v2.2.2

CuArrays v2.2.2

Diff since v2.2.1

v2.2.1

CuArrays v2.2.1

Diff since v2.2.0

Closed issues:

• Implement aliasing API (#133)
• CURAND_STATUS_PREEXISTING_FAILURE with v2.0.1 but not v1.7.3 (#682)

Merged pull requests:

• findfirst should return nothing if no matching element is found (#712) (@bsmiles32)
• Update manifest (#715) (@github-actions[bot])
• fix ambiguity for Random.seed!(rng, nothing) (#717) (@marius311)
• Actually put the workspace buffer back in the memory pool. (#718) (@maleadt)
• CompatHelper: bump compat for "CEnum" to "0.4" (#720) (@github-actions[bot])
• Update manifest (#723) (@github-actions[bot])

v2.2.0

CuArrays v2.2.0

Diff since v2.1.0

Closed issues:

• Kernel optimizations (#445)
• mul! failing for adjoint sparse matrices (#629)
• Deadlock during memory free (#685)
• Training Halts when Using CuArrarys (#691)
• CUBLAS initialization (#697)
• Performance issue with v2.1.0 compared with v1.7.3 (#701)

Merged pull requests:

• Implement a few sparse array basics (#572) (@janEbert)
• fix 1d convolution (#690) (@AStupidBear)
• Fix sparse mul! (#692) (@amontoison)
• Disable the GC after taking pool-related spinlocks. (#693) (@maleadt)
• Update manifest (#694) (@github-actions[bot])
• adds wrappers for syevjBatched/heevjBatched family of CUSOLVER functions (#695) (@electronsandstuff)
• Update manifest (#702) (@github-actions[bot])
• Repopulate the pool from freed blocks before allocating. (#704) (@maleadt)
• CEnum 0.3 compatibility. (#705) (@maleadt)
• Add direct BLAS calls for trmm! and trsm! (#707) (@aterenin)
• Add mul! for trmm! and tests. (#708) (@aterenin)
• Support and use broadcast with mapreduce. (#709) (@maleadt)
• Protect against recursive failures during init. (#710) (@maleadt)
• Perform threaded test in lock. (#711) (@maleadt)

v2.1.0

CuArrays v2.1.0

Diff since v2.0.1

Closed issues:

• Performance regression with mapreduce (#611)
• cufunc wrappers of NNlib activation functions (#614)
• rand(UInt32) gives low-quality results (#655)
• With CuArrays 2.0, multidimensional circshift fails for integer multiples of array size shift in one or more dimensions (#657)
• similar(PermutedDimsArray(::CuArray)) isa Array (#658)
• In CuArrays v2.0, GPU operation takes hours to run for the first time (#660)
• sum!(y::CuVector, x::CuMatrix) throws InvalidIRError error (#661)
• Where can I find (#667)
• Where can I find All the using instructions of CuArrays? (#668)
• add implicit float conversion to math functions (#671)
• Multiplication between mixed types doesn't drop leading dimensions (#673)
• Very slow 4D broadcast in 2.0.1 (#677)
• Sum function is slow (#679)
• Indexing CuArrays with Empty Ranges Errors (#687)
• Sum, any, etc. with function is no longer implemented (#688)

Merged pull requests:

• Actually test error throwing for lsvqr (#604) (@kshyatt)
• Wrapper functions for NNlib (#615) (@matsueushi)
• Improve mapreduce performance (#646) (@wongalvis14)
• Improved error output during finalizer exceptions. (#651) (@maleadt)
• Fix usage discrepancies (#654) (@maleadt)
• Add support for uniform random UInt32 (#656) (@chrstphrbrns)
• Update manifest (#659) (@github-actions[bot])
• Implement GPUArrays vararg mapreduce. (#663) (@maleadt)
• Test for implicit singleton dims with mapreducedim. (#665) (@maleadt)
• Update manifest (#674) (@github-actions[bot])
• Allow scalar indexing where necessary and add a few tests (#675) (@kshyatt)
• Update manifest (#681) (@github-actions[bot])
• Update manifest (#686) (@github-actions[bot])

v2.0.1

CuArrays v2.0.1

Diff since v2.0.0

Merged pull requests:

• Fix and test initialization without CUDA. (#650) (@maleadt)

v2.0.0

CuArrays v2.0.0

Diff since v1.7.3

Closed issues:

• Statistics functions (var,std) slow (#411)
• BinnedPool: key error (#520)

Merged pull requests:

• implement CuIterator for batching arrays to the GPU (#467) (@jrevels)
• Specialized _var and _std functions (#612) (@merckxiaan)
• NNlib batched_mul! (#619) (@mcabbott)
• Clean-up init code. (#620) (@maleadt)
• Thread safety fixes. (#621) (@maleadt)
• Update wrappers. (#622) (@maleadt)
• Use a library handle for each thread/device. (#623) (@maleadt)
• Better error upon use of missing libraries. (#625) (@maleadt)
• Lock memory allocator operations (#626) (@maleadt)
• Thread safety of memory allocator (#627) (@maleadt)
• Use a return value-based retry scheme for all APIs failing to allocate. (#633) (@maleadt)
• Provide 5-arg mul! (#634) (@haampie)
• Update manifest (#635) (@github-actions[bot])
• Add BLAS.axpby! (#636) (@amontoison)
• Let cublasXT use host memory (#639) (@kshyatt)
• Test Julia 1.4. (#640) (@maleadt)
• 5-arg mul! with 3-arg mul! support (#641) (@haampie)
• Avoid multiple mapreduce kernel launches (#642) (@maleadt)
• GC.preserve some arrays when unsafely accessing them. (#644) (@maleadt)
• Support for Julia's multitasking. (#645) (@maleadt)
• Don't depwarn if new env vars are used. (#648) (@maleadt)
• Avoid deadlock in debug mode. (#649) (@maleadt)

v1.7.3

v1.7.3 (2020-03-06)

Diff since v1.7.2

Closed issues:

• CI doesn't test old CUDA versions (#607)
• findmax slow (#606)
• import CuArrays always fails with CUDA 10.2.89 (but works fine with CUDA 10.0.130 and 10.1.105) (#601)
• Kernel exception in findfirst (#595)
• Test failures on Tesla K20c/CUDA 9.0 (#594)
• Linear indexing into arrays with two or more dimensions throws an obscure error during gradient calculation with Zygote (#590)
• Make CURAND seeding more consistent with Base (#589)
• mapreduce (sum, prod, etc.) fail in some cases when given a dims argument. (#583)
• Whitelist hypot (#442)
• polish mapreduce interface (#204)
• Use of mapreduce by Flux doesn't rewrite intrinsics (#154)
• Slow mapreduce compared to KnetArray (#141)
• Support for missing values (#125)

v1.7.2

v1.7.2 (2020-02-11)

Diff since v1.7.1

Closed issues:

• CuArrays looks for the wrong version of cuDNN on windows10 (#587)
• Gradient of the einsum not working with CuArrays, when it’s output is assigned to the field of a mutable struct (#586)

v1.7.1

v1.7.1 (2020-02-07)

Diff since v1.7.0

Closed issues:

• Only show debug timings during tests on CI (#573)
• Broadcasted setindex! triggers scalar setindex! (#571)
• ODE performance benchmark (#566)
• Invalid IR for reductions along trivial dimension (#542)
• cublas error with "--math-mode=fast" (#510)
• ArgumentError when displaying a view of CuArray (#506)
• Serial mapreduce kernel broken on -g2 (#418)
• Functions not available in CUDA 8.0 (#351)
• Rexport functions from AbstractFFTs (#341)

Merged pull requests:

• Exclude some large kernels from coverage. (#568) (maleadt)
• Fix CUBLAS math mode under fast math (#510). (#567) (maleadt)

v1.7.0

v1.7.0 (2020-01-21)

Diff since v1.6.0

Closed issues:

• Update dependecies (#562)
• Non-linear performance at larger batch sizes. (#555)
• Wrong return type for reductions of complex CuArrays (#550)
• ReshapedArray causes scalar indexing (#548)
• resize! broken (#547)
• exp.\(x\) is not type stable when x isa CuArray{Complex{Float32}} (#543)
• zero-size alloc before initialization errors (#538)
• CUDNN compatibility (#536)
• Error message on invalid adapt (#365)

Merged pull requests:

• CUTENSOR coverage fixes. (#565) (maleadt)
• Basic tests for CuTensor type (#564) (kshyatt)
• Fix and test seed methods. (#561) (maleadt)
• Revert to simpler GC.gc API. (#560) (maleadt)
• Preserve the pooled property when deriving a CuArray. (#558) (maleadt)
• Implement array resizing. (#557) (maleadt)
• Fix SplittingPool reclaim. (#556) (maleadt)
• Improve error display. (#554) (maleadt)
• Warn about version mismatches. (#553) (maleadt)
• Throw an error for unsupported element types (#552) (maleadt)
• Use the CUDAnative context getter. (#551) (maleadt)
• Add a test for #543. (#546) (maleadt)
• Avoid exponential of positive numbers in softplus implementation (#518) (vargonis)