[dependabot]

Updates the requirements on CUDACore to permit the latest version.

Release notes

Sourced from CUDACore's releases.

v6.4.2

CUDA v6.4.2

Diff since v6.4.1

Merged pull requests:

• Keep wrapped CPU arrays alive in unsafe_wrap(CuArray, ...) (#3308) (@​maleadt)
• Detect aliasing between contiguous views and wrapped arrays (#3312) (@​maleadt)
• Even more cooperative groups intrinsics tests (#3313) (@​kshyatt)
• Use GPUToolbox's cooperative_wait for synchronization (#3315) (@​maleadt)
• Fall back from cuBLAS in storage-level mul! where cuBLAS can't go (#3316) (@​aalquwayfili)
• Upgrade to LLVM.jl 10 (#3323) (@​maleadt)
• CI: run the test suite with 2 CPU threads again (#3324) (@​maleadt)

Closed issues:

• mightalias misses overlap between an offset CuArray view and a SubArray of the same buffer (#3286)

Changelog

Sourced from CUDACore's changelog.

CUDA.jl release history

This document lists the noteworthy changes in every release of CUDA.jl, newest first. It is a hand-written summary aimed at users; for the complete list of merged pull requests, see the GitHub releases.

CUDA.jl follows semantic versioning, so intentional breaking changes are reserved for major releases. Some minor releases nevertheless contain changes that are technically breaking, such as stricter validation of previously invalid uses, corrections to unsupported behavior, or changes to support floors; those are labeled separately below. Patch releases are listed as subsections of the minor release they belong to.

v6.5 (unreleased)

CUDA.jl now requires GPUArrays.jl 12, which implements sorting, scans, reductions, findall, logical indexing and reverse once for every GPU back-end on top of AcceleratedKernels.jl. CUDA.jl's own implementations of these are removed. The functions follow Base's semantics more closely: sort! and sortperm are stable by default, init is applied once and no neutral element needs to be registered for custom operators, and empty inputs give Base's results and errors. Large reductions along a dimension and sorting are faster; small reductions along a dimension can be slower.

Technically breaking changes:

• GPU resources are no longer released from finalizers, as many of the CUDA calls that release them wait for all running kernels to finish, which could stall or deadlock whichever thread the garbage collector runs on. Collected memory is now released the next time CUDA.jl allocates memory or synchronizes, or within a second, so GC.gc() by itself doesn't make it available anymore. Releases that may wait for the GPU, like unpinning memory or destroying some library objects, are deferred until memory is reclaimed: on an out-of-memory error, or when calling CUDA.reclaim(). Pinned arrays are kept alive until then.
• JULIA_CUDA_MEMORY_POOL=none now also disables the memory pools that are used to allocate host and unified memory.
• The stream that stream() returns in a task is handed to another task once the task has finished and the work on the stream has completed, instead of being kept alive until the GC collects the task. This keeps applications that spawn many short-lived GPU tasks from piling up thousands of streams, which each hold on to device memory and slow down memory allocation. Code that uses a task's stream after the task has finished should create its own stream with CuStream() instead.
• Pointers that are taken from GPU memory inside the function passed to CUDA.with_managed are assumed to be used by an operation on the stream passed to with_managed, and no longer wait on the CPU for other streams that used the memory before. Code that passes such pointers to libraries, other streams or the CPU should take them outside of with_managed

... (truncated)

Commits

• b01d91d Bump version.
• 3301021 Merge pull request #3323 from JuliaGPU/tb/llvm10
• d4927fc CI: run the test suite with 2 CPU threads again (#3324)
• 92363d5 Point the development instructions at LLVM.jl's main branch
• c783635 Inspect the compiled IR before its context is disposed of
• 9995d97 Don't overwrite the debug location of the kernel's first instruction
• c6d8018 Use LLVM.jl's intrinsic lookup, pointer casts and volatile loads
• 3fa1ab1 Generate the remaining IR templates with @​llvmgenerated
• 5bbdfeb Use LLVM.jl's scoped enumerations instead of the raw C API
• cedd1b5 Port to LLVM.jl 10
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)