About copy-time for cuda in TX1

Robotics & Edge Computing Jetson Systems Jetson TX1
1

Is there any other way to improve the performance CUDA program in TX1?
When I use cudaMemcpy or cudaMemcpyManaged to copy data from host to device (or device to host), it shows not very good.
When I use Zero-Copy, it can decrease copy time, but running rime in cuda increase, and the total is changeless.

In addition, how many cameras would be supported on TX1(We use v4l2 get buffer ) at the same time ? Is there any limitation ?

Any answer is welcome, thanks.

2

Hi,

It’s recommended to use unified memory(cudaMallocManaged()).
You don’t need to apply memory copy for it since CUDA driver will handler the synchronize for you.

Thanks.

3

Thanks.
But How should I use the cuda stream if I don’t use the cudaHostAlloc to specify a pinned memory? Or can I create cuda stream with cudaMallocManaged function ?(cudaStreamCreate,cudaMemcpyAsync, etc )

4

Hi,

You can create cuda stream with function like this:

cudaStream_t stream;
cudaStreamCreate(&stream);

CUDA stream works well for both memory type. Here is our tutorial for your reference:
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-unified-memory-programming-hd

Thanks.

5

That make sense, thanks a lot !