How to get the number of connected displays to a gpu on Linux?

Question

I need to determine whether a given CUDA device has displays connected or not. I know no CUDA function which does this.

On Windows, I can use NVAPI to get the number of connected displays and the PCI bus/slot id of each device. Using the latter, I can find the matching CUDA device (by calling cudaGetDeviceProperties).

How can I do the same on Linux, where NVAPI is not available?

Technically, what I need is a Linux alternative to the following code:

NvAPI_Initialize();

NvPhysicalGpuHandle gpuHandles[64];
NvU32 numOfGPUs;
NvAPI_EnumPhysicalGPUs(gpuHandles, &numOfGPUs);

for (int i = 0; i < numOfGPUs; i++)
{
    NvU32 connected_displays = 0;
    NvU32 busId = 0;
    NvU32 busSlotId = 0;

    NvAPI_GPU_GetConnectedDisplayIds(gpuHandles[i], NULL, &connected_displays, NULL);
    NvAPI_GPU_GetBusId(gpuHandles[i], &busId);
    NvAPI_GPU_GetBusSlotId(gpuHandles[i], &busSlotId);

    printf("Current device: %d\n", i);
    printf("Number of connected displays: %u\n", connected_displays);
    printf("Bus id: %u\tBus slot id: %u\n", busId, busSlotId);
}

NvAPI_Unload();

Perhaps use the `lsof` command. Or some deep thing under `/proc/`. — Basile Starynkevitch, Aug 01 '13 at 13:13
Can you detect if monito is on/off with cuda? So bassically when you have the id, you could call another function maybe ... !? — alap, Oct 14 '14 at 08:14
@Laszlo-AndrasZsurzsa I don't think that has anything to do with cuda or the GPU. — hthms, Oct 28 '14 at 07:53

Robert Crovella · Accepted Answer · 2013-08-01T14:31:01.060

The most analogous approach under Linux would be to use the NVCtrl API which is what nvidia-settings, the linux NVIDIA control panel application, provides.

How to download the nvidia-settings source package is documented in the linux driver release notes. Specifically, you can find various packages for specific driver versions here

Choose a package that is closest to your driver version.

Once you have downloaded and unzipped the nvidia-settings source, you will find a samples directory. In that directory is a sample program which has the necessary framework for what you want. Specifically, look in nv-control-targets.c. The following function in that file will do what you want:

    /* Connected Display Devices on GPU */

    ret = XNVCTRLQueryTargetAttribute(dpy,
                                      NV_CTRL_TARGET_TYPE_GPU,
                                      gpu, // target_id
                                      0, // display_mask
                                      NV_CTRL_CONNECTED_DISPLAYS,
                                      &display_devices);
    if (!ret) {
        fprintf(stderr, "Failed to query connected displays\n");
        return 1;
    }
    printf("   Display Device Mask (Connected) : 0x%08x\n",
           display_devices);

Note that there are some preparatory/setup function calls at the top of that program (nv-control-targets.c) that will need to be executed as well.

There is also a function (Display Mode) in NVML (nvidia-smi is based on NVML) that will inform you of whether the GPU is hosting a display or not, but I'm not sure it gives you the granularity you want.

Actually, upon re-reading your question, NVML Display Mode might be sufficient for what you want. Referring to the API documentation here, p46:

7.10.2.10 nvmlReturn_t DECLDIR nvmlDeviceGetDisplayMode (nvmlDevice_t device, nvmlEnableState_t 
display)
Retrieves the display mode for the device.
For Tesla ™and Quadro ®products from the Fermi and Kepler families.
This method indicates whether a physical display is currently connected to the device.
See nvmlEnableState_t for details on allowed modes.
Parameters:
device The identifier of the target device
display Reference in which to return the display mode
Returns:
• NVML_SUCCESS if display has been set
• NVML_ERROR_UNINITIALIZED if the library has not been successfully initialized
• NVML_ERROR_INVALID_ARGUMENT if device is invalid or display is NULL
• NVML_ERROR_NOT_SUPPORTED if the device does not support this feature

cudaDevAttrKernelExecTimeout is usually a pretty good guide. If there is a tiemout, that means X11 is managing the device. — talonmies, Aug 01 '13 at 14:58
Method 2 is only for Tesla an Quadro, so unfortunately I cannot use it. Method 1 works fine, but it seems I can only use from the machine where the X server is running and not through a remote terminal. — hthms, Aug 02 '13 at 08:30
For method 1, it's possible you are running into an X server permissions issue. Are you running the application as root user? — Robert Crovella, Aug 02 '13 at 12:11

How to get the number of connected displays to a gpu on Linux?

1 Answers1