v2.6.2

CUDA v2.6.2

Diff since v2.6.1

Closed issues:

• High allocations and getindex (#150)
• ResNet spending much time in CuArrays GC (#149)
• Broadcast inference failure results in scalar iteration (#145)
• Allocator very slow to reclaim memory after running for sufficiently long (#137)
• Assignment using logical indexing (#131)
• CUDNN convolution allocates outside of the memory pool (#111)
• Logical indexing per-dim (#106)
• Threading-related assertion failure in split allocator (#97)
• dims support for softmax (#226)
• Memory pinning needs more features (#242)
• External allocations fail under high memory pressure (#340)
• Incomplete CUDNN wrappers (#343)
• softmax(x) and logsoftmax(x) update their arguments (#592)
• Freeing large buffers takes a while (#594)
• softmax has problem with dim parameter (#599)
• gemmEx on sm_52 results in CUBLAS_STATUS_ARCH_MISMATCH (#609)
• about the document of conditional use (#689)
• GPU run out of memory if 2 workers use the same GPU (#692)
• CURAND handles are collected early (#699)
• cudnnConvolutionForward fails memory checking (#702)
• Deadlock during OOM (#706)
• Segfault during trampoline allocation when querying occupancy from multiple threads (#707)
• Ballot intrinsics should use .sync variety (#711)
• cfunction $shmem_cint use after free (#713)
• OOM when evaluating a small resnet (with both Flux and Knet) (#714)
• Supprt CUDA 11.2 Update 1 (#715)
• Base.mapreducedim returns wrong answer with non-zero target array (#720)
• CUBLAS_STATUS_ARCH_MISMATCH (#722)
• Test failures on linux (#727)
• Switching devices causes GC errors (#731)
• Pin CPU buffers when doing memory copies (#735)
• Memory free error with CUDA 11.2 and multi threads/GPUs (#737)
• Per-device memory pool (#742)
• Could not load library cudnn_ops_infer64_8.dll (#757)
• CUDA.lgamma(x) crashes Julia (#758)

Merged pull requests:

• New high level interface for cuDNN (#523) (@denizyuret)
• bilinear upsampling (#636) (@maxfreu)
• Use CUDA 11.2's stream-ordered allocator (#679) (@maleadt)
• Support an additional nvdisasm version. (#680) (@maleadt)
• Add fast getri_strided_batch (#682) (@cfranken)
• Fix race during multi-threaded init. (#687) (@maleadt)
• Update manifest (#690) (@github-actions[bot])
• Change to Buildkite v1 plugins. (#691) (@maleadt)
• CUPTI improvements for multithreading (#693) (@maleadt)
• Fix exception flag linkage for linking. (#694) (@maleadt)
• Update manifest (#695) (@github-actions[bot])
• Use simpler try/catch in show(CuError). (#696) (@maleadt)
• fix bug in CURAND.jl's set_stream function. (#698) (@norci)
• Update CUDNN to 8.1. (#701) (@maleadt)
• Remove special-cased algorithm selection for CUDNN convolution (#703) (@denizyuret)
• Keep track of active handles to avoid early collection. (#704) (@maleadt)
• Backports for Julia 1.5 (#705) (@maleadt)
• Support for cushow-ing multiple values, including LLVMPtrs. (#709) (@maleadt)
• Make CUDNN tests eagerly invoke at-test for better error reporting. (#710) (@maleadt)
• Report JIT error log with linker errors. (#712) (@maleadt)
• Update manifest (#716) (@github-actions[bot])
• Flip exception_flag filter! predicate (#717) (@S-D-R)
• Keep some memory reserved for external allocations. (#718) (@maleadt)
• Upgrade CUDA 11.2 to Update 1. (#719) (@maleadt)
• Add Abstract FFT compat (#721) (@DhairyaLGandhi)
• Add support for and switch test to warp-synchrnous vote intrinsics. (#723) (@maleadt)
• Specialize Base.to_index for AnyCuArray{Bool} (#724) (@pabloferz)
• Update manifest (#728) (@github-actions[bot])
• Eagerly dlopen cublasLt to prevent a system library getting picked up. (#729) (@maleadt)
• Switch tests over to compute-sanitizer. (#730) (@maleadt)
• Perform pool operations in the correct context. (#732) (@maleadt)
• Streamline use of retry_reclaim. (#733) (@maleadt)
• Threading fixes (#734) (@maleadt)
• copied the old rnn.jl->rnncompat.jl for Flux compatibility (#738) (@denizyuret)
• fix testmode batchnorm back (#739) (@CarloLucibello)
• Don't error out if failing to parse the local CUDA version. (#740) (@maleadt)
• Backport #739 (#741) (@maleadt)
• Add back an older artifact for CUDNN on PPC with CUDA 10.2. (#743) (@maleadt)
• Use the default memory pool. (#745) (@maleadt)
• Use a memory pool per device. (#746) (@maleadt)
• Test sort with at-test at the toplevel, for better reporting. (#749) (@maleadt)
• remove NNlib (#753) (@CarloLucibello)
• Update manifest (#754) (@github-actions[bot])
• Fix cuda-memcheck, don't use memory pools. (#756) (@maleadt)
• Update generated wrappers (#759) (@maleadt)
• Rework memory pinning and speed up async ops on unpinned memory (#760) (@maleadt)
• Improve context switching (#761) (@maleadt)
• Update manifest (#765) (@github-actions[bot])
• Docs on multitasking (#766) (@maleadt)
• Update to CUDA 11.2 Update 2. (#768) (@maleadt)
• Small backports for CUDA 2.4 / Julia 1.5 (#770) (@maleadt)
• Backports for CUDA 2.6 / Julia 1.6 (#771) (@maleadt)