Cg lerp() performance on Geforce FX 5600

[Cirrus]

Hello,

I started a thread on cgshaders.org about this issue and heard that I'm not the only one that noticed this.

The lerp function (which translates to an ADD and a MAD in arbfp1.0 assembly) significantly lowers performance on my Geforce FX 5600.

With the fragment program disabled I get anywhere from 300-330 fps while with the fragment program enabled it lowers to 70 fps with a single lerp and it goes down to 25 with 4 lerps (using 5 textures).

Does anyone know the reason why?

For reference (shader source code is there too), here's a link to the other thread: http://www.cgshaders.org/forums/viewtopic.php?t=2244

Thanks for reading.

[jwatte]

Just google for it -- apparently, if reports are to believed, the NV30 line (on which I believe 5600 is based) isn't as optimized for floating-point fragment programs as it is for fixed-point.

If you're using GC, try compiling for the nv30 profile, and using fixed variables rather than float or half.

[This message has been edited by jwatte (edited 11-19-2003).]

[Cirrus]

Thanks for the tip, that was definitely part of it. Could you post a link to where you found that?

I'm now getting from 50-70 fps which is much better than 25. There was no change switching to fixed point with the arb profile, while the fp30 definitely improved things.

However it's a long fall from 300 to 70, are these operations really so costly?

Thanks for the help.

[Cirrus]

The great benefit of Cg is that it will handle target specifics like lack of fixed or half datatypes by automatically up-casting. You could produce a shader using fixed types optimised for the FP30 profile (Nvidia+OpenGL+NV30) and the same Cg shader can be compiled, unchanged, to ARB OpenGL or DX9 PS2 targets.

Got that from http://www.pocketmoon.com/Cg/Cg.html

Has some interesting information.