llvm dev - Jul 2019 - [cfe-dev] RFC: change -fp-contract=off to actually disable FMAs

>  However, fp-contract is not a knob to control whether or not
abstract-machine operations generate a single arithmetic instruction

I think that makes sense, but the end result is the same. Wouldn't you
agree that -fp-contract=off still contracts floating point expressions with
the initial example I posted? That is the core of what I'm trying to
resolve here.

I still have some confusion of what FMAD is supposed to be. Is FMAD
actually MAD? Or is it something else? I am fine with leaving it alone if
FMAD is not actually contracting floating point operations.

On Fri, Jul 12, 2019 at 10:54 AM Stephen Canon <scanon at apple.com>
wrote:
> Echoing what everyone else has said, keying on the word “fused” is a red
> herring here.
>
> fp-contract refers to behavior governed by the STDC FP_CONTRACT pragma.
> “Contraction” has a formal definition in the C standard:
>
> > A floating expression may be contracted, that is, evaluated as though
it
> were a single operation, thereby omitting rounding errors implied by the
> source code and the way to disallow contracted expressions.
>
> Note that this definition is *purely* in terms of the rounding of
> arithmetic operations performed by the abstract machine; there is no notion
> of instructions generated. Formation of fused multiply-add instructions is
> one specific form of fusion licensed by this pragma, which happens to be
> the main one of interest from the standpoint of compiler performance
> optimization for FMA-based architectures.
>
> There’s some imprecision in the documentation caused by a mismatch between
> what’s interesting for compiler writers (where rounding changes due to FMA
> formation are allowed) and the abstract specification. That should be
> cleaned up. However, fp-contract is not a knob to control whether or not
> abstract-machine operations generate a single arithmetic instruction—it
> definitely does not, and should not, enable or disable MAD formation.
>
> – Steve
>
> > On Jul 10, 2019, at 6:14 PM, Scott Manley via cfe-dev <
> cfe-dev at lists.llvm.org> wrote:
> >
> > >  I think you have a different definition of fused then. Fused is
a
> description of how the operation is computed/rounded, not an instruction
> count.
> >
> > "Only fuse FP ops when the result won't be affected" is
what the
> existing comment says. So it can't be both a fused op and not a fused
op if
> it's only meant to imply a difference in rounding. I'm just
re-using the
> existing wording, and I agree it could be cleaned up if that's the
intent
> of the -fp-contract option -- which I why I was asking for context.
> >
> > >  For FMA, I think your example IR is correctly handled. The fast
> instruction flag should override the global FP option you’re providing. For
> the issue you are describing, this is more of a question of whether clang
> should be emitting the fast flag or not.
> >
> > I disagree. How does clang know what would ultimately form an FMA?  It
> would have to blanket remove 'fast' from all fadds.
> >
> >>
> >> On Wed, Jul 10, 2019 at 4:16 PM Matt Arsenault <arsenm2 at
gmail.com>
> wrote:
> >>
> >>> On Jul 10, 2019, at 16:56, Scott Manley <rscottmanley at
gmail.com>
> wrote:
> >>>
> >>> At any rate, I was only offering an additional reason.
Personally I
> think it's strange for an option to say "this will never fuse
ops" and then
> under the covers will fuse ops, regardless of how FMAD is defined. However,
> my primary concern is for FMAs. They have both numeric and performance
> implications and I do not think it's unreasonable that off means off.
> >>
> >> I think you have a different definition of fused then. Fused is a
> description of how the operation is computed/rounded, not an instruction
> count. The F in FMAD is not fused (I know this naming scheme is not great.
> Every other FP node besides FMA has an F prefix)
> >>
> >> For FMA, I think your example IR is correctly handled. The fast
> instruction flag should override the global FP option you’re providing. For
> the issue you are describing, this is more of a question of whether clang
> should be emitting the fast flag or not.
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190712/b8742796/attachment.html>

> On Jul 12, 2019, at 12:32, Scott Manley <rscottmanley at gmail.com>
wrote:
> 
> I still have some confusion of what FMAD is supposed to be. Is FMAD
actually MAD? Or is it something else? I am fine with leaving it alone if FMAD
is not actually contracting floating point operations.
Yes. FMAD is supposed to give the same result as the separate FMUL and FADD with
the intermediate rounding step. It exists because most of the combines for FMA
still apply, and AMDGPU has an instruction that does this (as long as denormal
flushing is OK). I don’t know of any other users.

-Matt
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190712/dafb90b9/attachment.html>

> On Jul 12, 2019, at 12:32 PM, Scott Manley <rscottmanley at
gmail.com> wrote:
> 
> I think that makes sense, but the end result is the same. Wouldn't you
agree that -fp-contract=off still contracts floating point expressions with the
initial example I posted? That is the core of what I'm trying to resolve
here.
Yes, absolutely. The `fast` flag licenses many transforms that are not allowed
in the default floating-point model, including contraction. So, yes, it’s being
contracted in your initial example, and that’s OK!
> I still have some confusion of what FMAD is supposed to be. Is FMAD
actually MAD? Or is it something else? I am fine with leaving it alone if FMAD
is not actually contracting floating point operations.

     /// FMAD - Perform a * b + c, while getting the same result as the
     /// separately rounded operations.
     FMAD
<https://llvm.org/doxygen/namespacellvm_1_1ISD.html#a22ea9cec080dd5f4f47ba234c2f59110a2415bea72c995bb9cf9f85bbbf90bcd7>,

It does exactly what it says on the tin. It represents an instruction that
produces exactly the same result as FADD(FMUL(a,b), c). No contraction is
occurring when such a node is formed.

– Steve
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190712/cf64ec33/attachment.html>

> On Jul 12, 2019, at 12:55 PM, Matt Arsenault <arsenm2 at gmail.com>
wrote:
> 
> Yes. FMAD is supposed to give the same result as the separate FMUL and FADD
with the intermediate rounding step. It exists because most of the combines for
FMA still apply, and AMDGPU has an instruction that does this (as long as
denormal flushing is OK). I don’t know of any other users.
armv7 also has a non-fused floating-point multiply-add (VMLA as opposed to the
fused operation VFMA).

– Steve
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190712/ebf50d5c/attachment.html>