fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-01 03:48:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir/serialize: make alu src deserialization consistent for unused swizzles

Browse source 
Currently we have 3 paths for ALU serialization/deserialization in NIR:

1. If the ALU is qualified as packed_src_ssa_16bit (identity swizzle)
  - The write will store an object index only.
  - The read will only load the swizzles actually used, the rest are 0.
2. If the ALU is not qualified as packed_src_ssa_16bit, we have two cases:
  2.1 Up to vec4:
    - The write stores all 4 swizzle components.
    - The read loads all 4 swizzle components.
  2.2 vec8/16
    - The write stores only swizzle components used, the rest are 0.
    - The read loads only swizzle components used, the rest are 0.

This inconsistency in how these paths encode/decode unsused swizzle components
can cause issues in some scenarios where a backend compiler may receive
functionally equivalent NIR shaders from Mesa that won't produce the same sha1,
leading to unnecessary cache misses.

This patch makes path 2.1 always encode and decode unused swizzle components
as 0, making it consistent with the other paths.

This fixes issues where sometimes backends need to compile a shader twice
before it is effectively retrieved from the disk cache. This has been
observed at least with V3d and Panfrost.

The problem occurs when an ALU src with unused swizzle components is serialized
in the Mesa frontend using path 1, but when it later hits the backend it is
serialized using path 2.1. The backend uses the sha1 of the serialized NIR for
the cache key. On the second execution the Mesa frontend has a cache hit and
when it deserializes the alu src, it sets its unused components to 0 but that
will cause the backend to have a cache miss since that NIR doesn't match the one
it cached on the first execution.

By always making unused swizzle components decode and encode consistently to 0
in all paths we ensure the issue never happens and that NIR variants that only
differ in swizzle components that are not used lead to cache hits.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Eric R. Smith <eric.smith@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37218>
This commit is contained in:
Iago Toral Quiroga 2025-08-05 23:13:15 +02:00
parent 2f7b1e8453
commit ac11e00b15
1 changed files with 12 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/compiler/nir/nir_serialize.c
View file

@ -751,9 +751,12 @@ write_alu(write_ctx *ctx, const nir_alu_instr *alu)
if (packed) {
src.alu.swizzle_x = alu->src[i].swizzle[0];
src.alu.swizzle_y = alu->src[i].swizzle[1];
src.alu.swizzle_z = alu->src[i].swizzle[2];
src.alu.swizzle_w = alu->src[i].swizzle[3];
if (src_channels >= 2)
src.alu.swizzle_y = alu->src[i].swizzle[1];
if (src_channels >= 3)
src.alu.swizzle_z = alu->src[i].swizzle[2];
if (src_channels >= 4)
src.alu.swizzle_w = alu->src[i].swizzle[3];
}
write_src_full(ctx, &alu->src[i].src, src);
@ -810,9 +813,12 @@ read_alu(read_ctx *ctx, union packed_instr header)
if (packed) {
alu->src[i].swizzle[0] = src.alu.swizzle_x;
alu->src[i].swizzle[1] = src.alu.swizzle_y;
alu->src[i].swizzle[2] = src.alu.swizzle_z;
alu->src[i].swizzle[3] = src.alu.swizzle_w;
if (src_channels >= 2)
alu->src[i].swizzle[1] = src.alu.swizzle_y;
if (src_channels >= 3)
alu->src[i].swizzle[2] = src.alu.swizzle_z;
if (src_channels >= 4)
alu->src[i].swizzle[3] = src.alu.swizzle_w;
} else {
/* Load swizzles for vec8 and vec16. */
for (unsigned o = 0; o < src_channels; o += 8) {