v5.6.0

CUDA v5.6.0

Diff since v5.5.2

CUDA.jl v5.6 is a relatively minor release, which the most important change being behind the scenes: GPUArrays.jl v11 has switched to KernelAbstractions.jl (#2524).

Features

• Update to CUDA 12.6.2 (#2512)
• CUSOLVER: support for Xgeev! (#2513), XsyevBatched (#2577), gesv! and gels! (#2406)
• CUBLAS: added multiplication of transpose / adjoint matrices by diagonal matrices (#2518, #2538)
• Improve handle cache performance in the presence of many short-lived tasks (#2583)
• CUFFT: Pre-allocate the buffer required for complex-to-real FFTs only once (#2578)
• Improved batched pointer conversion for very large batches (#2608)

Bug fixes

• Fix findall with an empty CuArray (#2554)
• CUBLAS: Fix use of level 1 methods with strided arrays (#2528)
• CUSOLVER: Fix Xgesvdr! (#2556)
• Preserve the array buffer type with more linear algebra operations (#2534)
  Work around LinearAlgebra.jl breakage in Julia 1.11.2 concerning generic triangular (l/r)mul! - (#2585)
• Fix ambiguity of LinearAlgebra.dot (#2569)
• Native RNG: Fixes when working with very large arrays (#2561)
• Avoid a deadlock due do union splitting in the mapreduce kernel (#2595)
• Fix pinning of resized CPU memory by automatically re-pinning (#2599)

Merged pull requests:

• [CUSOLVER] Interface gesv! and gels! (#2406) (@amontoison)
• Update wrappers for CUDA v12.6.2 (#2512) (@amontoison)
• [CUSOLVER] Interface Xgeev! (#2513) (@amontoison)
• Added multiplication of transpose / adjoint matrices by diagonal matrices (#2518) (@amontoison)
• CompatHelper: bump compat for GPUCompiler to 1, (keep existing compat) (#2521) (@github-actions[bot])
• Adapt to GPUArrays.jl transition to KernelAbstractions.jl. (#2524) (@maleadt)
• Switch CI to 1.11. (#2525) (@maleadt)
• CUTENSOR: Reduce amount of broadcasts compiled during tests. (#2527) (@maleadt)
• CUBLAS: Don't use BLAS1 wrappers for strided arrays, only vectors. (#2528) (@maleadt)
• Clarify the synchronize(ctx)/device_synchronize() docstrings (#2532) (@JamesWrigley)
• Issue #2533: Preserving the buffer type in linear algebra (#2534) (@kmp5VT)
• Clarify description of how LocalPreferences.toml is generated in the docs (#2535) (@glwagner)
• Adapt to JuliaGPU/GPUArrays.jl#567. (#2537) (@maleadt)
• Removed allocations for transpose/adjoint - diagonal multiplications (#2538) (@RedRussianBear)
• Consistent use of Nsight Compute (#2541) (@huiyuxie)
• Fix formatting in profiling docs page (#2543) (@efaulhaber)
• Fix typo in EnzymeCoreExt.jl (#2550) (@wsmoses)
• Enhance warning under a profiler (#2552) (@huiyuxie)
• Fix findall with an empty CuArray of Bool (#2554) (@amontoison)
• [CUSOLVER] Fix Xgesvdr! (#2556) (@amontoison)
• Test restore Enzyme.jl (#2557) (@wsmoses)
• Native RNG fixes for very large arrays (#2561) (@maleadt)
• [Enzyme] Mark launch_configuration as inactive (#2563) (@wsmoses)
• Update EnzymeCoreExt.jl (#2565) (@simenhu)
• Fix ambiguity of LinearAlgebra.dot (#2569) (@amontoison)
• [CUSOLVER] Add more tests for the dense SVD (#2574) (@amontoison)
• [CUSOLVER] Interface XsyevBatched (#2577) (@amontoison)
• [CUFFT] Preallocate a buffer for complex-to-real FFT (#2578) (@amontoison)
• Run the GC when failing to find a handle, but lots are active. (#2583) (@maleadt)
• Work around LinearAlgebra.jl breakage in 1.11.2. (#2585) (@maleadt)
• mapreduce: avoid deadlock by forcing the accumulator type. (#2596) (@maleadt)
• Switch to GitHub Actions-based benchmarks. (#2597) (@maleadt)
• Re-pin variable sized memory (#2599) (@jipolanco)
• Enzyme: add make_zero of cuarrays (#2600) (@wsmoses)
• Update cache.jl (#2604) (@jarbus)
• Enzyme: mark device_sync as non-differentiable [only downstream] (#2605) (@wsmoses)
• Move strided batch pointer conversion to GPU (#2608) (@THargreaves)
• Split linalg tests into multiple files (#2609) (@kshyatt)

Closed issues:

• Inference failure with sort(::CuMatrix) after loading MLDatasets (#2258)
• Kron Support for CuSparseMatrixCSC (#2370)
• Broadcasting a function returning an anonymous function with a constructor over CUDA arrays fails to compile, "not isbits" (#2514)
• CuArray view has different variable type outside x inside the cuda kernel (#2516)
• Can't build cuDNN on centos7.8 (#2517)
• Precompile errors (#2519)
• Precompile errors (#2520)
• Error returned from CUDA function in CUDA-aware MPI multi-GPU test (#2522)
• Broadcasting over random static array errors on Julia 1.11 (#2523)
• gemm_strided_batched only using strided CUDA kernel when first matrix is transposed (#2529)
• CUDA runtime libraries are loaded from a system path due to LD_LIBRARY_PATH being set (#2530)
• [Bug] UnifiedMemory buffer changes during LinearAlgebra operations (#2533)
• Improve system library warning when running under profiler (#2540)
• Local CUDA settings not propagated to Pkg.test (#2545)
• Out of Memory when working with Distributed for Small Matricies (#2548)
• findall is not working with an empty vector of bool (#2553)
• CUDA code does not return when running under VSC Debugging mode (#2558)
• dot is quite slow in multinest Arrays (#2559)
• UndefVarError: backend not defined in GPUArrays (#2564)
• view() returns CuArray instead of view for 1-D CuArrays (#2566)
• dot ambiguity (#2568)
• InvalidIRError thrown only if critical function is not previously compiled (#2573)
• circular dependency during precompilation (#2579)
• Sparse MatVec Is Nondeterministic? (#2582)
• CUDA triggers long Circular dependency list (#2586)
• Release v5.5.3 for GPUArray v11? (#2587)
• 'dot' gives different answers when viewing rather than slicing multidimensional arrays (#2589)
• Scalar indexing when performing kron on two CuVectors (#2591)
• Faster strided-batched to batched wrapper (#2592)
• Error when copying data to pinned and resized CPU array (#2594)
• mapreducedim! size-dependent fail when narrowing float element types (#2595)
• Missing Enzyme.make_zero in Enzyme extension leads to incorrect behaviour (#2598)
• 'ArgumentError: array must be non-empty' when attempting to pop idle handles from HandleCache (#2603)
• Do a release as current one doesn't support GPUArrays v11 (#2606)

v5.5.2

CUDA v5.5.2

Diff since v5.5.1

Merged pull requests:

• Fix type of AbstractFFTs.Plan for real-complex FFTs (#2504) (@jipolanco)
• Profiler: Demangle kernel names. (#2505) (@maleadt)
• Bump CUDNN. (#2507) (@maleadt)
• Restore Enzyme checks (#2508) (@wsmoses)

v5.5.1

What's Changed

• Update wrappers for CUDA v12.6.1 by @amontoison in #2499
• Enzyme: adapt to pending version breaking update by @wsmoses in #2490

Full Changelog: v5.5.0...v5.5.1

v5.5.0

CUDA v5.5.0

Blog post

Diff since v5.4.3

Merged pull requests:

• Add support for arbitrary group sizes in gemm_grouped_batched! (#2334) (@lpawela)
• Add kernel compilation requirements to docs (#2416) (@termi-official)
• Enzyme: reverse mode kernels (#2422) (@wsmoses)
• CUFFT: Support Float16 (#2430) (@eschnett)
• Updated compute-sanitizer documentation (#2440) (@alexp616)
• Add troubleshooting section for NSight Compute (#2442) (@efaulhaber)
• Correct typo in documentation (#2445) (@eschnett)
• Bump minimal Julia requirement to v1.10. (#2447) (@maleadt)
• fix compute-sanitizer typo (#2448) (@alexp616)
• Address a corner case when establishing p2p access (#2457) (@findmyway)
• Implementation of spdiagm for CUSPARSE (#2458) (@walexaindre)
• Update to CUDA 12.6. (#2461) (@maleadt)
• CompatHelper: bump compat for GPUCompiler to 0.27, (keep existing compat) (#2462) (@github-actions[bot])
• Bump CUDA driver JLL. (#2463) (@maleadt)
• CUSOLVER (dense): cache workspace in fat handle (#2465) (@bjarthur)
• Revert "Run full GC when under very high memory pressure." (#2469) (@maleadt)
• Fix a method deprecation. (#2470) (@maleadt)
• Add Enzyme sum derivatives (#2471) (@wsmoses)
• Re-use pre-converted kernel arguments when launching kernels. (#2472) (@maleadt)
• Bump LLVM compat (#2473) (@maleadt)
• Bump subpackage compat. (#2475) (@maleadt)
• Enzyme: Reversemode cudaconvert (#2476) (@wsmoses)
• Ignore Enzyme.jl CI failures (#2479) (@maleadt)
• Re-enable enzyme testing (#2480) (@wsmoses)
• Add missing GC.@preserves. (#2487) (@maleadt)
• [CUSPARSE] Implement a sparse GEMV for CuSparseMatrixCSC * CuSparseVector (#2488) (@amontoison)
• [CUSPARSE] Add conversions between CuSparseVector and CuSparseMatrices (#2489) (@amontoison)
• Update to LLVM 9.1. (#2491) (@maleadt)
• Use at-consistent_overlay for 1.11 compatibility. (#2492) (@maleadt)
• Rework NNlib CI. (#2493) (@maleadt)
• CUSPARSE: Fix sparse constructor with duplicate elements. (#2495) (@maleadt)

Closed issues:

• LinearAlgebra.norm(x) falls back to generic implementation for x::Transpose and x::Adjoint (#1782)
• dlclose'ing the compatibility driver can fail (#1848)
• Creating a sparse diagonal matrix of CuArray(u) (#1857)
• Support for Julia 1.11 (#2241)
• CUDA 12.4 Update 1: CUPTI does not trace kernels anymore (#2328)
• Adding CUDA to a PackageCompiler sysimage causes segfault (#2428)
• Error using CUDA on Julia 1.10: Number of threads per block exceeds kernel limit (#2438)
• Error when I load my model (#2439)
• Driver JLL improvements (#2446)
• Deadlock when callling CUDA.jl in an adopted thread while blocking the main thread (#2449)
• CUDA.Mem.unregister fails with CUDA.jl 5.4 (not with 5.3) (#2452)
• Segmentation Fault on Loading CUDA (#2453)
• Invalid instruction error when using CUDA (#2454)
• Missing adapt for sparse and CUDABackend (#2459)
• CUDA precompile cannot find/load "cupti64_2024.2.1.dll" during precompilation (juliaup 1.10.4, Windows 11) (#2466)
• Request: Option to disable the "full GC when under very high memory pressure". (#2467)
• copyto! ambiguous (#2477)
• NeuralODE training failed on GPU with Enzyme (#2478)
• issue with atomic - when running standard test, @atomic modify expression missing field access (#2483)
• Support for creating a CuSparseMatrixCSC from a CuSparseVector (#2484)
• Issue with compiling CUDA and cuTENSOR using local libraries (#2486)
• Memory Access error in sparse array constructor (#2494)
• Forwards-compatible driver breaks CURAND (#2496)
• CUDA 12.6 Update 1 (#2497)

v5.4.3

CUDA v5.4.3

Diff since v5.4.2

Merged pull requests:

• add cublasgetrsBatched (#2385) (@bjarthur)
• add two quirks for rationals (#2403) (@lanceXwq)
• Bump cuDNN (#2404) (@maleadt)
• Add convert method for ScaledPlan (#2409) (@david-macmahon)
• Conditionalize a quirk. (#2411) (@maleadt)
• Relax signature of generic matvecmul! (#2414) (@dkarrasch)
• Fix kron launch configuration. (#2418) (@maleadt)
• Run full GC when under very high memory pressure. (#2421) (@maleadt)
• Enzyme: Fix cuarray return type (#2425) (@wsmoses)
• CompatHelper: bump compat for LLVM to 8, (keep existing compat) (#2426) (@github-actions[bot])
• pre-allocated pivot and info buffers for getrf_batched (#2431) (@bjarthur)
• Profiler tweaks. (#2432) (@maleadt)
• Update the Julia wrappers for CUDA v12.5.1 (#2436) (@amontoison)
• Correct workspace handling (#2437) (@maleadt)

Closed issues:

• Legacy cuIpc* APIs incompatible with stream-ordered allocator (#1053)
• Broadcasted multiplication with a rational doesn't work (#1926)
• Incorrect grid size in kron (#2410)
• GEMM of non-contiguous inputs should dispatch to fallback implementation (#2412)
• Failure of Eigenvalue Decomposition for Large Matrices. (#2413)
• CUDA_Driver_jll's lazy artifacts cause a precompilation-time warning (#2415)
• Recurrence of integer overflow bug (#1880) for a large matrix (#2427)
• CUDA kernel crash very occasionally when MPI.jl is just loaded. (#2429)
• CUDA_Runtime_Discovery Did not find cupti on Arm system with nvhpc (#2433)
• CUDA.jl won't install/run on Jetson Orin NX (#2435)

v5.4.2

CUDA v5.4.2

Diff since v5.4.1

Merged pull requests:

• Fix and test the legacy memory pool. (#2402) (@maleadt)

v5.4.1

CUDA v5.4.1

Diff since v5.4.0

Merged pull requests:

• Fixup Enzyme: Mark CuArray as noalias (#2401) (@wsmoses)

v5.4.0

CUDA v5.4.0

Blog post

Diff since v5.3.5

Merged pull requests:

• Support CUDA 12.5 (#2392) (@maleadt)
• Mark cuarray as noalias (#2395) (@wsmoses)
• Update Julia wrappers for CUDA v12.5 (#2396) (@amontoison)
• Enable correct pool access for cublasXt. (#2398) (@maleadt)
• More fine-grained CUPTI version checks. (#2399) (@maleadt)

Closed issues:

• CUTENSOR breaks after device_reset! (#2319)
• cuBLASXt's xt_gemm! incompatible with stream-ordered allocated memory (#2320)
• Add helper function to recompile CUDA stack (#2364)

v5.3.5

CUDA v5.3.5

Diff since v5.3.4

Merged pull requests:

• Avoid constructing MulAddMuls on Julia v1.12+ (#2277) (@dkarrasch)
• CompatHelper: bump compat for LLVM to 7, (keep existing compat) (#2365) (@github-actions[bot])
• Enzyme: allocation functions (#2386) (@wsmoses)
• Tweaks to prevent context construction on some operations (#2387) (@maleadt)
• Fixes for Julia 1.12 / LLVM 17 (#2390) (@maleadt)
• CUBLAS: Make sure CUBLASLt wrappers use the correct library. (#2391) (@maleadt)
• Backport: Enzyme allocation fns (#2393) (@wsmoses)

Closed issues:

• Indexing a view uses scalar indexing (#1472)
• EnzymeCore is an unconditional dependency. (#2380)
• cuBLASLt wrappers ccall into cuBLAS (#2388)
• generic_trimatmul! error (#2389)

v5.3.4

CUDA v5.3.4

Diff since v5.3.3

Merged pull requests:

• Add Enzyme Forward mode custom rule (#1869) (@wsmoses)
• Handle cache improvements (#2352) (@maleadt)
• Fix cuTensorNet compat (#2354) (@maleadt)
• Optimize array allocation. (#2355) (@maleadt)
• Change type restrictions in cuTENSOR operations (#2356) (@lkdvos)
• Bump julia-actions/setup-julia from 1 to 2 (#2357) (@dependabot[bot])
• Suggest use of 32 bit types over 64 instead of just Float32 over Float64 [skip ci] (#2358) (@Zentrik)
• Make generic_trimatmul more specific (#2359) (@tgymnich)
• Return the currect memory type when wrapping system memory. (#2363) (@maleadt)
• Mark cublas version/handle as non-differentiable (#2368) (@wsmoses)
• Enzyme: Forward mode sync (#2369) (@wsmoses)
• Enzyme: support fill (#2371) (@wsmoses)
• unsafe_wrap: unconditionally use the memory type provided by the user. (#2372) (@maleadt)
• Remove external_gvars. (#2373) (@maleadt)
• Tegra support with artifacts (#2374) (@maleadt)
• Backport Enzyme extension (#2375) (@wsmoses)
• Add note about --check-bounds=yes (#2378) (@Zinoex)
• Test Enzyme in a separate CI job. (#2379) (@maleadt)
• Fix tests for Tegra. (#2381) (@maleadt)
• Update Project.toml [remove EnzymeCore unconditional dep] (#2382) (@wsmoses)

Closed issues:

• Native Softmax (#175)
• CUSOLVER: support eigendecomposition (#173)
• backslash with gpu matrices crashes julia (#161)
• at-benchmark captures GPU arrays (#156)
• Support kernels returning Union{} (#62)
• mul! falls back to generic implementation (#148)
• \ on qr factorization objects gives a method error (#138)
• Compiler failure if dependent module only contains a japi1 function (#49)
• copy!(dst, src) and copyto!(dst, src) are significantly slower and allocate more memory than copyto!(dest, do, src, so[, N]) (#126)
• Calling Flux.gpu on a view dumps core (#125)
• Creating CuArray{Tracker.TrackedReal{Float64},1} a few times causes segfaults (#121)
• Guard against exceeding maximum kernel parameter size (#32)
• Detect common API misuse in error handlers (#31)
• rand and friends default to Float64 (#108)
• \ does not work for least squares (#104)
• ERROR_ILLEGAL_ADDRESS when broadcasting modular arithmetic (#94)
• CuIterator assumes batches to consist of multiple arrays (#86)
• Algebra with UniformScaling Uses Generic Fallback Scalar Indexing (#85)
• Document (un)supported language features for kernel programming (#13)
• Missing dispatch for indexing of reshaped arrays (#556)
• Track array ownership to avoid illegal memory accesses (#763)
• NVPTX i128 support broken on LLVM 11 / Julia 1.6 (#793)
• Support for sm_80 cp.async: asynchronous on-device copies (#850)
• Profiling Julia with Nsight Systems on Windows results in blank window (#862)
• sort! and partialsort! are considerably slower than CPU versions (#937)
• mul! does not dispatch on Adjoint (#1363)
• Cross-device copy of wrapped arrays fails (#1377)
• Memory allocation becomes very slow when reserved bytes is large (#1540)
• Cannot reclaim GPU Memory; CUDA.reclaim() (#1562)
• Add eigen for general purpose computation of eigenvectors/eigenvalues (#1572)
• device_reset! does not seem to work anymore (#1579)
• device-side rand() are not random between successive kernel launches (#1633)
• Add EnzymeRules support for CUDA.jl (for forward mode here) (#1811)
• cusparseSetStream_v2 not defined (#1820)
• Feature request: Integrating the latest CUDA library "cuLitho" into CUDA.jl (#1821)
• KernelAbstractions.jl-related issues (#1838)
• lock failing in multithreaded plan_fft() (#1921)
• CUSolver finalizer tries to take ReentrantLock (#1923)
• Testsuite could be more careful about parallel testing (#2192)
• Opportunistic GC collection (#2303)
• Unable to use local CUDA runtime toolkit (#2367)
• Enzyme prevents testing on 1.11 (#2376)