## CUDA v2.6.1

[Diff since v2.6.0](JuliaGPU/CUDA.jl@v2.6.0...v2.6.1)


**Closed issues:**
- CUDA 11.2 (JuliaGPU#601)
- LLVM not found (JuliaGPU#681)

**Merged pull requests:**
- Automatic task-based concurrency using local streams (JuliaGPU#662) (@maleadt)
- add beta keyword to conv (JuliaGPU#672) (@jw3126)
- Protect the kernel closure from GC collection. (JuliaGPU#674) (@maleadt)
- Track external globals, use it to avoid needless exception flags (JuliaGPU#675) (@maleadt)
- Adapt to GPUCompiler changes. (JuliaGPU#676) (@maleadt)
- Minor improvements (JuliaGPU#677) (@maleadt)
- CompatHelper: add new compat entry for "Memoize" at version "0.4" (JuliaGPU#678) (@github-actions[bot])
- Use released GPUCompiler. (JuliaGPU#683) (@maleadt)
- v2.6.1 (JuliaGPU#684) (@maleadt)

## CUDA v2.6.0

[Diff since v2.5.0](JuliaGPU/CUDA.jl@v2.5.0...v2.6.0)


**Closed issues:**
- Invalid results due to shared memory + multiple function exits (?) mysteriously solved by @cuprintf (JuliaGPU#43)
- NVML-related segfault on Windows (JuliaGPU#610)
- @cuda with config keyword sometimes allocate lots of memory (JuliaGPU#643)
- Can someone with push access run the TagBot workflow? (JuliaGPU#644)
- Taking gradient with Flux results in NaNs when using CUDA arrays but not when using CPU arrays (JuliaGPU#657)
- Broadcasting fails in a special case (JuliaGPU#658)
- view causes KeyError in alias (JuliaGPU#661)
- PTXCompilerTarget error when creating a CuArray with Float64 (JuliaGPU#664)
- Complex dot product performance of CuArrays and of StructArrays of CuArrays (JuliaGPU#667)
- could not load cublas64_11.dll (JuliaGPU#670)

**Merged pull requests:**
- CUDA quicksort (JuliaGPU#431) (@xaellison)
- Bump Reexport to 1.0 (JuliaGPU#640) (@DhairyaLGandhi)
- Use newer NVML initialization method. (JuliaGPU#641) (@maleadt)
- README: add some information on viewing capabilities of your devices (JuliaGPU#642) (@DilumAluthge)
- Remove duplicate functions. (JuliaGPU#645) (@maleadt)
- Use released version of Adapt.jl (JuliaGPU#646) (@maleadt)
- Simplify list of tests to skip. (JuliaGPU#647) (@maleadt)
- Use a test-specific Project.toml. (JuliaGPU#648) (@maleadt)
- Use raw output for CUBLAS log message. (JuliaGPU#649) (@maleadt)
- Close the async condition used to call host functions. (JuliaGPU#650) (@maleadt)
- Backports for Julia 1.5 / CUDA 2.4 (JuliaGPU#651) (@maleadt)
- Allow running benchmarks outside of the master branch on other systems. (JuliaGPU#652) (@maleadt)
- Bump GPUCompiler. (JuliaGPU#653) (@maleadt)
- Reuse the compiler when generating SASS code. (JuliaGPU#654) (@maleadt)
- Run the tests from the current directory. (JuliaGPU#655) (@maleadt)
- Configure the PTX GPUCompiler codegen quirks. (JuliaGPU#656) (@maleadt)
- Update manifest (JuliaGPU#660) (@github-actions[bot])
- Support view on unmanaged arrays. (JuliaGPU#663) (@maleadt)
- Retry CuModule creation when OOM. (JuliaGPU#665) (@maleadt)
- Make fill async. (JuliaGPU#669) (@maleadt)
- Fix version lookups. (JuliaGPU#671) (@maleadt)
- Update manifest (JuliaGPU#673) (@github-actions[bot])

## CUDA v2.4.1

[Diff since v2.4.0](JuliaGPU/CUDA.jl@v2.4.0...v2.4.1)


**Closed issues:**
- `cudaconvert` for closures (JuliaGPU#67)
- Invalid results due to shared memory + multiple function exits (?) mysteriously solved by @cuprintf (JuliaGPU#43)
- NVML-related segfault on Windows (JuliaGPU#610)
- Update Reexport compat (JuliaGPU#629)
- Incomplete CUDA device attributes list  (JuliaGPU#637)
- @cuda with config keyword sometimes allocate lots of memory (JuliaGPU#643)
- Can someone with push access run the TagBot workflow? (JuliaGPU#644)
- Taking gradient with Flux results in NaNs when using CUDA arrays but not when using CPU arrays (JuliaGPU#657)
- Broadcasting fails in a special case (JuliaGPU#658)
- view causes KeyError in alias (JuliaGPU#661)
- PTXCompilerTarget error when creating a CuArray with Float64 (JuliaGPU#664)
- Complex dot product performance of CuArrays and of StructArrays of CuArrays (JuliaGPU#667)

**Merged pull requests:**
- CUDA quicksort (JuliaGPU#431) (@xaellison)
- cudaconvert captured values in closures. (JuliaGPU#625) (@maleadt)
- CompatHelper: only instantiate `/Manifest.toml` (the manifest file in the root of the repository) (JuliaGPU#631) (@DilumAluthge)
- CompatHelper: bump compat for "Reexport" to "1.0" (JuliaGPU#633) (@github-actions[bot])
- CompatHelper: bump compat for "AbstractFFTs" to "1.0" (JuliaGPU#634) (@github-actions[bot])
- Update wrappers (JuliaGPU#638) (@maleadt)
- Bump artifacts for Windows/Julia 1.6 compatibility. (JuliaGPU#639) (@maleadt)
- Bump Reexport to 1.0 (JuliaGPU#640) (@DhairyaLGandhi)
- Use newer NVML initialization method. (JuliaGPU#641) (@maleadt)
- README: add some information on viewing capabilities of your devices (JuliaGPU#642) (@DilumAluthge)
- Remove duplicate functions. (JuliaGPU#645) (@maleadt)
- Use released version of Adapt.jl (JuliaGPU#646) (@maleadt)
- Simplify list of tests to skip. (JuliaGPU#647) (@maleadt)
- Use a test-specific Project.toml. (JuliaGPU#648) (@maleadt)
- Use raw output for CUBLAS log message. (JuliaGPU#649) (@maleadt)
- Close the async condition used to call host functions. (JuliaGPU#650) (@maleadt)
- Backports for Julia 1.5 / CUDA 2.4 (JuliaGPU#651) (@maleadt)
- Allow running benchmarks outside of the master branch on other systems. (JuliaGPU#652) (@maleadt)
- Bump GPUCompiler. (JuliaGPU#653) (@maleadt)
- Reuse the compiler when generating SASS code. (JuliaGPU#654) (@maleadt)
- Run the tests from the current directory. (JuliaGPU#655) (@maleadt)
- Configure the PTX GPUCompiler codegen quirks. (JuliaGPU#656) (@maleadt)
- Update manifest (JuliaGPU#660) (@github-actions[bot])
- Support view on unmanaged arrays. (JuliaGPU#663) (@maleadt)
- Retry CuModule creation when OOM. (JuliaGPU#665) (@maleadt)
- Make fill async. (JuliaGPU#669) (@maleadt)

## CUDA v2.5.0

[Diff since v2.4.0](JuliaGPU/CUDA.jl@v2.4.0...v2.5.0)

## CUDA v2.4.0

[Diff since v2.3.0](JuliaGPU/CUDA.jl@v2.3.0...v2.4.0)


**Closed issues:**
- cublasXtStrmm test failures on Windows 10 Julia 1.1 (JuliaGPU#124)
- CUSPARSE tests broken (JuliaGPU#259)
- Make @cuda return a kernel object (JuliaGPU#341)
- Depend on CompilerSupportLibraries (JuliaGPU#359)
- CUBLAS and exceptions test failures on Windows (JuliaGPU#536)
- argmax(::CuArray) returns nothing with NaN-values (JuliaGPU#553)
- Multiple @cuDynamicSharedMem in kernel causes unexpected behavior (JuliaGPU#555)
- Illegal memory access with atomic shared memory (JuliaGPU#558)
- CUDA.sqrt will not found symbol "__nv_sqrt" (JuliaGPU#559)
- Exception with CUDA.exp (JuliaGPU#561)
- Use LazyArtifacts instead of Pkg (JuliaGPU#570)
- Test runner: early bail out (JuliaGPU#578)
- memory reporting issue (JuliaGPU#579)
- c[3:4]=0 leads to exception (JuliaGPU#580)
- Add math ops (including broadcast) for half types (JuliaGPU#581)
- Dot product of Array and CuArray fails with CPU address error. (JuliaGPU#586)
- Support for CUDA-capable GPU with compute capability 4.0 like GTX 1080 (JuliaGPU#587)
- mapreducedim! not threadsafe (JuliaGPU#588)
- Allow separate directories for cuda and cudnn (JuliaGPU#590)
- Difficulties installing CUDA on Julia 1.6.0 . (JuliaGPU#591)
- Bug in Initialisation Error (JuliaGPU#603)
- CUDA.jl initialisation fails after suspending Ubuntu 20.04 with CUDA 11.2 (JuliaGPU#605)
- CUDA 11.2 CUBLASError and "CUDA.jl does not yet support CUDA with nvdisasm 11.2.67" (JuliaGPU#607)
- This intrinsic must be compiled to be called (JuliaGPU#611)
- OpenGL interop (JuliaGPU#612)
- Add support for CuFFT callback functions (JuliaGPU#614)
-  I can’t multiply a CSR sparse matrix anymore (JuliaGPU#615)
- Julia version requirement (JuliaGPU#619)

**Merged pull requests:**
- Support all combinations of datatypes and transposes/adjoints in LinearAlgebra (JuliaGPU#535) (@cqql)
- Use structs for texture intrinsic return types. (JuliaGPU#554) (@maleadt)
- Backport some 1.6 fixes (JuliaGPU#557) (@maleadt)
- Update manifest (JuliaGPU#560) (@github-actions[bot])
- Correct dims error (JuliaGPU#562) (@DhairyaLGandhi)
- Lock `_shmem_cb` (JuliaGPU#564) (@vchuravy)
- Move to Julia 1.6 (JuliaGPU#566) (@maleadt)
- Adapt to JuliaLang/julia#38487. (JuliaGPU#568) (@maleadt)
- Support for 'delayed kernels' (JuliaGPU#569) (@maleadt)
- Run cuda-memcheck as part of CI (JuliaGPU#571) (@maleadt)
- Use at-sync instead of calls to synchronize in tests. (JuliaGPU#572) (@maleadt)
- Update artifacts to include cuda-memcheck (JuliaGPU#573) (@maleadt)
- Use LazyArtifacts instead of Pkg. (JuliaGPU#574) (@maleadt)
- Improve LinearAlgebra impl methods for triangular types (JuliaGPU#575) (@maleadt)
- New findmin/max implementation using single-pass reduction (JuliaGPU#576) (@maleadt)
- Fix synchronization before testing cublasXt calls. (JuliaGPU#577) (@maleadt)
- Fix used memory reporting. (JuliaGPU#582) (@maleadt)
- Implement Statistics.varm/stdm instead of Statistics._var (JuliaGPU#583) (@sdewaele)
- Test for JuliaGPU#558. (JuliaGPU#584) (@maleadt)
- Add a quick failure option to the test runner. (JuliaGPU#585) (@maleadt)
- Add lock around `cfunction` lookup (JuliaGPU#589) (@vchuravy)
- Catch all initialization errors. (JuliaGPU#593) (@maleadt)
- Update dependencies. (JuliaGPU#596) (@maleadt)
- Fix wrong initialisation error message (JuliaGPU#604) (@qin-yu)
- Fixes wrong spacing in docstring admonition (JuliaGPU#608) (@navidcy)
- Fix broadcasting with Base.angle (JuliaGPU#618) (@marius311)
- Test with the 1.6 nightly, not 1.7. (JuliaGPU#620) (@maleadt)
- Wrap cudaGL.h (JuliaGPU#621) (@maleadt)
- Initial compatibility with CUDA 11.2. (JuliaGPU#622) (@maleadt)
- 1.5 compatibility release (JuliaGPU#623) (@maleadt)
- Add CUDA 11.2 artifacts. (JuliaGPU#624) (@maleadt)

## CUDA v2.3.0

[Diff since v2.2.1](JuliaGPU/CUDA.jl@v2.2.1...v2.3.0)


**Closed issues:**
- Misaligned address on load from `Const` (JuliaGPU#548)

**Merged pull requests:**
- Allow `PermutedDimsArray` in `gemm_strided_batched` (JuliaGPU#539) (@mcabbott)
- Fix broken checkbounds for CuSparseMatrixCSR and tests (JuliaGPU#545) (@achuchmala)
- Emphasize rebooting option. (JuliaGPU#547) (@xanfus)
- fix address calculation for ldg (JuliaGPU#549) (@vchuravy)
- Don't use explicit per-stream threads. (JuliaGPU#551) (@maleadt)

Bump version again.

By having skipped CI on the previous tagged commit, we didn't get documentation.
[skip tests]

## CUDA v2.2.0

[Diff since v2.1.0](JuliaGPU/CUDA.jl@v2.1.0...v2.2.0)


**Closed issues:**
- cudnn missing after downloading artifact (JuliaGPU#521)
- Downloading artifact: CUDA110 when using DiffEqFlux (JuliaGPU#542)

**Merged pull requests:**
- Update manifest (JuliaGPU#520) (@github-actions[bot])
- Try out Buildkite. (JuliaGPU#522) (@maleadt)
- Update manifest (JuliaGPU#529) (@github-actions[bot])
- Support for / Upgrade to CUDA 11.1 update 1. (JuliaGPU#530) (@maleadt)
- Fix and test svd! (JuliaGPU#531) (@maleadt)
- Move more CI to Buildkite. (JuliaGPU#532) (@maleadt)
- Use type symbols to generate wrapper methods (JuliaGPU#534) (@cqql)
- Fully move to Buildkite. (JuliaGPU#537) (@maleadt)
- Add unit_diag option for sv2! functions (JuliaGPU#540) (@amontoison)
- Documentation fixes (JuliaGPU#543) (@maleadt)

## CUDA v2.1.0

[Diff since v2.0.2](JuliaGPU/CUDA.jl@v2.0.2...v2.1.0)


**Closed issues:**
- CUDNN convolution with Float16 always returns zeros  (JuliaGPU#92)
- axp(b)y! and mul! (scalar multiplication) with mixed argument types (JuliaGPU#144)
- Dispatching to generic matmul instead of CUBLAS (JuliaGPU#164)
- Support for Ints and Float16? (JuliaGPU#165)
- Subarrays/views support (JuliaGPU#172)
- Easy way to pick among multiple GPUs  (JuliaGPU#174)
- More prominently document JULIA_CUDA_USE_BINARYBUILDER (JuliaGPU#204)
- ERROR_COOPERATIVE_LAUNCH_TOO_LARGE during tests (JuliaGPU#247)
- Pkg.test error for cutensor test on Windows (JuliaGPU#422)
- Runtime build improvements (JuliaGPU#456)
- Fusing Wrappers (JuliaGPU#467)
- Could not find nvToolsExt (libnvToolsExt.dylib.1.0 or libnvToolsExt.dylib.1) in /Users/imac/.julia/artifacts/b502baf54095dff4a69fd6aba8667124583f6929/lib (JuliaGPU#482)
- mapreduce assumes commutative op (JuliaGPU#484)
- SubArray Broadcast Bug in 2.0 (JuliaGPU#488)
- Nested SubArray Scalar Indexing (JuliaGPU#490)
- Sparse matrix * view(vector) regression in 2.0 (JuliaGPU#493)
- Error transforming a reshaped 0-dimentional GPU array to a CPU array (JuliaGPU#494)
- test cuda FAILURE (JuliaGPU#496)
- Reshaped CuArray is not DenseCuArray (JuliaGPU#511)
- assignment failure when using array slicing. (JuliaGPU#516)

**Merged pull requests:**
- Use the correct CUDNN scaling parameter type. (JuliaGPU#454) (@maleadt)
- Fix versioned dylib discovery. (JuliaGPU#486) (@maleadt)
- Move inv from GPUArrays. (JuliaGPU#487) (@maleadt)
- Use dense array types in sparse wrappers. (JuliaGPU#495) (@maleadt)
- Update manifest (JuliaGPU#497) (@github-actions[bot])
- Revert array wrapper union changes (JuliaGPU#498) (@maleadt)
- Clean-up pointer field. (JuliaGPU#499) (@maleadt)
- mapreduce: change iteration for compatibility with non-commutative operators. (JuliaGPU#500) (@maleadt)
- Use versioned libcuda (JuliaGPU#502) (@maleadt)
- Dynamically choose versioned libcuda (JuliaGPU#503) (@mustafaquraish)
- Update multigpu.md (JuliaGPU#504) (@efmanu)
- Upgrade artifacts for CUDA 11 compatibility. (JuliaGPU#506) (@maleadt)
- Update dependencies. (JuliaGPU#507) (@maleadt)
- Convert unsigned short ints to Cint for printf. (JuliaGPU#508) (@maleadt)
- Update manifest (JuliaGPU#510) (@github-actions[bot])
- Fix reshape with missing dimensions. (JuliaGPU#512) (@maleadt)
- Don't return a pointer from 'alias'. (JuliaGPU#513) (@maleadt)
- Add some docs (JuliaGPU#514) (@maleadt)
- Fix CUDNN-optimized activation broadcasts (JuliaGPU#515) (@maleadt)
- Fix cooperative launch test. (JuliaGPU#517) (@maleadt)
- Fixes for Windows (JuliaGPU#518) (@maleadt)
- CUTENSOR fixes on Windows (JuliaGPU#519) (@maleadt)

## CUDA v2.0.2

[Diff since v2.0.1](JuliaGPU/CUDA.jl@v2.0.1...v2.0.2)


**Closed issues:**
- cu() behavior for complex floating point numbers (JuliaGPU#91)
- Error when following example on using multiple GPUs on multiple processes (JuliaGPU#468)
- MacOS without nvidia GPU is trying to download CUDA111 on julia nightly (JuliaGPU#469)
- Drop BinaryProvider? (JuliaGPU#474)
- Latest version of master doesn't work on Windows (JuliaGPU#477)
- `sum(CUDA.rand(3,3))` broken (JuliaGPU#480)
- copyto!() between cpu and gpu with subarrays (JuliaGPU#491)

**Merged pull requests:**
- Adapt to GPUCompiler changes. (JuliaGPU#458) (@maleadt)
- Fix initialization of global state (JuliaGPU#471) (@maleadt)
- Remove 'view' implementation. (JuliaGPU#472) (@maleadt)
- Workaround new artifact"" eagerness that prevents loading on unsupported platforms (JuliaGPU#473) (@ianshmean)
- Remove BinaryProvider dep. (JuliaGPU#475) (@maleadt)
- typo:  libcuda.dll -> libcuda.so on Linux (JuliaGPU#476) (@Alexander-Barth)
- NFC array simplifications. (JuliaGPU#481) (@maleadt)
- Update manifest (JuliaGPU#485) (@github-actions[bot])
- Convert AbstractArray{ComplexF64} to CuArray{ComplexF32} by default (JuliaGPU#489) (@pabloferz)