J784S4XEVM: some question about openvx pipeline depth

Hi,

rz liu said:

The TIOVX user guide has a description of the pipeline depth, but I don't understand why does it need multiple graph instance, it is only one node in an entity can run at any time?

Here, pipelining means that a second frame can not start graph execution on this same graph until the first graph execution is completed. 

Hence in order to increase the throughput, we would have to consider multiple instances of the graph so that once the node has processed the frame, it could take in the second frame and not wait until all the nodes completes the first frame.

You could refer the below documentation as well

The OpenVX Graph Pipelining, Streaming, and Batch Processing Extension to OpenVX 1.1 and 1.2 (khronos.org)

Regards,

Nikhil

Hi,

I did not understand:

    Here, pipelining means that a second frame can not start graph execution on this same graph until the first graph execution is completed. 

if there are 3 Nodes in graph, could it work like this:

On T1:
  +---------+      +---------+      +---------+
  |         |      |         |      |         |
  | frame 3 |----->| frame 2 |----->| frame 1 |
  |         |      |         |      |         |
  +---------+      +---------+      +---------+
   Node1(A72)       Node2(C71)       Node3(MCU)
   
On T2:
  +---------+      +---------+      +---------+
  |         |      |         |      |         |
  | frame 4 |----->| frame 3 |----->| frame 2 |
  |         |      |         |      |         |
  +---------+      +---------+      +---------+
   Node1(A72)       Node2(C71)       Node3(MCU)
   

if it could be work, it need one graph instance only, any problem?

Thanks.

Hi,

That doesn't sound particularly perfect! Will it waste memory space? Why not TI improve its implementation based on OpenVX function interface ? 

Thanks.

Hi,

you said:

To use the analogy of an assembly line, you could consider each factory worker as a separate node. In the case of 3 nodes, there is still only 3 factory workers whether you pipeline or not.  The pipeline depth can be something similar to a container of the item being worked on.  If you only work on 1 item at a time, you only need one container, and when the item is done, you can reuse it for the next item. If you want to work on 3 at a time, you need 3 containers, reusing the one that finished each time for the next one to start. This container is like a graph context that we call a graph instance.

But a graph instance looks more like a assembly lines, it has its own nodes and buffers , Maybe my understanding is wrong.

Thanks.

Hi,

This is exactly my question. There is only one worker busy on an assembly line, and the others are idle and waiting. Isn’t it a waste of resources?

Thansk.

Hi

Nikhil Dasan said:

the graph can be retriggered only once all the nodes are done with the execution

This is the current implementation of a graph as per the OpenVX standard where it cannot be retriggered unless it is completed all the nodes.
Hence, to extend this implementation to pipelining mode, (as node level triggering is not available), TI's implementation was to extend it to multiple graph instances. 

Hence, scaling from non-pipelining to pipelining by replicating the graph/node objects is a scalable and efficient way of using existing logic with minimal memory increase which is currently supported in the SDK.

Regards,

Nikhil