The commit 3484416b49 removed the
'rules_cc_toolchains' call from the bazel configuration. After that,
the build process is calling 'lld' to link several binaries, causing
problems on ppc64le arch.
This patch set the 'gold' linker as default in the configure script.
Eigen removed the dynamic dispatch by default. Now, it's necessary to add
EIGEN_ALTIVEC_ENABLE_MMA_DYNAMIC_DISPATCH=1 to enable this feature if
the linker version is 2.35 or greater.
7b10795e39591906477b
These checks were useful when we used to set the bazel version in the CI scripts but now that we have `.bazelversion` synced in the repo any user using Bazelisk or a recent version of Bazel will automatically get the exact Bazel version that we are using. Hence, this step in `configure.py` is no longer needed.
Furthermore, we are working on completely removing the need to run `configure.py`, at least for the vast majority of users.
PiperOrigin-RevId: 448033791
Since TF is in the progress of migrating to `cc_shared_library`, we are stress testing Bazel and uncovering a large number of bugs. We are in the state where some TF builds are broken with cascading failures and there is a long wait until a new LTS Bazel release. So, we're switching to rolling releases in an attempt to fix the builds.
However, if this causes more failures than it resolves, we will need to return back to stable releases.
PiperOrigin-RevId: 445259819
This is not actually used but to prevent confusion let's bring it up to date
PiperOrigin-RevId: 427769347
Change-Id: I1ebc7797a337b98135ee038bb8c5840c8c7c7b3c
`.bazelversion` specifies the Bazel version that TensorFlow uses. This changelist updates the Bazel version from 4.2.2 to 5.0.0. It also updates the CI environment.
PiperOrigin-RevId: 424897733
Change-Id: Idd427ee8b1fdc8212211fb22c25989cb47350ab4
This PR is to address the TF build failure reported in JIRA ticket:
http://ontrack-internal.amd.com/browse/SWDEV-241967
The valid set of values for the env var HIP_PLATFORM is being updated, and as a consequence `hcc` will no longer a valid value (staring with ROCm 3.7).
For ROCm 3.7, we will need to change the configure.py script such that the env var HIP_PLATFORM=amd is explicitly when invoked with TF_NEED_ROCM=1.
In the meantime (i.e. until we switch to ROCm 3.7), we need to provide a way for passing HIP_PLATFORM to bazel (to help QA in testing ROCm 3.7). This change provides that way.
If the HIP_PLATFORM env var is set in the shell that invokes the "configure.py" script, it will pass it through to bazel.
These are a holdover from the transition from Python 2.* to Python 3.*.
PiperOrigin-RevId: 399530502
Change-Id: I1c8b893fd7f5eeb5eb51b2e2c38c3b73ec5b9903
Explicitly listing --config=v2 incorrectly implied that building for TF2 was not the default.
PiperOrigin-RevId: 381038392
Change-Id: I6400a8edf40a7941d3b5132388d44f7aae698ea5
With the integration of rocsolver in the XLA cusolver pass, we need to set the ROCBLAS_TENSILE_LIBPATH env var. This seems to be required by rocsolver (which presumably calls rocblas internally). Not doing so leads to the following error in some of the XLA unit tests
```
test : //tensorflow/compiler/tests:dense_layer_test_gpu FAILED in 3 out of 3 in 4.7s
rocBLAS warning: No paths matched /root/.cache/bazel/_bazel_root/efb88f6336d9c4a18216fb94287b8d97/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/compiler/tests/dense_layer_test_gpu.runfiles/org_tensorflow/tensorflow/python/../../_solib_local/_U@local_Uconfig_Urocm_S_Srocm_Crocblas___Uexternal_Slocal_Uconfig_Urocm_Srocm_Srocm_Slib/library/*gfx900*co. Make sure that ROCBLAS_TENSILE_LIBPATH is set correctly.
rocBLAS error: Cannot read /root/.cache/bazel/_bazel_root/efb88f6336d9c4a18216fb94287b8d97/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/compiler/tests/dense_layer_test_gpu.runfiles/org_tensorflow/tensorflow/python/../../_solib_local/_U@local_Uconfig_Urocm_S_Srocm_Crocblas___Uexternal_Slocal_Uconfig_Urocm_Srocm_Srocm_Slib/library/TensileLibrary.dat: No such file or directory
```
Previously, XLA was enabled by default with --config=rbe_linux and when running ./configure.
PiperOrigin-RevId: 363433913
Change-Id: I9afa53496a9da2ae29b71a9ef38fc63b90dfba02
Use --config=avx_{linux,windows} instead of --config=opt in release builds. This is an NFC because release build scripts set build:opt to enable AVX.
PiperOrigin-RevId: 363193986
Change-Id: Ifb0788e29485873e35469d0936c1d33488315870
This PR introduces the support of Compute Library for the Arm
Architecture (ACL) as part of Bazel build.
Signed-off-by: Aleksandr Nikolaev <aleksandr.nikolaev@arm.com>
Remove TF_ENABLE_MLIR_GENERATED_GPU_KERNELS env trigger in ./configure for ROCm path. It has already been removed for CUDA, and ROCm currently doesn't yet support kernel gen.
PiperOrigin-RevId: 361635637
Change-Id: Id30c7c45a7fe033410457db9c9b8d51e573482b6
Before, the default PYTHON_LIB_PATH chosen was non-deterministically. This meant if ./configure was rerun and a different PYTHON_LIB_PATH was chosen, bazel would rebuild TF from scratch.
PiperOrigin-RevId: 356731041
Change-Id: Iefdebd0f6bb106a598f77e95cd36df87bb1c5d68
For GPU builds, the host compiler should get these compile flags as well.
PiperOrigin-RevId: 351470659
Change-Id: I8610ac871e370895f9234e3f047e7ee90557db86
Should resolve issue reported in #45744, #45866, #44701 and #45991 as well as multiple other issues from other ecosystem places.
We will patch 2.4 soon and release proper patches.
PiperOrigin-RevId: 350243794
Change-Id: I0aa814dedb44ac3ec992a0a432f117122337023f
Bug fixed: https://github.com/bazelbuild/bazel/issues/3828
We need this because we are working on integrating TF with a different project that should remove the breakages caused by LLVM integration
PiperOrigin-RevId: 348886284
Change-Id: Iad539321f1ba7590e4ee77b93d67eb5401d9fec8
When we switched to ROCm 3.3 (first ROCm release with "relocatable ROCm install" support), some parts of the ROCm toolchain relied upon the value of env var ROCM_ROOT to determine the location of the ROCm install dir. This has been fixed, and all of the ROCM toolchain now only uses the ROCM_PATH value.
This commit updates `configure.py` to no longer set ROCM_ROOT, when building with --config=rocm
Related JIRA ticket - http://ontrack-internal.amd.com/browse/SWDEV-226744
This flag was first made available in Visual Studio 16.4, (as of 2020-11-11,
VS is at version 16.8), and can now be considered generally available.
See also #38254
PiperOrigin-RevId: 342158585
Change-Id: Ice96be9a1392313357b314407262a005a2436a4d
This PR removes now-deprecated direct integration of nGraph in TF. nGraph
integration is still available via the external ngraph-bridge at:
https://github.com/tensorflow/ngraph-bridge/
or via `pip install ngraph-tensorflow-bridge`.
Signed-off-by: Abhishek Kulkarni <abhishek.kulkarni@intel.com>
Adds build option and supporting macros for oneDNN-reference builds on Aarch64.
For now, oneDNN will link to GCC's libgomp.so. See, related issue: https://github.com/tensorflow/tensorflow/issues/42858
Signed-off-by: cfRod <crefeda.rodrigues@arm.com>
MLIR generated gpu kernels will continue to be disabled by default.
This commit provides a way to enable them, if conifugre.py is run with the env `TF_ENABLE_MLIR_GENERATED_GPU_KERNELS=1`
Currently, if the version check fails, the error message is:
```
subprocess.CalledProcessError: Command '['bazel', '--batch', '--bazelrc=/dev/null', 'version']' returned non-zero exit status 1.
```
After this patch, it becomes:
```
Error checking bazel version: ERROR: The project you're trying to build requires Bazel 3.0.0 (specified in /usr/local/google/home/cheshire/code/opensource/docker_tf/tensorflow/.bazelversion), but it wasn't found in /usr/bin.
You can install the required Bazel version via apt:
sudo apt update && sudo apt install bazel-3.0.0
```
PiperOrigin-RevId: 312520687
Change-Id: I41523f7defa3db10aa34b6b313d6b65c792b2020
Avoid using `--enable_platform_specific_config` when cross-compiling for
iOS/Android, as this pulls in host build flags, which may not be
appropriate (e.g., when cross-compiling for Android on a Windows host).
Also fix an issue when building tensorflowlite_c for iOS.
Fixes#38525.
PiperOrigin-RevId: 311767770
Change-Id: I80b817fd89a6889dc78be50f1def8b899f091cb6
Use `bazel --version` instead of `bazel --batch version`.
`bazel --batch` fails on Windows with a "command line too long" error, if user's environment variable is too large.
Example: https://buildkite.com/bazel/bazelisk-plus-incompatible-flags/builds/496#f5fbd8db-7421-43f4-a018-555af9856be4
`bazel --version` can also print the Bazel version without starting Bazel server, it's even faster than `bazel --batch version`.
PiperOrigin-RevId: 310570885
Change-Id: Iafc4c90f0ff57610e5f77bee230e81e78d9f1289
Environment variables are strings, not integers.'0' does not evaluate to false.
PiperOrigin-RevId: 297916971
Change-Id: I9df143690b16a181fa5f0c4a60ac2f944b396672
Remove the prompt for XLA in configure. But still allow users to build without it for platforms like raspberry pi.
PiperOrigin-RevId: 297713015
Change-Id: I7a12b073389c1cf3d265e00952f28e98470c80ff
This is necessary to start using bazel's experimental cc_shared_library support.
This change is part of the refactoring described in https://github.com/tensorflow/community/pull/179
PiperOrigin-RevId: 294736824
Change-Id: I8d13c71545cf2f2edd47818b40fa413e3e9c3254
Today if you run the ./configure script without setting the
variable TF_NEED_CUDA, in the resulting .tf_configure.bazelrc
you get:
test:v1 --test_tag_filters=-benchmark-test,-no_oss,-gpu,-oss_serial
test:v1 --build_tag_filters=-benchmark-test,-no_oss,-gpu
test:v2 --test_tag_filters=-benchmark-test,-no_oss,-gpu,-oss_serial,-v1only
test:v2 --build_tag_filters=-benchmark-test,-no_oss,-gpu,-v1only
This is incorrect because -gpu means exclude the gpu test. It should be
-no_gpu.
Debugging the problem I found when the code was switched from using
os.env to using environ_cp, that the method system_specific_test_config
was never updated. So unless the environment variable TF_NEED_CUDA was
set before running ./configure then answering yes for CUDA would not
select the correct test filter for gpus.
With this change the test filters are correct:
test:v1 --test_tag_filters=-benchmark-test,-no_oss,-no_gpu,-oss_serial
test:v1 --build_tag_filters=-benchmark-test,-no_oss,-no_gpu
test:v2 --test_tag_filters=-benchmark-test,-no_oss,-no_gpu,-oss_serial,-v1only
test:v2 --build_tag_filters=-benchmark-test,-no_oss,-no_gpu,-v1only
This is necessary to start using bazel's experimental cc_shared_library support.
This change is part of the refactoring described in https://github.com/tensorflow/community/pull/179
PiperOrigin-RevId: 288808507
Change-Id: Ie78fb4ff8dad128ebef280037cce4d3c4f42addc
