v4.0.1

What's Changed

• Warn when using old devices by @maleadt in #1752
• Silence some errors to support conditional use. by @maleadt in #1754

Full Changelog: v4.0.0...v4.0.1

v4.0.0

CUDA v4.0.0

Diff since v3.13.1

Closed issues:

• Missing implementation of right multiply for QR decomposition (#1738)
• [CUSPARSE] Type error with mm! (#1743)

Merged pull requests:

• Implement rmul for qr. (#1739) (@maleadt)
• Update manifest (#1741) (@github-actions[bot])
• Update CUSPARSE for CUDA v12.0 (#1744) (@amontoison)
• Fix nvprof command (#1745) (@lucifer1004)
• Update manifest (#1747) (@github-actions[bot])
• Fix grammar (#1748) (@lucifer1004)

v3.13.1

CUDA v3.13.1

Diff since v3.13.0

Closed issues:

• CUDA.jl cuFFT underperforming against CuPy cuFFT (#1682)
• Is block-spmm supported？ (#1736)

Merged pull requests:

• Introduce cuFFT plan cache; switch to auto-managed memory. (#1734) (@maleadt)
• Stop pirating GPUArrays' RNG methods. (#1735) (@maleadt)

v3.12.2

CUDA v3.12.2

Diff since v3.12.1

Closed issues:

• CUDA.jl cuFFT underperforming against CuPy cuFFT (#1682)
• Error during CUDA test (#1718)
• Kernel error from bad broadcast (should be regular error?) (#1720)
• Freeze into StackOverflow when JULIA_DEBUG=CUDA set (#1721)
• Use of linear operators in CUDA.jl (#1727)
• Is block-spmm supported？ (#1736)

Merged pull requests:

• Allow copy(::RNG) (#1719) (@mcabbott)
• Update manifest (#1722) (@github-actions[bot])
• Simplify CuError rendering before library initialization. (#1723) (@maleadt)
• Simplify CuError rendering before library initialization (master branch version) (#1724) (@maleadt)
• Make device RNG test more robust. (#1725) (@maleadt)
• Rely on LLVM.jl's typed_ccall for more intrinsics. (#1728) (@maleadt)
• Backports for 3.13 (#1729) (@maleadt)
• Simplify CUBLAS and CUSPARSE wrappers, reducing code generated. (#1730) (@maleadt)
• Add Julia 1.9 CI. (#1731) (@maleadt)
• Use released dependencies. (#1732) (@maleadt)
• Remove NVTX. (#1733) (@maleadt)
• Introduce cuFFT plan cache; switch to auto-managed memory. (#1734) (@maleadt)
• Stop pirating GPUArrays' RNG methods. (#1735) (@maleadt)

v3.13.0

CUDA v3.13.0

Diff since v3.12.1

Closed issues:

• Error during CUDA test (#1718)
• Kernel error from bad broadcast (should be regular error?) (#1720)
• Freeze into StackOverflow when JULIA_DEBUG=CUDA set (#1721)
• Use of linear operators in CUDA.jl (#1727)

Merged pull requests:

• Allow copy(::RNG) (#1719) (@mcabbott)
• Update manifest (#1722) (@github-actions[bot])
• Simplify CuError rendering before library initialization. (#1723) (@maleadt)
• Simplify CuError rendering before library initialization (master branch version) (#1724) (@maleadt)
• Make device RNG test more robust. (#1725) (@maleadt)
• Rely on LLVM.jl's typed_ccall for more intrinsics. (#1728) (@maleadt)
• Backports for 3.13 (#1729) (@maleadt)
• Simplify CUBLAS and CUSPARSE wrappers, reducing code generated. (#1730) (@maleadt)
• Add Julia 1.9 CI. (#1731) (@maleadt)
• Use released dependencies. (#1732) (@maleadt)
• Remove NVTX. (#1733) (@maleadt)

v3.12.1

CUDA v3.12.1

Diff since v3.12.0

Closed issues:

• Accumulate doesn't work on >=4 dim Arrays with dims <= ndims(A) - 3 (#1039)
• CUSPARSE does not support dense-sparse matrix multiplication (#1403)
• Scalar indexing when comparing a CuArray to the identity matrix (#1557)
• CUBLAS_STATUS_NOT_INITIALIZED (#1567)
• LinearAlgebra./ and LinearAlgebra.\ breaks CuArray (#1568)
• Window size in grid-stride loop (#1573)
• Matrix multiplication works for primitive and non-primitive custom number types on the CPU, but it fails for primitive custom number types on the GPU. (#1574)
• CuIterator doesn't specify IteratorSize but has no length() (#1583)
• Garbage collection doesn't work as shown in the documentation (#1586)
• Adding sparse adjoint results in kernel error (#1591)
• sparse - sparse matrix multiplication partially missing (#1599)
• FastMath sincos(), cis(), exp(im..) aren't as fast as C++ (#1606)
• wrong type in wrapper of a cusolver function (#1621)
• Adding CUDNN support for 3D convolutions/cross-correlations (#1631)
• copyto! does not work between a CuArray and a view(Array) (#1634)
• Minor issue with sparse function (#1641)
• Scalar indexing when displaying Diagonal{Int64, CuSparseVector{Int64, Int32}} (#1645)
• Many errors running test suite on GTX 960 4GB (#1650)
• Driver discovery broken on platforms without compat driver (#1653)
• Aliasing/Polluted Result from rfftplan for Float32 2^n 3D array (#1656)
• Re-instate memory limit (#1670)
• Split libnvToolsExt from CUDA_Runtime_jll? (#1672)
• accumulate(op, a) causes scalar indexing (#1680)
• CUSPARSE CI failures (#1692)
• axpy! for nested base types (reshapedarray/adjoint/view) (#1696)
• copyto! between a PermutedDimsArray view and a CuArray doesn't work (#1697)
• WMMA test failure (#1700)
• UndefVarError when a binary is not found (#1701)
• Is CUSPARSELT supported? (#1702)
• Best practices to reduce startup time (#1707)
• 1.9 compatibility (#1710)
• WARNING: unused variadic paramters. (#1712)

Merged pull requests:

• Remove/rework CuDeviceArray constructors (#1308) (@maleadt)
• Add always_inline kernel parameter (#1554) (@lcw)
• Update manifest (#1564) (@github-actions[bot])
• Update manifest (#1569) (@github-actions[bot])
• Update manifest (#1571) (@github-actions[bot])
• Fix native RNG window calculation. (#1575) (@maleadt)
• Use Base.active_project. (#1576) (@maleadt)
• Fixes for and tests using JET. (#1577) (@maleadt)
• Update manifest (#1578) (@github-actions[bot])
• Docs, remove global variables in intro benchmark (#1580) (@SteffenPL)
• Update manifest (#1581) (@github-actions[bot])
• Update manifest (#1582) (@github-actions[bot])
• Bugfixes when using \ operator with non square matrices (#1584) (@GVigne)
• remove unbound type parameters (#1585) (@nsajko)
• added --openacc-profiling off to the nvprof (#1587) (@mbeltagy)
• Update manifest (#1588) (@github-actions[bot])
• Wrap at-cuda's code in a let block. (#1589) (@maleadt)
• Revert: Use JET during test suite. (#1590) (@maleadt)
• [CUSPARSE] Update mv! and mm! functions for CuSparseMatrixCOO and CuSparseMatrixCSC (#1592) (@amontoison)
• [CUSPARSE] Add sv! and sm! routines (#1593) (@amontoison)
• CompatHelper: bump compat for "BFloat16s" to "0.3" (#1594) (@github-actions[bot])
• Update wrap.jl (#1595) (@amontoison)
• Provide more useful explanation why an eltype is unsupported. (#1596) (@maleadt)
• CompatHelper: bump compat for "BFloat16s" to "0.4" (#1597) (@github-actions[bot])
• Improve eltype error reporting. (#1598) (@maleadt)
• Add () at the end of the library name in all ccall (#1600) (@amontoison)
• Define length for CuIterator (#1602) (@mcabbott)
• Added more sparse functions like: kron, tril, triu, reshape, adjoint, transpose, sparse-sparse multiplication (#1603) (@albertomercurio)
• Fix rotate! and reflect! for the generic fallback in GPUArrays.jl (#1604) (@amontoison)
• Update manifest (#1605) (@github-actions[bot])
• Update manifest (#1609) (@github-actions[bot])
• [CUSPARSE] Interface generic routines (#1611) (@amontoison)
• [CUSPARSE] Update sparse-sparse GEMM (#1613) (@amontoison)
• [CUSPARSE] Add sddmm! and gemvi! routines (#1615) (@amontoison)
• Update manifest (#1616) (@github-actions[bot])
• Don't use isbitsunion to support structs of union types. (#1617) (@maleadt)
• Update CUDA driver compatibility package to 11.8. (#1618) (@maleadt)
• Update CUDA artifacts to 11.7 Update 1. (#1619) (@maleadt)
• Update to CUDA 11.8 (#1620) (@maleadt)
• Update to CUDNN 8.6. (#1622) (@maleadt)
• Move CUDNN and CUTENSOR into separate packages (#1624) (@maleadt)
• Bump BFloat16s. (#1625) (@maleadt)
• fix #1621 (#1626) (@jemiryguo)
• Restore functionality of FastMath.sincos. (#1627) (@maleadt)
• Update manifest (#1628) (@github-actions[bot])
• Switch from manual artifact handling to automated JLLs (#1629) (@maleadt)
• [CUSPARSE] Add CuMatrix * CuSparseMatrix products (#1632) (@amontoison)
• Silence some test warnings. (#1635) (@maleadt)
• Update CUTENSOR to v1.6 (#1636) (@maleadt)
• [CUSPARSE] Add SparseMatrix * SparseVector products (#1637) (@amontoison)
• Upgrade CUSTATEVEC to v1.1 (#1638) (@maleadt)
• Upgrade CUTENSORNET to v1.1 (#1639) (@maleadt)
• [CUSPARSE] Add CuSparseVector ± CuSparseVector (#1640) (@amontoison)
• CompatHelper: add new compat entry for "Preferences" at version "1" (#1642) (@github-actions[bot])
• Fix #1641 (#1643) (@amontoison)
• Update manifest (#1646) (@github-actions[bot])
• [CUSPARSE] Add dot(CuSparseVector,CuVector) and vice-versa (#1647) (@amontoison)
• [CUSPARSE] Add ldiv! for CuSparseMatrixCOO and geam for CuSparseMatrixCSC (#1648) (@amontoison)
• Update autogenerated headers (#1649) (@maleadt)
• Remove deprecations (#1651) (@maleadt)
• Don't warn about the old JULIA_CUDA_USE_BINARYBUILDER env var when using preferences (#1652) (@maleadt)
• Update CUTENSORNET to use new slice group (#1654) (@kshyatt)
• [CUSPARSE] Fix conversions between CuSparseMatrixCOO and CuSparseMatrixCSC (#1655) (@amontoison)
• Include compiler options in error log. (#1657) (@maleadt)
• Discover the system driver when CUDA_Driver_jll isn't available. (#1658) (@maleadt)
• Preserve buffer type when adapting to CuArray. (#1659) (@maleadt)
• Update manifest (#1661) (@github-actions[bot])
• Extend conversion of QRPackedQ object to CuArray (#1662) (@GVigne)
• [CUSPARSE] Add CuSparseMatrixCSC * CuSparseMatrixCSC (#1663) (@amontoison)
• Update manifest (#1665) (@github-actions[bot])
• [CUSPARSE] Add more tests (#1668) (@amontoison)
• Update manifest (#1671) (@github-actions[bot])
• Update manifest (#1676) (@github-actions[bot])
• Fix eigen when using Hermitian or Symmetric matrices (#1677) (@GVigne)
• Update manifest (#1679) (@github-actions[bot])
• adding defaults for accumulate(op, a) with modified code from Base.accumulate (#1681) (@leios)
• Add right division operator for Diagonal matrices (#1683) (@GVigne)
• Update manifest (#1686) (@github-actions[bot])
• Bump CUQUANTUM libraries (#1688) (@maleadt)
• typo (#1689) (@ArnoStrouwen)
• Retry CUSOLVER handle creation when encountering an internal error. (#1691) (@maleadt)
• Fix #1692 (#1693) (@amontoison)
• Update manifest (#1694) (@github-actions[bot])
• [CUSPARSE] Support kron with Diagonal arguments (#1695) (@albertomercurio)
• Re-introduce memory limits. (#1698) (@maleadt)
• Adapt to GPUCompiler changes. (#1699) (@maleadt)
• WMMA: Don't wrap fragments of size 1 in a struct. (#1704) (@maleadt)
• Update manifest (#1708) (@github-actions[bot])
• Use plain llvmcall calling convention for WMMA intrinsics. (#1709) (@maleadt)
• Reclaim in cuDNN conv algorithm search (#1711) (@ToucheSir)
• CUBLAS: test against generic axp(b)y, not the BLAS-specific one. (#1713) (@maleadt)
• Fix LU getproperty invoke. (#1714) (@maleadt)
• Backports for 3.12.1 (#1715) (@maleadt)
• Specialize cholcopy to avoid scalar indexing. (#1716) (@maleadt)
• Fix handling of inline-allocated structures with unions. (#1717) (@maleadt)

v3.12.0

CUDA v3.12.0

Diff since v3.11.0

Closed issues:

• Implement Base.repeat (#177)
• repeat performs scalar indexing for multi-dimensional arrays (#1051)
• The GPU compiler fails on a call to maximum (#1548)
• versioninfo triggers artifact downloads (#1549)
• Error when broadcasting composed functions (#1550)
• overload Base.copy! for AbstractGPUArray{<:Any,1} (#1555)

Merged pull requests:

• Fix math quirk. (#1546) (@maleadt)
• Wrap cusolverRf.h and cusolverSp_LOWLEVEL_PREVIEW.h (#1547) (@frapac)
• Update manifest (#1551) (@github-actions[bot])
• tighten unsafe_wrap signature on scalar length (#1552) (@sjkelly)
• Update Documenter key. (#1553) (@maleadt)
• Update manifest (#1556) (@github-actions[bot])
• Import factorisation internal types from LinearAlgebra (#1558) (@theabhirath)
• Update manifest (#1560) (@github-actions[bot])
• add reshape for CuDeviceArray (#1561) (@omlins)

v3.11.0

CUDA v3.11.0

Diff since v3.10.1

Closed issues:

• CUSPARSE: Diagonal + CSC/CSR gives dense array (#1469)
• CUBLAS: Multiplication of UpperTriangular/LowerTriangular not supported (#1486)
• CUTENSOR tests consume lots of memory, breaking other tests (#1501)
• CUFFT doesn't work for ComplexF64 C2C in-place (#1519)
• Inconsistency of == and isequal for CuArray (#1524)
• Setting CUDA seed the first time changes Random's RNG non-deterministically (#1526)
• Undefined exported symbols (#1527)
• Could not load library libLLVMExtra-14.dll (#1535)
• Add an rrule for cholesky to CUDA.jl (#1541)

Merged pull requests:

• specialize +/- op for sparse diag (#1514) (@Roger-luo)
• Make sure instantiating RNGs doesn't affect the global CPU RNG. (#1530) (@maleadt)
• Update manifest (#1531) (@github-actions[bot])
• ldiv! for LU Decomposition (#1532) (@SBuercklin)
• Lower dmax for contraction tests (#1534) (@kshyatt)
• Fix convolution algorithm search (#1536) (@maxfreu)
• Update manifest (#1537) (@github-actions[bot])
• add specializations for some triangular-triangular multiplications (#1538) (@Red-Portal)
• Add a utility to download artifacts without a functional driver. (#1539) (@maleadt)
• Update manifest (#1543) (@github-actions[bot])
• Explicit tests for type conversion (#1544) (@kshyatt)
• Remove unused exports. (#1545) (@maleadt)

v3.10.1

CUDA v3.10.1

Diff since v3.10.0

Closed issues:

• Overflow in randn using CUDA.jl's native RNG (#1464)
• Segmentation fault with pre-compiled library importing CUDA (#1465)
• Julia freezes when using Polynomials with CuArray (#1497)
• Launch overhead regression (#1503)
• CUSOLVER: Matrix division requires identical types (#1512)
• Incorrect distribution for complex standard normals when using CUDA.default_rng() (#1515)
• loggamma (#1528)

Merged pull requests:

• CUSPARSE: Support mixed type mv (#1475) (@Roger-luo)
• Add method for LinearAlgebra.opnorm2 (#1516) (@danielwe)
• Promote to common eltype in matrix division (#1517) (@danielwe)
• Fix Box-Muller transformation for complex eltypes (#1518) (@danielwe)
• Update manifest (#1521) (@github-actions[bot])
• Use at-dispose for LLVM.jl resource cleanup. (#1523) (@maleadt)
• loggamma (#1529) (@cossio)

v3.10.0

CUDA v3.10.0

Diff since v3.9.1

Closed issues:

• Error while freeing DeviceBuffer-warning when using multiple GPUs (#1454)
• CUDNN cache locking prevents finalizers resulting in OOMs (#1461)
• EOFError from pool_cleanup when closing REPL (#1495)
• TypeError in compiler with custom kernel (#1496)

Merged pull requests:

• expose sparse mv/mm algo selection (#1201) (@Roger-luo)
• Always inspect the task-local context when verifying before freeing. (#1462) (@maleadt)
• support sparse opnorm (#1466) (@Roger-luo)
• Move CUSTATEVEC and CUTENSORNET into lib/ (#1478) (@vchuravy)
• Adapt to GPUCompiler 0.15 changes (#1488) (@maleadt)
• Limit time held by CUDNN locks. (#1491) (@maleadt)
• Docstring for cu (#1493) (@mcabbott)
• Update manifest (#1499) (@github-actions[bot])
• Silence EOFError in pool_cleanup (#1502) (@Octogonapus)
• Adapt to GPUCompiler changes (#1504) (@maleadt)
• Fixes for CUSPARSE 11.7.1. (#1505) (@maleadt)
• Update artifacts (#1507) (@maleadt)
• Update manifest (#1509) (@github-actions[bot])
• Add a new cache for HostKernel objects. (#1510) (@maleadt)