v0.3: Merge pull request #248 from raffenet/0.3-changes

Changes in 0.3

• Default to detecting the CUDA device capabilities at configure
  time. If no device is found on the build system, build all "major"
  CUDA capabilities to cut down on build time and library size. (thanks
  to Jeff Hammond for contributing)

• Add support for mixed memory types (thanks to ParTec AG for
  contributing)

• Add HIP backend for stream APIs

• Add automatic HIP SM detection

• Add automatic CUDA SM detection

• Add support for user-specified CUDA compiler

• Add support in --ze-native option to compile for multiple devices

• Add support for --pup-max-nesting < 2 in genpup.py

• Add support for --ze-revision-id to pass to ocloc compiler

• Other bug fixes and code cleanup

v0.2

Changes in 0.2

• Add support for reduction operations (e.g. sum, prod, min, max, ...)

• Add support for AMD GPUs via HIP backend

• Add "nogpu" info hint to avoid unnecessary pointer attribute queries

• Add stream-based pack/unpack APIs

• Add blocking pack/unpack APIs

• Add support for NVIDIA HPC SDK compilers

• Improve compile time for Level Zero kernels

• Extend tests to support subdevices (tiles) of Intel GPUs

• Many bug fixes and code cleanups