[cuda] Move to hal/drivers and wire up BUILD files #14620
Conversation
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
2 times, most recently
from
ff8ae89
to
fadf0c8
Compare
| option(IREE_HAL_DRIVER_CUDA "Enables the 'cuda' runtime HAL driver" ${IREE_HAL_DRIVER_CUDA_DEFAULT}) | ||
| option(IREE_HAL_DRIVER_CUDA2 "Enables the 'cuda2' runtime HAL driver" ${IREE_HAL_DRIVER_CUDA2_DEFAULT}) |
There was a problem hiding this comment.
Can we come up with a plan for switching from the old cuda to the new cuda2? Maybe post an RFC and an update to #13245, and could link to one or both of those here.
Some ideas:
- Drive to feature complete
- All CTS tests and execution tests we care about pass using
cuda2 - Figure out the story for flags like on Disable CUDA memset emulation when cuGraphAddMemsetNode fix is deployed. #13984
- All CTS tests and execution tests we care about pass using
- Switch default in-repo
- All in-repo benchmarks using
cuda2, without egregious regressions
- All in-repo benchmarks using
- Delete (old)
cudaand renamecuda2tocuda
Ideally we'd have a timeline in mind so we don't stay in this state of having two similar paths for long.
There was a problem hiding this comment.
SG. I've commented on #13245 with #13245 (comment). I'll send out a mail later too.
There was a problem hiding this comment.
Hard for me to give a precise timeline though; given I'm multiplexing over quite a few work streams. But yeah agreed we don't want to be in a hybrid state for too long; I'm hoping to get feature parity down soon.
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
3 times, most recently
from
c2a2284
to
5dcb81f
Compare
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
2 times, most recently
from
a5c70a0
to
5009f18
Compare
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
from
5009f18
to
c17e227
Compare
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
3 times, most recently
from
5903d84
to
104cbcd
Compare
This commit moves the CUDA HAL driver rewrite to the `hal/drivers` directory given it's functional and ready for normal usage. By this we can start run tests with CI to make sure it does not regress. Further improvements can happen directly in this directory. This provides an easy route for trying out the rewrite before eventually replace the existing HAL driver. Along the way wired up BUILD configurations.
antiagainst
force-pushed
the
cuda2-to-hal-drivers
branch
from
104cbcd
to
f441834
Compare
|
Hi @benvanik, I have addressed all known issues so this is ready for another review now. After moving I'd be switching the tests/benchmarks to use the new impl to test out. |
|
woo excited to get rid of the old impl! |
This commit moves the CUDA HAL driver rewrite to the
hal/driversdirectory given it's functional and ready fornormal usage. By this we can start run tests with CI
to make sure it does not regress. Further improvements
can happen directly in this directory.
This provides an easy route for trying out the rewrite before
eventually replace the existing HAL driver.
Along the way wired up BUILD configurations.
Progress towards #13245