Get Unique Identifier for Device about opencl-ruby HOT 5 CLOSED

Hello,

The answer is double.

From my experience, an implementation, on the same machine/reboot/configuration always returns it's devices in the same order, the enumerator is undefined but not random (in my experience). So I use the order of the devices returned by clGetDeviceIDs to differentiate them. In mpi a modulo on mpi rank usually does the trick (provided process on the same node have consecutive ranks).

Should you need to identify precisely the device (to track it across reboot, driver version and so on) you can easily dos so in opencl_ruby_ffi. Here is an irb example:

> require 'opencl_ruby_ffi'
=> true
> OpenCL::platforms.first.devices.map(&:name)
=> ["Tesla K40c", "Tesla K40c"]
> class OpenCL::Device
>   PCI_BUS_ID_NV = 0x4008
>   eval get_info("Device", :cl_uint, "PCI_BUS_ID_NV")
> end
=> :pci_bus_id_nv
> OpenCL::platforms.first.devices.map(&:pci_bus_id_nv)
=> [5, 130]
> OpenCL::platforms.first.devices.first.pci_bus_id_nv
=> 5

Which correspond to the list obtained through nvidia-smi:

Mon Nov 16 10:48:08 2015      
+------------------------------------------------------+                      
| NVIDIA-SMI 340.32     Driver Version: 340.32         |                      
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla K40c          Off  | 0000:05:00.0     Off |                    0 |
| 23%   34C    P8    20W / 235W |     23MiB / 11519MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla K40c          Off  | 0000:82:00.0     Off |                    0 |
| 23%   36C    P8    20W / 235W |     23MiB / 11519MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Compute processes:                                               GPU Memory |
|  GPU       PID  Process name                                     Usage      |
|=============================================================================|
|  No running compute processes found                                         |
+-----------------------------------------------------------------------------+

0x82 == 130

Note that they are returned in the same order that nvidia-smi does.

Don't hesitate if you have more questions.

Best regards,

Brice

from opencl-ruby.

Brilliant! Thanks for your help, this is exactly what I was trying to
achieve.

Best regards,
Jorge

On Mon, 2015-11-16 at 01:52 -0800, Brice Videau wrote:

Hello,

The answer is double.

From my experience, an implementation, on the same
machine/reboot/configuration always returns it's devices in the same
order, the enumerator is undefined but not random (in my experience).
So
I use the order of the devices returned by clGetDeviceIDs to
differentiate them. In mpi a modulo on mpi rank usually does the
trick
(provided process on the same node have consecutive ranks).

Should you need to identify precisely the device to track it across
reboot, driver version and so on you can easily dos so in
opencl_ruby_ffi. Here is an irb example:

require 'opencl_ruby_ffi'
=> true
OpenCL::platforms.first.devices.map(&:name)
=> ["Tesla K40c", "Tesla K40c"]
class OpenCL::Device
PCI_BUS_ID_NV = 0x4008
eval get_info("Device", :cl_uint, "PCI_BUS_ID_NV")
end
=> :pci_bus_id_nv
OpenCL::platforms.first.devices.map(&:pci_bus_id_nv)
=> [5, 130]

Which correspond to the list obtained through nvidia-smi:
Mon Nov 16 10:48:08 2015
+------------------------------------------------------+
| NVIDIA-SMI 340.32 Driver Version: 340.32 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile
Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util
Compute M. |
|===============================+======================
+======================|
| 0 Tesla K40c Off | 0000:05:00.0 Off
| 0 |
| 23% 34C P8 20W / 235W | 23MiB / 11519MiB | 0%
Default |
+-------------------------------+----------------------+----------------------+
| 1 Tesla K40c Off | 0000:82:00.0 Off
| 0 |
| 23% 36C P8 20W / 235W | 23MiB / 11519MiB | 0%
Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Compute processes: GPU
Memory |
| GPU PID Process name
Usage |
|
=============================================================================|
| No running compute processes
found |
+-----------------------------------------------------------------------------+

0x82 == 130

Note that they are returned in the same order that nvidia-smi does.

Don't hesitate if you have more questions.

Best regards,

Brice

On 16/11/2015 07:55, mtbikerjc wrote:

Hello,

I'm trying to create a load balancer that can dispatch tasks to
various GPUs if they're present. To do this, I need to launch one
process per GPU device, so that this process can dispatch jobs to
its
GPU. However, OpenCL itself doesn't have a way of providing a
unique
ID that is unique to a given GPU device. It only gives you
information
on the "type" of device or the OpenCL version number, which doesn't
help me. I did some searching around, and saw that for NVidia
devices,
the device-unique IDs need to be queried using the
CL_DEVICE_PCI_BUS_ID_NV token of the cl_nv_device_a ttribute_query
extension.

I found that here: https://anteru.net/2014/08/01/2483/

Is there a way to do this with opencl-ruby? I can see the
cl_nv_device_attribute_query extension when I run
"device.extensions"
but can't seem to find a way to query the
cl_nv_device_attribute_query
extension.

Thanks for your help!

• Jorge

—
Reply to this email directly or view it on GitHub
#8.

—
Reply to this email directly or view it on GitHub.

from opencl-ruby.

In next version ( > 1.3.0 ) cl_nv_device_attribute_query extension (and many others) will be directly supported. If you query a device that does not support it you will raise the following

> require 'opencl_ruby_ffi'
=> true
> OpenCL::platforms.last.devices.map(&:name)
=> ["Intel(R) Core(TM) i7-2760QM CPU @ 2.40GHz"]
> OpenCL::platforms.last.devices.map(&:pci_bus_id_nv)
OpenCL::Error::INVALID_VALUE: -30
    from /var/lib/gems/2.3.0/gems/opencl_ruby_ffi-1.2.2/lib/opencl_ruby_ffi/opencl_ruby_ffi_base.rb:237:in `error_check'
    from (eval):4:in `pci_bus_id_nv'
    from (irb):4:in `map'
    from (irb):4
    from /usr/bin/irb:11:in `<main>'

If you still need to add support for an extension the syntax would be slightly changed (using the previous and obsolete example):

> require 'opencl_ruby_ffi'
=> true
> OpenCL::platforms.first.devices.map(&:name)
=> ["Tesla K40c", "Tesla K40c"]
> module OpenCL
>   using OpenCLRefinements if RUBY_VERSION.scan(/\d+/).collect(&:to_i).first >= 2
>   class Device
>     PCI_BUS_ID_NV_NEW = 0x4008
>     eval get_info("Device", :cl_uint, "PCI_BUS_ID_NV_NEW")
>   end
> end
=> :pci_bus_id_nv_new
> OpenCL::platforms.first.devices.map(&:pci_bus_id_nv_new)
=> [5, 130]
> OpenCL::platforms.first.devices.first.pci_bus_id_nv_new
=> 5

Regards,

Brice

from opencl-ruby.

Major thanks for the heads up!  I'll implement these changes when I
update my gem.

Cheers,
JorgeOn Mon, 2016-05-09 at 09:15 -0700, Brice Videau wrote:

In next version ( > 1.3.0 ) cl_nv_device_attribute_query extension
(and many others) will be directly supported. If you query a device
that does not support it you will raise the following

require 'opencl_ruby_ffi'
=> true
OpenCL::platforms.last.devices.map(&:name)
=> ["Intel(R) Core(TM) i7-2760QM CPU @ 2.40GHz"]
OpenCL::platforms.last.devices.map(&:pci_bus_id_nv)
OpenCL::Error::INVALID_VALUE: -30
    from /var/lib/gems/2.3.0/gems/opencl_ruby_ffi-
1.2.2/lib/opencl_ruby_ffi/opencl_ruby_ffi_base.rb:237:in
error_check'     from (eval):4:inpci_bus_id_nv'
    from (irb):4:in map'     from (irb):4     from /usr/bin/irb:11:in

'
If you still need to add support for an extension the syntax would be
slightly changed (using the previous and obsolete example):
require 'opencl_ruby_ffi'
=> true
OpenCL::platforms.first.devices.map(&:name)
=> ["Tesla K40c", "Tesla K40c"]
module OpenCL
   using OpenCLRefinements if
RUBY_VERSION.scan(/\d+/).collect(&:to_i).first >= 2
   class Device
     PCI_BUS_ID_NV_NEW = 0x4008
     eval get_info("Device", :cl_uint, "PCI_BUS_ID_NV_NEW")
   end
end
=> :pci_bus_id_nv_new
OpenCL::platforms.first.devices.map(&:pci_bus_id_nv_new)
=> [5, 130]
OpenCL::platforms.first.devices.first.pci_bus_id_nv_new
=> 5
Regards,
Brice
—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub

from opencl-ruby.

Final version:

> require 'opencl_ruby_ffi'
=> true
> class OpenCL::Device
>   using OpenCLRefinements if RUBY_VERSION.scan(/\d+/).collect(&:to_i).first >= 2
>   PCI_BUS_ID_NV_NEW = 0x4008
>   get_info("Device", :cl_uint, "PCI_BUS_ID_NV_NEW")
> end
=> :pci_bus_id_nv_new
> OpenCL::platforms.first.devices.map(&:pci_bus_id_nv_new)
=> [5, 130]

eval is not needed anymore.

from opencl-ruby.