Announcement

Collapse
No announcement yet.

LLVM's Clang Lands More CUDA Improvements

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • LLVM's Clang Lands More CUDA Improvements

    Phoronix: LLVM's Clang Lands More CUDA Improvements

    Just days after writing about GPUCC as Google's open-source CUDA compiler built atop LLVM and how to compile CUDA code with LLVM, more improvements have landed...

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    So what's the end game with all this CUDA stuff?

    Comment


    • #3
      Originally posted by wodencafe View Post
      So what's the end game with all this CUDA stuff?
      It will be very easy to embed parallelized GPU code into your C++ application using C++.

      Comment


      • #4
        Originally posted by vadix View Post
        It will be very easy to embed parallelized GPU code into your C++ application using C++.
        Only when your code is using CUDA language extensions.

        Comment


        • #5
          Originally posted by log0 View Post

          Only when your code is using CUDA language extensions.
          That's not really true...

          CUDA is used to implement the Thrust library's CUDA backend. Thrust is the prototype of Parallelism TS, which is expected to be part of C++17. If Clang supports the CUDA language extensions, then that can be used to compile Thrust.

          Therefore, user code can be written in pure C++, without any language extensions, and be compiled for many devices, including CPUs, NVIDIA GPUs, and hopefully soon alternative GPUs (once SYCL or other efforts come to fruition and provide alternative Parallelism TS implementations). Internally, compilers have non-standard extensions, and if you view CUDA support as just an implementation detail of the standard C++17 parallel algorithms library, then you can write GPU code without CUDA language extensions.

          Comment


          • #6
            I have the feeling there is currently very little momentum around OpenCL ... AMD speaks about Cuda, Apple deprecates it, ... Except Beignet from Intel, nothing is goes on.
            Still little drivers for OpenCL2.0, not even speaking of 2.1 !

            Comment

            Working...
            X