spirv: Don't emit RMW for vector indexing in shared or global
Anything that fails the is_external_block check is getting the vtn_local_load/store path which does read-modify-write which isn't correct if the variable mode can be written cross-workgroup. Cc: mesa-stable@lists.freedesktop.org Reviewed-by: Jesse Natalie <jenatali@microsoft.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6372>
This commit is contained in:

committed by
Marge Bot

parent
f038b3a136
commit
b479de8537
@@ -124,6 +124,18 @@ vtn_mode_uses_ssa_offset(struct vtn_builder *b,
|
||||
mode == vtn_variable_mode_push_constant;
|
||||
}
|
||||
|
||||
static bool
|
||||
vtn_mode_is_cross_invocation(struct vtn_builder *b,
|
||||
enum vtn_variable_mode mode)
|
||||
{
|
||||
return mode == vtn_variable_mode_ssbo ||
|
||||
mode == vtn_variable_mode_ubo ||
|
||||
mode == vtn_variable_mode_phys_ssbo ||
|
||||
mode == vtn_variable_mode_push_constant ||
|
||||
mode == vtn_variable_mode_workgroup ||
|
||||
mode == vtn_variable_mode_cross_workgroup;
|
||||
}
|
||||
|
||||
static bool
|
||||
vtn_pointer_is_external_block(struct vtn_builder *b,
|
||||
struct vtn_pointer *ptr)
|
||||
@@ -1074,11 +1086,11 @@ _vtn_variable_load_store(struct vtn_builder *b, bool load,
|
||||
if (glsl_type_is_vector_or_scalar(ptr->type->type)) {
|
||||
/* We hit a vector or scalar; go ahead and emit the load[s] */
|
||||
nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
|
||||
if (vtn_pointer_is_external_block(b, ptr)) {
|
||||
/* If it's external, we call nir_load/store_deref directly. The
|
||||
* vtn_local_load/store helpers are too clever and do magic to
|
||||
* avoid array derefs of vectors. That magic is both less
|
||||
* efficient than the direct load/store and, in the case of
|
||||
if (vtn_mode_is_cross_invocation(b, ptr->mode)) {
|
||||
/* If it's cross-invocation, we call nir_load/store_deref
|
||||
* directly. The vtn_local_load/store helpers are too clever and
|
||||
* do magic to avoid array derefs of vectors. That magic is both
|
||||
* less efficient than the direct load/store and, in the case of
|
||||
* stores, is broken because it creates a race condition if two
|
||||
* threads are writing to different components of the same vector
|
||||
* due to the load+insert+store it uses to emulate the array
|
||||
|
Reference in New Issue
Block a user