[FFmpeg-devel] [PATCH] adpcm: Store trellis nodes in a heap structure

[Martin Storsjö]

On Wed, 3 Nov 2010, Martin Storsj? wrote:

> As pointed out by Michael when reviewing the G.722 trellis encoder, the 
> stored trellis nodes could be stored in a heap-like structure, instead of 
> in a straight sorted array.
> 
> Currently, when inserting a new trellis node, a linear search (which in 
> itself perhaps could be sped up by converting it to a binary search) is 
> used to find the spot where it should be inserted, and then all later 
> node pointers are moved back one step with memmove. Since only a subset of 
> all evaluated nodes are stored, the worst one is removed once the array is 
> full.
> 
> Instead of doing this, the attached patch set stored the node pointers in 
> a heap structure, by first adding all evaluated nodes to a heap, as long 
> as they all fit. Once they don't all fit, we check through all the 
> frontier/2 leaf nodes to find the worst one, replace that one with the 
> current and restore the heap property.
> 
> This doesn't give identical results to the initial version, since the 
> nodes from the previous round are used for doing the next step in the 
> order they're stored in the array, which is different.
> 
> Instead of chekcing all the frontier/2 leaf nodes to find the worst one, 
> patch #3 just picks one of the leaf nodes and tries replacing that one. By 
> picking a different one of the leaf nodes each time, we more or less 
> achieve the same thing. (If only one spot would be tested, only the path 
> from that node up to the root would be updated once the heap is full.)
> 
> The last patch removes the search for nodes with an equal sample value to 
> the one currently inserted, since it's an O(frontier) operation, which 
> speeds things up immensely, without notably affecting the quality.
> 
> Some numbers, runtime:
> Original: 16.1 s
> After patch #1: 14.8 s
> After patch #3: 10.9 s
> After patch #4: 5.6 s
> 
> Output from tiny-psnr:
> No trellis:
> stddev:  101.13 PSNR: 56.23 MAXDIFF: 7183 bytes:  4865398/  4865408
> -trellis 5, original code:
> stddev:   81.78 PSNR: 58.08 MAXDIFF: 4798 bytes:  4865398/  4865408
> After patch #1:
> stddev:   81.70 PSNR: 58.08 MAXDIFF: 4798 bytes:  4865398/  4865408
> After patch #3:
> stddev:   80.77 PSNR: 58.18 MAXDIFF: 4766 bytes:  4865398/  4865408
> After patch #4:
> stddev:   80.94 PSNR: 58.17 MAXDIFF: 4524 bytes:  4865398/  4865408
> 
> So even if patch #3 and #4 in theory should worsen the output slightly, 
> they actually seem to improve the result in this case, since other nodes 
> happen to be stored/thrown away. The main point is that it doesn't seem to 
> harm the output quality significantly while improving the runtime 
> performance massively.

Updated benchmarks of this code - the sample in question is a 30 second 
clip, 44 kHz mono. Now I've done the testing with adpcm_ima_wav as codec, 
which should use larger frame sizes and give more effect from the trellis 
encoding.

No trellis:
user    0m0.038s
stddev:   40.03 PSNR: 64.28 MAXDIFF: 2874 bytes:  2646016/  2649218

Trellis 8, original code:
user    1m29.470s
stddev:   31.30 PSNR: 66.42 MAXDIFF: 1539 bytes:  2646016/  2649218

After patch #1, using a min-heap:
user    1m1.199s
stddev:   31.25 PSNR: 66.43 MAXDIFF: 1834 bytes:  2646016/  2649218

After patch #3, replacing any of the leaf nodes at random:
user    0m56.906s
stddev:   31.16 PSNR: 66.46 MAXDIFF: 1512 bytes:  2646016/  2649218

After patch #4, no collapsing of paths ending up with equal sample values:
user    0m22.965s
stddev:   31.35 PSNR: 66.40 MAXDIFF: 1782 bytes:  2646016/  2649218

So, another argument for patch #4 is that even if it theoretically should 
give worse results (which it for some reason doesn't, in these tests), it 
speeds up the encoding so much that you can use a larger trellis size, 
giving even better results.


Also, for reference, the same input with different trellis sizes, after 
patch #4:

0:
user    0m0.037s
stddev:   40.03 PSNR: 64.28 MAXDIFF: 2874 bytes:  2646016/  2649218

1:
user    0m0.191s
stddev:   37.48 PSNR: 64.85 MAXDIFF: 3026 bytes:  2646016/  2649218

2:
user    0m0.433s
stddev:   35.22 PSNR: 65.39 MAXDIFF: 2681 bytes:  2646016/  2649218

3:
user    0m0.880s
stddev:   33.87 PSNR: 65.73 MAXDIFF: 2643 bytes:  2646016/  2649218

4:
user    0m1.684s
stddev:   32.96 PSNR: 65.97 MAXDIFF: 2650 bytes:  2646016/  2649218

5:
user    0m3.252s
stddev:   32.34 PSNR: 66.13 MAXDIFF: 2467 bytes:  2646016/  2649218

6:
user    0m6.044s
stddev:   31.69 PSNR: 66.31 MAXDIFF: 1907 bytes:  2646016/  2649218

7:
user    0m11.906s
stddev:   31.57 PSNR: 66.34 MAXDIFF: 2105 bytes:  2646016/  2649218

8:
user    0m22.960s
stddev:   31.35 PSNR: 66.40 MAXDIFF: 1782 bytes:  2646016/  2649218

// Martin