CUDA Interop

Overview

ovstream consumes raw video frames as BGRA8 CUDA buffers directly from device memory. No CPU copy, no upload — the SDK reads the GPU pointer you hand it and encodes via NVENC.

The frame descriptor (VideoFrame / ovstream_video_frame_t) holds:

• buffer — a CUDA device pointer.
• width, height — pixel dimensions.
• pitch_bytes — row stride in bytes (≥ width * 4).

Allocate with cudaMallocPitch

NVENC tolerates the buffer being row-padded for DMA alignment. Let CUDA pick the pitch:

Source: examples/c/basic_stream/main.cu snippet cuda-buffer-alloc

cudaMallocPitch(&buf, &pitch, widthBytes, height) returns a device pointer plus the actual pitch CUDA picked (often width * 4 rounded up to a 256 / 512-byte multiple). Pass that exact pitch into the VideoFrame.pitch_bytes field.