Is more or less involved..
Sources are at:
ftp://download.nvidia.com/CUDAOpen64/
you only compile Open64 backend so no nvcc changes (frontend)
they get only stable releases so no 3.0beta sources at this time..
you can compile for Linux, Windows and Mac..
Mac 32 bit only and Windows has Cygwin and Visual studio but Nvidia ships Cygwin one is the stable one.. for Win32 and Win64..
some time ago compiled for Windows cygwin32 and Linux x64 and tested also Mac x32
help for compiling and understanding compilation path:
1.package txts in src\doc..
most interesting nvopencc tutorial powerpoint helps adding instructions
2.NVIDIA's Experience with Open64
slides
3. For Visual Studio 2008 :
src\windows\ see txt's
Addin txt
Notes:
1.CUDA 3.0 compiles to standard ELF files no cubin embedded as source array of data (?).. has also some new exec.. so perhaps it's not true for it some steps..
but was correct for CUDA 2.2
2. Now Nvidia seems shifting some effort into LLVM for multicore as for OpenCL code gen..
search
PLANG: Translating NVIDIA PTX language to LLVM IR Machine
http://llvm.org/devmtg/2009-10/Grover_PLANG.pdf
in
http://llvm.org/devmtg/2009-10/
It's present in nvcc 3.0..
Also see
http://llvm.org/devmtg/2009-10/OpenCLWithLLVM.pdf
Some projects using it:
Exposing integer add with carry (has diff and Linux modified binaries)
http://www.mpi-inf.mpg.de/~emeliyan/cuda-compiler/
Improving the Open64 Backend for GPUs:
http://code.google.com/p/google-summer-of-code-2009-open64/updates/list
Wednesday, 9 December 2009
Subscribe to:
Post Comments (Atom)
0 comments:
Post a Comment