In implementing https://github.com/jax-ml/jax/pull/25787, I realized that while we lower `tridiagonal_solve` to cuSPARSE on GPU, we were using an explicit implementation of the Thomas algorithm on CPU. We should instead lower to LAPACK's `gtsv` on CPU because it should be more numerically stable and faster.
PiperOrigin-RevId: 714069225
A pivoted QR factorization is possible in `scipy.linalg.qr`, thanks
to the `geqp3` routine of LAPACK. To provide the same functionality
in JAX, we implement a new primitive `geqp3_p` which calls the LAPACK
routine via the FFI on CPU devices.
Both `jax.scipy.linalg.qr` and `jax.lax.linalg.qr` now support the
use of column-pivoting on CPU devices.
To provide a GPU implementation of `geqp3` may require using MAGMA,
due to the lack of a `geqp3` implementation in `cuSolver` - see
ccb331707e80b16d89de6e5c9f2f89b87c1682ed (`jax.lax.linalg.eig`) for
an example of using MAGMA in GPU lowerings. Such a GPU implementation
can be considered in the future.
Boolean fields in the descriptor struct led to padding, which let random
bytes in the string representation of the struct and variance in HLO
from run to run.
This change does not yet do the work necessary to make any tests pass with threading enabled, which will come in future changes.
This approach is broadly inspired by a6d205dd4c/testtools/testsuite.py (L113) and by unittest-ft.
We add a custom TestResult class that batches up any test result actions and applies them under a lock. We also add a custom TestSuite class that runs individual test cases in parallel using a thread-pool.
We need a reader-writer lock to implement a `@jtu.thread_hostile_test` decorator, which we do by adding bindings around absl::Mutex to jaxlib.
PiperOrigin-RevId: 713312937
This often lets us avoid ambiguities between selecting the (8, 128) and (16, 128) tiling,
by biasing the layout inference to prefer the latter.
PiperOrigin-RevId: 713270421
In our MSAN CI, the copy of LAPACK we use is not MSAN-instrumented, leading to false positives. Suppress those false-positives via annotations.
PiperOrigin-RevId: 712607044
Note that offsets outside of first tile are still disabled (for both infer and apply), and once we support it we will want to assign offsets differently, this is mostly to avoid assigning invalid layouts (that may not just be outside the first tile, but outside the vreg slice)
PiperOrigin-RevId: 709168368
We can enable them later but at least this way the support is available to build on
(e.g. in the new insert relayouts pass)
Reverts 05f3a701e769748ff1ec51d50324a3595c4aff0d
PiperOrigin-RevId: 708397219
We can use relayout-insertion pass to insert necessary ops and their layouts for relayout before unrolling in apply-vector-layout pass.
PiperOrigin-RevId: 708143852
For (1, 128) tiling 32-bit input, it assigns (1, 128) tiling at output, which can be invalid (e.g. it should be (1, 256) for bf16)
PiperOrigin-RevId: 708112341
This commit modifies the behavior of the build CLI when building jaxlib and GPU plugin artifacts together (for instance `python build --wheels=jaxlib,jax-cuda-plugin`.
Before, CUDA/ROCm build options were only passed when building the CUDA/ROCm artifacts. However, this leads to inefficient use of the build cache as it looks like Bazel tries to rebuild some targets that has already been built in the previous run. This seems to be because the GPU plugin artifacts have a different set of build options compared to `jaxlib` which for some reason causes Bazel to invalidate/ignore certain cache hits. Therefore, this commit makes it so that the build options remain the same when the `jaxlib` and GPU artifacts are being built together so that we can better utilize the build cache.
As an example, this means that if `python build --wheels=jaxlib,jax-cuda-plugin` is run, the following build options will apply to both `jaxlib` and `jax-cuda-plugin` builds:
```
/usr/local/bin/bazel run --repo_env=HERMETIC_PYTHON_VERSION=3.10 \
--verbose_failures=true --action_env=CLANG_COMPILER_PATH="/usr/lib/llvm-16/bin/clang" \
--repo_env=CC="/usr/lib/llvm-16/bin/clang" \
--repo_env=BAZEL_COMPILER="/usr/lib/llvm-16/bin/clang" \
--config=clang --config=mkl_open_source_only --config=avx_posix \
--config=cuda --action_env=CLANG_CUDA_COMPILER_PATH="/usr/lib/llvm-16/bin/clang" \
--config=build_cuda_with_nvcc
```
Note, this commit shouldn't affect the content of the wheel it self. It is only meant to give a performance boost when building `jalxib`+plugin aritfacts together.
Also, this removes code that was used to build (now deprecated) monolithic `jaxlib` build from `build_wheel.py`
PiperOrigin-RevId: 708035062
Adds an extra verification check. Since the source semaphore is used only for remote DMAs, we should check that device or core IDs are also provided when source semaphore is provided.
PiperOrigin-RevId: 707675228
