AMD GPU-PRO vs. NVIDIA Linux OpenCL Compute Performance

[bridgman]

Yeah, I guess I kind of confused AMDGPU with the previous open source effort from AMD. I still blame you for the whole naming "mess", though

That is an acceptable compromise

[jaguar]

Just for comparison
270x on win 10 latest drivers :
Luxball 8488
Michropone 4385
Hotel 1583

Around the 780 ti on linux,

[starshipeleven]

How about Vulkan compute performance?

Vulkan was designed as a gaming-oriented API, you mean SPIR-V perhaps? (the vulkan version of openCL)

[M1kkko]

Vulkan was designed as a gaming-oriented API, you mean SPIR-V perhaps? (the vulkan version of openCL)

On the Vulkan website, Vulkan is introduced as a "new generation graphics and compute API", so I don't think it's that far-fetched to use the name Vulkan when talking about the compute side.

Have any such benchmarks?

Sorry, no. I only assumed some would exist by now.

[smitty3268]

Vulkan was designed as a gaming-oriented API, you mean SPIR-V perhaps? (the vulkan version of openCL)

SPIR-V is the equivalent of GLSL or an OpenCL kernel. Vulkan does allow compute shaders.

[Pontostroy]

opencl from amdgpu pro works just fine with upstream amdgpu kernel module, so anyone with gcn >=1.1 may have two opencl realizations

http://www.gearsongallium.com/?p=2960

[bridgman]

First comment back from our HPC team is that the default size setting ("-s 1") used in the FFT SP test results in a workload that is too small to fill up a modern high-end GPU, so (a) not all the compute units are utilized, and (b) the hardware doesn't detect enough activity to justify raising clocks from their lowest state.

Recommendation is -s 3 or -s 4 (I think we use 4 for testing) in order to let larger GPUs work efficiently. I'll find out how that affects the run time (don't want to go from microseconds to weeks) then talk to Michael about whether changing defaults falls into the "patches welcome" category.

[Michael]

First comment back from our HPC team is that the default size setting ("-s 1") used in the FFT SP test results in a workload that is too small to fill up a modern high-end GPU, so (a) not all the compute units are utilized, and (b) the hardware doesn't detect enough activity to justify raising clocks from their lowest state.

Recommendation is -s 3 or -s 4 (I think we use 4 for testing) in order to let larger GPUs work efficiently. I'll find out how that affects the run time (don't want to go from microseconds to weeks) then talk to Michael about whether changing defaults falls into the "patches welcome" category.

Hmm, I see. Yeah I am doing a completely out-of-the-box build of SHOC (the raw test profile: http://openbenchmarking.org/innhold/...e363a013f3de9d)

I am not at all opposed to carrying patches to increase the size of the test. So patches welcome. However, might it be wiser to see if upstream SHOC is interested in setting a more sane default for modern GPUs? That's the ideal way I look at it is getting the upstream software to have better out-of-the-box defaults, but nevertheless happy to carry the patch otherwise.

[bridgman]

Thanks Michael. My initial impression (actually the initial impression of one of our HPC guys) was that the "-s 1" default came from the PTS test definition file, does that make sense ? He was looking in https://openbenchmarking.org/innhold...e363a013f3de9d

There are probably other defaults in SHOC itself, will take a look.

  <TestSettings>
    <Default>
      <Arguments>-s 1 </Arguments>
    </Default>

Just heard back that runtime should be <30s on a modern GPU even with large datasets, so looks like patching the test should work.

[faldzip]

Does anyone get this AMD GPU-PRO driver to work with R7 260X? It is GCN1.1 so should be working fine as R9 290 is, but I've installed this on Ubuntu 15.10 (4.2 kernel) and can't set any other resolution than 1024x768 or 800x600 (on my FHD monitor and TV too). Then ran The Talos Principle's built-in benchmark in this 1024x768 (Ultra-High settings) with OpenGL and Vulkan and I got ~7fps in both (while on the Win10 I have 32 and 36 respectively but in 1080p!). I've managed to add the 1080p mode to xrandr but after enabling it with xrandr everything on the screen looks really shitty (like much lower res upscaled - so the fonts are not quite readable). Does anyone faced such issues with AMD GPU-PRO?