 |
List of SVN commits: r9232 Improve image load times by parallelizing the load from disk and using glTexStorage in place of glTexImage r9233 Prefetch the entire LWO model file to avoid seeks in a ZIP file r9234 Actually support overriding parallelism in job queue submission r9235 Remove unneeded functionality from idFile_Memory r9236 Check for GL_ARB_texture_storage r9237 Use a separate utility job list for parallel image loading instead of reusing the one for frontend jobs. Also gets rid of warning about jobs taking too long r9238 Add cvar to potentially disable use of glTexStorage The actual improvements include: 1) Use glTexStorage instead of glTexImage --- if GL_ARB_texture_storage is supported and cvar image_useTexStorage is 1. 2) Load data for next 16 images in two background threads, while uploading the current 16 images to GPU simultaneously --- if cvar image_levelLoadParallel is 1. 3) When loading LWO model, unzip the whole file into memory buffer, and do LWO parsing from it. |
|
|
There are some concerns about how HDD handle parallel loading. For that reason the number of background threads for loading images is only 2, even if player has six-core CPU. Watch out for how HDDs handle this =) |
|
|
It turned out that AMD drivers crash on TexSubImage of compressed texture if its size is not divisible by 4. So here is one small fix: r9579. Do not use TexStorage + TexSubImage for compressed textures of weird sizes. See also: https://forums.thedarkmod.com/index.php?/topic/21073-a-house-of-locked-secrets-crashes-tdm-on-startup/ |
- Anonymous
