Allocator Proposal - Allocator Proposal - 6.3 English - XD056

If the allocation query already carries allocator parameters, vvas_xinfer preserves them and contributes only the pool. If the allocator is not a GstVvasAllocator, PPE-enabled mode copies each input frame into an internal VVAS pool before running the HW PPE kernel.

When HW PPE is enabled, vvas_xinfer proposes:

• GstVvasAllocator on the preprocess device;

• allocation on config in-mem-bank;

• DMA-BUF export enabled;

• stride-aligned GstVvasBufferPool;

• video alignment metadata for upstream stride handling.

When PPE is disabled and the backend is VART, vvas_xinfer proposes:

• GstVvasAllocator on the inference device;

• DMA-BUF export enabled;

• GstVvasBufferPool without PPE stride constraints.

For SW preprocessing and ONNX Runtime, vvas_xinfer proposes no custom allocator and uses a plain GstVideoBufferPool.

Additional query rules:

• proposed pool size is batch_size + 1;

• GST_VIDEO_META_API_TYPE is added to the query;

• PPE input zero-copy requires matching device index and memory bank;

• no-PPE inference uses the bank carried by the upstream allocator.