Option to set the number of threads for parallel compilation #3048
Conversation
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
src/Compiler/CompilerOptions.cpp
Outdated
| @@ -92,6 +92,7 @@ OptReport optReport; // onnx-mlir only | |||
| bool useOldBufferization; // onnx-mlir only | |||
| bool enableTiming; // onnx-mlir only | |||
| bool enableBoundCheck; // onnx-mlir only | |||
| int compilationNumThreads; // onnx-mlir only | |||
There was a problem hiding this comment.
Can we apply this option to onnx-mlir-opt as well so that we can check performance by calling onnx-mlir-opt --constprop-onnx -j8 for a single operation? It's ok if it needs another PR.
There was a problem hiding this comment.
@tungld I was thinking along the same lines, also for example to test that we get the same result for a test that uses parallelism.
@imaihal I would then add a test that tries the operation that you parallelized and try it with -j 4 for example. You may have to copy a large constant to do that... can lift it from a model.
There was a problem hiding this comment.
@tungld @AlexandreEichenberger Thanks for the comments.
I applied this option to onnx-mlir-opt, but parallelization of constprop is not included in this PR (It is included in PR #3042). So I want to add the tests in PR #3042.
There was a problem hiding this comment.
Absolutely, this PR set the options, and it will be tested once there is a new opt that can use parallelism.
There was a problem hiding this comment.
@imaihal I didn't see any change to onnx-mlir-opt.cpp (e.g. MLIRContext in onnx-mlir-opt.cpp). I am curious on how this option is effective inonnx-mlir-opt command line?
There was a problem hiding this comment.
@tungld New thread pool is created in MlirOptMain.cpp of MLIR. This is due to handle-split-input-file option (https://github.com/llvm/llvm-project/blob/main/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp#L609-L613). So, it seems to be difficult to set our own threadpool in onnx-mlir-opt.cpp.
In lit tests, how about using--mlir-disable-threading for single thread and using mlir default(use all cpus) for multi-thread? Or, using pass local option might be possible to specify our own thread pool.
There was a problem hiding this comment.
@tungld Sorry, it is possible just by modifying onnx-mlir-opt.cpp. I updated.
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
|
@imaihal Is your PR ready for review again, meaning do you feel you acted on all the questions/requests? |
|
@AlexandreEichenberger Yes. Could you review this again? |
There was a problem hiding this comment.
LGTM, thanks @imaihal for incorporating the feedback from Tong and I.
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
Signed-off-by: Haruki Imai <[email protected]>
|
Jenkins Linux amd64 Build #16186 [push] Option to set the number... started at 22:02 |
|
Jenkins Linux ppc64le Build #15215 [push] Option to set the number... started at 23:21 |
|
Jenkins Linux s390x Build #16188 [push] Option to set the number... started at 23:02 |
|
Jenkins Linux amd64 Build #16186 [push] Option to set the number... passed after 1 hr 25 min |
|
Jenkins Linux s390x Build #16188 [push] Option to set the number... passed after 1 hr 28 min |
|
Jenkins Linux ppc64le Build #15215 [push] Option to set the number... passed after 2 hr 26 min |
All available CPUs are used in parallel compilation in current main branch.
This PR creates an option to set the number of threads for parallel compilation. By using
-j <#threads>when compiling, thread pool with<#threads>threads are created, and used for parallel compilation inparallelFor()API.The default value is 0 and all the available CPUs are used as same with current main. When setting
-j 1, use single thread without setting thread pool.