diff options
| author | Eric Anholt <eric@anholt.net> | 2010-10-21 15:07:45 -0700 |
|---|---|---|
| committer | Eric Anholt <eric@anholt.net> | 2010-10-21 15:21:28 -0700 |
| commit | 4e7252510976d8d3ff12437ea8842129f24d88f5 (patch) | |
| tree | c2e50a075fd7ede483e7265f46567cdbc5370d2b /src/mesa/drivers/dri/i965/brw_wm_state.c | |
| parent | 0b77d57394a3712851ec271aa7ad353d56f302a1 (diff) | |
i965: Correct scratch space allocation.
One, it was allocating increments of 1kb, but per thread scratch space
is a power of two. Two, the new FS wasn't getting total_scratch set
at all, so everyone thought they had 1kb and writes beyond 1kb would
go stomping on a neighbor thread.
With this plus the previous register spilling for the new FS,
glsl-fs-convolution-1 passes.
Diffstat (limited to 'src/mesa/drivers/dri/i965/brw_wm_state.c')
| -rw-r--r-- | src/mesa/drivers/dri/i965/brw_wm_state.c | 5 |
1 files changed, 2 insertions, 3 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_wm_state.c b/src/mesa/drivers/dri/i965/brw_wm_state.c index 411204e704..433ccc66f0 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_state.c +++ b/src/mesa/drivers/dri/i965/brw_wm_state.c @@ -97,7 +97,7 @@ wm_unit_populate_key(struct brw_context *brw, struct brw_wm_unit_key *key) key->urb_entry_read_length = brw->wm.prog_data->urb_read_length; key->curb_entry_read_length = brw->wm.prog_data->curb_read_length; key->dispatch_grf_start_reg = brw->wm.prog_data->first_curbe_grf; - key->total_scratch = ALIGN(brw->wm.prog_data->total_scratch, 1024); + key->total_scratch = brw->wm.prog_data->total_scratch; /* BRW_NEW_URB_FENCE */ key->urb_size = brw->urb.vsize; @@ -184,7 +184,7 @@ wm_unit_create_from_key(struct brw_context *brw, struct brw_wm_unit_key *key, if (key->total_scratch != 0) { wm.thread2.scratch_space_base_pointer = brw->wm.scratch_bo->offset >> 10; /* reloc */ - wm.thread2.per_thread_scratch_space = key->total_scratch / 1024 - 1; + wm.thread2.per_thread_scratch_space = ffs(key->total_scratch) - 11; } else { wm.thread2.scratch_space_base_pointer = 0; wm.thread2.per_thread_scratch_space = 0; @@ -293,7 +293,6 @@ static void upload_wm_unit( struct brw_context *brw ) * bother reducing the allocation later, since we use scratch so * rarely. */ - assert(key.total_scratch <= 12 * 1024); if (key.total_scratch) { GLuint total = key.total_scratch * brw->wm_max_threads; |