oneAPI.jl Files

oneAPI v2.10.0

Diff since v2.9.2

Merged pull requests:

• Avoid boxing compiler jobs during kernel resolution (#635) (@maleadt)
• Set Intel oneAPI local build to 2026.1.0 (#636) (@michel2323)
• [CI] Julia 1.13 (#637) (@christiangnrd)
• Remove a quirk that never applies (#638) (@maleadt)
• Pass the exception type to @gputhrow (#639) (@maleadt)
• Use GPUToolbox's Float64-free math overrides on devices without fp64 (#640) (@maleadt)
• Stop excluding device code from coverage (#641) (@maleadt)
• Make oneMKL sparse matrices AbstractGPUSparseArrays (#642) (@michel2323)
• Use the oneMKL trmm/trsm variants for triangular multiply and solve (#643) (@michel2323)
• Add a compile-only test grid for Base math on the device (#644) (@michel2323)
• [CI] Restrict LTS coverage tracking to the checkout (#645) (@michel2323)
• Use GPUCompiler's workaround for nested insertvalues (#646) (@maleadt)
• Julia 1.14 compatibility; CI step selection (#647) (@maleadt)
• ALCF CI: trigger a fresh GitLab pipeline on re-runs after an unsuccessful pipeline (#648) (@michel2323)
• Synchronize before indexing host-accessible arrays from the CPU (#649) (@maleadt)
• Document unsafe_wrap(Array, ...) lifetimes, and support host buffers (#650) (@maleadt)
• Implement KernelInterface, and support KernelAbstractions 0.10 (#653) (@maleadt)
• Make synchronization cooperative, using GPUToolbox (#654) (@maleadt)
• Upgrade to LLVM.jl 10 (#656) (@maleadt)
• Report kernel lookups to the compile hook instead of recompiling (#659) (@maleadt)
• Declare the method tables instead of building the view (#660) (@maleadt)
• Declare indirect access on compiled kernels (#662) (@michel2323)
• ALCF CI: report GitLab pipeline results as a commit status (#664) (@michel2323)
• Use GPUArrays 12's reductions, scans, sorting and findall (#665) (@maleadt)
• Reduce CI pressure on pull requests (#666) (@maleadt)
• Fill GPUCompiler's SPIR-V atomics descriptor from the device (#668) (@maleadt)
• Re-enable the Validation CI step (#673) (@maleadt)
• Test Int16 reductions whose map overflows (#674) (@maleadt)
• Test branching reductions with loop-carried booleans (#679) (@maleadt)
• Free resized buffers with release, not free (#684) (@maleadt)
• Preserve @oneapi arguments behind a function barrier (#685) (@maleadt)
• Track outstanding GPU work per allocation (#686) (@maleadt)
• Specialize kernel launches on their argument types (#687) (@maleadt)
• Skip KernelAbstractions' private-memory tests on the LTS stack (#689) (@maleadt)
• Sub-allocate small device arrays (#692) (@maleadt)
• Submit kernel launches in batches (#693) (@maleadt)
• Test loops that exit when their counter wraps around (#696) (@maleadt)
• Allow CEnum 1.0 (#697) (@maleadt)
• Hash compiler configuration kwargs without boxing them (#699) (@maleadt)
• Bump NEO to 26.35, fix BFloat16 detection (#701) (@maleadt)
• Clear the output of sparse products when beta is zero (#702) (@maleadt)
• Make all Level Zero calls GC-safe (#704) (@maleadt)
• Cache per-device element-type support (#705) (@maleadt)
• Test multithreaded GC with NEO's SIGSEGV handler installed (#706) (@maleadt)
• Fix comparisons with Rationals (#707) (@maleadt)
• Test appending a batch to an idle stream after host syncs (#709) (@maleadt)
• Skip ComplexF32 accumulate tests on the LTS stack (#710) (@maleadt)
• Pass device pointers to kernels as global LLVMPtrs (#711) (@maleadt)
• Recycle command lists submitted to explicit queues (#713) (@maleadt)
• Test that kernels don't contain invalid integer widths (#715) (@maleadt)
• Destroy a context only once what was created in it is gone (#716) (@maleadt)
• Don't wait for the GPU when freeing memory (#720) (@maleadt)
• Fill unspecified type parameters when adapting to oneArray (#721) (@maleadt)
• Raise a clear error when passing CPU arrays to kernels (#723) (@maleadt)
• Fix BLAS calls on views oneMKL can't address (#724) (@maleadt)
• Remove documentation previews without rewriting gh-pages, and link them (#725) (@maleadt)
• ALCF CI: report results through a GitLab webhook, re-enable the bridge (#726) (@michel2323)
• Respect Symmetric/Hermitian wrappers in sparse products (#727) (@maleadt)
• Stage small host-to-device uploads instead of synchronizing (#728) (@maleadt)
• Add oneAPI.@profile external=true to restrict external profiling (#729) (@maleadt)
• Add an integrated profiler to oneAPI.@profile (#730) (@maleadt)
• Report device exceptions to the host (#731) (@maleadt)
• Test sub-group shuffles of a mixed-width tuple (#732) (@maleadt)
• Check the alignment of arrays passed to kernels (#733) (@maleadt)
• Test KernelAbstractions launches by their behavior, not the launch kind (#734) (@maleadt)
• Support arrays of isbits unions (#735) (@maleadt)
• Don't synchronize command lists and queues that were destroyed (#739) (@maleadt)
• Require SPIRVIntrinsics 1.5.1 for local union arrays (#740) (@maleadt)
• Don't check collected streams for kernel exceptions (#741) (@maleadt)
• Render code blocks with DocumenterCodeBlocks (#742) (@maleadt)
• Cache the kernel lookup of zefunction with GPUCompiler's cached_lookup (#743) (@maleadt)

Closed issues:

• Report device exceptions to the host (#65)
• Extremely slow reduction (#110)
• Interactive profiling (#111)
• Invalid bit width in input (#172)
• Memory leak when running GPU broadcast in a loop (#327)
• KernelAbstractions: add Atomix back-end (#336)
• Multiplication of StridedMaybeAdjOrTransMat broken for certain matrix sizes (#442)
• SYCL batching causes invalid results (#445)
• Copying HostBuffer is broken (#458)
• Support for logical indexing (#461)
• Host to device copies should be blocking (#469)
• Indexing oneArrays with a oneArray{Bool} fails for Julia 1.11 (#473)
• Correctness issue when indexing "big" oneArrays (#509)
• Error in oneAPI.jl tests (#541)
• Atomic operations on local memory are broken (#572)
• Internal compiler error when trying to run atomic operations (#575)
• Segfault in multithreaded code: NEO SIGSEGV handler race (intel/compute-runtime#1009) (#607)
• mkl sparse arrays aren't AbstractGPUSparseArray (#627)
• Track outstanding GPU work per allocation for host-accessible arrays (#652)
• Flaky: reductions of host-accessible arrays return 0 on Julia 1.13 (ALCF CI) (#661)
• resize! frees its new buffer with free instead of release (#667)
• Host JIT segfaults compiling @oneapi in a loop over a tuple of functions (#669)
• Wrong results for overflowing Int16 arithmetic in a reduction kernel (#670)
• Kernel launches are compiled for Vararg{Any} and dispatched at run time (#675)
• Wrong result from a reduction with a branching operator on Julia 1.12 with --check-bounds=yes (#678)
• Array construction queries device properties on every allocation (#690)
• IGC heap corruption compiling findmax's reduction kernel on Julia 1.12 with --check-bounds=yes (#691)
• Make blocking Level Zero calls GC-safe (#698)
• Storage-level mul! gives wrong results or aborts for SubArray operands (#700)
• BLAS-1 methods pass reversed views to oneMKL with a negative increment (#712)
• Collecting a context together with its memory segfaults (#714)
• Products with a Symmetric/Hermitian oneMKL sparse matrix ignore the wrapper (#718)
• Make host-to-device uploads non-blocking via a host-USM staging pool (#719)
• Segfault in zeCommandListHostSynchronize when reusing memory last used by a finished task (#736)
• Require SPIRVIntrinsics 1.5.1 for local arrays of isbits unions (#737)