Make GPU examples cmdstan 2.20.0

Developers
1

Hey all -

I’m trying to run the examples to make sure the GPU stuff is working. I’m running cmdstan 2.20.0 and the only make example is bernoulli.

Is this a mistake? Should cmdstand 2.20.0 or the current version be updated so that we can test out the GPU stuff?

I ran ./runTests.py test/unit -f opencl and the tests pass with flying colors. Does this mean the GPU should be good to go, even if I can’t run this test file?

Andre

EDIT: totes not G2G, just ran a model and the compilation is failing, but that’s cool I can figure it out.

2

Sorry which examples are you referencing?

3

The ones here: https://github.com/stan-dev/math/wiki/OpenCL-GPU-Routines

examples/GPU/gp3, but the only compilation examples are the bernoulli ones.

The bernoulli ones: https://github.com/stan-dev/cmdstan/tree/develop/examples/

The compilation fails, but just because I haven’t linked my amd/cuda library, whatever it is I need to point stan to

4

yeah, another n00b question. How do I find the correct opencl lib? The one that has to go here:

LDFLAGS_OPENCL= -L"$(CUDA_PATH)\lib\x64" -lOpenCL

The $(CUDA_PATH).

Here’s my lspci output, I have a GPU:

00:00.0 Host bridge: Intel Corporation Xeon E3-1200 v5/E3-1500 v5/6th Gen Core Processor Host Bridge/DRAM Registers (rev 07)
00:01.0 PCI bridge: Intel Corporation Xeon E3-1200 v5/E3-1500 v5/6th Gen Core Processor PCIe Controller (x16) (rev 07)
00:02.0 VGA compatible controller: Intel Corporation HD Graphics 530 (rev 06)
...

there’s the GPU on line 4.

And then I’m setting the bash variable able to my cuda’s lib64 directory:

/usr/local/cuda/lib64.

I think this is the right place, but it’s still not compiling.
How do I find the appropriate directory?

here’s my clinfo -l :

Platform #0: Clover
 `-- Device #0: AMD VERDE (DRM 2.50.0, 5.3.11-200.fc30.x86_64, LLVM 8.0.0)
Platform #1: Portable Computing Language
 `-- Device #0: pthread-Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz

And here’s my make local:

STAN_OPENCL=true
OPENCL_DEVICE_ID=0
OPENCL_PLATFORM_ID=0
CC = g++
LDFLAGS_OPENCL= -L"$/usr/local/cuda/lib64" -lOpenCL

the openCL stuff is here:

     27096 Aug 26  2018 libOpenCL.so.1.1
    16 Aug 26  2018 libOpenCL.so.1 -> libOpenCL.so.1.1
      14 Aug 26  2018 libOpenCL.so -> libOpenCL.so.1
5

yo forget it. I removed the $ at the beginning

6

OK, no idea what this means:

terminate called after throwing an instance of 'std::system_error'
  what():  compile_kernel: add input.cl:29:252: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
input.cl:29:272: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
input.cl:29:292: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
input.cl:29:459: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
input.cl:29:598: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
: Unknown error -11

any ideas? thanks for the help.

7

pinging @rok_cesnovar I don’t have access to a windows computer and Rok wrote the setup docs for Windows

Also I’m not sure cuda’s OpenCL will work with the AMD or Intel devices you have available

http://mc-stan.org/math/opencl_support.html

For windows AMD you have to use the old SDK. AMD took it off of their site for windows (linux users can use ROCm) @rok_cesnovar do you know a place to get that? I found a site but it’s not AMD affiliated so I don’t want to post it

You can get the windows one for intel from below

8

This is linux. There’s probably a mistake I’m making in my setup then.

9

On Linux you dont need to set the path to the OpenCL lib. If clinfo works you are all set.

10

Ah, for linux you should be able to ignore the LDFLAGS_OPENCL line. We should probably highlight that in the docs

http://mc-stan.org/math/opencl_support.html

11

Can you give some more info on what is the GPU you are using? 'double' requires cl_khr_fp64 extension to be enabled smells a bit like its an older GPU with no native double support, which seems a bit weird.

12

Yeah I’m pretty sure that’s the issue, idt the cape verde gpu supports doubles

13

Just tried this and my cmdstan build failed.

Weak, any way I can get it to work without a massive amount of work?

14

I would use your intel device