This change significantly improves the compression speed for both DXT and ETC encodings.
Explanation:
On each iteration of the vector quantization algorithm, the leaf with the highest variance is selected for splitting. If the leaf gets split, then two new leaves are created (while the leaves that can not be split will be ignored on the future iterations). There does not seem to be any simple way to compute or reliably predict the variances of the future leaves in advance, which means that there is no simple way to efficiently perform split operations in parallel.
And still, there is an interesting observation. Even though the order of the split operations depends on the previous iterations, the split operations performed in different subtrees are completely independent. So what if instead of solving the main quantization task we will first solve an alternative quantization task, which has a lot in common with the main task, but at the same time can be efficiently parallelized. Then the intermediate computation results of the alternative solution can be reused when solving the main task. Specifically, the idea is to efficiently compute an alternative split tree, which is more or less balanced, and has approximately the same number of nodes as the main tree. Then the overlapping part of the main and alternative trees can be reused while solving the main quantization task.
In order to achieve this, the initial root is first split normally until the number of splittable leaves reaches the number of available threads. Then each leaf is split in a separate thread, while the maximum number of split iterations for each subtree is defined as the maximum number of split iterations for the whole main tree divided by the number of used threads. This way the total number of nodes in the alternative tree will be approximately the same as the number of nodes in the main tree.
Note that in general, the alternative tree does not match the main tree, so some nodes of the alternative tree will never be reused. In practice however, the portion of such unnecessarily precomputed nodes is not very big. And considering that the nodes of the alternative tree are precomputed in parallel using multiple threads, in most cases the overall performance is significantly improved.
DXT Testing:
The modified algorithm has been tested on the Kodak test set using 64-bit build with default settings (running on Windows 10, i7-4790, 3.6GHz). All the decompressed test images are identical to the images being compressed and decompressed using original version of Crunch (revision ea9b8d8).
[Compressing Kodak set without mipmaps using DXT1 encoding]
Original: 1582222 bytes / 28.840 sec
Modified: 1468204 bytes / 6.303 sec
Improvement: 7.21% (compression ratio) / 78.14% (compression time)
[Compressing Kodak set with mipmaps using DXT1 encoding]
Original: 2065243 bytes / 36.955 sec
Modified: 1914805 bytes / 8.342 sec
Improvement: 7.28% (compression ratio) / 77.43% (compression time)
ETC Testing:
The modified algorithm has been tested on the Kodak test set using 64-bit build with default settings (running on Windows 10, i7-4790, 3.6GHz). The ETC1 quantization parameters have been selected in such a way, so that ETC1 compression gives approximately the same average Luma PSNR as the corresponding DXT1 compression (which is equal to 34.044 dB for the Kodak test set compressed without mipmaps using DXT1 encoding and default quality settings).
[Compressing Kodak set without mipmaps using ETC1 encoding]
Total size: 1607858 bytes
Total time: 13.322 sec
Average bitrate: 1.363 bpp
Average Luma PSNR: 34.050 dB
Alexander Suvorov
dbbef6a21f