v0.5.0

• build: don't rm(force) in catch as it obfuscates stack traces
• build: proper discovery of binary tools
• start migration towards LLVM.jl-based generated functions
• performance: internalize non-entry point functions, allowing for LLVM to discard calling conventions

v0.4.1

Hotfix: apparently the new AS-specific memory operations, which could trigger a known issue in LLVM, do really trigger that issue in real-life code and broke GPUArrays.

cc @SimonDanisch

v0.4.0

Major release:

• performance: introduce DevicePtr parameterized on address space, improves memory operations
• performance: generate automatic wrapper function for efficiently passing non-primitive types
• feature: fine-grained recompilation of dependent kernels (ie. 265)
• bugfix: package wouldn't get automatically precompiled again after a failed->successful build cycle
• bugfix: alignment calculation for memory operations
• compatibility with latest CUDAdrv

v0.3.1

• make it possible to load (& perform very limited testing) when the build failed, eg. when CUDAdrv.jl/LLVM.jl aren't available
• coarse version of #265 (ie. recompile everything as soon as the global world counter changed) for GPU kernels
• use a default, global context

v0.3.0

• fixes for 0.7/master (thanks @Keno)
• initial documentation
• adapt to new CUDAdrv features/deprecations
• generalized eltype support for at-cuShared/DynamicMem (thanks @SimonDanisch)

v0.2.0

• at-code reflection macros, with support for looking past at-cuda and converting types
• various fixes

v0.1.3

Hotfix: the upgrade could fail due to loading non-existing variables from a previous ext.jl.

v0.1.2

• Reduce launch overhead
• Adapt to changes in LLVM.jl and CUDAdrv.jl

v0.1.1

• Reflection: code_ptx, code_sass
• Reduction example
• Adapt to 0.6 changes wrt. new restrictions on generated functions