Janne Grunau <janne-li...@jannau.net> writes:
> diff --git a/libavcodec/arm/rv34dsp_neon.S b/libavcodec/arm/rv34dsp_neon.S
> index 1e8d4b4..715c242 100644
> --- a/libavcodec/arm/rv34dsp_neon.S
> +++ b/libavcodec/arm/rv34dsp_neon.S
> @@ -21,11 +21,11 @@
> #include "asm.S"
>
> .macro rv34_inv_transform
> - mov r1, #16
> - vld1.16 {d28}, [r0,:64], r1 @ block[i+8*0]
> - vld1.16 {d29}, [r0,:64], r1 @ block[i+8*1]
> - vld1.16 {d30}, [r0,:64], r1 @ block[i+8*2]
> - vld1.16 {d31}, [r0,:64], r1 @ block[i+8*3]
> + mov r1, #8
> + vld1.16 {d28}, [r0,:64], r1 @ block[i+4*0]
> + vld1.16 {d29}, [r0,:64], r1 @ block[i+4*1]
> + vld1.16 {d30}, [r0,:64], r1 @ block[i+4*2]
> + vld1.16 {d31}, [r0,:64], r1 @ block[i+4*3]
All of that can be replaced with a single vld1.16 {q14-q15}, [r0,:128]
This leaves r0 unchanged, so the code to copy/restore (forgot how it's
done) the old value can also be removed.
[...]
> /* void rv34_inv_transform_dc_noround_c(DCTELEM *block) */
> function ff_rv34_inv_transform_noround_dc_neon, export=1
> vld1.16 {d28[]}, [r0,:16] @ block[0]
> vmov.i16 d4, #251
> - vorr.s16 d4, #256 @ 13^2 * 3
> + vorr.s16 d4, #256 @ 13^2 *
Accident?
> - mov r1, #16
> + mov r1, #8
> vmull.s16 q3, d28, d4
> vshrn.s32 d0, q3, #11
> vst1.64 {d0}, [r0,:64], r1
Copy d0 to d1 and vst1.16 {q0}, [r0,:128]! twice instead.
--
Måns Rullgård
m...@mansr.com
_______________________________________________
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
