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)

v3.9.1

CUDA v3.9.1

Diff since v3.9.0

Closed issues:

• Issue with copy_cublasfloat (#1476)
• Errors when broadcasting random number generators (#1480)
• CPU version of linear algebra routine is dispatched when using Zygote.gradient (#1481)
• scan! fails on vectors of structs (#1482)
• InexactError when getting CUDA version info (#1489)

Merged pull requests:

• Allow more integer argument types for byte_perm (#1420) (@eschnett)
• support CuSparseMatrix(::Diagonal) (#1470) (@Roger-luo)
• Don't emit debug info until the next CUDA version. (#1473) (@maleadt)
• Update manifest (#1474) (@github-actions[bot])
• Update manifest (#1479) (@github-actions[bot])
• fix unsafe_wrap docstring and widen signature (#1483) (@piever)
• Update manifest (#1484) (@github-actions[bot])
• Check whether cudaRuntimeGetVersion succeeded. (#1490) (@maleadt)
• Update manifest (#1494) (@github-actions[bot])
• Fix #1476: Allow any container in copy_cublasfloat (#1498) (@danielwe)

v3.9.0

CUDA v3.9.0

Diff since v3.8.5

Closed issues:

• Tests for showing (#35)
• Support LU factorizations (#1193)
• Int8 WMMA not working in 3.8.4 and 3.8.5 despite merged PR. Add more unit tests? (#1442)
• Optional CPU cpu kernel call with @cuda (#1443)
• Add library/artifact management for NCCL (#1446)
• permutedims returns a lowertriangular matrix (#1451)
• New broadcast corrupts memory? (#1457)
• norm does not dispatch on CuSparseMatrixCSC (#1460)
• scalar * sparse multiplication (#1468)

Merged pull requests:

• CUTENSOR: axpy! and axpby! not mutating fixed (#1416) (@yapanuwan)
• Initial wrap of cuquantum (#1437) (@kshyatt)
• CompatHelper: bump compat for "GPUCompiler" to "0.14" (#1441) (@github-actions[bot])
• Fix return type of nrm2 for ComplexF16 (#1444) (@danielwe)
• Use a build matrix. (#1445) (@maleadt)
• Update manifest (#1447) (@github-actions[bot])
• Rework factorizations (#1449) (@maleadt)
• Add NCCL binaries. (#1450) (@maleadt)
• Support general eltypes in matrix division and SVD (#1453) (@danielwe)
• Update manifest (#1456) (@github-actions[bot])
• Look at more environment variables to find nsys. (#1459) (@maleadt)
• Fixes for 1.8 (#1463) (@maleadt)

v3.8.5

CUDA v3.8.5

Diff since v3.8.4

Merged pull requests:

• Update manifest (#1440) (@github-actions[bot])

v3.8.4

CUDA v3.8.4

Diff since v3.8.3

Closed issues:

• sparse-sparse and sparse-constant multiplication lose sparsity (output dense matrix) (#1264)
• LLVMExtra fails to load on Julia 1.8 and PPC (#1387)
• compute-sanitizer CUDA_ERROR_INVALID_VALUE on CUDA.jl 3.0+ (#1415)
• @cudnnDescriptor is not threadsafe (#1421)
• Precomplication of CUDA 3.8.3 broken on 1.7.1 due to changes in Random123.jl (#1422)
• OOM error should include memory status (#1427)
• WMMA kernel works with Julia 1.7.2 but fails with illegal memory access for Julia 1.8.0-beta1 (#1431)
• Non Int64 local memory size leads to dynamic function invocation (#1434)
• "initialization" test failing (#1435)
• cuda with julia 1.8 not working on windows (working fine(?) on wsl2) (#1436)

Merged pull requests:

• Add Int8 WMMA Support (#1119) (@max-Hawkins)
• Wrap generic sparse-sparse GEMM (#1285) (@kshyatt)
• Fix sparse COO to CSR conversion. (#1412) (@maleadt)
• Drop support for CUDA 10.1 and below (#1414) (@maleadt)
• Update manifest (#1417) (@github-actions[bot])
• Report the OOM memory status at the time of the error. (#1428) (@maleadt)
• Lock CUDNN descriptor cache lookups. (#1430) (@maleadt)
• Switch to new LLVM context management for 1.9 compatibility. (#1432) (@maleadt)
• Update manifest (#1433) (@github-actions[bot])
• Backports for 3.8.4 (#1438) (@maleadt)

v3.8.3

CUDA v3.8.3

Diff since v3.8.2

Closed issues:

• Sparse matrix addition not working (#528)
• Native implementation of sparse arrays (#829)
• CUSPARSE: Adding a value to the diagonal (#1372)
• Conversion by cu casts Float64 to Float32 but not Int64 to Int32 (#1388)
• CUDA.math_mode!(...; precision) option not working (#1392)
• cuIpcGetMemHandle failure resulting in CUDA-aware MPI to fail (#1398)
• axpby! support for BFloat16 (#1399)
• CUSPARSE does not support integer matrices, breaks printing (#1402)
• sparse(I, J, V) doesn't support unsorted inputs (#1407)

Merged pull requests:

• General purpose broadcast for sparse CSR matrices. (#1380) (@maleadt)
• Update manifest (#1389) (@github-actions[bot])
• Implement sparse operations with UniformScaling using broadcast. (#1390) (@maleadt)
• Prevent toplevel compilation. (#1391) (@maleadt)
• Fix and test math precision. (#1394) (@maleadt)
• Bump artifacts (#1397) (@maleadt)
• support BFloat16 for atomic_cas (#1400) (@bjarthur)
• Implement sparse broadcasting with CSC matrices. (#1401) (@maleadt)
• Always report issues with discovering CUDA. (#1404) (@maleadt)
• Fix sparse 1-argument broadcast output type. (#1405) (@maleadt)
• CUSPARSE BSR improvements (#1409) (@maleadt)
• Support limited sparse integer arrays by bitcasting to floating point. (#1410) (@maleadt)
• Support using sparse with unsorted inputs. (#1411) (@maleadt)
• Backports for 3.8.3 (#1413) (@maleadt)

v3.8.2

CUDA v3.8.2

Diff since v3.8.1

Closed issues:

• CuSparseMatrixCSC missing lu and interactions with UniformScaling (#79)
• CUSPARSE typo (#1231)
• similar(A::CuSparse,eltype) returns an Array (#1316)
• "errormonitor" undefined in julia1.6 (#1375)
• Pool free can switch tasks (#1384)

Merged pull requests:

• Define a compatibility shim for errormonitor (#1378) (@vchuravy)
• Backport #1361 to 3.8 (#1379) (@vchuravy)
• Backports for 3.8.2 (#1381) (@maleadt)
• Remove broken errormonitor implementation, just don't use it on 1.6. (#1382) (@maleadt)
• Memory pool improvements (#1383) (@maleadt)

v3.8.1

CUDA v3.8.1

Diff since v3.8.0

Closed issues:

• one(::CuMatrix) result on cpu (#142)
• Broadcasted setindex! triggers scalar setindex! (#101)
• OutOfGPUMemoryError With Available Memory (#1346)
• Distributions.jl with CuArrays (#1347)
• Views of Flux OneHotArrays (#1349)
• synchronize(blocking = false) hangs in julia 1.7 eventually (#1350)
• unsupported call through a literal pointer (call to log1pf) on Julia 1.6.5 (#1352)
• SpecialFunctions ^1.8 compat entry? (#1354)
• Performance deprecation using ^ on Float32 (#1358)
• Method definition setindex!(LinearAlgebra.Diagonal{T, V} ... overwritten in module CUDA (#1364)
• [PackageCompiler] Segmentation fault with CUDA.jl in multiversioning (#1365)
• Vectors in customary structs make julia stuck (#1366)
• sparseCSC-dense matrix multiplication yields unstable results (#1368)
• UndefVarError: parameters not defined on Windows10 (#1371)

Merged pull requests:

• Optimize memoization helpers. (#1345) (@maleadt)
• Update manifest (#1348) (@github-actions[bot])
• Update manifest (#1355) (@github-actions[bot])
• Fastmath improvements (#1356) (@maleadt)
• Make the default pool visible when doing P2P (#1357) (@maleadt)
• Fix resize of empty arrays. (#1359) (@maleadt)
• CUSPARSE: add COO ctors and similar with eltype. (#1360) (@maleadt)
• Add device_override for SpecialFunctions.gamma (#1361) (@vchuravy)
• Implement (limited) broadcast of sparse arrays (#1367) (@maleadt)
• Make nonblocking synchronization robust to errors. (#1369) (@maleadt)
• Update manifest (#1370) (@github-actions[bot])
• Backports for 3.8.1 (#1374) (@maleadt)

v3.8.0

CUDA v3.8.0

Diff since v3.7.1

Closed issues:

• Consider reserving memory (#1320)

Merged pull requests:

• Slight changes to pool management (#1344) (@maleadt)

v3.7.1

CUDA v3.7.1

Diff since v3.7.0

Closed issues:

• Moving data between devices (#1136)
• Repeated has_cuda_gpu errors when CUDA_VISIBLE_DEVICES is empty (#1331)
• Error when env var CUDA_VISIBLE_DEVICES is set but empty (#1336)

Merged pull requests:

• Wrap and test peer to peer memory copies (#1284) (@kshyatt)
• Update manifest (#1332) (@github-actions[bot])
• Have libcuda() fail repeatedly if anything (e.g. init) failed. (#1333) (@maleadt)
• Simplify workarounds. (#1334) (@maleadt)
• Properly detect a missing driver. (#1335) (@maleadt)
• Various small fixes (#1337) (@maleadt)
• Move CUDA.jl global state innto CUDAdrv wrapper "submodule" (#1338) (@maleadt)
• Add CUDA.return_type (#1339) (@tkf)
• Compute-sanitizer QOL improvements and docs (#1340) (@maleadt)
• Fix regression in backwards CUFFT plans. (#1341) (@maleadt)
• Don't assume host pointers are directly usable on the device. (#1342) (@maleadt)
• Backports for 3.7.1 (#1343) (@maleadt)