continuous build

Builds are built for CC 3.0 (real+virtual) by default. Linux "all CUDA architectures" build for all real and highest virtual architecture as supported by most recent CUDA toolkit (so the minimal CC is 5.0 as for CUDA 12). all can be used if application start-up time is critical (CUDA runtime doesn't need to compile the kernels for current architecture), otherwise it may not provide any advantage.

GPUJPEG 0.26.0

This release brings mainly much faster start-up (and reinitialization) times in Linux, which is significant especially when encoding small amount of images. The speed-up for first frame is as much as 20x (eg. 43 ms for 10000x10000.tst compared to original 810 ms). Subsequent images are still faster but this narrows the gap. The improvement is especially noticeable for big images.

Builds require CUDA compute capability at least 3.5.

API changes

• [changed] log level version numbers (verbose + debug +1)
• [added] log levels symbolic names
• [renamed] GPUJPEG_VERBOSE to GPUJPEG_INIT_DEV_VERBOSE
• [added] struct gpujpeg_encoder_input init functions returning the struct (like gpujpeg_encoder_input_gpu_image) to be able to directly initialize the variable
• if verbosity is at least GPUJPEG_LL_STATUS, gpujpeg_parameters.perf_stats
  doesn't need to be set to output coding duration

Other

• new patterns (noise, blank) to test images (.tst extension)
• cmake - add hint to enable the architecture to native (to speed up startup)
• improved logging of JPEG reader in debug mode to inspect JPEG structure
• print also coded image size and properties (useful if autodeduced)
• report also (re)initialization duration (in verbose mode)