Release Version 0.6.3 - Memory pool support, compatibility improvements, minor changes · eyalroz/cuda-api-wrappers

Changes since v0.6.2:

Memory pool support (#249) and related changes

Added a cuda::memory::pool_t proxy class
Memory pools are created using cuda::memory::pool::create() or via device methods
IPC: Can import (and export) pools and their allocations to/from other processes
#485 Moved the physical_allocation namespace up from memory::virtual_ into memory, as it is used also for memory pools

#508 With CUDA >= 11.3, we no longer give up on creating a module from a compiled program just because no CUBIN is available - and try to use the PTX like with earlier CUDA versions.
#493 cuda::launch_config_builder_t::overall_size() now takes a cuda::grid::overall_dimension_t rather than a cuda::grid::dimension_t (not the same size).
#492 Avoiding inclusion of cooperative_groups.h from within the API headers

#511 Can now create CUDA runtime errors with fully-user-overriden what() message (probably not very interesting outside the API's internals).
#510 When NVRTC/PTX compiler complains about an invalid option, we now include the options passed in the thrown exceptions
#499 No longer exposing deprecated surface-related API functions with CUDA 12 and later.
#498 The launch config builder class now supports num_blocks() as an alias for the grid_size() method.
#488 cuda::memory::host::allocate() now returns a cuda::memory::region_t, for better consistency.
#486 Some changes to cuda::kernel::wrap.
Renamed cuda::memory::attribute_value_type_t -> cuda::memory::attribute_value_t
#483 It's now easier to convert memory::region_t's into typed spans.
#482 Improvements to the built-in cuda::span (which is used when std::span is unavailable) - making it somewhat more compatible with std::span
Make more comparison operators constexpr and noexcept

Wrappers now build correctly (again) with --std=c++20.
#501 Added a new NVRTC error code introduced in CUDA 12.1
#500 When using CUDA 12, use the term "LTO IR" rather than "NVVM" as appropriate
#494 Work around an MSVC issue with variadic template-templates
#491 Avoiding some warnings issued by MSVC
#480 Add example program built with each C++ version after 11 supported by the compiler

Now requiring CMake version 3.25. You can download an up-to-date version from Kitware's website; it doesn't require any special installation.
#490 Switched from depending on CUDA::nvToolkitExt to depending on CUDA::nvtx, for CUDA versions 10.0 and above.