[Mesa-dev] [PATCH 6/7] gallium: Add TGSI_OPCODE_F2U to gallivm backend.

[Roland Scheidegger]

Am 14.06.2012 18:55, schrieb Paul Berry:
> On 14 June 2012 09:02, Roland Scheidegger <sroland at vmware.com
> <mailto:sroland at vmware.com>> wrote:
> 
>     Am 14.06.2012 15:07, schrieb Paul Berry:
>     > On 14 June 2012 05:10, Jose Fonseca <jfonseca at vmware.com
>     <mailto:jfonseca at vmware.com>
>     > <mailto:jfonseca at vmware.com <mailto:jfonseca at vmware.com>>> wrote:
>     >
>     >     lp_build_utrunc() doesn't exists yet.  For now, please use
>     >     lp_build_itrunc() and add a /* FIXME: implement and use
>     >     lp_build_utrunc() */ comment. It's not perfect, but it's no worse
>     >     what's there now.
>     >
>     >     Jose
>     >
>     >
>     > Ok, will do.  Thank you.
>     btw I'm wondering how you'd actually do that in some semi-efficient
>     manner?
>     llvm has LLVMBuildFPToUI but x86 has no such conversion instructions
>     they are all signed. I guess multiply by 0.5, convert to signed int,
>     cast to unsigned and multiply by 2 isn't really the solution :-).
> 
> 
> Playing around with gcc, it looks like it generally achieves
> float->unsigned conversions by first converting to a 64-bit integer
> (using the "fistpq" instruction on x86, or "cvttss2siq" on x86-64), then
> truncating the upper 32 bits.

I guess that's a reasonable option (unless you'd need 64bit uints...).
For vectorized llvm code though that option looks pretty bad.

Roland