fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-04 22:49:13 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

mesa/st: Lower to ALU scalar after fp64 subgroup lowering
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
The subgroup lowering may generate new fp64 vector operations, so
ensure that those are lowered before calling nir_lower_doubles().

Issue spotted by Georg Lehmann.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38003>
This commit is contained in:
Caio Oliveira 2025-10-13 15:37:27 -07:00 • committed by Marge Bot
parent fa6fa8c19e
commit e38491eb18
1 changed files with 11 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
src/mesa/state_tracker/st_glsl_to_nir.cpp
View file

@ -309,16 +309,23 @@ st_glsl_to_nir_post_opts(struct st_context *st, struct gl_program *prog,
subgroup_opts.lower_vote_feq = true;
subgroup_opts.lower_reduce = true;
if (nir->options->lower_doubles_options & nir_lower_fp64_full_software)
NIR_PASS(lowered_64bit_ops, nir, nir_lower_subgroups, &subgroup_opts);
/* nir_lower_doubles is not prepared for vector ops, so if the backend doesn't
* request lower_alu_to_scalar until now, lower all 64 bit ops, and try to
* vectorize them afterwards again */
if (!nir->options->lower_to_scalar) {
bool scalarize = !nir->options->lower_to_scalar;
if (nir->options->lower_doubles_options & nir_lower_fp64_full_software) {
NIR_PASS(lowered_64bit_ops, nir, nir_lower_subgroups, &subgroup_opts);
/* Subgroup lowering may generate 64-bit vector ALU operations. */
if (lowered_64bit_ops)
scalarize = true;
}
if (scalarize) {
NIR_PASS(revectorize, nir, nir_lower_alu_to_scalar, filter_64_bit_instr, nullptr);
NIR_PASS(revectorize, nir, nir_lower_phis_to_scalar, NULL, NULL);
}
/* doubles lowering requires frexp to be lowered first if it will be,
* since the pass generates other 64-bit ops. Most backends lower
* frexp, and using doubles is rare, and using frexp is even more rare