v3.2.0

CUDAnative v3.2.0

Diff since v3.1.0

Closed issues:

• Uncached JIT breaks recursion (#122)
• PTX JIT compilation issue: Call to gpu_report_oom has wrong number of parameters (#653)

Merged pull requests:

• Update manifest (#647) (@github-actions[bot])
• CompatHelper: bump compat for "CEnum" to "0.4" (#650) (@github-actions[bot])
• Update manifest (#651) (@github-actions[bot])

v3.1.0

CUDAnative v3.1.0

Diff since v3.0.4

Closed issues:

• AbstractGPUCompiler submodule (#442)
• 'Symbol "__nv_exp"not found' on Jetson Xavier NX (#632)
• Is there a @cushow macro? (#633)
• Initializing static shared memory with function argument (#638)
• Atomic operations on shared memory (#641)

Merged pull requests:

• Move compiler functionality to GPUCompiler.jl (#623) (@maleadt)
• Update manifest (#626) (@github-actions[bot])
• Propagate inbounds when using Const/ldg (#627) (@mwarusz)
• Update manifest (#628) (@github-actions[bot])
• Adapt to GPUCompiler changes. (#631) (@maleadt)
• Simplify use of Base.julia_cmd (#634) (@maleadt)
• Move execution helpers to GPUCompiler. (#636) (@maleadt)
• add @cushow macro (#637) (@jw3126)
• Use Base.julia_cmd better. (#639) (@maleadt)
• CEnum 0.3 compatibility. (#640) (@maleadt)
• Avoid address space casts. (#642) (@maleadt)
• Fixes for 1.5. (#643) (@maleadt)
• Adapt to GPUCompiler changes. (#644) (@maleadt)
• Protect against recursive failures during init. (#645) (@maleadt)

v3.0.4

CUDAnative v3.0.4

Diff since v3.0.3

Merged pull requests:

• Make sure array intrinsics propagate inboundness. (#624) (@maleadt)
• Allow NVTX.at-range with short function definitions. (#625) (@maleadt)

v3.0.3

CUDAnative v3.0.3

Diff since v3.0.2

Merged pull requests:

• Switch to ExprTools. (#617) (@maleadt)
• Deduplicate irgen. (#618) (@maleadt)
• Update manifest (#619) (@github-actions[bot])
• Update manifest (#621) (@github-actions[bot])

v3.0.2

CUDAnative v3.0.2

Diff since v3.0.1

Closed issues:

• Provide more informative error message if CUDA discovery fails (#447)
• WMMA tests fail on julia-debug (#587)

Merged pull requests:

• Use new non-recursive codegen (#162) (@maleadt)
• Fix return type of WMMA intrinsics (#591) (@thomasfaingnaert)
• Better test initialization without CUDA. (#615) (@maleadt)
• Bump CI to 1.4, and include debug build (#616) (@maleadt)

v3.0.1

CUDAnative v3.0.1

Diff since v3.0.0

v3.0.0

CUDAnative v3.0.0

Diff since v2.10.2

Closed issues:

• CUPTI-based SASS reflection (#219)
• Tests: cooperative groups not always available (#410)
• Implement atomicCAS? (#427)
• 265-fix is only partial (#459)
• Invalid device pointers cause segfault (#470)
• Use Base.datatype_alignment (#531)
• WMMA uses generic addressing (#548)
• CUDAnative.fma performance problems (#553)
• Better error if exclusive device is already in use (#555)
• C++ compatible name mangling (#556)
• WMMA examples always execute (#561)
• Better warning on CUPTI insufficient permissions (#563)
• shfl does not like Bool (#567)
• On Julia master: "Warning: Incompatibility detected between CUDA and LLVM 8.0+; disabling debug info emission for CUDA kernels" (#569)
• Performance regression with new mapreduce (#584)
• Errors in CUDAnative unit tests (#585)
• Cthulhu integration (#597)
• Racy initializing (#601)

Merged pull requests:

• Use BinaryBuilder (#492) (@maleadt)
• Run the inliner at the end of the pipeline. (#554) (@maleadt)
• Use AddrSpacePtr to call WMMA intrinsics (#557) (@thomasfaingnaert)
• Always run with debug logging on CI. (#558) (@maleadt)
• Mangle names like C++ does. (#559) (@maleadt)
• Improve debug output. (#560) (@maleadt)
• Fix and simplify device capability checks. (#562) (@maleadt)
• Fix the reduce example. (#564) (@maleadt)
• NVTX improvements (#565) (@maleadt)
• Update manifest (#568) (@github-actions[bot])
• Use MacroTools for more robust NVTX range macro. (#570) (@maleadt)
• Whitelist hypot from method substitution warning. (#571) (@maleadt)
• Replace nearest_warp with prev/nextwarp. (#572) (@maleadt)
• Add support for printing Bool values. (#573) (@maleadt)
• CompatHelper: add new compat entry for "MacroTools" at version "0.5" (#574) (@github-actions[bot])
• Avoid calling reinterpret(Nothing, nothing). (#575) (@maleadt)
• Fix tests on nightly. (#576) (@maleadt)
• Check the CUDA version on CI. (#577) (@maleadt)
• Better fix for Julia nightly. (#578) (@maleadt)
• Rework pointer/array interfaces, simplify shuffle implementation (#579) (@maleadt)
• use --startup=no in spawned tests (#580) (@MasonProtter)
• Fix method invalidation (#581) (@maleadt)
• Eagerly prune empty debug info during backtrace construction. (#582) (@maleadt)
• Capture type errors. (#583) (@maleadt)
• Update manifest (#586) (@github-actions[bot])
• Run CI with a Julia debug build. (#588) (@maleadt)
• CI tweaks (#589) (@maleadt)
• Simplify naming of wrapper kernels. (#592) (@maleadt)
• Defer initializing deps to actual run time. (#594) (@maleadt)
• Remove workaround. (#595) (@maleadt)
• Print compilation job details when doing reflection. (#596) (@maleadt)
• Update manifest (#598) (@github-actions[bot])
• Add interactive reflection utils based on Cthulhu. (#599) (@maleadt)
• Clean-up init code. (#600) (@maleadt)
• Thread safety fixes. (#602) (@maleadt)
• Update and regenerate wrappers. (#603) (@maleadt)
• Better error upon use of missing libraries. (#604) (@maleadt)
• Add a simple multi-threaded test. (#605) (@maleadt)
• Better error on CUPTI insufficient permissions. (#606) (@maleadt)
• Update manifest (#608) (@github-actions[bot])
• Support for Julia's multitasking. (#609) (@maleadt)
• Various fixes (#610) (@maleadt)
• Support for setting device flags. (#612) (@maleadt)
• CompatHelper: add new compat entry for "Cthulhu" at version "1.0" (#613) (@github-actions[bot])
• Reinstate, and test support for varargs kernels. (#614) (@maleadt)

v2.10.2

CUDAnative v2.10.2

Diff since v2.10.1

Closed issues:

• CUDAnative failed to initialize, no CUPTI provided (#549)

Merged pull requests:

• Bugfix for missing libraries. (#552) (@maleadt)

v2.10.1

CUDAnative v2.10.1

Diff since v2.10.0

Merged pull requests:

• Install TagBot as a GitHub Action (#547) (@JuliaTagBot)
• Update manifest (#550) (@github-actions[bot])
• Make the dependency on NVTX/CUPTI optional. (#551) (@maleadt)

v2.10.0

v2.10.0 (2020-02-07)

Diff since v2.9.1

Closed issues:

• PTX JIT compilation failure with CUDAnative v2.9.0 (#538)
• nvdisasm needs version CUDA 9.0 use cuobjdump instead if newer (#396)

Merged pull requests:

• Add a release getter. (#544) (maleadt)
• Don't prefix unsafe_load with Base. (#543) (maleadt)
• Promote the kernel during irgen. (#542) (maleadt)
• Use Base functionality to query datatype alignment. (#541) (maleadt)
• Fix handling of ghost values in unsafe_load/store. (#540) (maleadt)
• Use the Julia LLVM IR printer to support the debuginfo argument. (#539) (maleadt)
• Add autogenerated wrappers for CUPTI and NVTX (#473) (maleadt)
• Make the kernel transform mandatory (#400) (vchuravy)