NVIDIA's. More closed since controlled by NVIDIA. Also more popular for the same reason as they are the Top Dog now.
https://www.reddit.com/r/programming/comments/49uw97/cuda_reverse_engineered_to_run_on_nonnvidia/
https://en.wikipedia.org/wiki/CUDA NVIDIA's, currently only runs in NVIDIA hardware: http://stackoverflow.com/questions/12828268/is-it-possible-to-run-cuda-on-amd-gpus
TODO could AMD implement it legally without paying royalties to NVIDIA? Seems yes: https://www.reddit.com/r/MachineLearning/comments/4di4os/deep_learning_is_so_dependent_on_nvidia_are_there/d1rwuxb
Vulkan also adds the ability to run SPIR-V. So it is likely that SPIR-V will be well supported in NVIDIA and AMD cards, and we could then compile CUDA to SPIR-V. Result: AMD can run CUDA, and NVIDIA can run OpenCL.
https://www.reddit.com/r/vulkan/comments/3snrdv/vulkan_or_opencl/
If GPUCC generates SPIR-V, problem solved. TODO does it?
LLVM-based CUDA compiler. TODO output format?
Merged into LLVM.
http://wujingyue.github.io/docs/gpucc-talk.pdf https://www.youtube.com/watch?v=JHfb8z-iSYk
http://gpuocelot.gatech.edu/ is a precursor it seems https://code.google.com/archive/p/gpuocelot/
https://en.wikipedia.org/wiki/NVIDIA_CUDA_Compiler
--arch=sm thing: http://stackoverflow.com/questions/10330527/what-is-difference-between-arch-sm-13-and-arch-sm-20
TODO: compile OpenCL: failing with:
- http://stackoverflow.com/questions/13062469/compile-and-build-cl-file-using-nvidias-nvcc-compiler
- http://stackoverflow.com/questions/22730484/dont-know-what-to-do-with-nvcc-fatal-error
The kernel language, like OpenCL C.
Vs ISO: http://stackoverflow.com/questions/9846523/explanation-of-cuda-c-and-c