v1.6.0

v1.6.0 (2019-12-21)

Diff since v1.5.0

Closed issues:

• cu(x) always returns Float32 (#540)
• CUDNN Batchnorm fails (#531)
• Overhead of memory copies (#528)
• CUFFT_EXEC_FAILED when using Zygote (#524)
• Tutorial initial benchmark times/text don't make sense (#519)
• SplittingPool: assertion violated (#516)
• Split sublibraries into pure library wrappers and higher-level wrappers (#287)
• Require CUDA libraries unconditionally (#221)
• Intermittent CI failure: CUSPARSE switch2csr (#215)

Merged pull requests:

• Rework library handles for multithreading. (#544) (maleadt)
• Overhaul documentation and README. (#541) (maleadt)
• CompatHelper: bump compat for "Requires" to "1.0" (#535) (github-actions[bot])
• reenable complex tensor contraction tests (#533) (Jutho)
• CI improvements: ARM, and test several CUDA versions (#532) (maleadt)
• Clean-up copy constructors. (#530) (maleadt)
• Use 128 bits to represent blocks in the splitting memory pool. (#527) (maleadt)
• Take available memory into account when selecting a device. (#526) (maleadt)
• RFC: Allow cu to operate on Array types (#522) (baggepinnen)
• Upgrade Literate and Documenter. (#511) (fredrikekre)

v1.5.0

v1.5.0 (2019-11-29)

Diff since v1.4.7

Closed issues:

• Accessing CuArrays and CUDAnative through Flux gives WARNING: replacing module CUDA. and breaks precompilation (#509)
• 64-bit windows LoadError: Could not find libcublas (#507)
• GC preserve in generated ccall (#503)
• Conditional use of CUDA on Windows (#488)
• Memory allocator: @external_alloc (#426)
• setindex! fails with Array source CuArray target with different ndims or when broadcasting (#403)

Merged pull requests:

• Allocate workspace repeatedly right before calling into CUDNN. (#517) (maleadt)
• Use at-runtime_ccall's ability to delay the library lookup. (#514) (maleadt)
• Demote CUTENSOR version check to a warning. (#513) (maleadt)
• Don't always time allocations. (#512) (maleadt)
• API for manually reclaiming memory (#504) (maleadt)
• Updates for CUTENSOR 1.0 (#502) (maleadt)
• Add ND support for find functions. (#500) (maleadt)

v1.4.7

v1.4.7 (2019-11-15)

Diff since v1.4.6

Closed issues:

• CuArrays v1.4.6 fails to load, but v1.4.4 works fine (#498)
• wrong result on argmax (#304)

Merged pull requests:

• Don't needlessly populate the DL_LOAD_PATH. (#501) (maleadt)
• Extend findmin/findmax methods to cover argmin/argmax use. (#499) (maleadt)

v1.4.6

v1.4.6 (2019-11-15)

Diff since v1.4.5

Closed issues:

• Flux.jl works with CuArrays 1.4.4, but not 1.4.5 (#496)

Merged pull requests:

• Mark the package as initialized. (#497) (maleadt)

v1.4.5

v1.4.5 (2019-11-14)

Diff since v1.4.4

Closed issues:

• CuArrays 1.4.3 fails to initialize (#482)

Merged pull requests:

• Add some find functions (#493) (maleadt)
• Print a message if the package fails to load. (#489) (maleadt)
• Use CUDAapi.at-runtime_ccall for calling into potentially missing libraries (#486) (maleadt)
• CompatHelper: bump compat for "AbstractFFTs" to "0.5" (#484) (github-actions[bot])

v1.4.4

v1.4.4 (2019-11-08)

Diff since v1.4.3

Merged pull requests:

• Even more package initialization fixes (#483) (maleadt)

v1.4.3

Improve conditional use of the package.

v1.4.2

v1.4.2 (2019-11-06)

Diff since v1.4.1

Closed issues:

• performance regression (#478)
• FFTW vs AbstractFFTs (#419)

Merged pull requests:

• Fix full collection heuristic of binned allocator. (#480) (maleadt)
• Use the GPUArrays testsuite outside of its sources (#479) (maleadt)
• Delayed initialization using Base.inferencebarrier. (#477) (maleadt)
• Reinstate Julia 1.0 compatibility (#476) (maleadt)

v1.4.1

v1.4.1 (2019-11-05)

Diff since v1.4.0

Closed issues:

• CuArrays 1.4.0 fails to initialize on windows (#473)
• getrf! issue (#471)
• Bump package version more often? (#469)

Merged pull requests:

• Use CUDAapi-discovered paths to ccall CUDA libraries. (#475) (maleadt)
• Test fixes for 32-bit BLAS. (#474) (maleadt)

v1.4.0

v1.4.0 (2019-11-04)

Diff since v1.3.0

Merged pull requests:

• Use released GPUArrays. (#470) (maleadt)
• Init improvements (#468) (maleadt)
• Works towards precompilation (#466) (maleadt)
• Fix #463 (#464) (xukai92)
• Rename CuArray.own .pool to avoid confusion with unsafe_wrap. (#462) (maleadt)
• Add a strided batched method for getrf (#461) (maleadt)
• Improve type constraint for \ (#460) (xukai92)
• Implement matrix division using CUSOLVER (#459) (xukai92)
• CI and wrapper updates (#458) (maleadt)
• Adapt to CUDAdrv.jl#167 (#457) (maleadt)
• Prevent CPU to GPU copy in mapreduce. (#455) (maleadt)
• CI improvements. (#454) (maleadt)
• Upgrade to new CI templates. (#452) (maleadt)
• Accumulate improvements (#448) (maleadt)
• Optimize accumulate (#447) (maleadt)
• Implement findall (#446) (maleadt)
• More memory allocator improvements (#441) (maleadt)
• remove extraneous where and add some tests (#439) (kshyatt)
• Use a separate allocation timer. (#437) (maleadt)
• Use the memory pool for CUFFT workarea allocation. (#436) (maleadt)
• Update .gitlab-ci.yml (#435) (MikeInnes)
• CUDNN improvements (#404) (maleadt)