Why cudaHostRegister failed when pinned memory is larger than 4GB

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Xavier
1

In Xavier platform, I used cudaHostRegister to register pinned memory, but it failed when the size is over 4GB.
Why?
Somebody help me, thanks very much!

2

I have the same problem， someone help please！

3

Hi,

Thanks for the feedback.

We can reproduce this issue and is going to check this with our internal team.
Will update with you once we have further information.

4

Hi,

Sorry for keeping you waiting.
We are still checking this issue and will update information with you once we got any feedback.

Thanks.

5

Hi,

How is that going?
I have waiting for more than 3 months.

Please feedback.

6

Hi,

Thanks for your patience.
Unfortunately, we still need some time for this.

Sorry for the inconvenience.

7

This is an old ticket, but there was never any reply and this issue apparently still exists. I can only map about 2.7GB with cudaHostRegister on an AGX Xavier. After that, I get errors such as “nvgpu_vm_map_buffer:311 [ERR] buf size a0c7f000 < (offset(0) + map_size(200000000))” when attempting to map more memory.

Are there any workarounds or fixes such as kernel parameters, kernel module parameters, devicetree settings, API calls, etc. that could be used to increase the address space available to the GPU?

9

Hi

It’s recommended to file a new topic for your problem.
The original issue is fixed in 2021.

Thanks