TDA4VM: TIOVX Graph with 2 nodes that set the same output

Hi Jeremie,

The output of two nodes cannot be same reference. This can create cyclic dependency, which is not allowed in OpenVX. 

Regards,

Brijesh

Here's an example of our problem. We have one graph, which contains 2 nodes.

First node has an output parameter, xyz tensor. The node sets xyz tensor to 0.

Second node has two inputs parameter, abc and def tensors and one output parameter, xyz tensor. The node sets xyz tensor with the sum of abc and def tensors.

(Here we assume abc, def and xyz tensors are the same dimensions).

Hi Brijesh, is there a workaround to this? Having the same buffer pass through multiple nodes seems like a common usecase.

Hi FredC_LT,

But instead of xyz tensor, why can't we create abc tensor for the node-2's output? and then whatever the node uses these tensors as input, it can use both xyz and abc tensor as input parameters.

Regards,

Brijesh 

That's a very big tensor to duplicate for every node.

Here's the full flow

XYZ is really an input/output of Process A, and XYZ_T is really an input/output of Process B, but because TIOVX does not support VX_BIDIRECTIONAL for custom user kernels, we need to set XYZ as output for ProcessA et XYZ_T as input for ProcessB.

The graph works without the "Init to zero" node. When we add it is when it fails at validation.

If we change XYZ to be an input of ProcessA, then we get the risk of ProcessA and Transpose being executed out of order.

What would be the TIVX way of ensuring the right sequence?

Hi FredC_LT,

I think VX_BIDIRECTIONAL is supported in the OpenVX. I see this type is being checked in the core OpenVX framework. Could you help me understand where you see issue in using this type? 

Looking at the check again, there is a comment on this check, which says

/* check if any output of next node matches current node
* This would mean two nodes output to same data object which is not allowed
*/

The output of the next node cannot be same as the output of the current node, this is not allowed in the TIOVX.

Regards,

Brijesh

Brijesh Jadav said:

allowed

It's specifically for TIVX user kernels that TIVX does not support it. Look up

vx_kernel.c -> vxAddParameterToKernel() -> ownIsValidDirection()

ok, let me check with the team why bidirectional parameters are not supported. 

Hello FredC_LT,

The bidirectional parameter is not supported in the user kernel due to some issue especially in the pipeline mode. We are looking into it and it is going to be fixed in next or two release. 

Regards,

Brijesh

Great, thank you. In the meantime, can you suggest a workaround for our issue?

Yeah, that's what we went with.

Will look forward to next releases, thanks.

ok, then, let us know if there is any further questions or if we can close this ticket.