[python] Remove unused MLIR components #2443
Conversation
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
5 times, most recently
from
a8cef3f to
5b4059b
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
| registerAllDialects(registry); | ||
| auto *mlirContext = unwrap(context); | ||
| cudaq::registerAllDialects(registry); | ||
| MLIRContext *mlirContext = unwrap(context); | ||
| mlirContext->appendDialectRegistry(registry); | ||
| mlirContext->loadAllAvailableDialects(); |
There was a problem hiding this comment.
Does loadAllAvailableDialects actually load Quake, CC, and CodeGen?
Do we need an explicit test to verify that all the dialects we expect to be loaded are loaded?
There was a problem hiding this comment.
loadAllAvailableDialects() will load all dialects registered in the context. So, Quake and CC will be loaded. The CodeGen dialect is not registered in cudaq::registerAllDialects. I'm not sure why, but its being registered in other places and in different ways:
I can make it uniform in a follow-up PR.
There was a problem hiding this comment.
The codegen dialect was intended to be sort of second-class in that it would only be applicable at codegen time. The point being we don't want folks to start using it in ad hoc ways in random passes, etc. Adding a few landmines was sort of the idea.
OTOH, we could clean it up a bit and remove some of those hurdles. We still don't want that dialect used "in the wild" though...
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
2 times, most recently
from
f6a8985 to
a30eddd
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
from
a30eddd to
e04b6f8
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
from
e04b6f8 to
2010d99
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
2 times, most recently
from
992ba52 to
e62b6cf
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
2 times, most recently
from
754fffd to
c30f508
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
from
c30f508 to
99ad7a6
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
from
99ad7a6 to
a156bdc
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
We don't need to take everything from MLIR for our python bindings. This change cherry picks the upstream components our compiler depends on. The commit also cleans up some unnecessary code that ends up registering dialects more than once, and surfaces the `register_all_dialects` function to a less obscure location. Signed-off-by: boschmitt <7152025+boschmitt@users.noreply.github.com>
boschmitt
force-pushed
the
shrink_nvqpp_python
branch
from
a156bdc to
479db23
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
Description
We don't need to take everything from MLIR for our python bindings. This change cherry picks the upstream components our compiler depends on.
The commit also cleans up some unnecessary code that ends up registering dialects more than once, and surfaces the
register_all_dialectsfunction to a less obscure location.EDIT: The process of registering dialects can be further improved and made completely automatic. MLIR provides a global
_dialect_registrythat gets automatically populated with all upstream dialects whenRegisterEverythingis built. This global registry is loaded to allContext()objects on its__init__method. We can populate this global registry whenevercudaq.mliris loaded so that everyContext()gets all dialects that we need. This improvement will be left for later work.