better/faster encoding selection

[Antoine Martin]

The problem: we need lossless encoding (small regions, auto-refresh), but:

• PNG is slow
• RGB + zlib is inefficient
• webp has the potential, but the python API does not allow us to choose the trade off between speed and output size, though the underlying API does offer this option: see ​Advanced Encoding API:

  int lossless;           // Lossless encoding (0=lossy(default), 1=lossless).
  float quality;          // between 0 (smallest file) and 100 (biggest)
  int method;             // quality/speed trade-off (0=fast, 6=slower-better)
  

[Antoine Martin]

too late for 0.11 and fixing webp (#491) will have to be done first

[Antoine Martin]

re-scheduling all webp stuff

[Antoine Martin]

stub for a cython version of the webp encoder

[Antoine Martin]

The webp encoder in Cython was added in r5958, used ahead of the old python-webm encoder (which is still used as fallback for pixel modes it cannot handle).

r5959 adds some tests that make it easier to compare webp / webm, png and jpeg encoders. Some preliminary results:

• webp (on its medium speed setting) is faster than python-webm
• webp is very slow at low speeds..
• jpeg is still an order of magnitude faster than the rest

Conclusions:

• maybe we need to normalize the speed setting, so that webp won't go too slow unless we set the speed right down
• we should use webp instead of png in some cases
• rgb + lz4 does remarkably well - and when it doesn't compress well (small alpha test), it doesn't matter: the picture is so small that 6KB vs 3KB is not that big a deal
• maybe we need to change our API usage as lossless compression is way too slow
• lossless mode may need tweaking (as we may assume a specific encoding is lossless, as was the case with png - but webp can do both)

Some performance data using some sample images, showing both speed (in mega pixels per second - higher is better, and compression ratio - lower is better):

Mode					Desktop 2560x1600		Browser 1920x1080		Diagram 640x800		Small Alpha 64x48
Encoding	Compressor	Preset	Quality	Speed	MPixels/s	comp	MPixels/s	comp	MPixels/s	comp	MPixels/s	comp
png	PIL (old)	optimized		0	1.6	8.1
png	PIL (old)			100	6.5	8.5
png	PIL			100	11.6	10.0	16.1	3.1	14.6	5.0	6.0	30.1
png/P	PIL			100	21.8	4.7	17.2	9.0	23.9	3.9	2.2	14.0
png/L	PIL			100	24.3	3.8	31.1	1.4	29.7	2.6	7.6	4.7
webp	PIL		0		5.5	0.5	6.3	0.5	5.9	0.7	0.4	13.4
webp	PIL		50		5.1	2.1	6.0	1.1	5.6	1.6	0.4	16.7
webp	PIL		100		4.3	6.2	5.9	3.1	5.3	4.3	0.4	28.8
jpeg	PIL	optimized	0		71.0	0.5	68.5	0.4	80.6	0.5	16.6	3.2
jpeg	PIL		50		86.0	3.1	68.1	1.9	68.2	2.5	12.7	8.8
jpeg	PIL	optimized	100		48.0	10.8	56.8	5.7	56.0	7.7	9.5	26.1
jpeg	PIL		100		72.0	13.0	58.4	5.7	56.5	7.7	8.4	26.1
rgb	lz4			100	224.0	15.0	497.0	4.3	526.0	6.9	76.7	56.0
rgb	zlib			50	21.0	9.1	30.2	2.5	27.8	3.1	11.3	39.4
webp	python-webm		0		6.0	0.8	6.7	0.5	6.2	0.7	0.5	13.4
webp	python-webm		50		5.3	2.1	6.5	1.1	5.7	1.6	0.6	16.7
webp	python-webm		100		0.3	4.8	1.0	0.7	0.7	1.6	0.1	25.0
webp	Cython	TEXT	0	100	18.6	0.6	19.4	0.4	18.9	0.6	1.5	2.9
webp	Cython	TEXT	50	100	15.5	2.7	17.7	1.5	16.6	2.2	1.2	11.9
webp	Cython	TEXT	99	100	12.6	6.7	15.2	3.3	13.9	4.6	0.8	29.7
webp	Cython	TEXT + lossless	100	0	0.1	4.9	0.6	0.7	0.3	1.6	0.0	24.7
webp	Cython	TEXT + lossless	100	100	2.3	5.9	27.8	1.3	27.2	1.9	1.5	26.9

Notes and caveats:

• Low quality, low speed doesn't really make much sense, so some values have been omitted
• changing the webp preset does not affect speed or output size much, it probably does affect the perceived picture quality (seems best to stick with TEXT to ensure that text remains readable)
• The new webp encoder seems to perform about the same with speeds >50%, only low speed is really slow, and lossless is unbearably slow (it can take more than 30 seconds to encode a single frame using lossless + low speed!)
• quality is not the same for each encoder, some have lossless modes others not, etc..

[Antoine Martin]

After figuring out that libwebp wants us to use the quality attribute for storing the speed when in lossless mode... I kid you not! applied in r5962. The results are a lot better on "high speed" (s=100%), but still much slower than png... (table above updated)

[Antoine Martin]

Digging further shows that it is possible to make png faster instead.
r5966 tunes png and jpeg better.

I have also tested to see if changing the png compression_type attribute can give us better compression or performance (no: default is best):

		Desktop 2560x1600		Browser 1920x1080		Diagram 640x800		Small Alpha 64x48
Compression Type	Level	MPixels/s	comp	MPixels/s	comp	MPixels/s	comp	MPixels/s	comp
HUFFMAN_ONLY	n/a	9.8	19.0	11.4	15.8	10.8	17.3	6.6	29.9
FIXED	1	12.1	11.7	16.7	3.6	14.2	5.8	5.7	35.1
FIXED	3	10.7	10.7	16.6	3.5	13.8	5.5	5.0	33.0
RLE	1	12.1	11.5	16.6	4.6	15.2	6.5	7.1	27.3
RLE	3	10.8	10.7	16.6	4.6	14.7	6.5	6.8	27.3
FILTERED	1	12.1	10.0	16.7	3.1	14.1	5.0	5.8	30.1
FILTERED	3	10.7	9.4	16.4	3.0	12.0	4.8	5.0	29.1
DEFAULT (old code)	unset	6.8	8.4	12.5	2.7	10.3	4.9	3.4	26.2

Notes:

• Level 1 now corresponds to >75% speed and level 3 to 26% to 50% speed
• Lower speeds aren't very interesting, and we may want to prevent the level from going that low
• HUFFMAN_ONLY ignores the compression level
• DEFAULT is clearly the same as FILTERED (identical results - not shown)

[Antoine Martin]

(updating bug title to reflect the greater scope of this ticket)

[Antoine Martin]

updates the encoding heuristics based on the new data

[Antoine Martin]

Many improvements:

• r5987 (logging with -d compress)
• r5988 normalize speed for webp
• r5989 discard delta
• r5990 (BIG!) lots of changes, see commit log message
• r5994 avoid very small and very large areas with webp
• r6003 + r6001 + r5999: better video + csc pipeline selection
• r6008: saves getting a new xshm image when processing multiple regions

[Antoine Martin]

• added info to wiki/Encodings
• try to stick with the chosen encoding for non-video encodings (r6478)
• add ability to force the use of the selected encoding with XPRA_ENCODING_STRICT_MODE=1 (r6519). Note: with video encodings, we still need to send the edges of odd sized windows with rgb, when the picture is lossy we may still use another encoding for the lossless refresh..
• use BGRX directly with rgb encodings and mmap (already backported to 0.12.6): r6492, re-stride to save space: r6474, and associated fixes (GL: r6462)
• encoding selection bug fix for video encoding selection (ouch! r6477)
• ensure we don't use webp when it is too slow, even when selected as primary encoding (see r6500)
• webp encoder leak fixed in r6505, see also updates in #487

smo: please run this latest code on your test system so we can compare with previous runs and ensure the performance is at least as good as before for most tests

[Antoine Martin]

Important regression discovered and fixed in r6568 (see ​"Trouble with xpra 0.13" mailing list posts for the bug report), please test and verify that PNG, RGB, png/L and png/P never end up using lossy encodings, no matter what command line options or previous encodings were used. And no matter how bad the refresh rate gets when under pressure.

Raising to blocker as this should definitely go in 0.13.1. Backported in r6571.
Quite a few fixes have gone in, none of them quite as critical as the one above (transparency, opengl toggle, etc): r6577, r6578, r6583

[alas]

Testing with win client 0.14.0 r6910 against fedora 20 0.14.0 r6910 server.

With a client-side --encoding=png the encoding seems to stick to png, no matter how bad it gets.

2014-07-21 12:39:45,496 update_speed() info={'damage_latency.abs_factor': 0, 'damage_latency.avg': 6, 'decoding_latency.target': 8000000, 'damage_latency.target': 35, 'min_speed': 0, 'frame_delay': 8, 'decoding_latency.factor': 109$
2014-07-21 12:39:45,496 reconfigure(False) csc_encoder=None, video_encoder=None
2014-07-21 12:39:47,190 damage(WindowModel(0xa00022 - "jimador@zapopan:~"), 155, 301, 6, 13, {})
2014-07-21 12:39:47,190 damage(155, 301, 6, 13, {}) wid=1, sending now with sequence 59
2014-07-21 12:39:47,191 process_damage_regions: wid=1, adding png pixel data to queue, elapsed time: 1.5 ms, request rgb time: 0.6 ms
2014-07-21 12:39:47,192 make_data_packet: image=XShmImageWrapper(BGRX: 155, 301, 6, 13), damage data: (1, 155, 301, 6, 13, 'png')
2014-07-21 12:39:47,195 update_speed() info={'damage_latency.abs_factor': 0, 'damage_latency.avg': 7, 'decoding_latency.target': 8000000, 'damage_latency.target': 35, 'min_speed': 0, 'frame_delay': 4, 'decoding_latency.factor': $
2014-07-21 12:39:47,195 sending 6x13 BGRX as png, mode=RGB, options={'compress_level': 1}
2014-07-21 12:39:47,195 reconfigure(False) csc_encoder=None, video_encoder=None

Likewise with RGB & png/L.

With --encoding=png/P however, the encoder seems to use png/P for "existing delayed regions" (2014-07-21 14:32:10,285 damage(86, 427, 640, 54, {}) wid=3, using existing delayed png/P regions created 0.0ms ago), but seems to bounce around from rgb to webp to png/P overall.

Here's a bit of a clip of server-side logs (I'll include a full long log for this):

2014-07-21 14:31:32,441 update_speed() info={'damage_latency.abs_factor': 0, 'damage_latency.avg': 8, 'decoding_latency.target': 8000000, 'damage_latency.target': 35, 'min_speed': 0, 'frame_delay': 24, 'decoding_latency.factor':$
2014-07-21 14:31:32,442 damage(143, 2, 6, 13, {}) wid=1, sending now with sequence 2
2014-07-21 14:31:32,442 reconfigure(False) csc_encoder=None, video_encoder=None
2014-07-21 14:31:32,444 damage(WindowModel(0xc00022 - "jimador@zapopan:~"), 143, 2, 6, 13, {})
2014-07-21 14:31:32,444 damage(143, 2, 6, 13, {}) wid=2, scheduling batching expiry for sequence 3 in 24.0 ms
2014-07-21 14:31:32,446 process_damage_regions: wid=1, adding rgb24 pixel data to queue, elapsed time: 5.5 ms, request rgb time: 1.1 ms
2014-07-21 14:31:32,446 make_data_packet: image=XShmImageWrapper(BGRX: 143, 2, 6, 13), damage data: (1, 143, 2, 6, 13, 'rgb24')

...

2014-07-21 14:32:22,577 damage(86, 325, 640, 102, {}) wid=3, using existing delayed png/P regions created 0.0ms ago
2014-07-21 14:32:22,577 damage(WindowModel(0xe00001 - "John Oliver Is Psyched for the World Cup - Late Night with Seth Meyers - YouTube - Google Chrome"), 86, 427, 640, 54, {})
2014-07-21 14:32:22,577 damage(86, 427, 640, 54, {}) wid=3, using existing delayed png/P regions created 0.0ms ago
2014-07-21 14:32:22,618 send_delayed for wid 3, delaying again because of backlog: 3 packets, batch delay is 42.3135884506, elapsed time is 43.0 ms
2014-07-21 14:32:22,664 damage(WindowModel(0xe00001 - "John Oliver Is Psyched for the World Cup - Late Night with Seth Meyers - YouTube - Google Chrome"), 86, 481, 512, 6, {})
2014-07-21 14:32:22,664 damage(86, 481, 512, 6, {}) wid=3, using existing delayed png/P regions created 0.1ms ago
2014-07-21 14:32:22,704 send_delayed_regions: bytes_cost=238592, bytes_threshold=698084, pixel_count=233472
2014-07-21 14:32:22,707 process_damage_regions: wid=3, adding webp pixel data to queue, elapsed time: 132.2 ms, request rgb time: 2.5 ms
2014-07-21 14:32:22,707 make_data_packet: image=XShmImageWrapper(BGRX: 86, 121, 640, 366), damage data: (3, 86, 121, 640, 366, 'webp')

If I'm not mistaken, though, none of those are "lossy encodings" ... so that may be expected behavior.

Smo will try to arrange those tests for performance as soon as he can.

[alas]

ticket 419 (overlarge) server-side png/P -d encode log

[Antoine Martin]

I guess this should have been re-assigned to me - taking it back.

I thought that this was all working as expected until #645

I will try to clarify the logging:

• the encoding seen in the damage call is the encoding currently set as default (generally the one selected by the user - except when auto-refresh kicks in, or when specifically overriden..)
• process_damage_regions shows rgb.. which is not an encoding at all: in this case, that's just the raw pixel format we get from the x11 server
• the encoding actually used for compression is shown in make_data_packet

And so the encoding actually may well be different from the one that is set, especially when:

• with video modes, you can get pretty much any encoding out depending on many heuristics
• we prefer rgb for small regions (or when we have bandwidth to spare) - but not when using png/L (which is black and white, as this would look funny)
• auto-refresh will use a lossless encoding

etc.

[Antoine Martin]

Despite some fairly big bugs that went unnoticed until after the release (#645 mentioned above and also #651), I think this is now working as expected and also much faster than before (see #620).

r7424 removes 'rgb' from the logging in process_damage_regions.

I have added information to the wiki on debugging encodings here: wiki/Encodings. Closing this ticket, some of these issues will be followed up in #410.