Add explanation for subgroup vs dynamically uniform. #2118
Conversation
|
|
|
Would it be possible to instead "forbid" the repacking behavior in graphics? In essence only allow implementations to pack subgroups if the shader cannot observe the difference - I think this could be a more intuitive solution. |
Co-authored-by: Constantine Shablia <[email protected]>
HansKristian-Work
force-pushed
the
subgroup-vs-dynamic-uniform
branch
from
a7069b5
to
de61c3b
Compare
Not feasible since some implementations rely on this in shipping drivers. Only option I can think of would be an extension that lets you control pack boundary. |
| // | DrawID = 2 | DrawID = 1 | } | ||
| // Subgroup 1: { 2, 2, 2, 2, 1, 1, 1, 1 } | ||
|
|
||
| uint notActuallyDynamicallyUniformAnymore = |
There was a problem hiding this comment.
While everything in this PR seems technically correct, it's concerning because this part suggests that you can't use a peeling loop to remove the need for the nonuniform decoration, which I think existing content relies on. Maybe we should change VUID 06274 to say is not dynamically uniform or subgroup uniform? In the long term I'd prefer to get rid of the nonuniform decoration entirely (I think there are other, implementation-dependent cases where nonuniform is misleading/wrong), but I think that's a harder change to get through.
There was a problem hiding this comment.
Is it realistic for an application to be able to know whether they're subgroup-uniform or not in this case? Doesn't that depend on the hardware/driver and how it might pack draws?
There was a problem hiding this comment.
t's concerning because this part suggests that you can't use a peeling loop to remove the need for the nonuniform decoration, which I think existing content relies on.
Peeling loops to avoid nonuniformEXT are technically out of spec, since dynamically uniform is the requirement, not subgroup uniform. Not sure why it's specced like that, but it is what it is. Desktop content in the wild definitely relies on it because it works there.
I think a compiler could optimize away nonuniformEXT if the input is subgroup uniform and the hardware only cares about that kind of uniform for descriptor access.
|
|
||
| For implementations where subgroups are packed across draws, the implementation must | ||
| make sure to handle descriptor indexing correctly. From the specification's point of view, | ||
| a dynamically uniform index does not require code:NonUniform decoration, |
There was a problem hiding this comment.
Are you sure about this? My read of the spec says the opposite:
it is mandatory for certain operands to be decorated as NonUniform if they are not guaranteed to be dynamically uniform.
Fix #2112.