v2.5.0

CUDA v2.5.0

Diff since v2.4.0

v2.4.0

CUDA v2.4.0

Diff since v2.3.0

Closed issues:

• cublasXtStrmm test failures on Windows 10 Julia 1.1 (#124)
• CUSPARSE tests broken (#259)
• Make @cuda return a kernel object (#341)
• Depend on CompilerSupportLibraries (#359)
• CUBLAS and exceptions test failures on Windows (#536)
• argmax(::CuArray) returns nothing with NaN-values (#553)
• Multiple @cuDynamicSharedMem in kernel causes unexpected behavior (#555)
• Illegal memory access with atomic shared memory (#558)
• CUDA.sqrt will not found symbol "__nv_sqrt" (#559)
• Exception with CUDA.exp (#561)
• Use LazyArtifacts instead of Pkg (#570)
• Test runner: early bail out (#578)
• memory reporting issue (#579)
• c[3:4]=0 leads to exception (#580)
• Add math ops (including broadcast) for half types (#581)
• Dot product of Array and CuArray fails with CPU address error. (#586)
• Support for CUDA-capable GPU with compute capability 4.0 like GTX 1080 (#587)
• mapreducedim! not threadsafe (#588)
• Allow separate directories for cuda and cudnn (#590)
• Difficulties installing CUDA on Julia 1.6.0 . (#591)
• Bug in Initialisation Error (#603)
• CUDA.jl initialisation fails after suspending Ubuntu 20.04 with CUDA 11.2 (#605)
• CUDA 11.2 CUBLASError and "CUDA.jl does not yet support CUDA with nvdisasm 11.2.67" (#607)
• This intrinsic must be compiled to be called (#611)
• OpenGL interop (#612)
• Add support for CuFFT callback functions (#614)
• I can’t multiply a CSR sparse matrix anymore (#615)
• Julia version requirement (#619)

Merged pull requests:

• Support all combinations of datatypes and transposes/adjoints in LinearAlgebra (#535) (@cqql)
• Use structs for texture intrinsic return types. (#554) (@maleadt)
• Backport some 1.6 fixes (#557) (@maleadt)
• Update manifest (#560) (@github-actions[bot])
• Correct dims error (#562) (@DhairyaLGandhi)
• Lock _shmem_cb (#564) (@vchuravy)
• Move to Julia 1.6 (#566) (@maleadt)
• Adapt to JuliaLang/julia#38487. (#568) (@maleadt)
• Support for 'delayed kernels' (#569) (@maleadt)
• Run cuda-memcheck as part of CI (#571) (@maleadt)
• Use at-sync instead of calls to synchronize in tests. (#572) (@maleadt)
• Update artifacts to include cuda-memcheck (#573) (@maleadt)
• Use LazyArtifacts instead of Pkg. (#574) (@maleadt)
• Improve LinearAlgebra impl methods for triangular types (#575) (@maleadt)
• New findmin/max implementation using single-pass reduction (#576) (@maleadt)
• Fix synchronization before testing cublasXt calls. (#577) (@maleadt)
• Fix used memory reporting. (#582) (@maleadt)
• Implement Statistics.varm/stdm instead of Statistics._var (#583) (@sdewaele)
• Test for #558. (#584) (@maleadt)
• Add a quick failure option to the test runner. (#585) (@maleadt)
• Add lock around cfunction lookup (#589) (@vchuravy)
• Catch all initialization errors. (#593) (@maleadt)
• Update dependencies. (#596) (@maleadt)
• Fix wrong initialisation error message (#604) (@qin-yu)
• Fixes wrong spacing in docstring admonition (#608) (@navidcy)
• Fix broadcasting with Base.angle (#618) (@marius311)
• Test with the 1.6 nightly, not 1.7. (#620) (@maleadt)
• Wrap cudaGL.h (#621) (@maleadt)
• Initial compatibility with CUDA 11.2. (#622) (@maleadt)
• 1.5 compatibility release (#623) (@maleadt)
• Add CUDA 11.2 artifacts. (#624) (@maleadt)

v2.3.0

CUDA v2.3.0

Diff since v2.2.1

Closed issues:

• Misaligned address on load from Const (#548)

Merged pull requests:

• Allow PermutedDimsArray in gemm_strided_batched (#539) (@mcabbott)
• Fix broken checkbounds for CuSparseMatrixCSR and tests (#545) (@achuchmala)
• Emphasize rebooting option. (#547) (@xanfus)
• fix address calculation for ldg (#549) (@vchuravy)
• Don't use explicit per-stream threads. (#551) (@maleadt)

v2.2.1

CUDA v2.2.1

Diff since v2.2.0

v2.2.0

CUDA v2.2.0

Diff since v2.1.0

Closed issues:

• cudnn missing after downloading artifact (#521)
• Downloading artifact: CUDA110 when using DiffEqFlux (#542)

Merged pull requests:

• Update manifest (#520) (@github-actions[bot])
• Try out Buildkite. (#522) (@maleadt)
• Update manifest (#529) (@github-actions[bot])
• Support for / Upgrade to CUDA 11.1 update 1. (#530) (@maleadt)
• Fix and test svd! (#531) (@maleadt)
• Move more CI to Buildkite. (#532) (@maleadt)
• Use type symbols to generate wrapper methods (#534) (@cqql)
• Fully move to Buildkite. (#537) (@maleadt)
• Add unit_diag option for sv2! functions (#540) (@amontoison)
• Documentation fixes (#543) (@maleadt)

v2.1.0

CUDA v2.1.0

Diff since v2.0.2

Closed issues:

• CUDNN convolution with Float16 always returns zeros (#92)
• axp(b)y! and mul! (scalar multiplication) with mixed argument types (#144)
• Dispatching to generic matmul instead of CUBLAS (#164)
• Support for Ints and Float16? (#165)
• Subarrays/views support (#172)
• Easy way to pick among multiple GPUs (#174)
• More prominently document JULIA_CUDA_USE_BINARYBUILDER (#204)
• ERROR_COOPERATIVE_LAUNCH_TOO_LARGE during tests (#247)
• Pkg.test error for cutensor test on Windows (#422)
• Runtime build improvements (#456)
• Fusing Wrappers (#467)
• Could not find nvToolsExt (libnvToolsExt.dylib.1.0 or libnvToolsExt.dylib.1) in /Users/imac/.julia/artifacts/b502baf54095dff4a69fd6aba8667124583f6929/lib (#482)
• mapreduce assumes commutative op (#484)
• SubArray Broadcast Bug in 2.0 (#488)
• Nested SubArray Scalar Indexing (#490)
• Sparse matrix * view(vector) regression in 2.0 (#493)
• Error transforming a reshaped 0-dimentional GPU array to a CPU array (#494)
• test cuda FAILURE (#496)
• Reshaped CuArray is not DenseCuArray (#511)
• assignment failure when using array slicing. (#516)

Merged pull requests:

• Use the correct CUDNN scaling parameter type. (#454) (@maleadt)
• Fix versioned dylib discovery. (#486) (@maleadt)
• Move inv from GPUArrays. (#487) (@maleadt)
• Use dense array types in sparse wrappers. (#495) (@maleadt)
• Update manifest (#497) (@github-actions[bot])
• Revert array wrapper union changes (#498) (@maleadt)
• Clean-up pointer field. (#499) (@maleadt)
• mapreduce: change iteration for compatibility with non-commutative operators. (#500) (@maleadt)
• Use versioned libcuda (#502) (@maleadt)
• Dynamically choose versioned libcuda (#503) (@mustafaquraish)
• Update multigpu.md (#504) (@efmanu)
• Upgrade artifacts for CUDA 11 compatibility. (#506) (@maleadt)
• Update dependencies. (#507) (@maleadt)
• Convert unsigned short ints to Cint for printf. (#508) (@maleadt)
• Update manifest (#510) (@github-actions[bot])
• Fix reshape with missing dimensions. (#512) (@maleadt)
• Don't return a pointer from 'alias'. (#513) (@maleadt)
• Add some docs (#514) (@maleadt)
• Fix CUDNN-optimized activation broadcasts (#515) (@maleadt)
• Fix cooperative launch test. (#517) (@maleadt)
• Fixes for Windows (#518) (@maleadt)
• CUTENSOR fixes on Windows (#519) (@maleadt)

v2.0.2

CUDA v2.0.2

Diff since v2.0.1

Closed issues:

• cu() behavior for complex floating point numbers (#91)
• Error when following example on using multiple GPUs on multiple processes (#468)
• MacOS without nvidia GPU is trying to download CUDA111 on julia nightly (#469)
• Drop BinaryProvider? (#474)
• Latest version of master doesn't work on Windows (#477)
• sum(CUDA.rand(3,3)) broken (#480)
• copyto!() between cpu and gpu with subarrays (#491)

Merged pull requests:

• Adapt to GPUCompiler changes. (#458) (@maleadt)
• Fix initialization of global state (#471) (@maleadt)
• Remove 'view' implementation. (#472) (@maleadt)
• Workaround new artifact"" eagerness that prevents loading on unsupported platforms (#473) (@ianshmean)
• Remove BinaryProvider dep. (#475) (@maleadt)
• typo: libcuda.dll -> libcuda.so on Linux (#476) (@Alexander-Barth)
• NFC array simplifications. (#481) (@maleadt)
• Update manifest (#485) (@github-actions[bot])
• Convert AbstractArray{ComplexF64} to CuArray{ComplexF32} by default (#489) (@pabloferz)

v2.0.1

CUDA v2.0.1

Diff since v2.0.0

Closed issues:

• Can't update (#462)

Merged pull requests:

• Remove duplicate comment (#464) (@blegat)
• Add functionality to precompile the runtime library. (#465) (@maleadt)
• Update manifest (#470) (@github-actions[bot])

v2.0.0

CUDA v2.0.0

Diff since v1.3.3

Closed issues:

• Test failure during threading tests (#15)
• Bad allocations in memory pool after device_reset! (#16)
• CuArrays can lose Blas on reshaped views (#78)
• allowscalar performance (#87)
• Indexing with a CuArrays causes a 'scalar indexing disallowed' error from checkbounds (#90)
• 5-arg mul! for CUSPARSE (#98)
• copyto!(Device, Host) uses scalar iteration in case of type mismatch (#105)
• Array primitives broken for CUSPARSE arrays (#113)
• SplittingPool: CPU allocations (#117)
• error while concatenating to an empty CuArray (#139)
• Showing sparse arrays goes wrong (#146)
• Improve test coverage (#147)
• CuArrays allocates a lot of memory on the default GPU (#153)
• [Feature Request] Indexing CuArray with CuArray (#155)
• Reshaping CuArray throws error during backpropagation (#162)
• Match syntax and APIs against Julia 1.0 standard libraries (#163)
• CURAND_STATUS_PREEXISTING_FAILURE when setting seed multiple times. (#212)
• RFC: converts SparseMatrixCSC to CuSparseMatrixCSR via cu by default (#216)
• Add a CuSparseMatrixCOO type (#220)
• Test runner stumbles over path separators (#236)
• Error: Invalid bitcode signature when loading CUDA.jl after precompilation (#293)
• Atomic operations only work on global memory (#311)
• Performance: cudnn algorithm selection (#318)
• CUSPARSE is broken in CUDA.jl 1.2 (#322)
• Device-side broadcast regression on 1.5 (#350)
• API for fast math-like mode (#354)
• CUDA 11.0 Update 1: cublasSetWorkspace (#365)
• Can't precompile CUDA.jl on Kubuntu 20.04 (#396)
• CuPtr should be Ptr in cudnnGetDropoutDescriptor (#397)
• CUDA throws OOM error when initializing API on multiple devices (#398)
• Cannot launch kernel with > 5 args using Dynamic Parallelism (#401)
• Reverse performance regression (#410)
• Tag for LLVM 3? (#412)
• CUDA not working (#415)
• StatsBase.transform fails on CuArray (#426)
• Further unification of CUBLAS.axpy! and LinearAlgebra.BLAS.axpy! (#432)
• size(range), length(range) and range[end] fail inside CUDA kernels (#434)
• InitError: Cannot use memory pool 'binned' when CUDA.jl was precompiled for memory pool 'split'. (#446)
• Missing dispatch for matrix multiplication with views? (#448)
• New version not available yet? (#452)
• using CUDA or CUArray, output: UndefVarError: AddrSpacePtr not defined (#457)
• Unable to upgrade to the latest version (#459)

Merged pull requests:

• Performance improvements by calling cuDNN API (#321) (@gartangh)
• Use ccall wrapper for correct pointer type conversions (#392) (@maleadt)
• Simplify Statistics.var and fix dims=tuple. (#393) (@maleadt)
• Adapt to GPUArrays test change. (#394) (@maleadt)
• Default to per-thread stream semantics (#395) (@maleadt)
• Add a missing context argument for stateless codegen. (#399) (@maleadt)
• Keep track of package latency timings. (#400) (@maleadt)
• Update manifest (#402) (@github-actions[bot])
• Latency improvements (#403) (@maleadt)
• Fix bounds checking with GPU views. (#404) (@maleadt)
• Force specialization for dynamic_cudacall to support more arguments. (#407) (@maleadt)
• Fix some wrong pointer types in the CUDNN headers. (#408) (@maleadt)
• Refactor CUSPARSE (#409) (@maleadt)
• Fix typo (#411) (@yixingfu)
• Update manifest (#413) (@github-actions[bot])
• Simplify library wrappers by introducing a CUDA Ref (#414) (@maleadt)
• Simplify and update wrappers (#416) (@maleadt)
• GEMM improvements (#417) (@maleadt)
• CompatHelper: add new compat entry for "BFloat16s" at version "0.1" (#418) (@github-actions[bot])
• add CuSparseMatrixCOO (#421) (@marius311)
• Update manifest (#423) (@github-actions[bot])
• Global math mode for easy use of lower-precision functionality (#424) (@maleadt)
• Improve init error message (#425) (@maleadt)
• CUBLAS: wrap rot! to implement rotate! and reflect! (#427) (@maleadt)
• CUFFT-related optimizations (#428) (@maleadt)
• Fix reverse/view regression (#429) (@maleadt)
• Update packages (#433) (@maleadt)
• Introduce StridedCuArray (#435) (@maleadt)
• Retry curandGenerateSeeds when OOM. (#436) (@maleadt)
• Introduce DenseCuArray union (#437) (@maleadt)
• Array simplifications (#438) (@maleadt)
• Fix and test reverse on wrapped array. (#439) (@maleadt)
• Fixes after recent array wrapper changes (#441) (@maleadt)
• Adapt to GPUArrays changes. (#442) (@maleadt)
• Provide CUBLAS with a pool-backed workspace. (#443) (@maleadt)
• Fix finalization of copied arrays. (#444) (@maleadt)
• Support for/Add CUDA 11.1 (#445) (@maleadt)
• Update manifest (#449) (@github-actions[bot])
• Allow use of strided vectors with mul! (gemv! and gemm!) (#450) (@maleadt)
• Have convert call CuSparseArray's constructors. (#451) (@maleadt)

v1.3.3

CUDA v1.3.3

Diff since v1.3.2

Closed issues:

• Type changing Array conversions give error when allowscalar(false) (#344)
• getindex(::CuArray, ::Adjoint, ::Colon) fails (#345)
• View with array indices causes memory copy before broadcast (#384)
• Regression with Julia 1.5 (#390)

Merged pull requests:

• Replace DevicePtr with Core.LLVMPtr. (#199) (@maleadt)
• Make sure view indices reside on the GPU too. (#388) (@maleadt)
• CompatHelper: Update DataStructures to v0.18 (#389) (@ChrisRackauckas)