[Beignet] [PATCH] GBE: fix a src/dst register reuse bug.
Pan, Xiuli
xiuli.pan at intel.com
Wed Dec 21 02:43:32 UTC 2016
I got the point.
In some case, such as block read gemm. We may get 16 mad with g4, thus there may be some other mad using g4.2 or g4.3 later.
If we reuse the g4, the g4.2 or g4.3 may be overwrite. In this case may be g4.1 is just a scalar value, and g4.2 or g4.* can be some other scalar value. This fix check if the src and dst is all uniform. It should fix this case.
-----Original Message-----
From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of Song, Ruiling
Sent: Tuesday, December 20, 2016 8:28 PM
To: Yang, Rong R <rong.r.yang at intel.com>; beignet at lists.freedesktop.org
Cc: Yang, Rong R <rong.r.yang at intel.com>
Subject: Re: [Beignet] [PATCH] GBE: fix a src/dst register reuse bug.
> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf
> Of Yang Rong
> Sent: Tuesday, December 20, 2016 4:47 PM
> To: beignet at lists.freedesktop.org
> Cc: Yang, Rong R <rong.r.yang at intel.com>
> Subject: [Beignet] [PATCH] GBE: fix a src/dst register reuse bug.
>
> For case:
> mad(8) g4<1>:F g4.1<0,1,0>:F g127.7<0,1,0>:F g46:F
> src0 is uniform, dst is non-uniform, dst can't reuse the src0.
why g4.0 cannot be reused here? What's wrong here?
_______________________________________________
Beignet mailing list
Beignet at lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/beignet
More information about the Beignet
mailing list