[tensorflow-cc] fix build issues on Windows#12546
[tensorflow-cc] fix build issues on Windows#12546strega-nil merged 11 commits intomicrosoft:masterfrom
Conversation
…se customized bazel config is stored in wrong directory
|
@jgehw Please sign CLA first. Thanks. |
|
/azp run |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
The CI runs on windows fail due to some access denied issue while setting up MSYS2. No idea what's going on there. |
|
Will fix in #12603 |
|
When building x64-windows: |
… way it auto-detects include path avoiding mismatches between toolset and include files versions
|
Can you please re-run the failed x64_windows port? (I checked the build error logs and found a bazel package download error - bad timings because a server was down over the weekend. I had this error on my machine, too, and it disappeared after the server was back on.) |
|
/azp run |
|
Commenter does not have sufficient privileges for PR 12546 in repo microsoft/vcpkg |
|
Reruning... |
|
@jgehw The new error: |
… and tries to parse it), so fix the issue of inconsistency of toolset versions between auto-detection in vcpkg and bazel by letting vcpkg determine everything including toolset version
|
can you please re-run again? (the errors of the automated run seem to me to come from azure devops pipepline early stage, hopefully transient) |
Done. |
|
…into fix-ms-build-issues
It doesn't work if more than one VC toolkit is installed. Likely a bug in bazel. bazel is currently v3.4.1, however tensorflow v1.14 requires really outdated bazel v0.25. Going to upgrade tensorflow to v2.3 and then using up-to-date bazel, but will do this in a separate pull request, as it requires more work.
|
It turned out that tensorflow v1.14 requires really outdated bazel v0.25, so I reverted the bazel upgrade to 3.4.1 and set x64-windows CI to fail again. It only works if no more than one VC toolkit is installed. So the PR still has its value, but the CI won't succeed. As ultimate solution I'm going to upgrade tensorflow to v2.3 and then using up-to-date bazel, but will do this in a separate PR, as it requires more work. |
occurring from default bazel config being used because customized bazel config is stored in wrong directory:
Omitting the --workspace parameter of configure.py defaults not the the current working directory but to the directory where configure.py is extracted to. This results in the customized bazel config being stored in the wrong directory and thus the default bazel config being used. This, in turn leads to tons of build errors, mainly of two types:
.\tensorflow/stream_executor/rng.h(66): error C2589: 'constant': illegal token on right side of '::'.\tensorflow/stream_executor/rng.h(66): error C2059: syntax error: '::'.\tensorflow/stream_executor/rng.h(72): error C2589: 'constant': illegal token on right side of '::'.\tensorflow/stream_executor/rng.h(72): error C2059: syntax error: '::'Closes [tensorflow-cc:x64-windows] build failure #12506