Encode: do not perform rate control for single-tile lossless

[boxerab]

Currently, for single tile lossless, rate control is performed, although it is not needed.

This can lead to lossy encoding if there are code passes at the end of a code block that do not
contribute to a rate increase. The rate control algorithm will remove these code passes,
leading to potential lossy encode.

[boxerab]

Correct solution is to simply include all passes in a single layer for single-layer lossless

[rouault]

Do you have examples where that occurs ?

[boxerab]

No, I can't remember the exact case, but I saw this happening when I removed the limit of 10 resolution levels. Also, if you look at the logic in the code, you can see how this might happen.

But, even without a test image, there is no reason to perform rate control on single layer lossless image.
You just need a method to detect when this is the case, and skip the rate control.

[rouault]

Also, if you look at the logic in the code, you can see how this might happen.

Actually I don't see how that could happen. Lossless encoding implies cp->m_specific_param.m_enc.m_disto_alloc == 1 and tcd_tcp->rates[0] == 0. Which implies in opj_tcd_rateallocate() that goodthresh = min; which is the minimum of dd / dr for all passes of all code blocks (

openjpeg/src/lib/openjp2/tcd.c

Line 616 in a55c024

goodthresh = min;

). Then opj_tcd_makelayer() is called with this threshold and checks in

openjpeg/src/lib/openjp2/tcd.c

Lines 291 to 292 in a55c024

	if (thresh - (dd / dr) <
	DBL_EPSILON) { /* do not rely on float equality, check with DBL_EPSILON margin */

if (thresh - (dd / dr) < DBL_EPSILON), which should be always true. So all passes should be selected. Am I missing something ?

[boxerab]

hmmmm. Well, I guess that was a false alarm then :)

[boxerab]

How about lossless with multiple layers ? Could the final layer be missing some passes ? If so, that would be a problem.

[rouault]

The same logic will apply for multiple layers. For lossless encoding, the last layer should have tcd_tcp->rates[layno] == 0, in which case all remaining coding passes will be selected. I just committed a fix for opj_compress help which indicated 1 instead of 0 for this lossless terminating layer

[boxerab]

For -r flag, 1 indicates a final lossless layer, while for -q flag, 0 indicates a final lossless layer.

So, looks like everything is correct then. Closing.

[rouault]

For -r flag, 1 indicates a final lossless layer

Actually that doesn't match my experiments. You really need to set 0 to get all the remaining coding passes in the final layer. 1 gets you almost to lossless, but not completely

[boxerab]

Well, what I would do is keep the 1 for compression ratio flag, and then convert to 0 in the code. Compression ratio of 0 is a bit counter-intuitive. You could let the user know that 1 == final lossless layer.

[boxerab]

i.e. set rates[layer-1] to 0 if user has final ratio of 1.

[boxerab]

Also sanity check that ratios are decreasing and >= 1

[boxerab]

So, I think the key logic to look at is how distortion delta dd and rate delta dr are handled in rate control. When dd and dr are both zero (as is the case for passes that do not contribute to rate), these passes are ignored . I guess code assumes that a subsequent layer will then force them to be included. But, if this is the final layer, and the final passes, then these passes will be lost.

So, that's why there should be no rate control at all on final layer, if it is a lossless one. Just include everything and do not look at dd or dr.

[rouault]

OK, b428b8c should really make sure to include all passes for a lossless layer

[boxerab]

Looks good.

[rouault]

You might want to check that only the final layer is specified as lossless by the user.

Yes, 7aa071a will warn about that

[boxerab]

Cool.