topic Re: Bug in IppsSqrt_32f_I in Intel® Integrated Performance Primitives

Bug in IppsSqrt_32f_I

gol — Sun, 20 Jul 2008 06:56:29 GMT

IppsSqrt_32f_I doesn't like denormals.

In version 5.2, it all worked fine. Starting from 5.3 (still in 6 beta), denormals are turned into -infinite, even though the function returns 0 as no error.

We can of course filter out denormals first, but I think the function should be able to handle them for compatibility.

Re: Bug in IppsSqrt_32f_I

gol — Mon, 21 Jul 2008 04:05:30 GMT

This problem raises another, btw.

My thread where I call this Sqrt normally has FTZ & DAZ modes set. So if this Sqrt function returns -Inf for them, it can mean one of these:

-that function switches the FTZ or DAZ off (I doubt)

-that function performs in integer (most likely)

-that function has a special case for denormals (& it's then buggy)

-starting from 5.3, it would use a side thread (in which it wouldn't set the FTZ or DAZ), even though you'd tell IPP that you only want 1 thread (so I'd expect it to run in the same thread). Thiswould be strange, but in the 5.3 threaded version, I get a big performance hit as mentionned in another thread, which seems to show that it really does perform in *1* side thread.

Now, in the multithreaded version, do each thread get the SSE flags of the calling thread? That is, if the calling thread has FTZ & DAZ set, does it mean that each of the threads will have it to, OR do I have to call ippsSetFlushToZero to make sure that all of its thread are properly set up? This may explain the performance hit I get using the multithreaded version.

Re: Bug in IppsSqrt_32f_I

Ivan_Z_Intel — Mon, 21 Jul 2008 07:39:53 GMT

Thanks for you information!

The bugin ippsSqrt_32f_I will be fixed in next verison.

But in any case the functionperfomance drops if input dataare denozmalized.

Re: Bug in IppsSqrt_32f_I

gol — Mon, 21 Jul 2008 13:05:01 GMT

thanks

But in any case the functionperfomance drops if input dataare denozmalized

normally not if FTZ or DAZ are on, no?

Re: Bug in IppsSqrt_32f_I

Ivan_Z_Intel — Mon, 21 Jul 2008 13:19:40 GMT

You are right. Usually they helpto improve perfomanceof calculationfor denormalize data. But it isn't for this case.