David Kanter (dkanter@realworldtech.com) on 7/8/10 wrote:
---------------------------
>Ralf (tuiwz407q35z4785z34@morespam.de) on 7/8/10 wrote:
>---------------------------
>>Null Pointer Exception (lol@wut.com) on 7/8/10 wrote:
>>---------------------------
>>
>>
>>>
>>>Probably not relevant. If PhysX is written in C as suggested, proper compiler options
>>>alone can handle *most* of the conversion from x87 to SSE. Naturally, this excludes
>>>any hand-tuned assembly modules or other special sauce.
>>>
>>>While it takes a focused development effort to get the maximum performance improvement
>>>from vectorized FP, the "free" jump is several keystrokes and some basic QA---and
>>>history has precious little to do with it.
>>
>>NVIDIA also delivers PhysX source code to premium license partners, so isn't it
>>up to the game developers to enable SSE flags for their games?!
>>
>>http://developer.nvidia.com/object/physx_downloads.html
>
>It's also nvidia's responsibility to ensure that those who get binaries get ones
>that are sensibly compiled, since not all partners get source. NV's decisions
>are responsible for making their own products look good (or bad).
>
>David

Our game, quite FP intensive as they normally are, gains only about 10% in performance when compiled with SSE2 instructions (does not use intrinsics), and loses about 5% of current user base that has old (mostly AMD) CPUs without SSE2.

As this is likely the normal use scenario, correct choice for distribution is x87, not SSE2 and providing source for those who need the last mile out of the source. It is also what we have chosen for our products.

Another method of delivery would naturally be an Intel-compiled multi-architecture binary, but at least earlier that used to bias towards Intel CPUs quite heavily. But then again, that would not necessarily be altogether unwanted at NV these days.