v3.3.5

CUDA v3.3.5

Diff since v3.3.4

Closed issues:

• Integer division error for the product of sparse times empty matrices (#962)
• Bad conversion from QR to CuArray (#969)
• Errors during installation test (#1004)
• Be explicit about imports (#1028)
• Exponentiation with constants can produce bad GPU code compared to the CPU (#1031)
• rem uses wrong intrinsic (#1040)
• test Cuda fails on gpuarrays\reductions/minimum maximum (#1043)
• Broadcasted type conversion on literal value doesn't work (#1044)
• CUDA overrides somehow screwing up customized printing? (#1055)
• Is it possible to copy any data into GPU via recursive CuDeviceArray construction? (#1057)
• CUDA doesn't compile after upgrade to Julia 1.6.2 (#1065)
• Timing discrepancy between CUDA.@time and Benchmarktools for Flux model (#1067)
• cannot convert range to Curray (#1070)
• Thread safety issue with gemv! (#1072)
• CuSparseMatrixCSC conversion errors (#1075)
• cublasHgemmStridedBatched (#1076)
• ERROR: UndefKeywordError: keyword argument elements not assigned (#1077)
• Support for generating Float16 random numbers (#1081)
• Illegal memory access during complex exponential with large imaginary part as exponent (#1085)
• "Error: CUDA.jl does not yet support CUDA with ptxas 11.3.109" when using "JULIA_CUDA_USE_BINARYBUILDER=false" (#1089)

Merged pull requests:

• Add support for unified arrays. (#1023) (@maleadt)
• Look for libcuda in more places. (#1030) (@maleadt)
• Detect common integer exponentiations and handle them directly. (#1033) (@maleadt)
• Allow strided inputs to various library functions. (#1038) (@maleadt)
• Use correct intrinsics for rem (#1041) (@simonbyrne)
• update Package Manager link (#1052) (@ehgus)
• Update manifest (#1054) (@github-actions[bot])
• Add test for math_mode (#1056) (@kshyatt)
• Streamline atomics. (#1059) (@maleadt)
• Add support for device capability-dependent code. (#1060) (@maleadt)
• Adapt to GPUArrays changes. (#1061) (@maleadt)
• Add special constructors to work around Base AbstractQ size weirdness. (#1063) (@maleadt)
• Update manifest (#1064) (@github-actions[bot])
• Small allocator improvements (#1068) (@maleadt)
• Latency improvements (bis) (#1069) (@maleadt)
• lib: cusparse: fix #962 (#1073) (@thazhemadam)
• Make handle cache thread-safe. (#1074) (@maleadt)
• Bump GPUCompiler. (#1079) (@maleadt)
• add support for half-precision gemm (#1080) (@bjarthur)
• Extend and switch to the new CUDA RNG (#1082) (@maleadt)
• cusparse: fix conversion from sparse matrix to dense matrix (#1083) (@maleadt)
• Support/bump for CUDA 11.4.1 and CUDNN 8.2.2 (#1084) (@maleadt)
• Use sincos from libdevice to perform illegal global load. (#1086) (@maleadt)
• Bump GPUCompiler; use our own opt pipeline. (#1087) (@maleadt)
• Update manifest (#1090) (@github-actions[bot])
• Backports for 3.3.5 (#1091) (@maleadt)

v3.3.4

CUDA v3.3.4

Diff since v3.3.3

Closed issues:

• Cholesky on 1.8 doesn't dispatch correctly (#1046)

Merged pull requests:

• restore lost tests (#1042) (@vchuravy)
• Base.unsafe_lenght is deprecated on 1.8 (#1045) (@vchuravy)
• Update manifest (#1048) (@github-actions[bot])
• Fix cholesky on 1.8, fix #1046 (#1049) (@kshyatt)
• Backport changes for 3.3.4 (#1050) (@vchuravy)

v3.3.3

CUDA v3.3.3

Diff since v3.3.2

Merged pull requests:

• Adapt to LLVM changes. (#1022) (@maleadt)
• Update manifest (#1029) (@github-actions[bot])
• just some simple printing tests (#1032) (@kshyatt)
• Test for is_capturing (#1034) (@kshyatt)
• Tests for buffer printing (#1035) (@kshyatt)
• Make it possible to change the pool alloc and handle types. (#1036) (@maleadt)
• Backports for 3.3 (#1037) (@maleadt)

v3.3.2

CUDA v3.3.2

Diff since v3.3.1

Closed issues:

• Missing artifacts errors (#1003)
• Relax restriction on types allowed in kernels? (#1005)
• PPC: Atomic{Float64} is not supported (#1008)
• Unexpected result in combination with Zygote.gradient() (#1019)
• Both ExprTools and LLVM export "parameters"; uses of it in module CUDA must be qualified (#1025)

Merged pull requests:

• Fixes for artifact loading. (#1006) (@maleadt)
• dlopen CUBLAS before CUTENSOR. (#1007) (@maleadt)
• Use a plain integer to keep track of pool last use time. (#1009) (@maleadt)
• More fixes to artifact discovery. (#1010) (@maleadt)
• add custom structs tutorial (#1011) (@jw3126)
• big mapreduce performance (#1012) (@xaellison)
• Fixes for Julia 1.7 (#1013) (@maleadt)
• Update manifest (#1014) (@github-actions[bot])
• Remove memory pools (#1015) (@maleadt)
• Move refcounting to an array storage type (#1016) (@maleadt)
• Remove unneeded disambiguation method. (#1017) (@maleadt)
• Simplify context validity check. (#1018) (@maleadt)
• Improve LazyInitialized (#1020) (@maleadt)
• More allocator clean-ups (#1021) (@maleadt)
• CUDA 11.4 (#1024) (@maleadt)
• Only import from ExprTools what we need. (#1026) (@maleadt)
• Backports release 3.3 (#1027) (@maleadt)

v3.3.1

CUDA v3.3.1

Diff since v3.3.0

Closed issues:

• Reclaim with stream-ordered allocator (#952)
• possible hanging with CUDA.@profile? (#961)
• Upgrading from v3.2.1 to v3.3.0 broke my installation (#970)
• Calls to has_cudnn running on wrong CuDevice? (#978)
• Test does not run on MIT Supercloud after upgrading to 3.3.0 (#980)
• Performance issue with complicated loops in function (#984)
• Is it possible to set cache config in CUDA.jl? (#988)
• @atomic should perform type conversions (#989)
• Compatible NVIDIA driver but still got compatibility warning (#1001)

Merged pull requests:

• Update manifest (#971) (@github-actions[bot])
• Fix disambiguation of CUDA 11.1 using CUSOLVER. (#972) (@maleadt)
• Simplify initialization helper macro. (#973) (@maleadt)
• Move at-typed_ccall to LLVM.jl. (#976) (@maleadt)
• Replace workspace macro with function (#981) (@maleadt)
• Implement and improve reclaim for the stream-ordered allocator (#983) (@maleadt)
• Bump GPUCompiler to fix WMMA test issue. (#985) (@maleadt)
• Rework memoization (#986) (@maleadt)
• Fixes for CUBLAS/CUDNN logging (#987) (@maleadt)
• Perform type conversions in at-atomic. (#990) (@maleadt)
• Don't initialize the API when setting log callbacks. (#992) (@maleadt)
• Create a helper for lazy, thread-safe initialization. (#993) (@maleadt)
• Optimize library handles (#996) (@maleadt)
• Optimize PerDevice for abstract element types. (#997) (@maleadt)
• Update manifest (#999) (@github-actions[bot])
• Replace PerDevice with context-keyed dictionaries. (#1000) (@maleadt)
• Improve launch latency (#1002) (@maleadt)

v3.3.0

CUDA v3.3.0

Diff since v3.2.1

Closed issues:

• PTX code missing DWARF debug information (#72)
• Suggestion - Disable AbstractArray indexing fallback by default (#178)
• Support isbits Union Arrays (#103)
• Missing norm(x, p) kernel (#84)
• CUDA enhanced compatibility (#832)
• Support for CuSparseMatrixCSC{Float16} x CuVector{Float16} (#849)
• CuArray to zeroth power returns Matrix (#897)
• Fatal errors during sorting tests (#916)
• Error when computing reductions into a view with reduce_blocks > 1 (#919)
• CUDA FFT plan application runs Out of Memory in Pluto (#926)
• has_cuda() errors in CPU-only environments on master (#928)
• Race condition when computing mean! of large arrays? (#929)
• Supporting union bits types (#934)
• test failing in device/intrinsics (#942)
• Memory allocation fails for multi-GPU (#943)
• Scalar operations when using output of cu(::OffsetArray) (#954)
• Quicksort kernel does not cope with reduced threads (#955)
• CUDA.jl cannot find installed CUPTI libraries with local installation on linux (#956)
• Error for complex sparse-dense Matrix-vector multiplication (#958)
• "using CUDA" gives error in type inference of Ref{Bool} (#965)

Merged pull requests:

• Override outlined throw functions. (#874) (@maleadt)
• Enable location and debug info. (#891) (@maleadt)
• Compile using the toolkit, not the driver. (#892) (@maleadt)
• Rework timings (#898) (@maleadt)
• Fix #849, allow CUSPARSE to use F16 (#904) (@kshyatt)
• Add Windows CI. (#907) (@maleadt)
• Split test for better parallelization. (#908) (@maleadt)
• Update manifest (#909) (@github-actions[bot])
• Improve package latency. (#910) (@maleadt)
• Just some missing tests for CUBLAS (#911) (@kshyatt)
• Fix bug and add tests for iamax/iamin (#913) (@kshyatt)
• Fix profiler initialization and exception handling. (#914) (@maleadt)
• Add a show method for devices(). (#915) (@maleadt)
• Fix update of CUFFT handle. (#921) (@maleadt)
• Update manifest (#922) (@github-actions[bot])
• Reinstate compatibility with Kepler GPUs. (#923) (@maleadt)
• Use multiple GPUs on CI when available. (#924) (@maleadt)
• Fix two-step mapreduce with wrapped output. (#925) (@maleadt)
• Eagerly free the CUFFT workspace when generating a new one. (#927) (@maleadt)
• Fix CUDA.function without throwing. (#930) (@maleadt)
• Fix the REPL synchronization hook. (#931) (@maleadt)
• Re-initialize the random seed every time. (#932) (@maleadt)
• Protect against race in iterating compute processes. (#933) (@maleadt)
• Helper function to get the device given a cu ptr. (#935) (@akashkgarg)
• Implement CUDA's Enhanced Compatibility when selecting a toolkit. (#936) (@maleadt)
• Update manifest (#939) (@github-actions[bot])
• Re-introduce specialization of cufunction. (#940) (@maleadt)
• Support isbits union element types with CuArray. (#941) (@maleadt)
• Try generating code with unreachable control flow. (#944) (@maleadt)
• Upgrade to CUDA 11.3 Update 1. (#945) (@maleadt)
• Always use exit instead of trap. (#947) (@maleadt)
• Select devices without NVML. (#948) (@maleadt)
• Fixes for Julia 1.7. (#949) (@maleadt)
• Query the CUBLAS version without requiring a handle. (#951) (@maleadt)
• Improve CUBLAS and CUDNN logging. (#953) (@maleadt)
• Update manifest (#957) (@github-actions[bot])
• Enable sorting with reduced block sizes (#959) (@xaellison)
• Adapt to GPUCompiler changes, bump GPUArrays. (#963) (@maleadt)
• Adapt to change in allowscalar. (#964) (@maleadt)
• Don't disable the CUDNN log callback on Windows. (#966) (@maleadt)
• Use released dependencies. (#968) (@maleadt)

v3.2.1

CUDA v3.2.1

Diff since v3.2.0

Closed issues:

• adding constant to an array: performance regression compared to CUDAdrv (#838)
• CUDA.abs() on vector input: performance regression compared to CUDAdrv (#839)
• CUDA.@sync seems to be using a lot of CPU while waiting (#893)
• Memory leaks with repeated use of fft of a CUDA Array (#894)
• CUDA.jl v3.2 seems to download wrong version of CUDNN and CUTENSOR (#899)

Merged pull requests:

• Rework synchronization: first spin, then yield, and finally block. (#896) (@maleadt)
• Make cusolvermg really optional. (#900) (@maleadt)
• Rebuild artifacts. (#901) (@maleadt)
• Take back control over the CUFFT work area. (#902) (@maleadt)

v3.2.0

CUDA v3.2.0

Diff since v3.1.0

Closed issues:

• Explore CUDA graph API (#65)
• Runtime functions are missing debug information (#53)
• Native RNGs do not pass SmallCrush (#803)
• Remaining threads/FFT/mult-gpu error (#876)

Merged pull requests:

• Add wrappers for the CUDA graph API. (#877) (@maleadt)
• Use the profiler API to start capture. (#878) (@maleadt)
• Duplicate RNG state across block to avoid need for synchronization (#879) (@maleadt)
• Support for printing tuples. (#880) (@maleadt)
• Support unsigned inputs to integer intrinsics. (#881) (@maleadt)
• Switch to Philox2x32 for device-side RNG (#882) (@maleadt)
• Update manifest (#884) (@github-actions[bot])
• Treat CartesianIndices in views as scalars. (#886) (@maleadt)
• Robustly get variables from the environment during init. (#887) (@maleadt)
• Move Statistics functionality to GPUArrays. (#888) (@maleadt)
• Update artifacts and use sources from unified JLLs. (#889) (@maleadt)
• Lazy initialization of CUDNN and CUTENSOR (#890) (@maleadt)
• Update manifest (#895) (@github-actions[bot])

v3.1.0

CUDA v3.1.0

Diff since v3.0.3

Closed issues:

• GPU Implementation of partialsort! (#93)
• Document associativity requirements of scan/reduce operators (#819)
• Problem in reduce_block? (#843)
• CUDNN convolution incorrect for small images (#848)
• Newly-spawned tasks should re-set the device (#851)
• sort!(CUDA.zeros(2^25)) throws invalid configuration argument (code 9, cudaErrorInvalidConfiguration) (#852)
• Type-preserving upload about cu in doc may be wrong (#855)
• Memory corruption / segfault with Threads.@async and planned FFTs (#859)
• Don't call nvmlErrorString (during init?) to prevent crashes on WSL (#860)
• unsafe_copy3d! does not work with stream-ordered allocations (#863)
• CUDA3 seems to have memory leak (#866)

Merged pull requests:

• Implement statistics functions: correlation and covariance (#509) (@berquist)
• @atomic support * and / (#842) (@yuehhua)
• CUDNN docstring revisions. (#844) (@GunnarFarneback)
• Sorting perf (again) (#845) (@xaellison)
• Update manifest (#846) (@github-actions[bot])
• Remove extraneous apostrophe (#847) (@kshyatt)
• reduce_block fixes. (#853) (@maleadt)
• Fix sorting large arrays. (#854) (@maleadt)
• Remove unsupported config launch keyword. (#856) (@maleadt)
• Identify the buffer during unsafe_wrap to support unified free. (#857) (@maleadt)
• Add support for CUDA 11.3. (#858) (@maleadt)
• Work around buggy NVML initialization on WSL (#861) (@maleadt)
• ae/partialsort (#864) (@xaellison)
• Update manifest (#865) (@github-actions[bot])
• Improve multitasking with CUFFT. (#867) (@maleadt)
• Introduce a HandleCache type. (#868) (@maleadt)
• Improve multitasking with CURAND (#869) (@maleadt)
• Document associativity requirement of accumulate (#870) (@HenriDeh)
• Half-Precision Intrinsics (#871) (@iyaja)
• Work around offset calculation bug in cuMemcpy3DAsync. (#872) (@maleadt)
• fix #848: CUDNN convolution incorrect for small images (#873) (@denizyuret)

v3.0.3

CUDA v3.0.3

Diff since v3.0.2

Closed issues:

• CUDA.jl init error in the REPL without using a CUDA feature (#841)

Merged pull requests:

• Only synchronize the REPL when CUDA is configured. (#840) (@maleadt)