[Mesa-dev] [PATCH 1/5] i965: Hard code scratch_ids_per_subslice for Cherryview

Jordan Justen jordan.l.justen at intel.com
Sun Mar 11 03:22:22 UTC 2018


On 2018-03-09 09:51:31, Mark Janes wrote:
> Could this be the reason that BSW systems never reliably passed all unit
> tests?  Up to now, we re-execute each failing test, and mark it as a
> pass if it succeeds a second time.
> 
> I'd like to remove that crutch if possible.

It is possible. We basically had memory corruption happening outside
the scratch buffer. The corruption was happening a bit passed the end
of the buffer we had allocated. It can be difficult to predict the
outcome of such corruption. :)

-Jordan

> Jordan Justen <jordan.l.justen at intel.com> writes:
> 
> > Ken suggested that we might be underallocating scratch space on HD
> > 400. Allocating scratch space as though there was actually 8 EUs
> > seems to help with a GPU hang seen on synmark CSDof.
> >
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104636
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105290
> > Cc: Kenneth Graunke <kenneth at whitecape.org>
> > Cc: Eero Tamminen <eero.t.tamminen at intel.com>
> > Cc: <mesa-stable at lists.freedesktop.org>
> > Signed-off-by: Jordan Justen <jordan.l.justen at intel.com>
> > ---
> >  src/mesa/drivers/dri/i965/brw_program.c | 44 ++++++++++++++++++++-------------
> >  1 file changed, 27 insertions(+), 17 deletions(-)
> >
> > diff --git a/src/mesa/drivers/dri/i965/brw_program.c b/src/mesa/drivers/dri/i965/brw_program.c
> > index 527f003977b..c121136c439 100644
> > --- a/src/mesa/drivers/dri/i965/brw_program.c
> > +++ b/src/mesa/drivers/dri/i965/brw_program.c
> > @@ -402,23 +402,33 @@ brw_alloc_stage_scratch(struct brw_context *brw,
> >        if (devinfo->gen >= 9)
> >           subslices = 4 * brw->screen->devinfo.num_slices;
> >  
> > -      /* WaCSScratchSize:hsw
> > -       *
> > -       * Haswell's scratch space address calculation appears to be sparse
> > -       * rather than tightly packed.  The Thread ID has bits indicating
> > -       * which subslice, EU within a subslice, and thread within an EU
> > -       * it is.  There's a maximum of two slices and two subslices, so these
> > -       * can be stored with a single bit.  Even though there are only 10 EUs
> > -       * per subslice, this is stored in 4 bits, so there's an effective
> > -       * maximum value of 16 EUs.  Similarly, although there are only 7
> > -       * threads per EU, this is stored in a 3 bit number, giving an effective
> > -       * maximum value of 8 threads per EU.
> > -       *
> > -       * This means that we need to use 16 * 8 instead of 10 * 7 for the
> > -       * number of threads per subslice.
> > -       */
> > -      const unsigned scratch_ids_per_subslice =
> > -         devinfo->is_haswell ? 16 * 8 : devinfo->max_cs_threads;
> > +      unsigned scratch_ids_per_subslice;
> > +      if (devinfo->is_haswell) {
> > +         /* WaCSScratchSize:hsw
> > +          *
> > +          * Haswell's scratch space address calculation appears to be sparse
> > +          * rather than tightly packed. The Thread ID has bits indicating
> > +          * which subslice, EU within a subslice, and thread within an EU it
> > +          * is. There's a maximum of two slices and two subslices, so these
> > +          * can be stored with a single bit. Even though there are only 10 EUs
> > +          * per subslice, this is stored in 4 bits, so there's an effective
> > +          * maximum value of 16 EUs. Similarly, although there are only 7
> > +          * threads per EU, this is stored in a 3 bit number, giving an
> > +          * effective maximum value of 8 threads per EU.
> > +          *
> > +          * This means that we need to use 16 * 8 instead of 10 * 7 for the
> > +          * number of threads per subslice.
> > +          */
> > +         scratch_ids_per_subslice = 16 * 8;
> > +      } else if (devinfo->is_cherryview) {
> > +         /* For Cherryview, it appears that the scratch addresses for the 6 EU
> > +          * devices may still generate compute scratch addresses covering the
> > +          * same range as 8 EU.
> > +          */
> > +         scratch_ids_per_subslice = 8 * 7;
> > +      } else {
> > +         scratch_ids_per_subslice = devinfo->max_cs_threads;
> > +      }
> >  
> >        thread_count = scratch_ids_per_subslice * subslices;
> >        break;
> > -- 
> > 2.16.1
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list