Re: [MPlayer-dev-eng] [PATCH 4/6] libvo: optimize fastmemcpy for loongson with mmi
- Date: Mon, 7 Sep 2015 11:17:18 +0800 (GMT+08:00)
- From: 周晓勇 <zhouxiaoyong@xxxxxxxxxxx>
- Subject: Re: [MPlayer-dev-eng] [PATCH 4/6] libvo: optimize fastmemcpy for loongson with mmi
we have do some work in glibc about memcpy, strcmp, memset, etc
but for some reason the patchs have not been pushed into the official
as cpu 3A1000 with old GS464 core has poor memory unit,
using 128bit copy instruction may improve the VO efficiency evidently
libc: CPU clocks=276897us (361.145fps) 541.7MB/s
MMI: CPU clocks=201757us (495.646fps) 743.5MB/s
i think the fastmemcpy in mplayer is used only for large blocks,
even using optimized glibc-memcpy, when doing copy large blocks,
it will finally call the branch of memcpy written by 128bit mem-instr.
> On Wed, Sep 02, 2015 at 06:07:10PM +0800, 周晓勇 wrote:
> > From b2709e8c8034977b21b239aaf41f291878b8199c Mon Sep 17 00:00:00 2001
> > From: ZhouXiaoyong <zhouxiaoyong@xxxxxxxxxxx>
> > Date: Thu, 20 Aug 2015 17:26:57 +0800
> > Subject: [PATCH 4/6] libvo: optimize fastmemcpy for loongson with mmi
> Isn't it possible to just make sure the libc memcpy is optimized
> well enough to make this unnecessary?
> I'd prefer to get rid of this code long-term, as having our own
> memcpy implementation isn't really maintainable.
> MPlayer-dev-eng mailing list
MPlayer-dev-eng mailing list