NVJPEG Decode acceleration, libnvjpeg, and NVJPEG_BACKEND_HARDWARE

Ostensibly, the Jetson AGX Orin supports JPEG decode acceleration, according to the data sheet Section 4.7. We are seeing insufficient performance using the libnvjpeg library with NVJPEG_BACKEND_DEFAULT, so we tried NVJPEG_BACKEND_HARDWARE but this yields a NVJPEG_STATUS_ARCH_MISMATCH error. What gives?

Hi,
Please refer to the sample for using hardware JPEG decoding:

/usr/src/jetson_multimedia_api/samples/06_jpeg_decode
Jetson Linux API Reference: 06_jpeg_decode (JPEG decode) | NVIDIA Docs

Wow, thank you, that is really key. For those playing along at home: There are two libnvjpeg.so libraries, one in CUDA, which I was attempting to use, and one specific to AGX Orin/Tegra which employs acceleration (/usr/lib/aarch64-linux-gnu/libnvjpeg.so), which has completely different entry points and semantics. This library appears to be undocumented beyond the sample code.

Hi,
We are trying to unify the interface between dGPU and Jetson platforms. This is a long term plan and may take some time. For now please user the interface for Jetson platforms.