Image & Video Processing SDK on NVIDIA Tegra X2
Updated:
NVIDIA Tegra X2 pairs a Pascal-class mobile GPU with low power draw, which lets PC-class imaging run on embedded devices. We ported the Fastvideo Image & Video Processing SDK to Tegra, so the same GPU debayer, JPEG, JPEG2000, resize and denoise modules run on the mobile GPU for real-time imaging and video.
Tegra X2 benchmarks (selected SDK kernels)
Averaged kernel times at 2K (1920×1080) and 4K (3840×2160). This is a small subset of the SDK:
| Operation | 2K | 4K |
|---|---|---|
| HQLI debayer (16-bit) | 0.66 ms | 2.6 ms |
| DFPD debayer (16-bit) | 1.9 ms | 7.5 ms |
| MG debayer, no enhance (16-bit) | 6.1 ms | 25 ms |
| JPEG encode (24-bit, 4:2:0, q90) | 2.0 ms | 7.9 ms |
| JPEG decode (24-bit, 4:2:0, q90) | 5.3 ms | 19.5 ms |
| Resize (24-bit, Lanczos3, to quarter area) | 4.9 ms | 17.7 ms |
| Denoiser (24-bit, wavelet 9/7) | 11.3 ms | 43 ms |
| JPEG2000 encode (24-bit, lossy, single) | 57 ms | 184 ms |
These modules are part of the Fastvideo SDK; the same code base also runs on newer Jetson devices (Xavier, Orin).
Roadmap
- Multi-camera embedded system on Tegra X2 – done
- Fastvideo SDK benchmarks on Tegra X2 – done
- CUDA JPEG2000 encoder and decoder – done