=1 part of load_acc_store
The same applies to the 16x16 transforms
Janne
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
