On 2013-09-10, at 2:21 , Pekka Jääskeläinen <[email protected]> wrote:
> On 09/09/2013 03:51 PM, Erik Schnetter wrote: >> There are also performance issues. ABI differences are usually introduced >> because they give better performance. In this case, we do want the "best >> possible" ABI for a particular CPU, not the "safe default choice" one would >> use in a binary Linux distribution. > > If by ABI you mostly mean the calling convention here (or what else of it is > visibile in the bitcode?), I do not think it should impact the performance > a lot as we try to inline everything to the kernel in the kernel compiler. > And, like I said, for the kernel itself, we should use a special calling > convention (like the CUDA_KERNEL one or perhaps the SPIR's kernel one). Yes, I refer to the calling convention. Yes, this may not impact performance much. Of course, having said this, my astrophysics research uses a compute kernel that is exceptionally large, and where code size is the main limitation (doesn't fit into instruction cache any more). Thus "inline as much as possible" is a very bad idea there... > The intrinsics is a problem if one needs to use target-specific ones > which cannot be reverted in llc in case of a switched off target feature. > The sqrt example, though, seems to have a generic intrinsic: > http://llvm.org/docs/LangRef.html#llvm-sqrt-intrinsic Yes, there is __builtin_sqrt(float). -erik -- Erik Schnetter <[email protected]> http://www.perimeterinstitute.ca/personal/eschnetter/ My email is as private as my paper mail. I therefore support encrypting and signing email messages. Get my PGP key from http://pgp.mit.edu/.
signature.asc
Description: Message signed with OpenPGP using GPGMail
------------------------------------------------------------------------------ How ServiceNow helps IT people transform IT departments: 1. Consolidate legacy IT systems to a single system of record for IT 2. Standardize and globalize service processes across IT 3. Implement zero-touch automation to replace manual, redundant tasks http://pubads.g.doubleclick.net/gampad/clk?id=51271111&iu=/4140/ostg.clktrk
_______________________________________________ pocl-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/pocl-devel
