| Re: [eigen] SSE square root |
[ Thread Index |
Date Index
| More lists.tuxfamily.org/eigen Archives
]
- To: eigen@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [eigen] SSE square root
- From: Gael Guennebaud <gael.guennebaud@xxxxxxxxx>
- Date: Fri, 27 Mar 2009 11:36:46 +0100
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:mime-version:received:in-reply-to:references :date:message-id:subject:from:to:content-type :content-transfer-encoding; bh=0TMOhzN9YJwdjIE8kK/bfvWBvxTUcXiDF3AZ9NxwGS8=; b=YOvHfoKUUIo4cDxP79riDkVMlt2goukMAANBgRonMeUBvyHOZbASnoQESukITAmtaw sXvUVqzq2ILvBElkQV2gDAdMc6M/juuta27/kWt/hPtrfvsBa9KUL7GdMNqtUme5f47r Sh6yas3HzYf3joIbsWB0aKiTAaHghSI8VE1Mo=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=mime-version:in-reply-to:references:date:message-id:subject:from:to :content-type:content-transfer-encoding; b=lC9BsJqkmXRHp8oHMvXyjnEOb8kBx98ybQANfhIH21YAmY6eIoqf6Limt3xNN1qrnx iWXp3yNeNBaM6oj1930oqqpor6lERQtyQuh3ExRhU4OBJqpJnwOi16BI+K3qlmvpoMy1 ExYe5SLDzbP1MlZ3X5iRfo7X2LtzoOQvL6GVs=
On Fri, Mar 27, 2009 at 11:28 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
> On Fri, Mar 27, 2009 at 8:32 AM, Rohit Garg <rpg.314@xxxxxxxxx> wrote:
>> This file has my sse float implementation for square root. The SSE
>> square root instruction has only 12 bits of precision so extra
>
> where did you find sqrtss or sqrtps has only 12 bits of precision ?
>
> this is right for the reciprocal functions (rcpps and rsqrtps) but I
> thought sqrtps was ieee754 compliant. For instance I found that in the
> glibc:
>
> float
> __ieee754_sqrtf (float x)
> {
> float res;
>
> asm ("sqrtss %0, %1" : "=x" (res) : "x" (x));
>
> return res;
> }
>
> anyway, sqrtps is very slow and I'm sure there exist a faster
> alternative. Perhaps a rsqrtps followed by a div and Newton Raphson
> iterations..
hm.. not a div but a mul:
x = _mm_mul_ps(_mm_rsqrt_ps(x),x);
// iterations
also, on my architecture, sqrtf(x) is compiled as a SSE sqrtss instruction.....
>
>> iterations of Newton Raphson may be neccessary. How many of the are
>> neccessary, I don't know. the max error I was getting was O(1e-8) in
>> [0,1]. The cephes implementation has square root only for limited
>> range. They do some other hacks to take care of range. I'll look into
>> implementing those later. For now, this should be an acceptable for
>> the fast implementations of square root atleast.
>
>
>
>> Regards,
>>
>> --
>> Rohit Garg
>>
>> http://rpg-314.blogspot.com/
>>
>> Senior Undergraduate
>> Department of Physics
>> Indian Institute of Technology
>> Bombay
>>
>