How much faster it went?
It seems 0.94’s AI is much faster.
Here’s apk of 0.94: a.pomf.se/faklhi.apk . Sadly VCMI crashes with SIGBUS when entering castle so this version is not really playable either.
F/libc (22066): Fatal signal 7 (SIGBUS) at 0x00000000 (code=128), thread 22097 (CGuiHandler::ru)
I/DEBUG ( 114): *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
I/DEBUG ( 114): Build fingerprint: 'samsung/espresso10wifixx/espresso10wifi:4.0.3/IML74K/P5110XWALD2:user/release-keys'
I/DEBUG ( 114): Revision: '10'
I/DEBUG ( 114): pid: 22066, tid: 22097, name: CGuiHandler::ru >>> is.xyz.vcmi.nightly <<<
I/DEBUG ( 114): signal 7 (SIGBUS), code 128 (SI_KERNEL), fault addr 00000000
I/DEBUG ( 114): r0 2e7644a9 r1 2d17eab8 r2 2e764170 r3 2e6326a8
I/DEBUG ( 114): r4 2e7644a1 r5 6096faa8 r6 2a2338a0 r7 00000000
I/DEBUG ( 114): r8 5ec7bb28 r9 2d25c890 sl 2a2338a0 fp 6096fbac
I/DEBUG ( 114): ip 00000000 sp 6096fa40 lr 5e9659e5 pc 5e965a02 cpsr 600f0030
I/DEBUG ( 114): d0 32545845434e4254 d1 ffffffffffffffff
I/DEBUG ( 114): d2 4144414430c22882 d3 0000000031031882
I/DEBUG ( 114): d4 401c000000000000 d5 401bf8d6d33ea848
I/DEBUG ( 114): d6 408f400000000000 d7 0062e08000000000
I/DEBUG ( 114): d8 4082c00000000320 d9 44a0000000000000
I/DEBUG ( 114): d10 00000000443c0000 d11 0000000000000000
I/DEBUG ( 114): d12 0000000000000000 d13 0000000000000000
I/DEBUG ( 114): d14 0000000000000000 d15 0000000000000000
I/DEBUG ( 114): d16 0000000000000000 d17 0000000000000000
I/DEBUG ( 114): d18 41bf93e1ae000000 d19 3fd99b7c3731d10b
I/DEBUG ( 114): d20 3f974c074a362e68 d21 3fe5568b9401f367
I/DEBUG ( 114): d22 3fd24aaf828e1369 d23 3fcc745e0ab0ad85
I/DEBUG ( 114): d24 3fc748e72055a505 d25 c050800000000000
I/DEBUG ( 114): d26 3fc2f112df3e5244 d27 40034eeae0000000
I/DEBUG ( 114): d28 0000000000000000 d29 0000000000000000
I/DEBUG ( 114): d30 0000000000000000 d31 40034eeae0000000
I/DEBUG ( 114): scr 20000013
I/DEBUG ( 114):
I/DEBUG ( 114): backtrace:
I/DEBUG ( 114): #00 pc 0031ea02 /data/app-lib/is.xyz.vcmi.nightly-2/libapplication.so (void CDefFile::loadFrame<CompImageLoader>(unsigned int, unsigned int, CompImageLoader&) const+53)
I/DEBUG ( 114): #01 pc 0031e9e1 /data/app-lib/is.xyz.vcmi.nightly-2/libapplication.so (void CDefFile::loadFrame<CompImageLoader>(unsigned int, unsigned int, CompImageLoader&) const+20)
I/DEBUG ( 114):
I/DEBUG ( 114): stack:
I/DEBUG ( 114): 6096fa00 0000a228
I/DEBUG ( 114): 6096fa04 00000004
I/DEBUG ( 114): 6096fa08 2e764170 [heap]
I/DEBUG ( 114): 6096fa0c 4003766f /system/lib/libc.so (dlmalloc+4278)
I/DEBUG ( 114): 6096fa10 5ec854f8
I/DEBUG ( 114): 6096fa14 6096fa1c
I/DEBUG ( 114): 6096fa18 4007203c
I/DEBUG ( 114): 6096fa1c 5ec854f8
I/DEBUG ( 114): 6096fa20 00000006
I/DEBUG ( 114): 6096fa24 0000a21f
I/DEBUG ( 114): 6096fa28 6096fad8
I/DEBUG ( 114): 6096fa2c 001367cc
I/DEBUG ( 114): 6096fa30 00000018
I/DEBUG ( 114): 6096fa34 2d0d2440 [heap]
I/DEBUG ( 114): 6096fa38 df0027ad
I/DEBUG ( 114): 6096fa3c 00000000
I/DEBUG ( 114): #00 6096fa40 2d0d8390 [heap]
I/DEBUG ( 114): 6096fa44 00000000
I/DEBUG ( 114): 6096fa48 2d17eab8 [heap]
I/DEBUG ( 114): 6096fa4c 4003766f /system/lib/libc.so (dlmalloc+4278)
I/DEBUG ( 114): 6096fa50 6096fad8
I/DEBUG ( 114): 6096fa54 77ffffd0
I/DEBUG ( 114): 6096fa58 2d07da20 [heap]
I/DEBUG ( 114): 6096fa5c 00000018
I/DEBUG ( 114): 6096fa60 2d0d2440 [heap]
I/DEBUG ( 114): 6096fa64 001367cc
I/DEBUG ( 114): 6096fa68 6096fb00
I/DEBUG ( 114): 6096fa6c 00000000
I/DEBUG ( 114): 6096fa70 00000024
I/DEBUG ( 114): 6096fa74 2e636a88 [heap]
I/DEBUG ( 114): 6096fa78 4007204c
I/DEBUG ( 114): 6096fa7c 2e6326a8 [heap]Generally it runs slower, as algorithms are more complex. But it’s negligible in typical situations.
However, that shouldn’t be an issue, as AI should behave better in extreme situations. In general the issue is not to choose one of multiple solutions. The problem arises when there are no options left and AI needs to check entire map over and over with no success. That part was improved in numerous ways (mostly exploration and clearing the way). The goal is to lower upper bound of AI time.
The remaning issues are some loopholes in rare situations and incorrect conditions, like what Povelitiel shows.
Also, if Android device runs 100 times slower on more than my PC, there’s not much chance for improvement, either.
One thing I know is this doesn’t belong to this thread.
Okay, I found a solution to stop AI from checking exploration possibilities many times with same hero, without loosing funcionality. This limits most extreme turn times, however does nothing to improve average AI speed.
The conditions are listed in r3819:
I optimized most expensive function in r3834. Apparently pathfinder was crippled by recalculating hero bonuses for every tile.
Huh… such cases should be covered bonus system caching. Wasn’t it working at all or was it just too slow?
Well, actually getting catched bonus was the function that caused significant CPU usage. Reading catched string is of course faster than navigating bonus tree, but still can take time if called thousands of times. Streams are slow.
I found another bottleneck specific to water-based maps only. Fixed in r3830.
Is it possible to change the enemy movement speed?
In the in-game settings menu there is an option, to change your own movement speed, but the movement speed for the enemy is disabled.
Additionally in Heroes II is an option the change the whole ai speed.
This option works for AI heroes as well.
Okay, are you happy with AI speed in 0.95b?
Nope. In case AI generates a lot of heroes with 1-4 creatures on a small area, it produces a lot of warnings like: Another allied hero stands in our way
and it becomes extremely slow. Makes me annoyed.
But this is exactly how it worked since 0.95 and no one reported such issue.
The AI is much faster than in previous version and this is definitely noticeable. But there is one thing that really bother me - heroes which belong to different players don’t fight each other.
+1 for Kantor
About AI, it definitely became better, but it becomes very slow in such cases I’ve described.
In scenarios where you play with Gelu everything is ok, in scenarios where you play with Sandro the area is quite limited and it becomes such a pain to wait 5-6 mins for 1 turn.