The Tegra X1 provides a hardware module dedicated to JPEG encoding/decoding (NVJPG).
Using this library, images can be rendered to an RGB or YUV (triplanar) surface. YUV➝RGB conversion is handled in hardware. In addition, images can be downscaled to up to 8, also done in hardware.
Note: only baseline JPEGs are supported. Progressive and arithmetic coded files will return an error.
Decoding times are largely faster than those obtained by software rendering, even with SIMD optimizations. The results below were obtained on an Aarch64 Cortex-A57 clocked at max 2.091GHz, by decoding the same image to an RGB surface, averaging 1000 iterations:
| Library | 70Kib 720x1080 | 1.6Mib 3200x1800 |
|---|---|---|
| NVJPG | 2.036ms | 15.997ms |
| libjpeg-turbo | 6.127ms | 66.341ms |
| stb_image (with NEON routines) | 16.512ms | 158.474ms |
| Python-pillow | 11.688ms | 114.918ms |
| nanoJPEG | 48.299ms | 531.575ms |
Requires C++20 support.
meson build && meson compile -C buildAdditionally, run meson compile -C build examples to build the examples.
make -j$(nproc)Additionally, run make examples to build the examples.
- The Ryujinx project for extensive documentation of the Host1x interface