scrolling improvements

[Antoine Martin]

follow up from #1232.

Issues:

• it can take a lot of CPU to figure out the scroll encoding (figures given here are made worse by the instrumentation and profiling tools enabled):

  scroll encoding total time: 89ms
  

So high that we might has well not bother!

• we should add the ability to enable it even when there are no video encodings available (ie: for html5 client)
• we make a copy of the xshm backing for each subimage we send to the non-scroll compression - these could be handled as memoryviews

[Antoine Martin]

add logging for scroll timing within encode_scrolling function

[Antoine Martin]

Using the patch above, we can see that there are 2 main areas that take too long to complete:

2017-02-02 02:01:12,113 updated scroll data, previously set: True
2017-02-02 02:01:12,116 best scroll guess took 4ms, matches 76% of 1485 lines: -80
2017-02-02 02:01:12,118 encode_scrolling(XShmImageWrapper(BGRX: 0, 0, 1923, 1485), {..}, [], [], {'scroll': True}) window-dimensions=(1923, 1485)
2017-02-02 02:01:12,175  will send scroll packets for yscroll=[-80, -79, -81, -78, -82, -77, -83, -76, -84, -75, -85, -74, -86, -73, -87, -72, -88, -71, 0, -70, -89, -69, -90, -68, -67, -91, -66, -92, -65, -64, -93, -63, -62, -94, -61, -95, -60, -96, -97, -59, -97, -59, -98, -58, -98, -58, 11, -57, 11, -57], elapsed=56ms
2017-02-02 02:01:12,180  scroll groups for distance=-80 : [(71, 429), (599, 167), (865, 540)]
2017-02-02 02:01:12,183  scroll groups for distance=-79 : [(1405, 1)]
2017-02-02 02:01:12,186  scroll groups for distance=-78 : [(1406, 1)]
2017-02-02 02:01:12,189  scroll groups for distance=-77 : [(1407, 1)]
2017-02-02 02:01:12,192  scroll groups for distance=-76 : [(1408, 1)]
2017-02-02 02:01:12,195  scroll groups for distance=-75 : [(1409, 1)]
2017-02-02 02:01:12,198  scroll groups for distance=-74 : [(1410, 1)]
2017-02-02 02:01:12,201  scroll groups for distance=-73 : [(1411, 1)]
2017-02-02 02:01:12,204  scroll groups for distance=-72 : [(1412, 1)]
2017-02-02 02:01:12,208  scroll groups for distance=-71 : [(1413, 1)]
2017-02-02 02:01:12,211  87 lines have not changed
2017-02-02 02:01:12,215  scroll groups for distance=-70 : [(1414, 1)]
2017-02-02 02:01:12,218  scroll groups for distance=-69 : [(1415, 1)]
2017-02-02 02:01:12,221  scroll groups for distance=-68 : [(1416, 1)]
2017-02-02 02:01:12,224  scroll groups for distance=-67 : [(1417, 1)]
2017-02-02 02:01:12,226  scroll groups for distance=-66 : [(1418, 1)]
2017-02-02 02:01:12,229  scroll groups for distance=-65 : [(1419, 1)]
2017-02-02 02:01:12,232  scroll groups for distance=-64 : [(1420, 1)]
2017-02-02 02:01:12,234  scroll groups for distance=-63 : [(1421, 1)]
2017-02-02 02:01:12,237  scroll groups for distance=-62 : [(1422, 1)]
2017-02-02 02:01:12,240  scroll groups for distance=-61 : [(1423, 1)]
2017-02-02 02:01:12,243  scroll groups for distance=-60 : [(1424, 1)]
2017-02-02 02:01:12,248  scroll groups for distance=11 : [(784, 9), (1450, 2)]
2017-02-02 02:01:12,251  distance matching took 73ms
2017-02-02 02:01:12,253  non scroll: 6 packets: [(500, 99), (766, 18), (793, 25), (820, 45), (1439, 11), (1452, 33)]
2017-02-02 02:01:12,257 compress: 137.4ms for 1923x1485 pixels at    0,0    for wid=1     using scroll as  25 rectangles  (11154KB)           , sequence    90, client_options={'scroll': True, 'flush': 6}
2017-02-02 02:01:12,260  splitting took 6ms
2017-02-02 02:01:12,262 sub_image 190377 pixels: 0.1ms
2017-02-02 02:01:12,268 compress:   6.0ms for 1923x99   pixels at    0,500  for wid=1     using   jpeg with ratio   5.4%  (  743KB to    40KB), sequence    91, client_options={'scroll': True, 'flush': 5}
2017-02-02 02:01:12,271 sub_image 34614 pixels: 0.1ms
2017-02-02 02:01:12,276 compress:  14.1ms for 1923x18   pixels at    0,766  for wid=1     using   jpeg with ratio   6.1%  (  135KB to     8KB), sequence    92, client_options={'scroll': True, 'flush': 4}
2017-02-02 02:01:12,279 sub_image 48075 pixels: 0.1ms
2017-02-02 02:01:12,285 compress:  22.5ms for 1923x25   pixels at    0,793  for wid=1     using   jpeg with ratio   1.3%  (  187KB to     2KB), sequence    93, client_options={'scroll': True, 'flush': 3}
2017-02-02 02:01:12,288 sub_image 86535 pixels: 0.1ms
2017-02-02 02:01:12,296 compress:  33.8ms for 1923x45   pixels at    0,820  for wid=1     using   jpeg with ratio   5.0%  (  338KB to    17KB), sequence    94, client_options={'scroll': True, 'flush': 2}
2017-02-02 02:01:12,299 sub_image 21153 pixels: 0.0ms
2017-02-02 02:01:12,304 compress:  41.5ms for 1923x11   pixels at    0,1439 for wid=1     using   jpeg with ratio   1.8%  (   82KB to     1KB), sequence    95, client_options={'scroll': True, 'flush': 1}
2017-02-02 02:01:12,306 sub_image 63459 pixels: 0.1ms
2017-02-02 02:01:12,313 compress:  50.4ms for 1923x33   pixels at    0,1452 for wid=1     using   jpeg with ratio   4.1%  (  247KB to    10KB), sequence    96, client_options={'scroll': True}
2017-02-02 02:01:12,315 non-scroll encoding using jpeg (quality=42, speed=90) took 52ms for 6 rectangles
2017-02-02 02:01:12,318 scroll encoding total time: 199ms

• elapsed=56ms is the time spent figuring out which scroll values to send
• distance matching took 73ms
• the rest is the region encoding (already improved in r14952 by using jpeg instead of png - quality / speed settings permitting)

The whole thing should be cythonized.
For the v1.0.x branch, maybe we can just limit the size of the search space (<1K vertical?)

[Antoine Martin]

this cythonized version of the code is super fast, but also not visually correct (yet)

[Antoine Martin]

almost done - almost renders correctly