v1.3.0

New features:

• #330: wrappers for CUTENSOR 0.2.2
• #414: support for multiple memory pools, and an initial split memory allocator. activate by running with CUARRAYS_MEMORY_POOL=split -- feedback is appreciated.
• #421: headers are now autogenerates, and expose all of the underlying APIs
• #416: optimized ND-reverse kernel
• #447: optimized ND-accumulate kernel (also findall, cumsum, etc)

Important bug fixes:

• #362: specialized reduce/minimum/maximum now dispatch to CuArrays
• #380: culiteral_pow fixes
• #433: CuArrays.@time now synchronizes the GPU while doing time measurements

v1.2.1

Hotfix: avoid errors when CUDNN is not available and its functionality isn't used.

v1.2.0

v1.2.0 (2019-08-23)

Diff since v1.1.0

Closed issues:

• size(collect(F.Q)) != size(Matrix(F.Q)) when size(F, 1) > size(F, 2) for F :: CuQR (#393)
• CuArrays.randn(odd_number, another_odd_number) throws CURANDError (#392)
• Pkg.test("CuArrays") fails on julia-1.1.1, linux, sm35, sm70, cuda10.1, cudnn7.0.6 (#387)
• Assigning a number to a CuArray slice via broadcasting fails (#386)
• using CuArrays fails (#384)
• Invalid versions with Adapt (#374)
• Allow CuArrays to initialise if there is no CUDA/GPU on the system (#371)
• LLVM error when using abs function to TrackedArray + CuArray (#346)

Merged pull requests:

• Improve CURAND: support for non-pow2 dims, and other fixes (#399) (maleadt)
• Define det(::CuQRPackedQ) (#397) (tkf)
• Only set the CUBLAS math mode on CUBLAS 9.0 or higher (#396) (maleadt)
• FFT test failure workarounds (#395) (maleadt)
• Avoid ugly rounding issues when printing pool stats. (#394) (maleadt)
• Define det(::CuQRPackedQ) (#391) (tkf)
• Add ldiv! and tests (#383) (willtebbutt)
• Don't use map to convert SubArray indices. (#382) (maleadt)
• Warn about missing libraries. (#381) (maleadt)
• Support for limiting total memory usage. (#379) (maleadt)
• Replace Pkg.build with runtime initialization (#375) (maleadt)
• Add more methods for seed! (#344) (findmyway)

v1.1.0

v1.1.0 (2019-07-17)

Diff since v1.0.2

Closed issues:

• Slow permute(view) (#367)
• view of CuArray with CPU indices yields CPU array (#360)
• Sudden performance drop after many matrix multiplications (#359)
• Avoid scalar fallback for non-contiguous views (#354)
• unsupported call through a literal pointer on float power broadcast (#349)
• Using CuArrays with complex arrays (#347)
• CuArrays test fail: an illegal memory access (#342)
• Fail to precompile CuArrays (#329)
• Wrong NNlib bounds wiith CuArrays 1.0.2 (#327)
• Register v1.0.2 (#326)
• RFC: Rename CuArrays.CURAND.curand -> CuArrays.CURAND.rand (#310)
• reverse(::CuArray) (#299)
• Support for cufftSetStream() (#256)
• Package is not loadable if e.g. CUBLAS cannot be opened (#255)
• Mimic Base exports (#224)
• Reduction to a tuple (or any other complex type) (#28)

Merged pull requests:

• Add 1.2 testing. (#370) (maleadt)
• Fix ND GPUArrays.LocalMemory allocation. (#369) (maleadt)
• Allow combined use of SubArray and PermutedDimsArray (#368) (maleadt)
• Add dims support to similar for broadcasted arrays (#358) (lcw)
• remove redundant global annotation on TimerOutput in function (#357) (KristofferC)
• Various small tests (#353) (kshyatt)
• More tests for BLAS and improve error msg (#352) (kshyatt)
• precompile is the default now. (#337) (maleadt)
• Use more efficient / nondeprecated array ctors. (#335) (maleadt)
• Support for reversing arrays. (#334) (maleadt)
• Wrap cufftSetStream. (#332) (maleadt)
• Don't export cu-prefixed functions, module prefix instead. (#331) (maleadt)
• Always require the libraries that are part of CUDA. (#328) (maleadt)
• broadcast for ^ (#325) (chengchingwen)

v1.0.2

Use released versions of dependencies.

v1.0.1

Hotfix: add missing dependencies to REQUIRE

v1.0.0

Minor release:

• adapt to CuPtr changes
• performance improvements to pool allocator, and unsafe_free! to bypass the GC
• initial support for logical indexing, filter and accumulate by @dpsanders
• initial low-level wrappers for cublasxt by @kshyatt

v0.9.1

Bug fix release:

• fix loading the package when build failed
• fix CuArrays.@sync not properly synchronizing
• fix svd wrappers and high level interface (@andreasnoack)
• improve broadcast rewrite of custom functions (@chengchingwen)

v0.9.0

Major release, some highlights:

• use Adapt.jl for better support of array wrappers
• make CURAND wrappers behave more like Base
• wrap CUSPARSE

v0.8.1

Deprecate old module names. (#159)