VART HW Backend - VART HW Backend - 6.3 English - XD056

The input buffer must already contain the model tensor in the required format, dimensions, and layout. No resize, color conversion, layout conversion, or normalization is performed by vvas_xinfer.

Supported input memory:

• compatible GstVvasMemory from an upstream scaler or tensor producer;

• generic DMA-BUF that can be imported on the inference device;

• system memory - vvas_xinfer_ensure_xrt_input_buffer() lazily allocates an internal VVAS pool on the inference device (XDNA_DEVICE_IDX, vart_info.mbank_idx, stride_align = 1) and vvas_xinfer_copy_input_buffer() memcpy’s the sysmem frame into a buffer drawn from that pool. VVAS_SYNC_TO_DEVICE is set and gst_vvas_memory_sync_bo() flushes before the VART runner reads. Same mechanism as the PPE-enabled sysmem path, just on the inference device. The buffer payload must already match the model tensor format, dimensions, and layout; no resize, conversion, or normalization is applied on this path.

For VVAS-memory input, prepare_infer_input_frame() extracts the allocator context and memory bank from the input buffer. The bank is passed to the inference path; vart-config.in-mem-bank does not gate this path.

For generic DMA-BUF input, vvas_videoframe_from_gstbuffer() imports the DMA-BUF as an XRT BO on the inference device.