Minutes_2018_03_22
Stan Seibert edited this page Mar 22, 2018
·
1 revision
Attendees: Todd, Ehsan, Siu, Stuart, Stan
- #2841: Misconfigured CUDA driver prevents test discovery from working
- Fixed with merged PR
- #2837: np.transpose() support
- Being addressed in an open PR
- #2834: Disable FMA from @cuda.jit() decorator?
- Can expose some fine grained options to control this
- #2830: List inference in loop with variable re-binding
- Workaround available
- Reveals an issue with our backpropagation of type information when dealing with empty list in loop
- #2826: index method does not work on tuples
- Addressed in open PR
- #2825: Document setting up a dev stack with VS build tooling 2017
- Know what to write, need to write it
- #2823: Interop with C struct
- probably use cffi to represent struct type
- llvmlite#341: Increment build number.
- Fixed
- llvmlite#340: This moves llvmlite to use LLVM 6.0.0 as its backend.
- Reverify with SVML patch
- Basically ready to go
- llvmlite#337: Add support for optimization remarks in pass managers.
- Wait until LLVM 6 merged, then re-test
- llvmlite#330: Add support for LLVM fence instruction
- Confirm is still OK for LLVM 6
- #2843: [WIP] Custom pipeline
- Needs docs
- Ehsan will review
- #2842: Fix issue with test discovery and broken CUDA drivers.
- Merged
- #2840: [WIP] Support list of refcounted types
- Needs docs
- Nested listed comprehensions "work"?
- Some expected-fail tests now pass
- #2839: Support for np.transpose and np.reshape.
- Needs review
- #2836: Support `index` method on tuples
- Original PR used low-level interface
- Contributor has rewritten with higher level interface
- Needs final review
- #2835: Emit selp Instructions to Avoid Branching
- Ready to merge
- #2832: Fixups for llvmlite 0.23/llvm 6
- Ready to merge pending llvmlite update
- #2831: Attempt to fix #2473
- Fixes one of our unit tests when run as root
- Merged already
- #2824: Add CUDA Primitives for Population Count
- Needs final review
- #2817: [WIP] Emit LLVM optimization remarks
- Pending llvmlite PR and LLVM 6
- This approach is very proof of concept, may need to reimplement
- #2813: Fix #2812: gufunc scalar output bug.
- Needs minor update, then merge
- #2807: Expose `import_cython_function`
- Almost ready. Needs one final review pass
- #2793: Simplify and remove javascript from html_annotate templates.
- Needs some work to make more general (other Python versions, maybe Jupyter versions)
- #2780: PowerPC reference counting fences & minor fixes
- Pending revisit after LLVM 6
- #2779: Implement np.random.permutation
- Needs re-review
- #2777: Add support for np.correlate and np.convolve
- Needs review
- #2748: Added Intel SVML optimizations as opt-out choice working by default
- Pending LLVM6 switch
- Needs unit tests
- Needs to autodetect SVML
- Lowering numpy scalar functions different than math module. Need to fix.
- Docs explaining how to get SVML
===========================
- Review status of current critical bugs:
- viewcfg test segfaults on all platform
- test disabled for now
- viewcfg produces incorrectly formatted .dot file on OS X (memory corruption caused by compiler flags?)
- Old Theory: Anaconda clang had a problem?
- Current Discovery: System clang also has this problem when same compiler flags are used as are default with Anaconda clang.
- Flags: security flags + -O2
- Problem fixed if only remove -O2 and add debug flags
- Two options:
- LLVM source code has undefined that O2 "exploits"
- -O2 has a bug that has been around for a while
- Confirmed that problem is unrelated to conda-build
- Next steps: Verify system clang result, try gcc, determine which -O level triggers this, use sanitizer to find out if it is LLVM bug
- Race condition with registry initialization
- Race condition with parfor
- Global counter should be protected by compiler lock
- Is it somehow not covered through some code path
- Numba thread pool add_task / initialization needs locking
- Multiprocessing test suite failing with gil reacquire test
- NULs in snazzy strings in llvm-config
- Workaround by stripping trailing nulls from output. Issue caused by path relocation in conda.
- Drop CUDA 7.5?
- Yes
- Test against Numpy 1.10, 1.13, 1.14.
- Sure
- viewcfg test segfaults on all platform
- First gufunc improvements
- Better SIMD generation (SVML + parfor fixes)
- LLVM 6.0
- Better debug/troubleshooting tools
- Improve docs and information as per community feedback
- Keep working through backlog of bugs and minor feature requests
- Pipeline manipulation (sklam)