Re: First stage of Quesa changes now complete

[Roger H. <rog...@mi...>]

On Wednesday, January 26, 2005, at 07:44  pm, James W. Walker wrote:

> Roger Holmes <rog...@mi...> wrote:
>
>> 9.7% of the time in sqrt. I have switched my renderer to use 
>> std::sqrtf as
>> it only ever handle floats, never doubles. Hopefully will speed 
>> things up.
>
> Here's a quote from a recent thread on Apple's performance 
> optimization mailing list:
>
> At 4:44 PM +0000 1/23/05, Jerry wrote:
>>>>  This should be faster, since it need compute fewer
>>>>  bits of precision, but it is still returning the correct result,
>>>>  rounded to a float.
>>>
>>>  On some architectures. On others the calculations will be in double
>>>  precision and shortened to float on return.
>>
>> This is exactly what libmx does on OS X (try stepping into it in the 
>> debugger), so switching to the "f" functions may actually slow you 
>> down.
>
> Moral:  Don't just hope, test.

Later on I did, and the times were identical.

>
> When Shark sees often-used square root operations, it suggests that if 
> you don't need full precision, you can use a special PowerPC 
> instruction, frsqrte, floating reciprocal square root estimate.  I 
> don't know if we want to get into platform-dependent optimizations 
> though.

Do they suggest how to generate such code from a high level language ?

>
>> I think in the release version we are checking out inputs too often, 
>> for instance
>> in E3Shared_Acquire we check the parameter is of type shared. We then 
>> call
>> Q3Shared_GetReference which again checks its parameter is of type 
>> shared.
>> It also calls E3System_Bottleneck even though we must have come into
>> E3Shared_Acquire from a Q3 routine which will have already called
>> E3System_Bottleneck.
>>
>> My question is this, can I change the call to Q3Shared_GetReference 
>> into a
>> call of E3Shared::GetReference (my preferred option) or can I remove 
>> the
>> check in E3Shared_Acquire  or should I leave it alone?
>
> It's OK with me if you change the Q3Shared_GetReference to 
> E3Shared::GetReference.
>
> As long as you're messing with E3Shared_Acquire, how about combining 
> the two if (theObject != NULL) statements into one?

Done.

>
>> Can I apply you answer to all similar situations or should I ask 
>> every time?
>
> You may apply my answer in similar situations.


Thanks.