Can the R300 issue a 4-component vector operation in its PS?

[Luminescent]

So let me get this straight, there are 3 fmads, specifically designed for vector ops, but there is also a scalar unit which carries its own fmad and can operate along with the vectors for a 4-component vector op. Now the scalar unit also has a simplified fmad for the table lookups of the complex functions. Thus, in total there are 4 complete fmads and one simplified fmad, along with the complex operation units in the scalar processor. Is this right? Sounds to me like the scalar unit is like a mini dsp. At what precision level do the complex scalar units operate?

Thankyou for all the information sireric, the simple facts are appreciated. Usually all the cpu internals are expressed and described explicitly but gpu's are left a mystery. I just wanted to clear some of the smog.

 

[sireric]

Every single part operates at 24b FP precision (the thought for that was that for up at least 2kx2k texture lookup, that's all that's required, even with 4b subprecision; basically, errors will come from the lack of resolution of the textures, not the texture coordinates). Anyway, you can actually look at it as 4 FMADs, though the instruction for 3 of them comes from a SIMD, while the 4th unit has its own instruction set. The 4th one also has some extra functions available.

 

[sireric]

Also, nearly forgot, there's the LERP, input modifier and output mod (though that's not much) HW on top of the FMADs & scalar.

 

[Zephyr]

sireric, thank you for your clear description.