CmdStan OpenCL GPU problems and wiki page

ping @rok_cesnovar @Stevo15025 (you have been the last ones editing that wiki page)

I followed the instructions at https://github.com/stan-dev/math/wiki/OpenCL-GPU-Routines, and I get tests and a model to compile, but trying to run them I get error
clBuildProgram CL_OUT_OF_HOST_MEMORY: Unknown error -6

I tried to search what this means, with not much luck, but it might be related to version problems.
What is the required version of OpenCL? It would be nice to mention that also on that wiki page.
Is there some other minimal requirements?

Instructions say " In cmdstan, an example model is provided in examples/GP/ , which uses OpenCL Cholesky decomposition. You can check if your OpenCL configuration works by trying to build it.", but I couldn’t find that example in cmdstan develop or any obvious branch. Where can I find it?

1 Like

Hey,

the example link needs to be fixed, this was copied from when we had an experimental branch with a gp example way back when. We should instead make a GPU GLM and GP examples. Will get on that in the following days. For now I removed the text that talks about the example that isnt there.

But seeing as you already have a model you want to try this with, that maybe isnt needed at this moment.

Can you give me some info on the device and OS you are using.

Linux vdiubuntu103 4.15.0-74-generic #84-Ubuntu SMP Thu Dec 19 08:06:28 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux
Tue Jan 28 11:16:06 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.92       Driver Version: 410.92       CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GRID P40-2Q         On   | 00000000:02:01.0  On |                  N/A |
| N/A   N/A    P0    N/A /  N/A |   1883MiB /  2048MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
/usr/lib/x86_64-linux-gnu/libOpenCL.so.1

Thanks, GPU looks fine, OpenCL installation looks fine.

And presuming your make/local has the following right:
STAN_OPENCL=true
OPENCL_DEVICE_ID=0
OPENCL_PLATFORM_ID=0

Can you run clinfo to check if you have any other OpenCL-enabled devices on your system and the 0-0 index points to some other device. This is unlikely but just to make sure.

Regarding the version, OpenCL 1.2 is required, I do think that is mentioned somewhere, but not on the wiki (I added it now, thanks for the suggestion) and that should not be a problem as I have not seen any devices with 1.0 and 1.1 for at least 8 years.

A basic model example that works with OpenCL is the following bernoulli GLM example (I will post some fake data to go with this later):

data {
  int<lower=1> k;
  int<lower=0> n;
  matrix[n, k] X;
  int y[n];
}

parameters {
  vector[k] beta;
  real alpha;
}

model {
  target += bernoulli_logit_glm_lpmf(y | X, alpha, beta);
}

Yes. Btw, I would be great that in context of Stan, it would be nice to always mention which of the different make directories is the one where this local should be edited (cmdstan/make/local, cmdstan/stan/make/local, cmdstan/stan/lib/stan_math/local???)

Did some time to figure how I can install it without admin rights…

Number of platforms                               1
  Platform Name                                   NVIDIA CUDA
  Platform Vendor                                 NVIDIA Corporation
  Platform Version                                OpenCL 1.2 CUDA 10.0.246
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_khr_gl_event cl_nv_create_buffer
  Platform Extensions function suffix             NV

  Platform Name                                   NVIDIA CUDA
Number of devices                                 1
  Device Name                                     GRID P40-2Q
  Device Vendor                                   NVIDIA Corporation
  Device Vendor ID                                0x10de
  Device Version                                  OpenCL 1.2 CUDA
  Driver Version                                  410.92
  Device OpenCL C Version                         OpenCL C 1.2 
  Device Type                                     GPU
  Device Topology (NV)                            PCI-E, 02:00.1
  Device Profile                                  FULL_PROFILE
  Device Available                                Yes
  Compiler Available                              Yes
  Linker Available                                Yes
  Max compute units                               30
  Max clock frequency                             1531MHz
  Compute Capability (NV)                         6.1
  Device Partition                                (core)
    Max number of sub-devices                     1
    Supported partition types                     None
  Max work item dimensions                        3
  Max work item sizes                             1024x1024x64
  Max work group size                             1024
  Preferred work group size multiple              32
  Warp size (NV)                                  32
  Preferred / native vector sizes                 
    char                                                 1 / 1       
    short                                                1 / 1       
    int                                                  1 / 1       
    long                                                 1 / 1       
    half                                                 0 / 0        (n/a)
    float                                                1 / 1       
    double                                               1 / 1        (cl_khr_fp64)
  Half-precision Floating-point support           (n/a)
  Single-precision Floating-point support         (core)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
    Correctly-rounded divide and sqrt operations  Yes
  Double-precision Floating-point support         (cl_khr_fp64)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
  Address bits                                    64, Little-Endian
  Global memory size                              2147483648 (2GiB)
  Error Correction support                        No
  Max memory allocation                           536870912 (512MiB)
  Unified memory for Host and Device              No
  Integrated memory (NV)                          No
  Minimum alignment for any data type             128 bytes
  Alignment of base address                       4096 bits (512 bytes)
  Global Memory cache type                        Read/Write
  Global Memory cache size                        491520 (480KiB)
  Global Memory cache line size                   128 bytes
  Image support                                   Yes
    Max number of samplers per kernel             32
    Max size for 1D images from buffer            134217728 pixels
    Max 1D or 2D image array size                 2048 images
    Max 2D image size                             16384x32768 pixels
    Max 3D image size                             16384x16384x16384 pixels
    Max number of read image args                 256
    Max number of write image args                16
  Local memory type                               Local
  Local memory size                               49152 (48KiB)
  Registers per block (NV)                        65536
  Max number of constant args                     9
  Max constant buffer size                        65536 (64KiB)
  Max size of kernel argument                     4352 (4.25KiB)
  Queue properties                                
    Out-of-order execution                        Yes
    Profiling                                     Yes
  Prefer user sync for interop                    No
  Profiling timer resolution                      1000ns
  Execution capabilities                          
    Run OpenCL kernels                            Yes
    Run native kernels                            No
    Kernel execution timeout (NV)                 Yes
  Concurrent copy and kernel execution (NV)       Yes
    Number of async copy engines                  2
  printf() buffer size                            1048576 (1024KiB)
  Built-in kernels                                
  Device Extensions                               cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_khr_gl_event cl_nv_create_buffer

NULL platform behavior
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  No platform
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   No platform
  clCreateContext(NULL, ...) [default]            No platform
  clCreateContext(NULL, ...) [other]              Success [NV]
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT)  No platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  No platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  Invalid device type for platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  No platform

g

Its always the make/local of cmdstan as that propagates down to Stan Math. I never had to edit any lower level make/local files whenever I worked with Cmdstan (OpenCL related or not). I do realize that is maybe not obvious to everyone. I added that note (thanks again for the suggestion).

Ok, so its the only device, which means 0 and 0 is fine. OpenCL is fine, double precision is also fine… Argh. I apologize this is wasting your time. I have never seen this error before :/. Your system most definitely does not have resource problems.

The next thing to try would be to eliminate that its a bug with the exact model you are trying this on. So compile the model above and I will post some fake data to go with it.

Shouldn’t the tests work first?

./runTests.py test/unit -f opencl
...
make: 'test/unit/math/opencl/rev/triangular_transpose_test' is up to date.
make: 'test/unit/math/opencl/rev/zeros_test' is up to date.
------------------------------------------------------------
test/unit/math/opencl/assign_event_test --gtest_output="xml:test/unit/math/opencl/assign_event_test.xml"
terminate called after throwing an instance of 'std::system_error'
  what():  neg_binomial_2_log_glm: clBuildProgram CL_OUT_OF_HOST_MEMORY: Unknown error -6
Aborted
test/unit/math/opencl/assign_event_test --gtest_output="xml:test/unit/math/opencl/assign_event_test.xml" failed
exit now (01/28/20 10:48:34 EET)

Ah I misunderstood, sorry. Yep, those tests should pass.

Does ./runCmdStanTests.py src/test/interface/opencl_test.cpp fail in the same way?

test/interface/opencl_test --gtest_output="xml:test/interface/opencl_test.xml"
terminate called after throwing an instance of 'std::system_error'
  what():  neg_binomial_2_log_glm: clBuildProgram CL_OUT_OF_HOST_MEMORY: Unknown error -6
Aborted

Is this a virtual machine or anything like that by any chance? Probably not, but just want to get that out of the way.

Yes, it’s virtual machine. The purpose of that instance is to make it easy to test GPU computing. It’s maintained by my university IT, so I can ask changes to get things to work, but I need to first know what to ask.

image

1 Like

CL_OUT_OF_HOST_MEMORY would mean that there isnt enough resources to compile the OpenCL kernels (OpenCL kernels are compiled just-in-time at the start of any OpenCL-enabled Stan/Stan Math program).

4GB should be enough in general, but I do have to admit that I never observed how much RAM is used during tests. But given that non-OpenCL Stan alone uses a few GB of RAM to compile this might be it. Is there an easy option to request a GB or two of additional RAM?

There is an option to request. Let’s see what they answer.

Can you also check the size of swapfile on the system? If it is large more RAM will not help.

KiB Mem :  4039712 total,  1949048 free,  1251464 used,   839200 buff/cache
KiB Swap:  7911420 total,  6808316 free,  1103104 used.  2474640 avail Mem 

Ok, it’s a memory problem. I closed emacs and R, and the test passes!

Running main() from stan/lib/stan_math/lib/gtest_1.8.1/src/gtest_main.cc
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from StanUiCommand
[ RUN      ] StanUiCommand.opencl_ready
[       OK ] StanUiCommand.opencl_ready (0 ms)
[----------] 1 test from StanUiCommand (0 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (0 ms total)
[  PASSED  ] 1 test.
1 Like

Wow, I have never seen kernel compilation to take that much RAM. Maybe there are some issues in compiler. You can try requesting an update for GPU driver and hope for the best.

1 Like

I tried running the exact same test and was also monitoring the RAM usage and it barely makes a dent, under 100MB added to the baseline usage. Upgrading the NVIDIA driver might be the best way yes.