Re: glTexSubImage2D long download time with 1x1 texel level on IR

New Message Reply Date view Thread view Subject view Author view

Celeste Fowler (celeste++at++obelix.engr.sgi.com)
Fri, 12 Mar 1999 10:04:10 -0800 (PST)


Mark --

After consulting with local texturing experts (in the person of Mark Peercy) it
appears that this is a hardware "feature" that we can't get around. The IR has
to replicate coarse LOD's -- 2x2 is replicated twice and 1x1 is replicated 4
times. The cost is higher because the setup overhead is repeated 4 times (in
the 1x1 case this will be much more time-consuming than actually transferring
the data).

As for more performance, the solution we suggest is to avoid downloading the
coarse levels of detail if at all possible (are they really changing that
much?)

Hope this helps (or at least informs). Please let us know if you have other
questions.

-Celeste

On Mar 9, 9:08am, Mark Lewis wrote:
> Subject: glTexSubImage2D long download time with 1x1 texel level on IR
>
> Dear Performers,
>
> Although not strictly a Performer issue, I would like to hear from
> anyone who can shed light on a problem I have with glTexSubImage2D (and
> glTexSubImage2DEXT). This function can be used to download specific
> sub-areas of specific LODs of a textures in TRAM. It works fine for me
> for all but the highest LOD (the 1x1 texel level), when it takes an
> unusually long time to complete. The sort of times I am measuring are
> shown below:
>
> 3 components: 512 x 512 : 3.53295 ms
> 3 components: 256 x 256 : 1.24431 ms
> 3 components: 128 x 128 : 0.658368 ms
> 3 components: 64 x 64 : 0.410448 ms
> 3 components: 32 x 32 : 0.277744 ms
> 3 components: 16 x 16 : 0.267592 ms
> 3 components: 8 x 8 : 0.30048 ms
> 3 components: 4 x 4 : 0.354984 ms
> 3 components: 2 x 2 : 0.87924 ms
> 3 components: 1 x 1 : 3.0324 ms
>
> You can see that the 1x1 level takes 3.0324 milliseconds, compared to
> only 0.87924 ms for the 2x2 level.
>
> It is also worth noting that the same program run on an O2 behaves much
> better, in that the time to download the 1x1 level is about the same as
> the 2x2 level.
>
> Is there a way of getting more sensible performance out of the IR in
> this respect (perhaps a patch) ?
>
> My machine details are attached.
>
> Thanks in advance for any help offered.
>
> --
> Mark C Lewis
>
> Software Development Manager
> Equipe Electronics Ltd.
>
> "Visualising Your Imagination"
>
> Phone +44 (0) 1903 216 622
> FAX +44 (0) 1903 216 633
>
> mailto:markl++at++equipe.ltd.uk
> http://www.equipe.ltd.uk
>
> 13 Liverpool Gardens
> Worthing, W. Sussex
> BN11 1RY
> ENGLAND
>
> [ Attachment (text/plain): "sade.info" 52107 bytes
> Character set: us-ascii ]
>-- End of excerpt from Mark Lewis

-- 
------------------------------------------------------------
Celeste Fowler
celeste++at++sgi.com
------------------------------------------------------------

New Message Reply Date view Thread view Subject view Author view

This archive was generated by hypermail 2.0b2 on Fri Mar 12 1999 - 10:04:30 PST

This message has been cleansed for anti-spam protection. Replace '++at++' in any mail addresses with the '@' symbol.