SQRT FP16 performance on the NV35 - What the heck?!

[Arun]

Hey everyone,

I just did a few tests... And I got some VERY strange results ( still running 45 Series driver here, BTW )

If I put SQRT alone, thus loads of SQRT instructions in a row, I get 35% faster performance in FP16 than in FP32.
But, if I do a lot of SQRTs with MULs, ADDs and MADs surrounding these SQRTs, I get 2% swloer performance in FP16 than in FP32

I did these tests by modifying the .fx file of RightMark3D ( the "Marble" shader ) , thus it's DX9 HLSL.

I've yet to try it with the Det50s - but it does seem strange! Also, with the Det45s, COS/SIN doesn't seem to be done in parallel to the other stuff. COS/SIN also doesn't have the strange behavior SQRT has.


Uttar

 

[Xmas]

Are those sqrt (RSQ in assembly actually) dependent or independent instructions? I.e. are you testing latency or throughput? 2% seems to fall into the inaccurate measurement range. How many temp values are you using?

 

[Arun]

Hmm, strange stuff, might have done some mistakes.
Det50s do give rather different results too.

I'll try redoing some tests in the coming days.


Uttar

 

[rwolf]

That doesn't sound strange that sounds like a pipelined architecture.