HAL9000

HAL9000
"It just isn't conceivable that you can design a program strong enough to beat players like me."

May 31, 2015

Stockomodos: 6 x Stockfish against 5 x Komodo // PART-II

Yep. That becomes much fun with further rounds played. I'm glad to see 11 different versions find their rank more and more accurately.

Compared to the first version posted here, i've added Stockfish DD but couldn't do the same with Komodo 3. This version doesn't work with Rockchip 3188 in contrast to Exynos 4412. I don't know the reason of this. Both processors are quad core and both compiles are arm7 but no way. Therefore the tourney contained 11 engines.

I've just finished the 6th opening of TCEC-6 to reach 120 games per engine. Below chart is based on manual ELO calculation, not Elostat or Bayes. Four engines are currently rated in Rapidroid, so i use them as anchors to calculate the others.

The results are already speaking given that all versions follow the logical chronological progress. Thay are telling that Komodo 9 has a great jump forward but it seems it's still not enough to catch up Stockfish 6.

I see why as per my experience with Android compiles:
* Android compile of Stockfish is close to perfect, better than Komodo's.
* Komodo is somehow vulnerable vs Stockfish compared to other engines. Proof is that Stockfish 6 is more successful to its anchestors than Komodo. I think in an experiment open to various other engines, Komodo would have more chances to overtake Stockfish. For example in Rapidroid!
* Komodo is known to perform better with more threads and longer time control. My experiment is based on 15'+2" on a RK3188 which is equivalent to 10'+2" on an Exynos running at 1.5GHz only.

I want to continue some (!) more rounds to reach 300 games per engine and will decide to stop it or not later on.

For the moment, here is the status on Android as shown below. The difference column shows how the engines performed so far vs their Rapidroid ELO. And the last column shows the ELO they reached in this closed tourney. Finally, blue cells indicate the estimated ELOs of previous versions not included in Rapidroid anymore.

May 21, 2015

Stockomodos: 5 x Stockfish against 5 x Komodo

While waiting for the N7100 to finish the games of the top divisions in Rapidroid, i've tought it may be a good idea to organise a parallel battle including all versions of Stockfish vs all versions of Komodo on one of the devices on hold.

Slightly different from regular Rapidroid configuration, this specific tournament is played on Rockchip 3188 processor only.

As most of the competing engines use multi threads and because it would be insane to keep a 10 inch tablet in a refrigerator during all summer, i could find no other way than reducing the clock speed from 1.4 Ghz to 1.0 Ghz but i didn't want to balance the loss with longer time controls.

This safety against throttling sacrifies a little bit from the quality of the games but still equals an Exynos 4412 running at 1.5 Ghz with time control of 10'+1". Barely acceptable.

Despite low specs, first two rounds using 8 moves openings of TCEC-6 didn't deliver any surprise. It makes 36 games per engine so far and due to huge error margins, i post only the standings for the moment.

It takes about two days to finish one round of 18 games per engine. Therefore i expect to get accurate results in one week and share here.

Maybe i should add Stockfish DD and Komodo 3 as well to reach 6 x 2 = 12 engines.

180 GAMES PLAYED BY 10 PROGRAMS
Stockfish 5      26.0 72.2%
Stockfish 6      26.0 72.2%
Komodo 8         23.5 65.3%
Komodo 9         23.5 65.3%
Stockfish 4      22.0 61.1%
Stockfish 3      18.5 51.4%
Stockfish 2.3.1  17.5 48.6%
BeeKay           13.0 36.1%
Komodo 2.03       7.0 19.4%
Doch32 1.3.4      5.0 13.9%

May 20, 2015

New Android UCI engine: Cheese 1.7

It feels good to see more and more engines are compiled for Android. What feels slightly bad is to follow Jim Ablett every day to check whether there's something new from him and meanwhile to skip the compiles provided by the authors.

I'd skipped Komodo 9 and got on with it lately last week. Cheese seems to be another case of deep sleep for me. Hopefully its author Patrice Duhamel notified me instantly after today's update of Rapidroid about the availability of a compile for arm7. Then i woke up!

Cheese 1.7 was released in March-2015. I'm two months late to discover it.

The engine looked healty to me using one of four cores of my Rockchip 3188 without glitches. First speed test at 180 seconds/move revealed 345 kNps which is very good for a single thread engine.

I hope Cheese will work well under Rapidroid conditions and will reach near 2500 ELO vs 2676 of the CCRL ranking.

Those who wanna taste a slice of cheese may visit Cheese 1.7 download page or alternatively download it from my engines repository HERE

May 19, 2015

RAPIDROID RELOADED: May-2015

Let me introduce the first Rapidroid ranking featuring KOMODO 9!

This is good news from the silicone reptile but there are bad ones too.

Due to unexpected failures caused by engines which didn't like continous tourney environment, that release took a lot of delays. At least i'm glad i could finish 3 rounds of ~880 games each with 107 Android engines after post-analyzing, replaying and deleting a lot of games until disqualifying many unstable engines.

I don't wanna speculate too much about these problem boys because i'm not sure whether the errors come from their code or the compiler or the GUI itself. One thing is true: They don't play until checkmate, often stop playing and worse than this, they may hang to freeze the tournament, causing hours of wait without any move played. Now, all gone for different kind of issues keeping Rapidroid from progressing. Below engines are mostly trustable and present the least ratio of termination defects.

Statistical improvements forecasted last time couldn't become true either but i didn't miss the chance to rearrange better pairings in groups and start better. There are more games played at incremented time controls instead of fixed time per move. If there's less score % variation with less games played, that's statistically better.

Anyway, don't take below ranking much to serious for the moment. It needs 100 games per engine to speak well. Let's say in July...

PS-1: GreKo 12.6 not yet validated. 12.5 remains in list.
PS-2: K2 0.75 has just been updated and remains to be tested. 0.71 remains in list.

Games in PGN compressed with 7z (2.60MB): HERE

BAYES RATINGS AFTER 2668 GAMES PLAYED BY 109 PROGRAMS
Rnk Name                     T O/S Elo    +   - Ga Sco Oppo Dra
001 Stockfish 6              4 A32 3227  83  77 54 78% 3041 41%
002 Komodo 9                 4 A32 3217  84  78 54 75% 3042 35%
003 Critter 1.6a             4 A32 3116  74  73 54 59% 3053 48%
004 Firenzina 2.4.1 xTreme   4 A32 3096  74  73 54 56% 3055 48%
005 BeeKay                   4 A32 3067  73  73 54 51% 3059 50%
006 BlackMamba 2.0           4 A32 3061  72  72 54 50% 3059 56%
007 Texel 1.05               4 A32 3029  81  79 54 57% 2971 33%
008 Sting SF 4.8.4 JA        4 A32 3016  80  78 54 56% 2971 35%
009 DeepSaros ver.2.3f       1 A32 3006  79  78 54 53% 2969 39%
010 Senpai 1.0               4 A32 2960  83  80 54 61% 2882 30%
011 RobboLito 0.085e4l       1 A32 2929  78  79 54 43% 2981 41%
012 cheng4 0.38              4 A32 2858  78  79 54 45% 2897 35%
013 Gaviota v1.0-d           4 A32 2836  81  83 54 42% 2896 28%
014 Arasan 17.5              4 A32 2811  77  79 54 40% 2897 39%
015 DiscoCheck 5.2.1         1 A32 2798  79  78 54 56% 2747 35%
016 Grapefruit 1.0           4 A32 2769  75  76 54 44% 2811 41%
017 Doch32 1.3.4 JA          1 A32 2746  78  80 52 40% 2820 31%
018 Cyclone 3.4              4 A32 2743  74  74 54 50% 2753 44%
019 Toga II 3.0              1 A32 2730  75  76 54 46% 2758 41%
020 Bobcat 6.4b              4 A32 2725  80  79 52 53% 2703 29%
021 Deuterium v14.3.34.130   1 A32 2720  76  76 54 52% 2698 41%
022 Deep Saros 0.9           4 A32 2712  76  77 54 44% 2756 35%
023 Fruit reloaded 2.1       1 A32 2698  75  75 50 49% 2709 54%
024 Hakkapeliitta 120115     1 A32 2681  81  79 54 56% 2642 19%
025 GNU Chess 5.60           1 A32 2651  77  74 54 64% 2553 43%
026 Strelka 5                1 A32 2641  79  77 54 61% 2557 30%
027 IvanHoe 9.46b            4 A32 2627  77  76 54 54% 2597 33%
028 Scorpio_2.7.7.JA_xb      1 A32 2626  81  84 50 40% 2704 24%
029 Crafty_24.1.JA_xb        1 A32 2593  78  79 54 46% 2614 22%
030 Murka 3 ARM UCI          1 A32 2590  74  76 54 41% 2652 37%
031 TheMadPrune 1.7.04       4 A32 2585  88  87 46 52% 2566 22%
032 Tucano_5.00.JA_xb        1 A32 2570  82  82 44 53% 2545 48%
033 Rodent 1.7 build 1       1 A32 2564  80  80 54 53% 2550 24%
034 Gull 1.2 JA              1 A32 2558  75  75 54 48% 2566 41%
035 Rhetoric 1.4.1           1 A32 2530  77  77 54 48% 2553 33%
036 RedQueen 1.1.97          4 A32 2513  80  82 54 44% 2555 20%
037 Bison 15.1               1 A32 2488  79  78 54 54% 2467 30%
038 Alfil 12.10              1 A32 2454  78  78 54 46% 2471 33%
039 GarboChess 3             1 A32 2412  80  81 54 44% 2454 22%
040 Daydreamer 1.75 JA       1 A32 2408  78  80 54 43% 2465 30%
041 Danasah_5.07.JA_xb       1 A32 2395  82  81 50 51% 2385 26%
042 Rotor 0.8                1 A32 2392  76  77 54 47% 2413 35%
043 BBChess 1.3b JA          4 A32 2379  79  80 54 47% 2394 24%
044 Glaurung Mainz           1 A32 2377 108 105 30 58% 2315 23%
045 DanasahZ_0.4.JA_xb       1 A32 2369  81  82 52 45% 2407 25%
046 Dirty_030411.JA_xb       1 A32 2358  81  81 50 53% 2340 26%
047 GreKo_12.5.JA_xb         1 A32 2310  79  78 52 52% 2293 31%
048 Pepito v1.59             1 A32 2309  76  77 54 44% 2348 31%
049 Sloppy_0.23.JA_xb        1 A32 2303  78  77 52 52% 2290 35%
050 Pawny_1.0.JA_uci2xb      1 A32 2285  83  80 54 62% 2199 20%
051 BetsabeII_1.47.JA_xb     1 A32 2280  88  86 50 54% 2246 12%
052 Phalanx_XXIV.JA_xb       1 A32 2267  83  84 52 47% 2294 13%
053 TJchess 1.1U             1 A32 2253  91  86 54 65% 2138  7%
054 Ifrit_m1.8.JA_uci2xb     1 A32 2237  79  80 54 46% 2261 30%
055 Diablo 0.5.1b JA         1 A32 2236  79  79 54 48% 2249 26%
056 Maverick 1.0 JA          1 A32 2234  79  80 54 46% 2257 22%
057 Typhoon_1.0.r358.JA_xb   1 A32 2230  89  85 52 64% 2114 13%
058 Olithink_5.3.2.JA_xb     1 A32 2147  78  78 54 50% 2142 30%
059 MangoPaola_1.1.JA_xb     1 A32 2140  82  83 54 49% 2143 17%
060 Amy_0.8.JA_xb            1 A32 2109  83  84 54 44% 2154 19%
061 Natwarlal_0.14.JA_xb     1 A32 2107  84  85 52 46% 2128 19%
062 Myrddin_0.86.JA_xb       1 A32 2085  84  84 52 50% 2090 19%
063 Sungorus 1.4 JA          1 A32 2048  85  89 52 38% 2135 15%
064 KmtChess_1.21.JA_xb      1 A32 2037  83  82 52 54% 2010 23%
065 Rattate_Nosferatu.JA_xb  1 A32 2036  84  83 52 54% 2010 19%
066 Resp_0.19.JA_xb          1 A32 1994  77  78 54 46% 2012 30%
067 Scidlet_2.61b2.JA_xb     1 A32 1979  84  86 52 45% 2017 13%
068 NGplay_9.86.JA_xb        1 A32 1970  81  82 50 50% 1970 32%
069 zurichess                1 A32 1939 129 118 28 71% 1779 21%
070 DanChess_1.04.JA_xb      1 A32 1926  85  84 52 54% 1895 15%
071 robocide                 1 A32 1878  77  78 56 47% 1896 27%
072 Witz_Alpha21.JA_xb       1 A32 1876  85  85 48 48% 1898 29%
073 Clubfoot_r2293.JA        1 A32 1857  85  85 52 48% 1874 12%
074 Kurt 0.9.2.2 JA          1 A32 1840  79  78 54 53% 1823 24%
075 Woodpecker_2.11.JA_xb    1 A32 1818  90  87 46 60% 1741 20%
076 ZCT-0.3.2500             1 A32 1814  84  86 52 41% 1879 13%
077 Knightcap_3.7F.JA_xb     1 A32 1797  86  87 50 46% 1827 12%
078 Cilian_4.14.JA_xb        1 A32 1792  76  76 56 50% 1786 32%
079 Leonidas_r83.JA_xb       1 A32 1790  86  85 50 52% 1771 12%
080 BikJump v1.8             1 A32 1788  86  84 52 59% 1728 17%
081 Samchess_JA_xb           1 A32 1767  86  85 46 55% 1731 24%
082 Sjeng_1.12.JA_xb         1 A32 1746  86  91 52 33% 1887 12%
083 Ecce rev. 508            1 A32 1733  81  82 54 47% 1756 17%
084 Faile_1.44.JA_xb         1 A32 1731  83  83 46 49% 1736 33%
085 AdroitChess0.4 JA        1 A32 1726  88  86 50 60% 1641 16%
086 Surprise_4.3.b13.JA_xb   1 A32 1724 112 114 28 46% 1751 14%
087 Colchess_8.0.JA_xb       1 A32 1687  85  86 46 49% 1695 24%
088 Bismark_1.1              1 A32 1663  79  80 50 48% 1679 32%
089 Zzzzzz_3.5.1.JA_xb       1 A32 1657  92  87 40 64% 1563 33%
090 Hoichess_0.12.1.JA_xb    1 A32 1637  98  94 48 69% 1437 21%
091 Rocinante 2.0 JA         1 A32 1610  83  84 50 48% 1622 24%
092 Kitteneitor_060513.JA_xb 1 A32 1601  84  87 44 40% 1667 34%
093 Claudia v. 0.5           1 A32 1601 106 104 32 55% 1563 28%
094 Gunborg_1.0.JA_uci2xb    1 A32 1598  83  83 50 52% 1572 24%
095 Jester_0.84.JA_xb        1 A32 1594 108 104 40 64% 1444  8%
096 smash                    1 A32 1582  86  89 50 39% 1659 18%
097 Tscp_1.8.1.AB_xb         1 A32 1555  93  97 42 39% 1641 12%
098 Pulse 1.5-cpp            1 A32 1476  87  89 52 49% 1441 29%
099 Chenard_2014.07.11.JA_xb 1 A32 1425 103 112 42 38% 1512 10%
100 Colossus 4.0 100X        1 C64 1422 286 256  6 83% 1081 33%
101 VIRUTOR CHESS 1.1.1      1 A32 1373 103 101 46 60% 1273 11%
102 Chess for Android        1 A32 1350 101 100 46 57% 1276 13%
103 K2 v.071                 1 A32 1130 127 137 30 27% 1350 13%
104 Trappy_Beowulf_2.0.JA_xb 1 A32 1123 103 104 44 47% 1151  7%
105 Xadreco_5.7.JA_xb        1 A32 1074 121 131 28 32% 1238  7%
106 Colossus 4.0             1 C64 1069 288 237  6 50% 1081  0%
107 Byak 8.10.14.JA          1 A32 1049 109 120 32 27% 1244 16%
108 Joanna_2002.JA_xb        1 A32  986 140 148 28 27% 1221 11%
109 OliveChess 0.2.7         1 A32  717 208 -68 32  2% 1275  3%

Rapidroid test platform:
* Samsung GT-N7100 1.7 x4, 256 hash: All Android progs,
* Codegen Novatab tablet 1.4 x 4, 256 hash: All Android progs,
* Polypad 1010IPS tablet 1.61 x 2, 128 hash: Single thread Android progs
* HTC Diam 528Mhz, 16MB hash: Windows Mobile,
* i7 M620 2.67 Ghz dual + Arena 3.5 + 2GB hash: Windows 64
* iPhone5S A7 1.3Ghz dual: iOS programs
* DosBox 1.74: DOS programs,
* WinVICE 2.24: Commodore-64 progs,
* Messtiny UCI adapters or CB-Emu2014: Mephisto programs,
* Openings: 20 ply from Adam Hair or 16 ply from TCEC, no Q exchange, +0.15 to +0.40 by Stockfish, Komodo @ d20 min, played twice each color.
* Repeating openings and twin games not allowed between two programs
* Tablebases and pondering off
* Time control: 600+0 to 1800+5 or 10 to 30 sec/move or closest known by both programs.

Android UCI engine update: K2 0.75 JA

K2 by Sergey Meus from Russia has been updated to 0.75. It's an engine in continous development, so we should expect more updates in near future. Thanks to Jim Ablett for bringing it to Android.

Previous compile was based on v0.71. After 30 games played, it collected 1130 ELO in the Rapidroid ranking i expect to release today.

This is disappointingly far below its 2414 ELO level at CCRL 40/40 on a PC. There must have been some incompatibility somewhere which caused that.

We'll see what happens with the new compile soon.

Download K2 0.75 for Android: HERE

May 12, 2015

Is chess meant to finish in seconds?

I'm afraid i'll never manage to understand what is the purpose of testing engines at extremely low time controls like 15 secs/game and issuing results of shoot outs kind of "compile of the day" vs "compile of the day" continously.

One may get more samples in given duration and quickly obtain some rankings or some kind of results stating that engine X is better than engine Y by Z ELO points in the conditions used. Right.

So what's next? Is there yet a concrete answer to what level of accuracy does such experiment can lead us?

I have two permanent reasons to defend that bullets and even blitz levels are far behind the real world. I believe rapid chess is the optmal simulation acceptable by a serious tester, no matter whether the developers of the top engines tend to verify ELO gains obtained with new patches, for instance based on thousands of games at 60"+0.05", the "LTC" definiton of the Stockfish test network, for LONG TIME CONTROL.

Time is relative but i relatively suggest that one minute is not long at all.

My reasons are:

1) A suitable time control must be "watchable" by human eyes and should not extract games faster than humans can follow live. 15 seconds x 2 sides = 30 seconds for a game is really a lightning in the sky. Don't we need to see what's going on? If i can't watch something happening, i'd better forget about it!

2) Surrounding components such as hard disks and memory are still slow compared to the processor speeds. Impact of relatively slower components are always suspected to deviate results.

I've been active in Stockfish test network for a while for hundreds of hours, testing their patches. My latest opinion is that what's called LTC there, is exactly what i'd been teached to call "bullet" time control for years. Impatience and hunger for quick samples matters here. But real life differs often and sometimes it's too late to discover that on extreme hardware with FIDE time controls things  may change unexpectedly. Therefore you may lose the crown to a reptile. Surprise! You are TCEC'ed :-)

My verdict and my commitment is to keep 15 sec/move or 15 mins/game as recommended setting and 10 sec/move or 10 min/game as minimum acceptable time control for all my experiments. As practiced in Rapidroid...

And about other lists based on blitz or bullet time controls, i can simply congratulate for the effort but my respect is not a sign of confirmation.

To my taste, i still prefer less samples based on the simulation of a "watchable" chess game which requires no less than a Rapid chess time control, the way it's defined by FIDE. Shortly said: "Each side can use a base time for the whole game from 10 minutes to 60 minutes, be it without increment or with an increment per move to add in such a way that the total of the base time and 60 increments does not exceed 60 minutes".

That rule looks totally healthy and 600"+0" is the minimum acceptable, at least for me. Therefore if my tourneys take time, so be it. I'd better wait.

May 9, 2015

Android UCI engine update: Komodo 9

Ooops!!! It seems i was sleeping calm n' quiet in my corner waiting for Komodo 9 to replace v8 as an update. No! The Komodo team released the ninth separately. Bad news is that we need to pay separately as well.

Now it's time to wake up and realize that Komodo 9 is available at: GOOGLE PLAY

The bad news is that i've been too late to notice the new version is released. Sorry for that but there are good news too. The good news is that we can have Komodo 8 and 9 on our devices at the same time and that it's possible to let them perform against other engines in parallel to conclude a verdict.

I still think the price vs performance ratio of Komodo for Android is way higher than the PC version. The price is highly attractive and the strength obtained is competititive enough.

Let's wait for the new version compete in Rapidroid gauntlets to record a new rating. Meanwhile i'm more than interested to hear from the followers whether we should keep Komodo 8 and 9 together in the same ranking or choose v9 to replace v8.

The theory tells it's better to kick out the old and focus on the newest to avoid multiple versions of an engine and prevent eventual rating distortions.

In constrast, the public interest requests a comparison between two versions which is not representative enough when based only on head to head matches. It's generally recommended to have two versions playing against same other opponents.

So, you tell me what is better and i'll follow:
* Introduce v9 an keep v8 too in the ranking > Tolerate rating distortion.
* Replace v8 with v9 and forget the elder brother.
* Conduct a small tourney between top 5 including two Komodo's and after the verdict, continue with v9 in Rapidroid for further rounds.

May 8, 2015

RAPIDROID: May-2015 list to be released on May 19th

Too much delay... paying back many improvements:
* Engines causing interruption of the tourneys are elliminated.
* Engines terminating games early without reason are considered as buggy and elliminated.
* All previous games played with fixed time per move are replaced by Fischer clocked games.
* All previous games played at 10 sec/move are replaced by 15 sec/move.
* All bad pairings reducing accuracy are replaced by better paired groups

I have now around 1800 games after 2 rounds vs 3300 games after 4 rounds. One would think it's stupid to re-reload everything but the improvement speaks with numbers: Average scores variation and number of games variation are both better than before with almost half of the games played. That means the list develops "much quicker".

With 5 devices running full, the theoritical speed of the experiment is 3600 / (65 moves * 2 plies * 15 seconds) = 9,2 games / hour which is still not sky high but i won't lower 15 seconds anyway. With the best of the luck, it's possible to reach 9.2 * 50% efficiency * 24  = 110 games / day, as long as i remain a working guy.

Without any doubt, more usage efficiency requires to be stationary at home. Frequently checking what's happening in the devices helps keeping them loaded with tasks. But i'm not yet retired and the devices often wait for me after finishing.

Since an acceptable list needs 100 games per engine, for 107 current Android engines, i would need: (10700-1800) / 110  = 81 calendar days from now on. If i calculate a target date for the first "GOOD" Rapidroid release, July-28th is in the horizon. Then there will be May and June releases not well cooked.

Again i say, so many engines, so little time. As expected...

May 3, 2015

Robin is not a pawn


Stockfish Android under Windows!

Last time i'd attempted to emulate Android in Windows, it was almost one year ago. I'd thought it could be an alternative way to increase the number of devices in Rapidroid experiment, therefore to increase the games played per hour.

Another fantasy was to repeat what comes true with C64 emulation on a PC. If it could run ten times faster than the original, it could open door to simultaneous tourneys, many divisions playing at the same time.

Though it wasn't a stupid idea, it had totally failed with that old version of Bluestacks emulator, mainly due to:
* No multi core cpu emulation
* Very low nodes per second
* Low RAM forcing lower hash

Yesterday i've made another test with a "MODDED" new version of Bluestacks running a rooted Kitkat 4.4 with possibility to set RAM size. It looked yummy at first, as always.
After downloading a 260MB msi file from this page, the installation took around 10 minutes. Thanks to a bundled utility it was a piece of cake to transfer some apk's and chess engine binaries into that Windowroid.
The desktop with CfA and Droidfish both ready to go

But, i quickly bumped to "engine exits" error under CfA using the official Stockfish binary. I've soon realized the whole thing is emulating an Android x86 device only and it's impossible to run an Arm app, none of the arm5 or 7 chess engine binaries either!

When i've installed Droidfish and verified that it runs well with its own Stockfish engine, i've checked the engine file and discovered it was the x86_nopie version. Theory confirmed...

I've gone back to CfA and installed the x86_nopie version of Stockfish. As expected, it worked without problem the same way it works on an x86 tablet with an Intel cpu. What shocked me most was 677 kNps performance, not slow at all, given that my three quad core devices CodeGen RK3188, Asus Z3745 and Galaxy Note II run consecutively at 645, 750 and 810 kNps.
Stockfish 6 x86 nopie running on Bluestacks

Shortly said, there is no cpu instruction translation anymore in Bluestack. The developers (or modders?) must have prefered performance over compatibility. Indeed, it's a perfectly emulated x86 tablet on PC but we must forget everything specifically written for ARM.

My opinion is that somebody is slowly taking away Android from ARM domination.

Regarding 3D games on Bluestacks, i know nothing and i can't say anything. All i know is chess.