Hi Ken, I think this never got reviewed. Could you have a quick look at least to patch 1? That looks like something we want to fix since we currently have a hard-coded array size for 16 MRFs registers which won't do for SNB.
Iago On Tue, 2015-09-22 at 14:00 +0200, Iago Toral Quiroga wrote: > Originally, these could conflict with our spills, but now that we moved the > latter to MRFs 21..23, that is no longer the case. Still, in gen6 we > now use MRFs 1..15 for URB writes, so we probably want to make our pull > constant loads out of that MRF space (currently, they start at MRFs 13/14). > > Even if we do not want to do this for some reason, I still think we should > at least apply the first patch, since that plugs a hardcoded array size of > 16 MRF registers. For some reason this only became a problem when I tried > to move pull constant loads to MRFs > 15 and not when I did the same for > spills, but it looks like the right thing to do in any case. > > I tested this on SandyBridge and IvyBridge and did not observe any > regressions in piglit. > > Iago Toral Quiroga (3): > i965: Fix remove_duplicate_mrf_writes so it can handle 24 MRFs in gen6 > i965: make pull constant loads in gen6 start at MRFs 16/17 > i965: Define FIRST_SPILL_MRF and FIRST_PULL_LOAD_MRF only once and in > one place > > src/mesa/drivers/dri/i965/brw_fs.cpp | 6 +++--- > src/mesa/drivers/dri/i965/brw_fs_reg_allocate.cpp | 2 -- > src/mesa/drivers/dri/i965/brw_inst.h | 6 ++++++ > src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 4 +--- > 4 files changed, 10 insertions(+), 8 deletions(-) > _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev