October 2017 version of Kaldi #8
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… mini_librispeech (#4388)
* Make BatchedThreadedNnet3CudaOnlinePipeline::CorrelationID public (fallout from #4490). * Replace Win32-specific Sleep with portable C++ std::chrono classes in src/base/kaldi-utils.*. Change `Sleep(float)` arg type to `double` because this is what these classes take and return when converting to/from time in seconds. * Use C++ constants for timers instead of sesquipedalian macros. * Move includes under `HAVE_CUDA` conditionals in files without code that would make no sense without CUDA. * Issue an #error if attempting to compile CUDA-dependent code with HAVE_CUDA not set. * Add, reword some and reformat commentary and help strings in cudadecoderbin/batched-wav-nnet3-cuda-online.cc Accidental touch-paint fixes: * Replace usleep() with kaldi::Sleep in util/kaldi-table-test.cc. * Fix escape non-sequence '\%' where '%' was intended in strings in src/nnet3/nnet-utils.cc.
* KALDI_ASSERT enough space in a matrix column before memcpy()'ing into it. * [chore] rearrange include guard to make including the file invariant w.r.t. HAVE_CUDA defined, and rename the guard preprocessor var to the coding conventions.
* Add two static member functions to CuDevice which may be called before Initialize() to set boolean options that currently settable only from config file: - EnableTensorCores() - EnableTf32Compute() Chores: * Remove unneeded inline keywords and mark const functions const. * Rearrange includes, fix spacing, wrapping and commentary to the coding standard, and for Doxygen. * Remove extra vertical and all EOL whitespace.
Not need to check is $data/feats.scp exist
Change all uses of path variables to consistently use CMAKE_CURRENT_* variables. Previously a combination of CMAKE_* and CMAKE_CURRENT_* variables was used, which lead to error when Kaldi was added to another cmake project using add_subdirectory.
Co-authored-by: Daniel Galvez <[email protected]>
) Also add several defensive CU_SAFE_CALL guards that weren't there before. Co-authored-by: Daniel Galvez <[email protected]>
Before C++11 we declared the same members private, usually at the very end of a class declaration. With standardization of the delete keyword and adoption of the pattern, it is now idiomatic to make deleted members public. The rationale is that deletion of copy members is part of the public behavior of a type, namely its uncopyability, and should be expressed in and readable from the public: section alone. Google coding style, which we are generally following, does not mention the 'DISALLOW_COPY_AND_ASSIGN' macro any more, and recommends using the '= delete;' construct directly. From https://google.github.io/styleguide/cppguide.html#Copyable_Movable_Types: > Every class's public interface must make clear which copy and move > operations the class supports. This should usually take the form of > explicitly declaring and/or deleting the appropriate operations in > the public section of the declaration. Examples from the guide use the delete keyword straight without any macro, but I think we should make an exception and continue using the macro, which is more self-explanatory. clang-tidy gets angry at non-public deleted members, too.
1. Order SUBDIRS alphabetically. 2. Obtain MEMTESTDIRS by set subtraction: there are only 3 members of SUBDIRS not in it. 3. Add a phony 'libs' target to build libs only. Just because. 4. Abbreviate a common dependency 'BMU = base matrix util', and order all listed dependencies of each target alphabetically (e.g., 'decoder: $(BMU) fstext gmm hmm lat transform tree'). 5. Use true and false as values for the variable 'with_cudadecoder' in ./configure, as all other variables do. Change Makefiles in related subdirs to use the new value. 6. Bump 'CONFIGURE_VERSION' to 14 because of this. 7. Never set 'WITH_CUDADECODER = true' in kaldi.mk if CUDA is false or undefined. The previous default disregarded the detection of CUDA, the root cause of issue #4544. 8. Declare the 'depend' target phony. 9. Demote sgmm2 and sgmm2bin to EXTRA_SUBDIRS.
* Unpin memory of h_all_waveform_, deallocated but left pinned upon instance destruction. * In the same added destructor, invoke WaitForLatticeCallbacks() to prevent a short race between the destroying of this instance and completion of callback threads, which decrement the pending callback count _after_ the callback has returned and deleted. * Assert the quiescent object state the destructor.
BatchedThreadedNnet3CudaOnlinePipeline never freed a CudaFst object it initialized. Freeing it caused an error returned from cudaFree (with our allocator disabled). The root cause of this problem was copying of the CudaFst in CudaDecoder's constructor. The code has been changed so that CudaDecoder only stores a reference to CudaFst, and the object itself is uniquely owned by the pipeline. This allowed folding of separate CudaFst::Initialize and CudaFst::Finalize methods into its respective constructor and destructor. There is also a proof-of-concept unique_device_ptr, a specialization of std::unique_ptr for device-allocated memory. The type is declared privately in, and only used by CudaFst. The Finalize method had been removed entirely, and the memory is freed in the object's destructor, as the members are destroyed.
Co-authored-by: Hossein Hadian <[email protected]>
When python3 and numpy are available, verify that the generated file is successfully loaded by numpy without exceptions by invoking 'python3 -c ...' via a system() call. Don't leave behind the temporary file. If python3 and numpy aren't both available, print a warning and kill the file.
* Took heed from static analysis about an unpaired delete. * Remove an unused data member and a GetPathName overload. * Move the remaining used GetPathName() helper into the .cc file, as it is a static helper function which doesn't semantically belong in the class. * Adjust formatting to the coding style (not exhaustively).
Remove 'template<bool B> class KaldiCompileTimeAssert' because 'static_assert' is now a language declaration. Before C++17 a second argument for the message is required (lame, lame!), which is, lacking anything better, is obtained by stringizing the function-like macro argument. For C++17, the second argument is optional, and the semantics of 'KALDI_COMPILE_TIME_ASSERT' and 'static_assert' is identical when the second argument to the latter is not provided.
* Initialize CuDevice::curand_handle_ to NULL in constructor. * Remove stray include guards from cu-common.cc. * Rearrange #include's per coding guidelines.
Also reformat string literals that ended up chopped into too many pieces over too many lines, and add commentary on the importance of data member ordering within the class.
Other minor tweaks: * Convert runtime asserts to static asserts in template. * Remove ThreadPoolLight::nworkers_ as it's equal to the .size() of the worker vector, and was used only once. * Replace pointer indirection with std::move and make ThreadPoolLightWorker::thread_ a direct class member. * Update to coding guidelines and IWYU.
…4569) * Wait for D2H then H2H transfers to dry out before stopping the pump. * Assert no H2H tasks after it stops. * Assert that an offset into vector is within bounds (in the #4556 scenario, it's not). * Add CU_SAFE_CALL to every CUDA call. * Use compile time asserts on constants instead of runtime asserts * The rest is reformatting comments, IWYU and updates to coding conventions.
Fix missing FLT_MAX in some CUDA installation scenarios.
Fix reported issues w.r.t python2.7 and some apple silicone quirks
Support for both OpenFST 1.7.3 and 1.8.2
* upload the error logs to artifact repository * fix yaml error
* upgrade the checkout action on build pipeline * upgrade the checkout action on build pipeline
Make it buildable under Fedora41 using their board tools
* Support OpenBLAS with WASM * Update README * Force g_num_threads = 0 in WASM
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.