Fix MXNet R package build#13952
Conversation
|
@xinyu-intel could you help take a review? Does this PR conflict with yours #13867 ? |
|
@rajeshii Hi, thanks for your contribution! Can you also help add R unittest with MKL-DNN to the CI(./ci/jenkins/Jenkins_steps.groovy)? |
|
@xinyu-intel @pengzhao-intel Added and passed the CI:) |
|
@hetong007 could you help take a review? |
|
LGTM. @marcoabreu do you have any comments? |
| custom_steps.test_unix_python3_tensorrt_gpu(), | ||
| custom_steps.test_unix_perl_gpu(), | ||
| custom_steps.test_unix_r_gpu(), | ||
| custom_steps.test_unix_r_mkldnn_gpu(), |
There was a problem hiding this comment.
I'm not sure whether we need GPU+MKLDNN for R since we already test it in python. Could you elaborate what concerns you are specifically having?
|
Ideally, we could do some smoketests for different compilation options by running a simple test to ensure everything is working. Considering we test different backends in Python already, I don't think there's much benefit of now testing the full spectrum across the different frontend languages. As far as I understand, the goal is to ensure everything gets compiled and linked properly, right? |
|
Agree that we only need run minimal tests to make sure libraries can be found and properly linked during runtime. |
|
@marcoabreu @TaoLv could you help to take a review again? |
|
My feedback regarding the amount of testing is still unchanged.
Am Fr., 25. Jan. 2019, 04:15 hat PatricZhao <[email protected]>
geschrieben:
… ***@***.**** approved this pull request.
LGTM
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13952 (review)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ARxB6_mAvdZCdFElOe915hWGp3nAVQYIks5vGnbMgaJpZM4aMBrU>
.
|
pengzhao-intel
left a comment
There was a problem hiding this comment.
@marcoabreu sorry, I mixed up the different PR.
@rajeshii could you update the PR based on the change request?
|
@marcoabreu Sorry for the late response. I'm confused that whether I should skip the rpkg test for mkldnn backend and leave only the rpkg build in CI or I need to add a lite test for it. For the later one, I don't know how to design the test. What's your opinion @hetong007 ? |
|
Think about a small task that just creates a 5x5 matrix of ones. Just to
make sure everything is properly linked at runtime. Or pick a random
existing test and run that one.
Am So., 27. Jan. 2019, 14:05 hat rajeshii <[email protected]>
geschrieben:
… @marcoabreu <https://github.com/marcoabreu> Sorry for the late response.
I'm confused that whether I should skip the rpkg test for mkldnn backend
and leave only the rpkg build in CI or I need to add a lite test for it.
For the later one, I don't know how to design the test. What's your opinion
@hetong007 <https://github.com/hetong007> ?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13952 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ARxB6xF4isZUoIAIGayi2OhWgjGkKRUpks5vHaQGgaJpZM4aMBrU>
.
|
|
@marcoabreu addressed your comments. It seems that all tests have passed but the CI is pending... |
|
Rebase three times with different kinds of CI failures:( |
|
@pengzhao-intel @xinyu-intel @marcoabreu All tests pass now:) |
| R_LIBS=/tmp/r-site-library | ||
|
|
||
| R CMD INSTALL --library=/tmp/r-site-library R-package | ||
| R_LIBS=/tmp/r-site-library Rscript -e "library(mxnet); as.array(mx.nd.ones(c(2,3)))" |
There was a problem hiding this comment.
Could we add a more meaningful test, like convolution, FC or relu which will hit the real MKLDNN calculation?
We need to guarantee R can execute at least in the CI.
* fix mxnet r package build * add ci * remove mkldnn-gpu test for R * add minimal test for MKLDNN-R * pick mlp as minimal R test
* fix mxnet r package build * add ci * remove mkldnn-gpu test for R * add minimal test for MKLDNN-R * pick mlp as minimal R test
* fix mxnet r package build * add ci * remove mkldnn-gpu test for R * add minimal test for MKLDNN-R * pick mlp as minimal R test
Description
As mentioned in #13859 & #13936 , below error will occur when building R-Package on my ubuntu:
It seems that some dependent libraries of mxnet should also be copied to R directory.
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
Comments