[INTEL MKL] Setting default value for KMP_BLOCKTIME

[vishakha-nervana]

To improve overall user experience and performance, this PR sets KMP_BLOCKTIME to 1, if not set already by user.

[penpornk]

Thank you for the PR!

[penpornk]

@alenik01 This PR has changes that could affect --config=mkl_aarch64 build. Could you please help check if things are okay on your end? (I plan to cherry-pick this into TF 2.5.)

[alenik01]

Good morning, @penpornk Many thanks for the heads up. Mmain problem which I encounter with this commit is that it doesn't build even with --config=mkl, because Bazel for some reason fails to find external/llvm_openmp/include/omp.h. The similar situation is also for --config=mkl_aarch64. Did you manage to build the master after merging this commit?

Execution platform: @local_execution_config_platform//:platform
tensorflow/core/common_runtime/threadpool_device.cc:19:10: fatal error: external/llvm_openmp/include/omp.h: No such file or directory
19 | #include "external/llvm_openmp/include/omp.h"
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.
Target //tensorflow/tools/pip_package:build_pip_package failed to build

[penpornk]

Good morning, @alenik01! We decided not to cherry-pick this PR into 2.5 for now.

Did you manage to build the master after merging this commit?

I didn't try building it manually, but the oneDNN CI test for this PR was fine.

[alenik01]

Good afternoon, @penpornk

Could you please help check if things are okay on your end?

I have checked --config=mkl_aarch64 with the current TF master and it works correctly on AArch64 server. The only disadvantage for us is that additional --cxxopt="-DDNNL_AARCH64_USE_ACL=1" has to be added now in the build command, i.e. defined globally for all the build instead of the necessary files only, I will try to find a more elegant solution.

[haitong]

we should not merge this PR as tensorflow build failed with --config=mkl. it is surprising that no one complained about it except @alenik01 alenik01

[haitong]

For example, tfserving build is failing due to this tensorflow/serving#1958 (comment)

[penpornk]

cc: @TensorFlow-MKL @agramesh1

[agramesh1]

HI @penpornk we can revert this PR, it does not work the way it was intended. We will submit a PR to revert. Thanks.