Support nvcc --device-debug flag
#2384
Conversation
|
cc: @robertmaynard for review |
|
Can you take a look into the CI tests. This might have some overlap with the CI failures in #2382 |
|
@drahnr the integration tests are failing due to this: https://github.blog/changelog/2025-03-20-notification-of-upcoming-breaking-changes-in-github-actions/#decommissioned-cache-service-brownouts, which is being addressed in #2385 |
|
@trxcllnt could you rebase your PRs since the CI fixup landed in |
…ts can run in parallel
These changes ensure cache hits for compilations which are subsets of previously cached compilations * Normalize cudafe++, ptx, and cubin names regardless of whether the compilation flag is `-c`, `-ptx`, `-cubin`, or whether there are one or many `-gencode` flags * Include the compiler `hash_key` in the output dir for internal nvcc files to guarantee stability and uniqueness * Fix cache error due to hash collision from not hashing all the PTX and cubin flags
add more multi-arch tests to ensure combining cached/new PTX and cubins doesn't produce corrupted objects
…nding deprecation
trxcllnt
force-pushed
the
fea/nvcc-device-debug
branch
from
1960f9b to
29cc48d
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2384 +/- ##
==========================================
- Coverage 71.41% 69.50% -1.92%
==========================================
Files 65 45 -20
Lines 36349 26611 -9738
==========================================
- Hits 25960 18496 -7464
+ Misses 10389 8115 -2274 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
src/compiler/cicc.rs
Outdated
| @@ -118,7 +119,7 @@ where | |||
| let mut take_next = false; | |||
| let mut outputs = HashMap::new(); | |||
| let mut extra_dist_files = vec![]; | |||
| let mut gen_module_id_file = false; | |||
| // let mut gen_module_id_file = false; | |||
There was a problem hiding this comment.
Could you delete the comment?
src/compiler/cicc.rs
Outdated
| @@ -146,7 +161,7 @@ where | |||
| } | |||
| Some(GenModuleIdFileFlag) => { | |||
| take_next = false; | |||
| gen_module_id_file = true; | |||
| // gen_module_id_file = true; | |||
There was a problem hiding this comment.
Please delete commented code
| ); | ||
| } else { | ||
| extra_dist_files.push(module_id_path); | ||
| if module_id_path.exists() { |
There was a problem hiding this comment.
If the generated file doesn't exist, does this exhibit an issue or is it ok to silently ignore?
| @@ -158,24 +173,6 @@ where | |||
| take_next = false; | |||
| &mut unhashed_args | |||
| } | |||
| Some(UnhashedOutput(o)) => { | |||
There was a problem hiding this comment.
Any particular reason to rename as part of this PR?
| m.update(cwd.join(output).as_os_str().as_encoded_bytes()); | ||
| m.finish() | ||
| }); | ||
| fs::create_dir_all(&out_dir).ok(); |
There was a problem hiding this comment.
The above line is odd. Do we not want to bail if the IO operation fails?
|
|
||
| #[allow(unused)] | ||
| impl SccacheClient { | ||
| pub fn new_no_cfg() -> Self { |
There was a problem hiding this comment.
Could use a comment or two
|
The changes in this PR are related to the Do you want to review and merge each of these separately, or do everything in this PR? |
trxcllnt
force-pushed
the
fea/nvcc-device-debug
branch
2 times, most recently
from
4535e5a to
63e9d72
Compare
trxcllnt
force-pushed
the
fea/nvcc-device-debug
branch
from
63e9d72 to
4765772
Compare
trxcllnt
force-pushed
the
fea/nvcc-device-debug
branch
from
4765772 to
9d4e3c9
Compare
This PR extracts the changes to support
nvcc --device-debugfrom #2356 per @drahnr's request.This PR branched from/is a follow-up to #2383. See the diff of the two branches here.
This PR adds support for
nvcc -G/nvcc --device-debug. Presently-Gis interpreted as a GCC argument that expects a numeric value. nvcc doesn't support that, so instead it should be intercepted and passed to thenvcc --dryruncommands.