fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-08 11:18:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

radeonsi: don't subtract max_verts_per_prim from hw_max_esverts on gfx10.3

Browse source 
gfx10.3 does it properly. This change enables the last 2 lanes in a workgroup
on gfx10.3.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7542>
This commit is contained in:
Marek Olšák 2020-10-21 14:39:26 -04:00 • committed by Marge Bot
parent e29e41a3cd
commit ea90d8a744
1 changed files with 19 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

24
src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
View file

@ -2037,8 +2037,12 @@ retry_select_mode:
max_esverts = max_esverts =
MIN2(max_esverts, (max_lds_size - max_gsprims * gsprim_lds_size) / esvert_lds_size); MIN2(max_esverts, (max_lds_size - max_gsprims * gsprim_lds_size) / esvert_lds_size);
max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim); max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
/* Hardware restriction: minimum value of max_esverts */ /* Hardware restriction: minimum value of max_esverts */
max_esverts = MAX2(max_esverts, min_esverts - 1 + max_verts_per_prim); if (gs_sel->screen->info.chip_class == GFX10)
max_esverts = MAX2(max_esverts, min_esverts - 1 + max_verts_per_prim);
else
max_esverts = MAX2(max_esverts, min_esverts);
max_gsprims = align(max_gsprims, wavesize); max_gsprims = align(max_gsprims, wavesize);
max_gsprims = MIN2(max_gsprims, max_gsprims_base); max_gsprims = MIN2(max_gsprims, max_gsprims_base);
@ -2056,10 +2060,16 @@ retry_select_mode:
} while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims); } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
/* Verify the restriction. */ /* Verify the restriction. */
assert(max_esverts >= min_esverts - 1 + max_verts_per_prim); if (gs_sel->screen->info.chip_class == GFX10)
assert(max_esverts >= min_esverts - 1 + max_verts_per_prim);
else
assert(max_esverts >= min_esverts);
} else { } else {
/* Hardware restriction: minimum value of max_esverts */ /* Hardware restriction: minimum value of max_esverts */
max_esverts = MAX2(max_esverts, min_esverts - 1 + max_verts_per_prim); if (gs_sel->screen->info.chip_class == GFX10)
max_esverts = MAX2(max_esverts, min_esverts - 1 + max_verts_per_prim);
else
max_esverts = MAX2(max_esverts, min_esverts);
} }
unsigned max_out_vertices = unsigned max_out_vertices =
@ -2077,12 +2087,16 @@ retry_select_mode:
prim_amp_factor = gs_sel->info.base.gs.vertices_out; prim_amp_factor = gs_sel->info.base.gs.vertices_out;
} }
/* The GE only checks against the maximum number of ES verts after /* On gfx10, the GE only checks against the maximum number of ES verts after
* allocating a full GS primitive. So we need to ensure that whenever * allocating a full GS primitive. So we need to ensure that whenever
* this check passes, there is enough space for a full primitive without * this check passes, there is enough space for a full primitive without
* vertex reuse. * vertex reuse.
*/ */
shader->ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1; if (gs_sel->screen->info.chip_class == GFX10)
shader->ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1;
else
shader->ngg.hw_max_esverts = max_esverts;
shader->ngg.max_gsprims = max_gsprims; shader->ngg.max_gsprims = max_gsprims;
shader->ngg.max_out_verts = max_out_vertices; shader->ngg.max_out_verts = max_out_vertices;
shader->ngg.prim_amp_factor = prim_amp_factor; shader->ngg.prim_amp_factor = prim_amp_factor;