HMM support in driver

Graphics / Linux Linux
1

[ The CUDA Setup forum might be more appropriate for this question. Please comment there: https://devtalk.nvidia.com/default/topic/1015146/cuda-setup-and-installation/hmm-support-in-linux-driver/post/5172050/#5172050 ]

I have installed Jerome Glisse’s kernel with HMM support (https://cgit.freedesktop.org/~glisse/linux/?h=hmm-v23) (see also https://lkml.org/lkml/2017/5/24/731).

I’m using driver 381.22. Should I expect this driver to be able to use the HMM capabilities of the kernel?

float *gpuData = static_cast<float*>( malloc( sizeof(float)*size));
CUDA_CHECK_RETURN(cudaMemcpy(gpuData, data, sizeof(float)*size, cudaMemcpyHostToDevice));

The above code throws an error. I think if the driver were HMM-aware the above code would succeed. Is that correct?

And assuming the answer is “the 381.22 driver is not HMM-aware”, is there a HMM-aware driver I might use?

Thank you!

Edit: I’m using a GTX 1080Ti

2

Thanks for reporting this issue. We would like to reproduce for investigation. OR Track this support or bug.
https://devtalk.nvidia.com/default/topic/1015146/cuda-setup-and-installation/hmm-support-in-linux-driver/post/5172573/#5172573

What option did you used to compile this test program? Is it .c , .cpp , .cu program?

>>It succeeds if you set HMM = 0; it fails if HMM = 1.
What is expected and unexpected results when compiling and executing this test program?

Can I get nvidia bug report of your system?

3

Hi,

Did anyone successfully run HMM in NVIDIA driver and Linux with the latest NVIDIA driver on Pascal or other GPUs? If you did, which GPU/ driver version/ Linux version did you use?

Thanks,
Kevin