v5.0.0

CUDA v5.0.0

Blog post: https://info.juliahub.com/cuda-jl-5-0-changes

This is a breaking release, but the breaking changes are minimal (see the blog post for details):

• Julia 1.8 is now required, and only CUDA 11.4+ is supported
• selection of local toolkits has changed slightly

---

Diff since v4.4.1

Merged pull requests:

• Added support for more transform directions (#1903) (@RainerHeintzmann)
• Add some performance tips to the documentation (#1999) (@Zentrik)
• Re-introduce the 'blocking' kwargs to at-sync. (#2060) (@maleadt)
• Adapt to GPUCompiler#master. (#2062) (@maleadt)
• Batched SVD added (gesvdjBatched and gesvdaStridedBatched) (#2063) (@nikopj)
• Use released GPUCompiler. (#2064) (@maleadt)
• Fixes for Windows. (#2065) (@maleadt)
• Switch to GPUArrays buffer management. (#2068) (@maleadt)
• Update CUDA 12 to Update 2. (#2071) (@maleadt)
• Update manifest (#2076) (@github-actions[bot])
• Test improvements (#2079) (@maleadt)
• Update manifest (#2082) (@github-actions[bot])

Closed issues:

• StaticArrays.SHermitianCompact not working in kernels in Julia 1.10.0-beta2 (#2069)
• Support for LinearAlgebra.pinv (#2070)

v4.4.1

CUDA v4.4.1

Diff since v4.4.0

Closed issues:

• CUDA driver device support does not match toolkit (#70)
• Launching kernels should not allocate (#66)
• sync_threads() appears to not be sync'ing threads (#61)
• Exception when using CuArrays with Flux (#129)
• Kernel using MVector fails to compile or crashes at runtime due to heap allocation (#45)
• Performance regression on matrix multiplication between CUDA.jl 1.3.3 and 2.1.0/master (#538)
• Improve 'VS C++ redistributable' error message (#764)
• CUSPARSE does not support reductions (#1406)
• CUDA test failed (#1690)
• Type constructor in broadcast doesn't compile (#1761)
• accumulate(+) gives different results for CuArray compared to Array. (#1810)
• Compat driver: preload all libraries (#1859)
• Stream synchronization is slow when waiting on the event from CUDA (#1910)
• cuDNN: Store convolution algorithm choice to disk. (#1947)
• Disable 'No CUDA-capable device found' error log (#1955)
• CUDNN_STATUS_NOT_SUPPORTED using 1D CNN model (#1977)
• Memory allocations during in-place sparse matrix-vector multiplication (#1982)
• CUSPARSE.sum_dim1 sums the absolute values of elements (#1983)
• Update to CUDA 12.2 (#1984)
• unsafe_wrap fails on zero element CuArrays (#1985)
• rand in kernel works in a deterministic way (#2008)
• Scalar indexing with CuArray * ReshapedArray{SubArray{CuArray}}} (#2009)
• volumerhs performance regression (#2010)
• CuSparseMatrix constructors allocate too much memory? (#2015)
• Native profiler using CUPTI (#2017)
• libLLVM-15jl.so (#2018)
• "symbol multiply defined" error (#2021)
• Confusion on row major vs column major (#2023)
• Printing of CuArrays gives zeros or random numbers (#2033)
• sortperm! fails when output is UInt vector (#2046)
• Re-introduce spinning loop before nonblocking synchronization (#2057)

Merged pull requests:

• Check mathType only if not Float32 (#1943) (@RomeoV)
• 1.10 enablement (#1946) (@dkarrasch)
• Implement reverse lookup (Ptr->Tuple) for CUDNN descriptors. (#1948) (@RomeoV)
• Wrapper with tests for gemmBatchedEx! (#1975) (@lpawela)
• Add wrappers for gemv_batched! (#1981) (@lpawela)
• Update CUSPARSE.sum_dim<n> to allow for arbitrary function on elements (#1987) (@lpawela)
• Update manifest (#1988) (@github-actions[bot])
• Add vectorized cached loads (#1993) (@Zentrik)
• Update manifest (#1995) (@github-actions[bot])
• Fix typo in captured macro example (#1996) (@Zentrik)
• Adapt Type call broadcasting to a function (#2000) (@simonbyrne)
• [CUSPARSE] Added support for generalized dot product dot(x, A, y) = dot(x, A * y) without allocating A * y (#2001) (@albertomercurio)
• Update manifest (#2002) (@github-actions[bot])
• Support for printing types. (#2003) (@maleadt)
• Fix accumulate bug (#2005) (@chrstphrbrns)
• Update manifest (#2013) (@github-actions[bot])
• Add a raw mode to code_sass. (#2019) (@maleadt)
• Update manifest (#2022) (@github-actions[bot])
• Add a native profiler. (#2024) (@maleadt)
• Perform synchronization on a worker thread (#2025) (@maleadt)
• Remove broken video link in docs (#2028) (@christiangnrd)
• When freeing memory, use the high-level device getter. (#2029) (@maleadt)
• Add support for @cuda fastmath (#2030) (@maleadt)
• Make "CUDA.jl" a link on the doc entry page (#2031) (@carstenbauer)
• Add support for CUDA 12.2. (#2034) (@maleadt)
• rand: seed kernels from the host. (#2035) (@maleadt)
• Update wrappers for CUDA 12.2. (#2039) (@maleadt)
• On CUDA 12.2, have the memory pool enforce hard memory limits. (#2040) (@maleadt)
• Delay all initialization errors until run time. (#2041) (@maleadt)
• JLL/CI/Julia changes. (#2042) (@maleadt)
• Add support for NVTX events to the integrated profiler. (#2043) (@maleadt)
• Update cuStateVec to cuQuantum 23.6. (#2044) (@maleadt)
• Add some more fastmath functions (#2047) (@Zentrik)
• Fixup wrong key lookup. (#2048) (@RomeoV)
• Update manifest (#2049) (@github-actions[bot])
• Make sortperm! resilient to type mismatches. (#2051) (@maleadt)
• Disable tests that cause GC corruption on 1.10. (#2053) (@maleadt)
• enable dependabot for GitHub actions (#2054) (@ranocha)
• Bump actions/checkout from 2 to 3 (#2055) (@dependabot[bot])
• Bump peter-evans/create-pull-request from 3 to 5 (#2056) (@dependabot[bot])
• Rework how local toolkits are selected. (#2058) (@maleadt)
• Busy-wait before doing nonblocking synchronization. (#2059) (@maleadt)

v4.4.0

CUDA v4.4.0

Diff since v4.3.2

Closed issues:

• Unreachable control flow leads to illegal divergent barriers (#1746)
• CUBLAS fails on new CUDA.jl v4 (#1852)
• Sort fails on Lovelace (sm8.9) GPUs (#1874)
• gesvd! crashes on Pascal and v12.0 (#1932)
• No effect for calling "nsys launch" (#1938)
• Basic math operations with nested adjoint and transpose (#1940)
• CPU and GPU implementations return results at dissimilar scales, even in double precision arithmetics (#1950)
• Failed CUDA.jl initialization breaks Flux? (#1952)
• Recent mul! changes break multiplication with matrices that have StaticArray elements (#1953)
• Test infrastructure: define test groups (#1961)
• Strange rand errors when sampling large matrices (#1963)
• Add aqua tests (#1964)
• Support of Orin GPU from Nvidia ? (#1966)
• Crash in LLVM (#1971)
• Warning cuDNN Convolution (#1972)
• Strange behaviour when installed at system level (#1973)

Merged pull requests:

• Update benchmarks for 1.8 and 1.9 (#1933) (@maleadt)
• CUSOLVER: Explicitly pass NULL when not requesting svd outputs. (#1934) (@maleadt)
• Detect and complain about loading system libraries. (#1935) (@maleadt)
• Update manifest (#1936) (@github-actions[bot])
• Avoid stack overflow with eary OOM reporting. (#1937) (@maleadt)
• [CUSPARSE] Improved support for UniformScaling ad Diagonal (#1941) (@albertomercurio)
• Update manifest (#1949) (@github-actions[bot])
• Update GPUCompiler to fix unreachable control flow. (#1951) (@maleadt)
• Allow StaticArray eltype in matmat{vec,mul} (#1954) (@lcw)
• Bump CUDNN to v8.9. (#1959) (@maleadt)
• Bump CUTENSOR to v1.7. (#1960) (@maleadt)
• Add and fix some aqua tests (#1965) (@charleskawczynski)
• Fix compatibility of CUDA 11.4 to support Orin. (#1967) (@maleadt)
• Don't use Int32 indices in rand kernels. (#1969) (@maleadt)
• CI simplifications (#1970) (@maleadt)
• Use Base.pkgversion on 1.9. (#1974) (@maleadt)
• Update to LLVM.jl 6. (#1976) (@maleadt)
• fix launch config bug in bitonic sort (#1979) (@xaellison)
• Update manifest (#1980) (@github-actions[bot])

v4.3.2

CUDA v4.3.2

Diff since v4.3.1

Merged pull requests:

• Reduce load time by shifting mul! definition (#1904) (@dkarrasch)

v4.3.1

CUDA v4.3.1

Diff since v4.3.0

Closed issues:

• Array testsuite compiles kernel with large types (#1902)
• CUDA.jl v4 installs CUDA runtime despite version=local (#1922)
• Occaisonal "CUSOLVERError: an internal operation failed (code 7, CUSOLVER_STATUS_INTERNAL_ERROR)" (#1924)
• Does [email protected] need [email protected]? (#1929)

Merged pull requests:

• Simplify libdevice linking. (#1927) (@maleadt)
• Add a show method for kernel objects. (#1928) (@maleadt)
• Update manifest (#1930) (@github-actions[bot])
• Pass a higher capability to ptxas. (#1931) (@maleadt)

v4.3.0

CUDA v4.3.0

Diff since v4.2.0

Closed issues:

• Multidimensional reverse (#1126)
• Test errors on master (#1866)
• Integer overflow error with svd for large matrix (#1880)
• Erratic behaviour of CUDA.jl if used in the REPL of VSCode. (#1892)
• QR decomposition requires scalar indexing (#1893)
• BSOD during package tests (#1898)
• Insufficient coverage of CuArrays in the documentation (#1901)
• Failed to compile with Julia v1.9 on PowerPC (#1911)
• CUDA test failed in wmma.jl (#1914)
• Fix deprecation warnings (#1920)

Merged pull requests:

• CUSOLVER: Fix workspace size passing. (#1890) (@maleadt)
• Lovelace fixes (#1894) (@maleadt)
• Update manifest (#1897) (@github-actions[bot])
• Reverse with multiple dimensions (#1899) (@RainerHeintzmann)
• Restrict number of test jobs based on available memory. (#1900) (@maleadt)
• Avoid unneeded macros to cut down on generated code (#1905) (@maleadt)
• Avoid unneeded macros to cut down on generated code (#1906) (@maleadt)
• Update manifest (#1907) (@github-actions[bot])
• Bump GPUCompiler. (#1908) (@maleadt)
• Don't use Float64 atomics on unsupported platforms. (#1912) (@maleadt)
• Report package versions as part of versioninfo(). (#1913) (@maleadt)
• Align variables in constant memory by 256 bit (#1915) (@Zentrik)
• Add norm functions for 3 floats (#1916) (@Zentrik)
• cuDNN: only choose conv algorithms if they match descriptor mathType (#1917) (@ToucheSir)
• Update manifest (#1918) (@github-actions[bot])
• Skip Integer WMMA tests on older devices. (#1919) (@maleadt)

v4.2.0

CUDA v4.2.0

Diff since v4.1.4

Closed issues:

• NVTX: consider using Start/End for ranges (#1485)
• Limitations of CuIterator (#1768)
• Testing fails on unsupported devices. (#1815)
• Local runtime discovery does not work for external libraries (CUDNN, CUTENSOR) (#1850)
• Passing tests using Github CI workflow errors with libcuda not defined (#1867)
• Cannot precompile GPU code with SnoopPrecompile (#1870)
• Incorrect kernel execution with bounds checking using Julia 1.9.0-rc2 (#1875)
• Fake CUDA library (#1879)
• Error thrown when launching Julia with Nsight systems or compute. (#1886)
• Cannot construct CuDeviceArray (#1887)
• Incorrect colVal array when using CuSparseMatrixCSR command on sparse matrix (#1888)

Merged pull requests:

• Use adapt symmetrically in CuIterator (#1769) (@mcabbott)
• Allow but warn when testing on not fully-supported devices. (#1818) (@maleadt)
• Support runtime discovery for non-toolkit libraries (CUTENSOR, CUDNN, CUQUANTUM) (#1858) (@mloubout)
• Add KernelAbstractions.jl unsafe_free! (#1863) (@pxl-th)
• Allow precompiling CUDA code. (#1865) (@maleadt)
• Assert CUDA.jl is functional when creating the TLS. (#1868) (@maleadt)
• Update manifest (#1871) (@github-actions[bot])
• Don't collect AbstractQ objects in tests (#1872) (@dkarrasch)
• Add compatibility entry for Lovelace (#1873) (@xaellison)
• remove some type-piracy from cusparse (#1876) (@vtjnash)
• Remove more unneeded ndims methods. (#1878) (@maleadt)
• Guard the initialization-time CUDA driver check in a try/catch. (#1881) (@maleadt)
• Update manifest (#1882) (@github-actions[bot])
• Update CUDA 12.1 to 12.1.1. (#1883) (@maleadt)
• Use atomics for allocation statistics. (#1884) (@maleadt)
• Fix atomic increment of alloc stats. (#1885) (@maleadt)
• Update manifest (#1889) (@github-actions[bot])

v4.1.4

CUDA v4.1.4

Diff since v4.1.3

Closed issues:

• Buggy precompilation of init-defined symbols can break CUDA_Driver_jll initialization (#1798)
• Calling CUDA.set_runtime_version!() with float parameter makes CUDA.jl unusable. (#1831)
• Unexpexted memory allocation when using randn! (#1856)
• The memory copy speed seems to exceed the hardware limit (#1860)
• PCG produces different output on GPU (via Krylov.jl) (#1864)

Merged pull requests:

• Fix system_driver_version on platforms not supported by CUDA_Driver_jll. (#1854) (@maleadt)
• Update manifest (#1861) (@github-actions[bot])

v4.1.3

CUDA v4.1.3

Diff since v4.1.2

Closed issues:

• CUDA.versioninfo() triggers download of lazy artifacts (#1844)

Merged pull requests:

• Choose parallel tests based on CPUs, not threads. (#1842) (@maleadt)
• Adapt to LLVM.jl 5 and GPUCompiler.jl 0.19. (#1847) (@maleadt)

v4.1.2

CUDA v4.1.2

Diff since v4.1.1

Closed issues:

• Flux's gradient differentiatingrfft leads to non-bit error (#1835)

Merged pull requests:

• switch to using defined globals (#1832) (@simonbyrne)
• Update manifest (#1837) (@github-actions[bot])