ir3: Only use per-wave pvtmem layout for compute
The blob seems to do this since a630, and it fixes spec@glsl-1.30@execution@fs-large-local-array on a650. Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10922>
This commit is contained in:
parent
0ab01f4215
commit
9350900fcd
|
@ -176,10 +176,13 @@ void * ir3_shader_assemble(struct ir3_shader_variant *v)
|
|||
if (compiler->gpu_id >= 400)
|
||||
v->constlen = align(v->constlen, 4);
|
||||
|
||||
/* Use the per-wave layout by default on a6xx. It should result in better
|
||||
* performance when loads/stores are to a uniform index.
|
||||
/* Use the per-wave layout by default on a6xx for compute shaders. It
|
||||
* should result in better performance when loads/stores are to a uniform
|
||||
* index.
|
||||
*/
|
||||
v->pvtmem_per_wave = compiler->gpu_id >= 600 && !info->multi_dword_ldp_stp;
|
||||
v->pvtmem_per_wave =
|
||||
compiler->gpu_id >= 600 && !info->multi_dword_ldp_stp &&
|
||||
v->type == MESA_SHADER_COMPUTE;
|
||||
|
||||
fixup_regfootprint(v);
|
||||
|
||||
|
|
Loading…
Reference in New Issue