J784S4XEVM: YUV422 to YUV420 Conversion with Zero-Copy GPU Rendering

You can probably use DSS in m2m mode to do this conversion.

You can refer to the following faq for using display_m2m node in multi_cam demo,
[FAQ] TDA4VM: How to use DSS M2M for format conversion?
Patch for sdk 10.1  e2e.ti.com/.../displaym2m_5F00_multi_5F00_cam_5F00_demo.patch

Omkar Bhandari 说：

We have a GPU-based render node. What is the recommended approach to share YUV buffers using zero-copy(no CPU memcpy) between the capture/convert node and the GPU render node?

You can use the dma-buf-fd to share the buffers in render-node. We have a surround view demo that does 360 degree rendering from 4 cameras,
https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/11_01_00_04/exports/docs/vision_apps/docs/user_guide/group_apps_srv_demos_app_srv_camera.html
Node implementation can be found inside $(psdkra)/vision_apps/kernels/srv/gpu/vx_gl_srv_target.c

Omkar Bhandari 说：

Additionally, if we want to run a computer vision algorithm in parallel using another node ,how can we design the pipeline so that both the render node and the CV node run in parallel, preferably still using zero-copy buffer sharing?

Can you give more info about the cv node, is it running on a72 or other hardware accelerators ?

You can create openvx graph to run both nodes in parallel provided they process different frames, 
for example when render node is processing frame 0, CV node can process frame 1 at same time.
Refer to the vision_apps demos from rtos sdk.