Announcement

Collapse
No announcement yet.

Radeon's ROCm OpenCL Runtime Finally Open-Sourced

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

  • Marc.2377
    replied
    I have been reading Phoronix for quite a while now, including the forums, and I felt the need to register in order to manifest how happy I am for this.

    Thank you AMD for keeping your promise, and huge thanks to everyone's hard work in making this possible.

    I'm switching to Linux as main OS pronto.

    Leave a comment:


  • torsionbar28
    replied
    Originally posted by dungeon View Post
    Instead of "Yes" on "Driver status" and under "Free/Libre" it should say "No(firmware required)"
    This news from AMD is fantastic and pure goodness. What do you gain from dropping a turd into the punchbowl? A modern x86 PC, both intel and AMD, has binary blobs in the CPU, in the motherboard, in the add-in cards for sound, RAID, etc. That's not the point here. Having a 100% open source driver stack for a tier-1 GPU is truly a milestone in Open Source history. Don't play the turd delivery boy - This is a historic event, in case that hasn't sunk in for you yet, not to mention the announcement's comment about AMD continuing working towards open sourcing the blobs. Intel isn't doing this, and neither is NVidia. AMD is to be commended for pursuing an aggressively open-source friendly strategy.
    Last edited by torsionbar28; 05-12-2017, 10:27 PM.

    Leave a comment:


  • Marc Driftmeyer
    replied
    Sure seems to have some bugs in the repo script.

    Correction: Saving the repo from the link doesn't corrupt the script. Now it works.

    Close but no cigar:

    [ 97%] Built target ocml_bc
    In file included from /home/mdriftmeyer/bin/opencl/runtime/platform/command.hpp:26:0,
    from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocdevice.hpp:11,
    from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.hpp:11,
    from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.cpp:15:
    /home/mdriftmeyer/bin/opencl/runtime/platform/perfctr.hpp:27:50: warning: ignoring attributes on template argument ‘cl_perfcounter_property {aka long unsigned int}’ [-Wignored-attributes]
    typedef std::map<cl_perfcounter_property, ulong> Properties;
    ^
    In file included from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocdevice.hpp:11:0,
    from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.hpp:11,
    from /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.cpp:15:
    /home/mdriftmeyer/bin/opencl/runtime/platform/command.hpp:61:33: warning: ignoring attributes on template argument ‘amd::Event::CallBackFunction {aka void (*)(_cl_event*, int, void*)}’ [-Wignored-attributes]
    std::atomic<CallBackFunction> callback_; //!< callback function pointer.
    ^
    /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.cpp: In member function ‘void roc::Memory::syncCacheFromHost(roc::VirtualGPU&, device::Memory::SyncFlags)’:
    /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.cpp:384:5: error: ‘wait’ was not declared in this scope
    wait(gpu);
    ^~~~
    /home/mdriftmeyer/bin/opencl/runtime/device/rocm/rocmemory.cpp:384:5: note: suggested alternative: ‘write’
    wait(gpu);
    ^~~~
    write
    runtime/device/rocm/CMakeFiles/oclrocm.dir/build.make:240: recipe for target 'runtime/device/rocm/CMakeFiles/oclrocm.dir/rocmemory.cpp.o' failed
    make[2]: *** [runtime/device/rocm/CMakeFiles/oclrocm.dir/rocmemory.cpp.o] Error 1
    CMakeFiles/Makefile2:23207: recipe for target 'runtime/device/rocm/CMakeFiles/oclrocm.dir/all' failed
    make[1]: *** [runtime/device/rocm/CMakeFiles/oclrocm.dir/all] Error 2
    Makefile:151: recipe for target 'all' failed
    make: *** [all] Error 2
    [email protected]:~/bin/build$

    Ideas?
    Last edited by Marc Driftmeyer; 05-13-2017, 01:06 AM. Reason: compile fail.

    Leave a comment:


  • Spazturtle
    replied
    Originally posted by dungeon View Post

    Instead of "Yes" on "Driver status" and under "Free/Libre" it should say "No(firmware required)"
    1) Firmware isn't part of the driver.

    2) Would you rather the firmware was on a read-only chip on the GPU?

    Leave a comment:


  • dungeon
    replied
    Originally posted by illwieckz View Post
    So, there is definitely no one reason to buy something else than an AMD discrete GPU if your workstation is running Linux:
    http://dl.illwieckz.net/b/linux-gpu-...bility-matrix/
    Instead of "Yes" on "Driver status" and under "Free/Libre" it should say "No(firmware required)"

    Leave a comment:


  • Nille_kungen
    replied
    \o/ Thanks AMD i have been waiting on this since the announcement September 2015
    https://www.x.org/wiki/Events/XDC201...hou_amdgpu.pdf

    Does this also mean that AMD vulcan support is a step closer now? We got radv until it's released.
    Last edited by Nille_kungen; 05-12-2017, 08:31 PM.

    Leave a comment:


  • smitty3268
    replied
    So since I'm lazy, can anyone list what the requirements to run this are?

    Do you need the -pro kernel driver? I'm assuming upstream llvm isn't good enough, right?

    Leave a comment:


  • nevion
    replied
    Originally posted by illwieckz View Post
    So, there is definitely no one reason to buy something else than an AMD discrete GPU if your workstation is running Linux:
    http://dl.illwieckz.net/b/linux-gpu-...bility-matrix/

    The complete stack is now opened, from kernel driver (amdgpu) to OpenGL 4.5 (radeonsi), Vulkan (radv), DirectX9 (galliumnine) rendering and OpenCL (rocm) computing.
    No, I'm afraid competition doesn't end. NVidia really made the game much harder hardware wise and especially software wise as we saw in GTC last week. Even in raw cuda enhancements hip already needs some work, but I think OpenCL is the most screwed by the disparity of feature support. Unfortunately they still got their work cut out for themselves and they'll have to be pioneers and carry OpenCL into a competitive age with where CUDA 9 is going.. But they're building a solid platform to fight from.

    I'm a huge fan of this work but the business case is still difficult. Will have to see Vega 10 and any other surprises they have waiting, if any. TPUs, multi level synchronization, and what seems to be the end of warp divergence (which also cures lockless algorithm problems) are some of the biggest things they have to work on now. I wish them luck.

    Leave a comment:


  • jrch2k8
    replied
    Amazing job, Thank You AMD for the best support ever, waiting to retire my old damn good r9-280 once Vega is out(prolly going to Polaris price wise but until Vega I'm not sure, that damn HBM looks sexy)

    Leave a comment:


  • illwieckz
    replied
    So, there is definitely no one reason to buy something else than an AMD discrete GPU if your workstation is running Linux:
    http://dl.illwieckz.net/b/linux-gpu-...bility-matrix/

    The complete stack is now opened, from kernel driver (amdgpu) to OpenGL 4.5 (radeonsi), Vulkan (radv), Direct3D9 (galliumnine) rendering and OpenCL (rocm) computing.

    If you're running Linux, there is no point buying something else than an AMD Radeon and there is no point using something else than free software. No other vendor provides all these features, and AMD only provides that completeness with free software stack. Free software strikes again.
    Last edited by illwieckz; 05-13-2017, 02:51 AM.

    Leave a comment:

Working...
X