Hi Alexandre,
in fact the img size is only reduced when the bit stream length in bytes in shorter. The new algo will prefer the trick, so if you are right and the trick costs runtime in the device I should change that. The effect is probably very small though. Gerd ________________________________ Von: mkgmap-dev <[email protected]> im Auftrag von Alexandre de Menezes <[email protected]> Gesendet: Mittwoch, 20. Juli 2016 17:55:30 An: Development list for mkgmap Betreff: Re: [mkgmap-dev] [Patch v1] improve LinePreparer to reduce img size Hi Gerd, I am not familiar with the format, but from your explanation I am guessing this will be at some point stored in full bytes (or words or dwords). Assuming bytes, there will be indeed a size gain going from 30bits to 24bits as in your example, but not in going from 32 to 26, for example. In this case, wouldn't the original encoding be simpler (faster) to decode? Does this make sense? Are you taking this into account when choosing the best encoding? Best regards, Alexandre On 18/07/2016 04:22, Gerd Petermann wrote: Let's try with a real world example: https://www.openstreetmap.org/way/245112355 <https://www.openstreetmap.org/way/245112355> This highway has 6 points, it starts in the north, the x-delta values (in resolution 24) are 4, -19, -5, 0, 3 The largest absolute value is 19 which requires 5 bits: 10011. Since both positive and negative values exist, we have to encode the sign for each single delta value, so we have to use 6 bits (the leftmost gives the sign) So, the first iteration gives: 6 bit words : 4 = 000100 , -19 = 110011, -5 = 111011, 0 = 000000, 3 = 000011 --> 30 bits Now, the optimization process starts trying to use the "trick" with 5 bits. WIth the trick, the value 19 doesn't fit into 5 bits, so the trick is used (it stands for the binary value 1111) and the result for -19 is 10000 11100 (15 + 4 made negative) All other values fit into 5 bits, so the 2nd iteration gives: 5 bit words: 4 = 00100 , -19 = 10000 11100, -5 = 11011, 0 = 00000, 3 = 00011 --> 30 bits The next step is to try with only 4 bits. Again, only the value -19 needs the trick, this time we get - 19 = 1000 1000 1011 (7 +7 + 5 made negative) 4 bit words : 4 = 0100 , -19 = 1000 1000 1011, -5 = 1011, 0 = 0000, 3 = 0011 --> 24 bits The next step is to try with only 3 bits. This time, more values need the trick (4 , -5 and -19). For -19 we get 100 100 100 100 100 100 111 (3 + 3 + 3 + 3 + 3+ 3 + 1 made negative) The result: 3 bit words : 4 = 100 001 , -19 = 100 100 100 100 100 100 111, -5 = 100 110, 0 = 000, 3 = 011 --> 39 bits So, we see that the trick is used too often now and the resulting bit stream is much longer than the one before. So, using the trick with 4 bit words saves 6 bits for the x values. ciao, Gerd P.S. To be complete: The y-deltas are all negative: -3, -7,- 3,- 5, -4 This allows to encode only the absolute values which all fit into 3 bits, the trick doesn't help here. ________________________________ Von: mkgmap-dev <[email protected]><mailto:[email protected]> im Auftrag von nwillink <[email protected]><mailto:[email protected]> Gesendet: Sonntag, 17. Juli 2016 16:59:19 An: [email protected]<mailto:[email protected]> Betreff: Re: [mkgmap-dev] [Patch v1] improve LinePreparer to reduce img size Many thanks for that Gerd I hope this makes it a bit clearer? The adjective 'crystal' does not immediately spring to mind but I understand what you are getting at. Presumably the patch now checks every line for the need to apply the trick. -- View this message in context: http://gis.19327.n5.nabble.com/Patch-v1-improve-LinePreparer-to-reduce-img-size-tp5878545p5878569.html Sent from the Mkgmap Development mailing list archive at Nabble.com. _______________________________________________ mkgmap-dev mailing list [email protected]<mailto:[email protected]> http://www.mkgmap.org.uk/mailman/listinfo/mkgmap-dev _______________________________________________ mkgmap-dev mailing list [email protected]<mailto:[email protected]> http://www.mkgmap.org.uk/mailman/listinfo/mkgmap-dev
_______________________________________________ mkgmap-dev mailing list [email protected] http://www.mkgmap.org.uk/mailman/listinfo/mkgmap-dev
