fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-08 02:00:21 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir/lower_subgroups: Mask off unused bits in ballot ops

Browse source 
Thanks to VK_EXT_subgroup_size_control, we can end up with
gl_SubgroupSize being as low as 8 on Intel.

Fixes: d10de25309 "anv: Implement VK_EXT_subgroup_size_control"
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4694>
(cherry picked from commit fdf9b674ee)
This commit is contained in:
Jason Ekstrand 2020-04-22 21:35:48 -05:00 • committed by Dylan Baker
parent 74e0db6171
commit 109caeeb43
2 changed files with 27 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.pick_status.json
View file

@ -1147,7 +1147,7 @@
"description": "nir/lower_subgroups: Mask off unused bits in ballot ops",
"nominated": true,
"nomination_type": 1,
"resolution": 0,
"resolution": 1,
"master_sha": null,
"because_sha": "d10de2530976ed3aba9c5d077e2edb141f71e7dd"
},

26
src/compiler/nir/nir_lower_subgroups.c
View file

@ -446,6 +446,32 @@ lower_subgroups_instr(nir_builder *b, nir_instr *instr, void *_options)
assert(intrin->src[0].is_ssa);
nir_ssa_def *int_val = ballot_type_to_uint(b, intrin->src[0].ssa,
options->ballot_bit_size);
if (intrin->intrinsic != nir_intrinsic_ballot_bitfield_extract &&
intrin->intrinsic != nir_intrinsic_ballot_find_lsb) {
/* For OpGroupNonUniformBallotFindMSB, the SPIR-V Spec says:
*
* "Find the most significant bit set to 1 in Value, considering
* only the bits in Value required to represent all bits of the
* group’s invocations. If none of the considered bits is set to
* 1, the result is undefined."
*
* It has similar text for the other three. This means that, in case
* the subgroup size is less than 32, we have to mask off the unused
* bits. If the subgroup size is fixed and greater than or equal to
* 32, the mask will be 0xffffffff and nir_opt_algebraic will delete
* the iand.
*
* We only have to worry about this for BitCount and FindMSB because
* FindLSB counts from the bottom and BitfieldExtract selects
* individual bits. In either case, if run outside the range of
* valid bits, we hit the undefined results case and we can return
* anything we want.
*/
int_val = nir_iand(b, int_val,
build_subgroup_mask(b, options->ballot_bit_size, options));
}
switch (intrin->intrinsic) {
case nir_intrinsic_ballot_bitfield_extract:
assert(intrin->src[1].is_ssa);