[PATCH] [OpenCL] fix the various -cl-* flags to relax FP arith

[Hal Finkel]

----- Original Message -----
> From: "Pekka Jääskeläinen" <pekka.jaaskelainen at tut.fi>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "Matt Arsenault" <Matthew.Arsenault at amd.com>, cfe-commits at cs.uiuc.edu
> Sent: Monday, December 8, 2014 11:10:17 AM
> Subject: Re: [PATCH] [OpenCL] fix the various -cl-* flags to relax FP arith
> 
> Any further comments, or shall I commit this?

LGTM, go ahead and commit.

 -Hal

> 
> On 12/01/2014 11:33 PM, Hal Finkel wrote:
> > ----- Original Message -----
> >> From: "Pekka Jääskeläinen" <pekka.jaaskelainen at tut.fi>
> >> To: "Matt Arsenault" <Matthew.Arsenault at amd.com>,
> >> cfe-commits at cs.uiuc.edu
> >> Sent: Monday, December 1, 2014 3:19:40 PM
> >> Subject: Re: [PATCH] [OpenCL] fix the various -cl-* flags to relax
> >> FP arith
> >>
> >> On 12/01/2014 06:13 PM, Matt Arsenault wrote:
> >>> I think the test should also check for the function target
> >>> attribute.
> >>
> >> Not sure of this as Hal said that those are being phased out.
> >> I added a test for the attributes that were there already anyways,
> >> but didn't add a new attribute for the no-signed-zeros.
> >
> > Sounds reasonable to me.
> >
> >   -Hal
> >
> >>
> >>> What about the other flags? I think this patch should handle
> >>> (maybe
> >>> just
> >>> test?) all of the CL fast math flags instead of just this one (
> >>> -cl-no-signed-zeros, -cl-unsafe-math-optimizations,
> >>> -cl-finite-math-only, and -cl-fast-relaxed-math)
> >>
> >> Done. The OpenCL flags seem to be supersets of the
> >> flags of LLVM, so there might be more optimization
> >> opportunities to extract from those than these flags
> >> enable.
> >>>
> >> For example:
> >>
> >> "Allow optimizations for floating-point arithmetic that (a) assume
> >> that
> >> arguments and results are valid, (b) may violate IEEE 754 standard
> >> and
> >> (c) may violate the OpenCL numerical compliance requirements as
> >> defined
> >> in section 7.4 for single-precision floating-point, section 9.3.9
> >> for
> >> double-precision floating-point, and edge case behavior in section
> >> 7.5.
> >> This option includes the -cl-no-signed-zeros and -cl-mad-enable
> >> options."
> >> https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clBuildProgram.html
> >>
> >> Now it just sets the nnans flag even though (b) and (c) allow
> >> pretty much anything (if read with "lawyer's eyes").
> >>
> >> --
> >> --Pekka
> >>
> >> _______________________________________________
> >> cfe-commits mailing list
> >> cfe-commits at cs.uiuc.edu
> >> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
> >>
> >
> 
> 
> --
> --Pekka
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory