[Mesa-dev] [PATCH v2] nir: delete magic number

[tournier.elie]

On 9 February 2017 at 02:48, Timothy Arceri <tarceri at itsqueeze.com> wrote:

> On Wed, 8 Feb 2017 15:54:46 -0800
> Jason Ekstrand <jason at jlekstrand.net> wrote:
>
> > On Wed, Feb 8, 2017 at 2:20 PM, Elie Tournier
> > <tournier.elie at gmail.com> wrote:
> >
> > > Signed-off-by: Elie Tournier <tournier.elie at gmail.com>
> > > ---
> > >  src/compiler/nir/nir_opt_loop_unroll.c | 10 +++++++++-
> > >  1 file changed, 9 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/src/compiler/nir/nir_opt_loop_unroll.c
> > > b/src/compiler/nir/nir_opt_loop_unroll.c
> > > index 37cbced43d..035a030239 100644
> > > --- a/src/compiler/nir/nir_opt_loop_unroll.c
> > > +++ b/src/compiler/nir/nir_opt_loop_unroll.c
> > > @@ -26,6 +26,14 @@
> > >  #include "nir_control_flow.h"
> > >  #include "nir_loop_analyze.h"
> > >
> > > +
> > > +/* This limit is chosen fairly arbitrarily.  The GLSL IR limit is
> > > 25.
> > > + * However, due to slight differences in the way the two IRs count
> > > + * instructions, some loops that would unroll with GLSL IR fail to
> > > unroll
> > > + * if we set this to 25 so we set it to 26.
> > >
> >
> > Ok, I lied in my comment.  It's not 25, it's 32.  Tim, can you
> > explain the discrepancy?
> > --Jason
>  
> Max iteration is 32 this is number of instructions, in GLSL IR it was
> counting nodes and the magic number was 5. There is no 1:1 mapping 25
> was picked because it seemed to give about the same results. Around 5
> instructions per node. 
>
>
So to resume. Is this comment correct for you?

This limit is chosen fairly arbitrarily.
GLSL IR max iteration is 32 instructions. (Counting nodes by magic number 5).
But there is no 1:1 mapping between GLSL IR and NIR so 25 was picked because
it seemed to give about the same results. Around 5 instructions per node.
But some loops that would unroll with GLSL IR fail to unroll if we set
this to 25 so we set it to 26.

>
> >
> > > + */
> > > +#define LOOP_UNROLL_LIMIT 26
> > > +
> > >  /* Prepare this loop for unrolling by first converting to lcssa
> > > and then
> > >   * converting the phis from the loops first block and the block
> > > that follows
> > >   * the loop into regs.  Partially converting out of SSA allows us
> > > to unroll
> > > @@ -460,7 +468,7 @@ is_loop_small_enough_to_unroll(nir_shader
> > > *shader, nir_loop_info *li)
> > >        return true;
> > >
> > >     bool loop_not_too_large =
> > > -      li->num_instructions * li->trip_count <= max_iter * 26;
> > > +      li->num_instructions * li->trip_count <= max_iter *
> > > LOOP_UNROLL_LIMIT;
> > >
> > >     return loop_not_too_large;
> > >  }
> > > --
> > > 2.11.0
> > >
> > >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170213/5f5580ce/attachment.html>