Make _get_program_from_buffer work for bundled programs take 2 #14503
Conversation
Add some cmake to only do this if executorch is built with bundleio. codegen/tools subdirectory needs to be moved in top-level CmakeLists.txt to have access to the bundled_program target. Signed-off-by: Erik Lundell <[email protected]> Change-Id: Ic953bdfd5223eb835e46150bde19b08b082e8444
Signed-off-by: Erik Lundell <[email protected]> Change-Id: Ib7fbdb7190312320d73a937bc285db7140d2bb1a
Erik-Lundell
requested review from
larryliu0820,
lucylq,
kirklandsign,
digantdesai and
mergennachin
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14503
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New Failures, 2 Cancelled Jobs, 3 Unrelated FailuresAs of commit 390840c with merge base df5bfd5 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
Merge should wait until #14488 to make sure Meta's internal build is not broken |
Erik-Lundell
added
partner: arm
| ) | ||
|
|
||
| # Link against required libraries | ||
| if(TARGET bundled_program) |
There was a problem hiding this comment.
Should you check EXECUTORCH_BUILD_DEVTOOLS?
Something like
if (EXECUTORCH_BUILD_DEVTOOLS and TARGET bundled_program)
...
There was a problem hiding this comment.
I went with this since it is done that way on L754 in top-level CMakeLists.txt. Is there a benefit in additionally checking the flag or just to be on the safe side?
| target_compile_definitions(selective_build PRIVATE -DET_BUNDLE_IO) | ||
| target_link_libraries(selective_build PRIVATE bundled_program) |
There was a problem hiding this comment.
You also wanna change the targets.bzl file too
I added a CI to build buck target for selective_build
For future reference, here's how you can download buck:
python tools/cmake/resolve_buck.py --cache_dir==/tmp/
# it will download and output a buck executable file within /tmp/ file
Once you have the buck file, you can do something like this:
/tmp/buck2 build //codegen/tools/...
# build all targets within //codegen/tools/...
/tmp/buck2 build codegen/tools:selective_build
# build all targets
There was a problem hiding this comment.
I had a look and the bundled_program seems to have multiple targets runtime and runtime_aten, depending on some get_aten_mode_options(). I need some assistance to figure this out. From what I can tell, if I add the different targets to selective_build that target is also split which means that targets depending on it will also split and so on... That doesn't sound right.
There was a problem hiding this comment.
@Erik-Lundell -- no need to bifurcate on aten vs not. You can add dependency on "//executorch/devtools/bundled_program:runtime
Historically, runtime_aten is mainly for internal customers when portable ops coverage was low. That's not the case, so you should be able to just depend on "//executorch/devtools/bundled_program:runtime directly
|
@mergennachin has imported this pull request. If you are a Meta employee, you can view this in D83148788. |
This is a re-upload of #14435, with variables moved inside #ifdef to not get unused variables when building without devtools.
Add some cmake to only do this if executorch is built with bundleio.
codegen/tools subdirectory include needs to be moved in top-level CmakeLists.txt
to have access to the bundled_program target.
Follow-up patch to enable the fix in arm backend.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218