Re: R: [RFC PATCH v2] target/ppc: Enable hardfloat for PPC

[Alex Bennée]

G 3 <address@hidden> writes:

> Accuracy is an important part of the IEEE 754 floating point standard. The
> whole purpose of this standard is to ensure floating point calculations are
> consistent across multiple CPUs. I believe referring to this patch as
> inaccurate is itself inaccurate. That gives the impression that this patch
> produces calculations that are not inline with established standards. This
> is not true. The only part of this patch that will produce incorrect values
> are the flags.

As I stated further up the thread I'd be happy to take a patch that
turns this on without the messing about with the FPU flags which AFAICT
breaks the architectural compliance of those instructions. The ability
to detect inexact results is part of the IEEE spec even if it is
considerably looser about when the flag should be reset.

> There *may* be a program or two out there that depend on
> these flags, but for the majority of programs that only care about basic
> floating point arithmetic this patch will produce correct values. Currently
> the emulated PowerPC's FPU already produces wrong values for the
> flags.

Those are bugs that should be fixed. The state of the flags after a
calculation should be considered part of the "values" generated by the
FPU.

> This patch does set the Inexact flag (which I don't like), but since I have
> never encountered any source code that cares for this flag, I can let it
> go. I think giving the user the ability to decide which option to use is
> the best thing to do.

Giving the user the option to break things is a poor flag, most of QEMUs
configurable knobs are about trading of size/speed without affecting
correctness.

If the PPC maintainers are happy that hardfloat's speed trade-offs are
worth it for usual workloads (whatever they may be) then I have no
objection to making defaulting hardfloat to on - perhaps* even with an
option to force softfloat if it faster for some workloads.

>
> On Wed, Feb 26, 2020 at 10:51 AM Aleksandar Markovic <
> address@hidden> wrote:
>
>>
>>
>> On Wed, Feb 26, 2020 at 3:29 PM Alex Bennée <address@hidden>
>> wrote:
>> >
>> >
>> > Dino Papararo <address@hidden> writes:
>> >
>> > > Please let's go with hardfloat pps support, it's really a good feature
>> to implement.
>> > > Even if in a first step it could lead to inaccuracy results, later it
>> > > could solved with other patches.
>> >
>> > That's the wrong way around. We have regression tests for a reason.
>>
>> I tend to agree with Alex here, and additionally want to expand more on
>> this topic.
>>
>> In my view: (that I think is at least very close to the community
>> consensus)
>>
>> This is *not* a ppc-specific issue. There exist a principle across all
>> targets
>> that QEMU FPU calculation must be accurate - exactly as specified in any
>> applicable particular ISA document. Any discrepancy is an outright bug.
>>
>> We even recently had several patches for FPU in ppc target that handled
>> some fairly obscure cases of inaccuracies, I believe they were authored
>> by Paul Clarke, so there are people in ppc community that care about
>> FPU accuracy (as I guess is the case for any target).
>>
>> There shouldn't be a target that decides by itself and within itself
>> "ok, we don't need accuracy, let's trade it for speed". This violates
>> the architecture of QEMU. Please allow that for any given software
>> project, there is an architecture that should be respected.
>>
>> This doesn't mean that anybody's experimentation is discouraged. No-one
>> can stop anybody from forking from QEMU upstream tree and do whatever
>> is wanted.
>>
>> But, this doesn't mean such experimentation will be upstreamed. QEMU
>> upstream should be collecting place for the best ideas and implementations,
>> not for arbitrary experimentations.
>>
>> Best regards,
>> Aleksandar
>>
>>
>> > I'll happily accept patches to turn on hardfloat for PPC if:
>> >
>> >  a) they don't cause regressions in our fairly extensive floating point
>> >  tests
>> >  b) the PPC maintainers are happy with the new performance profile
>> >
>> > The way forward would be to:
>> >
>> >  1. patch to drop #if defined(TARGET_PPC) || defined(__FAST_MATH__)
>> >  2. audit target/ppc/fpu_helper.c w.r.t chip manual and fix any unneeded
>> >  splatting of flags (if any)
>> >  3. measure the before/after performance effect and decide if on balance
>> >  it's worth keeping
>> >
>> > > I think it's important for qemu to as global as possible and don't
>> > > target only recent hardware.
>> >
>> > Are you referring to guests or hosts? For guests we will always favour
>> > accuracy of speed of emulation. For hosts we need to have IEEE compliant
>> > FPU HW to even stand a chance of using hardfloat.
>> >
>> > --
>> > Alex Bennée
>> >
>>


-- 
Alex Bennée