[llvm] r279995 - AMDGPU/SI: Implement a custom MachineSchedStrategy

[Hal Finkel via llvm-commits]

Hi Tom,

...

----- Original Message -----
> From: "Tom Stellard" <tom at stellard.net>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "Tom Stellard" <thomas.stellard at amd.com>, llvm-commits at lists.llvm.org
> Sent: Thursday, September 1, 2016 11:39:06 AM
> Subject: Re: [llvm] r279995 - AMDGPU/SI: Implement a custom MachineSchedStrategy
> 
> On Mon, Aug 29, 2016 at 03:01:00PM -0500, Hal Finkel via llvm-commits
> wrote:
> > Hi Tom,
> > 
> > This is interesting. I see the following note:
> > 
> > +  // If two instructions increase the pressure of different
> > register sets
> > +  // by the same amount, the generic scheduler will prefer to
> > schedule the
> > +  // instruction that increases the set with the least amount of
> > registers,
> > +  // which in our case would be SGPRs.  This is rarely what we
> > want, so
> > +  // when we report excess/critical register pressure, we do it
> > either
> > +  // only for VGPRs or only for SGPRs.
> > 
> > Is this the main difference? Could you explain why this matters?
> > This seems like a lot of boilerplate just to make that change. Is
> > there some other hook we should have?
> > 
> 
> There are a few differences between this custom scheduler and
> GenericScheduler.

In general, it would be great if the outcome of this thread were captured in comments in the code.

> 
> The first and most important is that the GCN scheduler only considers
> two register pressure sets one that includes all VGPRs and one that
> includes all SGPRs.  There are other pressure sets that are generated
> by TableGen, that are either subsets or intersections of these sets,
> but we don't need these sets to compute accurate register pressure,
> and they usually end up just confusing the scheduler into making bad
> choices.
> 
> The second difference is the computation of the Excess register sets,
> which
> is what this comment refers to.  There is already a hook to allow
> backends to break ties between register sets that increase pressure
> equally: TRI->getRegPressureSetScore(MF, TryPSet), however, there is
> one other thing we need to consider and that is if both VGPR and SGPR
> sets are determined to be Excess sets, we want the scheduler to only
> consider VGPR pressure, because VGPR spills are more expensive.

Are they so much more expensive than the scalar spills that it is not even worth modeling the cost, or is something you might want to model more carefully in the future?
 
> We also compute our own Excess/Critical limits based on current
> register pressure and other properties of the program.

Can you explain a little more about this?

Thanks again,
Hal

> 
> -Tom
> 
> > Thanks again,
> > Hal
> > 
> > ----- Original Message -----
> > > From: "Tom Stellard via llvm-commits"
> > > <llvm-commits at lists.llvm.org>
> > > To: llvm-commits at lists.llvm.org
> > > Sent: Monday, August 29, 2016 2:42:52 PM
> > > Subject: [llvm] r279995 - AMDGPU/SI: Implement a custom
> > > MachineSchedStrategy
> > > 
> > > Author: tstellar
> > > Date: Mon Aug 29 14:42:52 2016
> > > New Revision: 279995
> > > 
> > > URL: http://llvm.org/viewvc/llvm-project?rev=279995&view=rev
> > > Log:
> > > AMDGPU/SI: Implement a custom MachineSchedStrategy
> > > 
> > > Summary:
> > > GCNSchedStrategy re-uses most of GenericScheduler, it's just uses
> > > a different method to compute the excess and critical register
> > > pressure limits.
> > > 
> > > It's not enabled by default, to enable it you need to pass
> > > -misched=gcn
> > > to llc.
> > > 
> > ...
> > > 
> > > 
> > > _______________________________________________
> > > llvm-commits mailing list
> > > llvm-commits at lists.llvm.org
> > > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> > > 
> > 
> > --
> > Hal Finkel
> > Assistant Computational Scientist
> > Leadership Computing Facility
> > Argonne National Laboratory
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> 

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