Skip to content

v0.0.310

Choose a tag to compare

@vchuravy vchuravy released this 15 Nov 12:36
· 13 commits to main since this release
ea4cdcd

What's Changed

  • Preserve barrier ID and thread count for NVVM CTA barrier adjoints by @wsmoses in #3746
  • [MLIR] Replace originalToNewFnOps with originalToNewFn by @wsmoses in #3722
  • Fix clang-format and LLVM 15/16 CI failures on main by @wsmoses in #3749
  • ActivityAnalysis: treat jl_eqtable_get from an inactive table as inactive by @wsmoses in #3752
  • MLIR: support enzyme.disable_loop_invariant_cache metadata attribute by @wsmoses in #3755
  • Use getCacheAlignment for loop-invariant cache stores and memcpys by @wsmoses in #3754
  • Fix reverse-mode freeze addingType and non-inbounds GEP index activity by @wsmoses in #3761
  • Add LLVM and MLIR coding standard skills and CI linters by @wsmoses in #3547
  • [MLIR] Add complex pow and powi rules by @Antipath1 in #3771
  • Fix clang-format in GradientUtils.cpp by @wsmoses in #3780
  • Link Clang libraries into ClangEnzyme on Windows with LLVM dylib by @wsmoses in #3727
  • [MLIR] Make tblgen broadcast emission order compiler independent by @wsmoses in #3786
  • [CI] Make the MLIR clang-tidy step run under POSIX sh by @wsmoses in #3790
  • Root bundled clang plugin headers at an absolute path on Windows by @wsmoses in #3787
  • Pass the LLVMContext when querying inline params on LLVM 24 by @ftynse in #3800
  • Do not assume a freed shadow allocation is an instruction by @vchuravy in #3801
  • Fix erfinv, tgamma and lgamma derivatives and add missing libm names by @wsmoses in #3791
  • Restore -Werror=unused-variable by @wsmoses in #3802
  • Emit inline matrix-scaling helper for CBLAS lascl when EnzymeLapackCopy is disabled by @wsmoses in #3760
  • Add CUDA/NVVM math derivatives and PreserveNVVM coverage for rsqrt, rcbrt, erfcx, erfcinv, normcdf, normcdfinv, fmin, copysign, and nvvm.fabs by @wsmoses in #3807
  • Fix ConstantExpr GEP recursion, loose memcpy fallback, and constant alloc indirect calls by @wsmoses in #3811
  • Cache whole allocations per iteration of their lifetime loop by @wsmoses in #3810
  • Support lvalue lambdas and Duplicated/Const-wrapped callables in C++ sugar by @wsmoses in #3627
  • Fix C++ sugar return tuples for non-Active arguments and void forward calls by @wsmoses in #3626
  • Type the CHARACTER arguments of the flang I/O runtime by @vchuravy in #3367
  • [MLIR] Register all passes in enzymeRegisterPasses by @agarret7 in #3879
  • Derivatives of the LLVM flang runtime MOD and MODULO for reals by @vchuravy in #3878
  • Use the original GEP's source type when shadowing a constant GEP by @vchuravy in #3877
  • Keep the cache allocation when cacheForReverse removes its frees by @vchuravy in #3374
  • Add Metal AIR math mappings for sinpi, cospi, asinh, acosh, atanh, rsqrt, f16, and fast variants by @wsmoses in #3866
  • Keep Julia allocation zeroing bundles on shadow allocations by @vchuravy in #3426

Full Changelog: v0.0.309...v0.0.310