TDA4VM: Custom Node

Dear Sir, 

Thanks for the response!

X Node logic is the same for all cameras. Only the different LUTs will be making changes in the output.

As per the understanding, 

the object_array  0th index generally has all num_channel input/output.

Query 1: Is it possible to manually fill the LUT Object array at the 0th index for all 4 cameras LUT?

Query 2:

If my input(LUT) is an Object array, filled with respective LUTs at index(0) for all 4 channels(num_channel) and then its 0 index is passed to X node including the TIDL output [0]( as done normally for all replicated nodes), will my X node will be replicated?

Please let me know your opinion on this.

Kindly confirm the same.

Thanks and Regards,

Vyom Mishra

Hi Vyom,

Your response somehow got deleted from your end. 

Pasting the same here 

Dear Sir,

I have integrated the X Node with the application, but I am not able to get the Process function running of the custom Node (X Node).

I have kept a print inside the kernel process function, but it is not printing on the console.

What could be the possible reason for Node not running?

Note:

- No errors on the logs

- Application running in real-time with no issue 

Thanks and Regards,

Vyom Mishra

May I know which core is X Node running ON?

Is the application running?

Do you get logs from the create function of X node?

Regards,

Nikhil

Dear Sir,

I was able to resolve the issue, so I deleted the query.

Thanks for your time.

However, I have a query related to the timing of the custom kernel.

My custom kernel is running on A72 and I want to optimize the same.

Input to the kernel is TIDL output and LUTs for each channel.

I have stored LUTs using copyImagePatch 

vxCopyImagePatch(mesh_img_in_arg_right,
	               	                         &rect, 0,
	               	                         &image_addr,
	               	                         (vx_uint32*)lutTable_right,
	               	                         VX_WRITE_ONLY,
	               	                         VX_MEMORY_TYPE_HOST);
	               	    

I assume the host memory is in DDR and the kernel is running on A72 so memory reading might consume more time.

Please let me know if my understanding is correct, if yes could you please suggest how to write it to A72 memory or the other best possible way to reduce the total time of the custom Replicated Node?

Thanks and Regards,

Vyom Mishra

Dear Sir,

Thanks for the response!

Your suggestions help a lot!

My custom node profiling time is 34ms approx, I would like to run the Custom Node on GPU.

I would like to know, how(API similar to vxCopyImagePatch),  I will be uploading the LUTs(once) so that the custom node can access the data while running on GPU.

If any reference is present, please let me know.

Thanks and Regards,

Vyom Mishra

Hi,

Vyom Mishra1 said:

I would like to know, how(API similar to vxCopyImagePatch),  I will be uploading the LUTs(once) so that the custom node can access the data while running on GPU.

By running a node on GPU, do you mean to use OpenGL APIs and access the GPU and render the buffers using the same?

If yes, tivxGlSrvNode() which uses OpenGL APIs internally. You could refer the target implementation of this node in vision_apps/kernels/srv/gpu/vx_gl_srv_target.c

This node is being utilized in the surround view demos in vision_apps/apps/srv_demos/app_srv_camera

Regards,

Nikhil