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 CL adds very simple support for int4 to the MHLO dialect.
It doesn't worry too much about integrations (with frameworks or compilers)
or semantics (e.g. verifiers is a known improvement area for the dialect)
and simply aims to expand the representation power of MHLO to unblock experiments.
Figuring out integrations and semantics is also on the roadmap, but there is
consensus that we should unblock forward progress right away and figure out
the rest in parallel.
PiperOrigin-RevId: 445198171
CHANGELOG
=========
b02c384ef - Add fused multiply functions for PowerPC - pmsub, pnmadd and pnmsub
3de96caea - Fix HouseholderSequence.h
f845a8bb1 - Fix cwise NaN propagation for scalar input.
a4bb513b9 - Update HouseholderSequence.h
fc1d88841 - Remove AVX512VL dependency in trsm
07db964bd - Restrict new AVX512 trsm to AVX512VL, rename files for consistency.
67eeba6e7 - Avoidable heap allocation in applyHouseholderToTheLeft
3342fc7e4 - Allow all tests to pass with `EIGEN_TEST_NO_EXPLICIT_VECTORIZATION`
efb08e0bb - Revert "Fix ambiguous DiagonalMatrix constructors."
53eec53d2 - Fix Power GEMV order of operations in predux for MMA.
a81bba962 - Fix ambiguous DiagonalMatrix constructors.
f7b31f864 - Revert "Replace call to FixedDimensions() with a singleton instance of"
f3ba220c5 - Remove EIGEN_EMPTY_STRUCT_CTOR
5ed7a86ae - Fix MSVC+CUDA issues.
734ed1efa - Fix ODR issues in lapacke_helpers.
2c45a3846 - Fix some max size expressions.
edc822666 - Fix navbar scroll with toc.
df87d40e3 - constexpr reshape helper
403fa3340 - Performance improvements in GEMM for Power
e1df3636b - More constexpr helpers
64909b82b - static const class members turned into constexpr
2c0ef43b4 - Added Scaling function overload for vector rvalue reference
ba2cb835a - Add back std::remove* aliases - third-party libraries rely on these.
0c859cf35 - Consider inf/nan in scalar test_isApprox.
1ddd3e29c - fixed order of arguments in blas syrk
2c5644280 - Don't include .cpp in lapack.
73b2c13bf - Disable f16c scalar conversions for MSVC.
9bc9992dd - Eliminate trace unused warning.
e22d58e81 - Add is_constant_evaluated, update alignment checks
f0a91838a - Enable Aarch64 CI
b9d2900e8 - added a missing typename and fixed a unused typedef warning
0611f7fff - Add missing explicit reinterprets
cd3c81c3b - Add a NNLS solver to unsupported - issue #655
0699fa06f - Split general_matrix_vector_product interface for Power into two macros - one ColMajor and RowMajor.
19a6a827c - Optimize visitor traversal in case of RowMajor.
f2a3e03e9 - Fix usages of wrong namespace
4451823fb - Fix ODR violation in trsm.
9a14d91a9 - Fix AVX512 builds with MSVC.
7b10795e3 - Change EIGEN_ALTIVEC_ENABLE_MMA_DYNAMIC_DISPATCH and EIGEN_ALTIVEC_DISABLE_MMA flags to be like TensorFlow's...
3ca1228d4 - Work around MSVC compiler bug dropping `const`.
40eb34bc5 - Fix RowMajorBit <-> RowMajor mixup.
c06298346 - Completed a missing parenthesis in tutorial.
9deaa1912 - Work around g++-10 docker issue for geo_orthomethods_4.
e34db1239 - Fix missing pound
591906477 - Fix up PowerPC MMA flags so it builds by default.
518fc321c - AVX512 Optimizations for Triangular Solve
01b5bc48c - Disable schur non-convergence test.
421cbf086 - Replace Eigen type metaprogramming with corresponding std types and make use of alias templates
514f90c9f - Remove workarounds for bad GCC-4 warnings
9ad566148 - Revert "Fix up PowerPC MMA flags so it builds by default."
65eeedf96 - Fix up PowerPC MMA flags so it builds by default.
cb1e8228e - Convert bit calculation to constexpr, avoid casts.
baf9a985e - Fix swap test for size 1 inputs.
788240885 - Temporarily disable aarch64 CI.
2a6be5492 - Fix construct_at compilation breakage on ROCm.
a3b64625e - Remove ComputeCpp-specific code from SYCL Vptr
9296bb4b9 - Fix edge-case in zeta for large inputs.
cd2ba9d03 - Add construct_at, destroy_at wrappers. Use throughout.
dfa517678 - make SparseSolverBase and IterativeSolverBase move constructable
9883108f3 - Remove copy_bool workaround for gcc 4.3
3a9d404d7 - Add support for Apple's Accelerate sparse matrix solvers
PiperOrigin-RevId: 445164768
When iterating through the tensor to extract shape values, an underlying missing kernel
(`StridedSlice` for quantized types) causes an error, which then results in a `nullptr`
being passed to `ParseDimensionValue()`, causing a segfault.
The `nullptr` check allows the missing kernel error to propagate.
Adding the missing kernel registrations allows the shape values
to be extracted successfully.
PiperOrigin-RevId: 445045957
This is similar to a fix to a similar problem that we encountered in Dec 2021
when we introduced a static limit for folding scatters. There's a ticket to
follow up on this with a general solution.
PiperOrigin-RevId: 445030081
