Unified Memory Access using Jetson TX2

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Hi, I’m testing cuda program using Jetson TX2

But there is problem when I try to use ‘unified memory access’
(It occurs segmentation fault)

So, I query the property of Jetson TX2’s gpu, and I found that It doesn’t support ‘concurrentManagedAccess’ even though it is Pascal architecture.

How can I use unified memory access with ‘concurrentManagedAccess’ property

I’m using Jetpack 3.2 and cuda 9.0

2

Hi,

For TX2, unified memory requires exclusive access or will lead to a segmentation fault.
Please avoid concurrent CPU/GPU accesses with cudaDeviceSynchronize().

You can find more information in our document here:
http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-gpu-exclusive

Thanks.

3

Hello,
I’m also having this issue (Although with a Jetson TX1 rather than TX2).

The documentation states that cudaStreamSynchronize can also be used for synchronisation.

I still get the bus error using cudaStreamSynchronize, but I don’t get it using cudaDeviceSynchronize.

Unfortunately cudaDeviceSynchronize is not suitable for my application as I wish to update the CUDA buffer from a CPU thread while the GPU works on other streams.

(Note: I’m using the Driver API so the calls for me are actually cuCtxSynchronize and cuStreamSynchronize.)

Are there any other gotchas to be aware of?

Thanks
Russell

4

Just to update, I’ve managed to get it working with cuStreamSynchronize now.
I think it was just my call to cuMemAllocManaged needed the flag CU_MEM_ATTACH_HOST to be accessible from CPU.

5

Hi PeartreeStudios,

Could you please clarify whether you are able to access cuMemAllocaManaged buffer when GPU is still active ?

I am looking to understand meaning of what you said, vs what document says. For example

https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__MEM.html#group__CUDA__MEM_1gb347ded34dc326af404aa02af5388a32

says -

"If CU_MEM_ATTACH_HOST is specified, then the allocation should not be accessed from devices that have a zero value for the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS; an explicit call to cuStreamAttachMemAsync will be required to enable access on such devices. "

Did you also do a cuStreamAttachMemAsync() to enable concurrent access ?

Thanks