In the future, we'd like to have all drivers only ingest unified atomics, and
all frontends only produce unified atomics, and garbage collect the existing
non-unified atomics. To get to that future, it's a lot nicer to convert drivers
one-by-one. Add a pass to translate old-style atomics to new-style atomics so
drivers can opt-in to the new form one-by-one. Once all drivers are converted,
we can convert producers one-by-one. Finally, we can just drop the calls to the
pass and garbage collect this pass and the old atomics. That's probably a while
out, though, so this will be out bridge to get there.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22914>
Currently, we have an atomic intrinsic for each combination of memory type
(global, shared, image, etc) and atomic operation (add, sub, etc). So for m
types of memory supported by the driver and n atomic opcodes, the driver has to
handle O(mn) intrinsics. This makes a total mess in every single backend I've
looked at, without fail.
It would be a lot nicer to unify the intrinsics. There are two obvious ways:
1. Make the memory type a constant index, keep different intrinsics for
different operations. The problem with this is that different memory types
imply different intrinsic signatures (number of sources, etc). As an
example, it doesn't make sense to unify global_atomic_amd with
global_atomic_2x32, as an example. The first takes 3 scalar sources, the
second takes 1 vector and 1 scalar. Also, in any single backend, there are a
lot more operations than there are memory types.
2. Make the opcode a constant index, keep different intrinsics for different
operations. This works well, with one exception: compswap and fcompswap
take an extra argument that other atomics don't, so there's an extra axis of
variation for the intrinsic signatures.
So, the solution is to have 2 intrinsics for each memory type -- for atomics
taking 1 argument and atomics taking 2 respectively. Both of these intrinsics
take an nir_atomic_op enum to describe its operation. We don't use a nir_op for
this purpose, as there are some atomics (cmpxchg, inc_wrap, etc) that don't
cleanly map to any ALU op and it would be weird to force it.
The plan is to transition to these new opcodes gradually. This series adds a
lowering pass producing these opcodes from the existing opcodes, so that
backends can opt-in to the new forms one-by-one. Then we can convert backends
separately without any cross-tree flag day. Once everything is converted, we can
convert the producers and core NIR as a flag day, but we have far fewer
producers than backends so this should be fine. Finally we can drop the old
stuff.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22914>
Serious preprocessor voodoo here. There are two tricks here.
1. Iterating only phis. We know that phis come only at the beginning of a block,
so all over the tree, we open-code iteration like:
nir_foreach_instr(instr, block) {
if (instr->type != phi)
break;
/* do stuff */
}
We can express this equivalently as
nir_foreach_instr(instr, block)
if (instr->type != phi)
break;
else {
/* do stuff */
}
So, we can define a macro
#define nir_foreach_phi(instr, block)
if (instr->type != phi)
break;
else
and then
nir_foreach_phi(..)
statement;
and
nir_foreach_phi(..) {
...
}
will expand to the right thing.
2. Automatically getting the phi as a phi. We want the instruction to go to some
hidden variable, and then automatically insert nir_phi_instr *phi =
nir_instr_as_phi(instr_internal); We can't do that directly, since we need to
express the assignment implicitly in the control flow for the above trick to
work. But we can do it indirectly with a loop initializer.
for (nir_phi_instr *phi = nir_instr_as_phi(instr_internal); ...)
That loop needs to break after exactly one iteration. We know that phi
will always be non-null on its first iteration, since the original
instruction is non-null, so we can use phi==NULL as a sentinel and express a
one-iteration loop as for (phi = nonnull; phi != NULL; phi = NULL).
Putting these together gives the macros implemented used.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22967>
Up until now, we have been initializing MCS with fast clears. This is
mostly safe, but there's a corner case that can be an issue.
The issue is with a workaround for MCS that requires the sampler not see
any fast-cleared blocks for certain surfaces (14013111325). Even though
we have been initializing MCS with fast clears, we expect most
applications to be safe because we expect that they would only sample
the samples they've rendered to previously (and the render would've
removed the fast-cleared blocks). In other words we don't expect that
apps would transition from VK_IMAGE_LAYOUT_UNDEFINED to
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL and start sampling immediately.
If an application took the unexpected path of sampling undefined
samples, it's possible they'd hit the issue described in the workaround.
Fix this corner case by using an ambiguate to initialize MCS.
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22545>
Alyssa Rosenzweig