Some users do not want or need this support and it significantly increases
their dependencies and binary sizes. After this change, users must explicitly
depend on :gcs_file_system to enable this support.
Change: 143129406
Additionally:
- change single quotes to double quotes to make path rewriting easier
- guard windows lib reference with PLATFORM_WINDOWS
- fixed failing kmeans test
Change: 141515942
DirectSession support for TensorFlow Debugger (tfdbg). Binaries that do not
want debugging support can avoid this dependency and its transitive deps. This
replaces the previous approach that was based on a preprocessor flag
(-DNOTFDBG).
Change: 141321165
1. Created open-source target libdevice_root that wraps all libdevice files.
2. platform/posix/cuda_libdevice_path depends on the libdevice_root target.
3. Added cuda_libdevice_path_test that verifies libdevice files exist in the
computed libdevice directory.
Change: 141237087
On macOS, poll(2) may return POLLNVAL for a valid file descriptor that was
closed on its remote side (by the other process) before this process calls
poll() the first time. Treat this case as an error.
Re-enable the SubProcess tests, one of which (StdinChildExit) appeared
flaky on macOS depending on the timing of the child process exiting and
the parent process first calling poll().
Change: 139850066
It currently does not (0.11.0rc2), however, with this unittest
future changes to the TF_VERSION_STRING should follow the
semantic versioning 2.0 spec described at www.semver.org
Change: 139404692
Add to TensorFlow a subset of the functionality of SubProcess to support the
open-sourceable parts of XLA and support TensorFlow's own testing.
On google3 let it use the internal SubProcess from google3. In the open-source
world, use the TF implementation of the feature subset.
Modify XLA and some TF tests to use the TF SubProcess.
Change: 139114815
Mostly drop-in replacements for std::unordered_map and std::unordered_set,
but much faster (does not do an allocation per entry, and represents
entries in groups of 8 in a flat array, which is much more cache efficient).
Benchmarks not included in this cl show about 3X to 5X performance
improvements over the std::unordered_{set,map} for many kinds of
common maps e.g. std::unordered_mapmap<int64, int64> or
std::unordered_map<string, int64>.
Change: 137401863
rule for lib/framework, instead of manually curating the filegroup.
Note that this has to expose "internal" headers, such as
array_slice_internal.h, because array_slice.h includes array_slice_internal.h
and thus needs to be in the package to build. This does *not*
mean that array_slice_internal is public, just that it must be
in the package.
If users start depending on internal headers, we can break them.
The policy is that only those headers that are visible in bazel
using "//visibility:public" are actually public (that is what
bazel can enforce). Furthermore, this policy extends to
what we do for python: only documented headers/functions are
public, even if they are 'includable'.
Fixes#4996.
Change: 137219132
