Is Max slower
BlitzMax Forums/BlitzMax Programming/Is Max slower
| ||
| http://www.deviantmachine.com/link/sieve.zip Please download the sieve Test, re-name the MaxTest to somethinge else "Max Sieve2.exe" for example, then rebuild the test from the sorce. Are they the same speed? On mine, when I compile it, its concistantly 5% slower no matter what settings. Does this mean that Bmax is now 5% slower than march 2006? |
| ||
| hmmm. I get 25% slower and a lot bigger. old - 1768 new - 2227 size: old - 21.5KB new - 81.0KB |
| ||
| More features being added to modules and whatnot explains the size, but not necessarily speed. |
| ||
| old - 4175 new - 4522 size: old - 21,5 KB new - 81 KB Athlon64 3500+ WinXP Pro sp2, Max 1.24 |
| ||
| We need support for the latest mingw in the next version. There are many optimization which can't be used with BlitzMax because of that. |
| ||
| About 33% slower here, which would be a huge discrepancy if it held true in broader tests. Old one : 2482 New one : 3265 Did we change FASM versions in there somewhere? We need support for the latest mingw in the next version. There are many optimization which can't be used with BlitzMax because of that. Sure, but that's completely irrelevant to this test. |
| ||
| Anyone noticed how slow the purebasic one is? |
| ||
| About 33% slower here, which would be a huge discrepancy if it held true in broader tests. True, but as this bechmark has been often towted as a trueish comparison of the different Languages, (That why I downloaded it), it must hold that its still a trueish comparison of the BMax versions.If you have some more General tests, and still have older Bmaxes knocking about, could you run some tests? |
| ||
| Old Max : 4406 , size 21.5 New Max 4397, size 81kb |
| ||
| Old: 2419 New: 3053 (using latest MinGW) |
| ||
| True, but as this bechmark has been often towted as a trueish comparison of the different Languages, (That why I downloaded it), it must hold that its still a trueish comparison of the BMax versions. Well it might, depending on how accurate the people doing the touting were. I can't say as I've spotted a similarly massive drop in the speed of my game project, but I haven't been monitoring the old FPS counter too fanatically, so it's possible. I'd certainly like to see some more tests if indeed anyone does have an old version. |
| ||
| how did tonyg get faster? |
| ||
| how did tonyg get faster? I think its safe enough to say 4406 and 4397 are statisticly the same. But yep, how come I wonder? |
| ||
| Old: 5177 New: 5914 |
| ||
| I think it might be related to the GC. I was able to get much faster execution just by moving the Local declarations outside the loop. Possibly the older MAX has a faster GC? original 4716ms 22K New 4786ms 84K Locals outside of loop 4004ms 84K |
| ||
| No the older Max has a worse handling of local scope which might be the reason for this behavior. It was optimized to give the "i want control over memory usage" yellers more control. But makes sense that a local outside a loop is by far faster as it does not enforce cache reaction on every loop iteration (as locals go into cpu registers) |
| ||
| I just tested with the locals moved and got 3065, still 25% slower than the old versions of Blitzmax with the locals *inside* the loop. Even assuming that the old version would not be any faster with the locals moved outside the loop, that still leaves a pretty huge 25% speed loss unaccounted for. |
| ||
| Are you on a stoneage AMD with 256kb Cache or a P4? |
| ||
| No, I'm on an Intel Core Duo 6600 ( you probably could have surmised I'm not on an ancient machine, since my numbers are lower than everyone else but two. ) |
| ||
| 1. Don't see why this should be a GC issue at all, as the GC isn't used at all (only ints are used, they are not reference handled) 2. The main difference between old and new blitz is that the new one has modified math code. Thats the reason that people with newer MingW versions should most likely outperform all with stock BM and MingW 3.1 My times: Old: 2050 (+- same as C++) New: 2450 (+- same as C#) the rest is 5 seconds upward Core 2 Duo E6600 |
| ||
| I was able to get much faster execution just by moving the Local declarations outside the loop I did all of that messing about trying to speed it up by moveing things before I posted the question. And yes it is faster moving the loacal declarations arround, but without a previous Bmax to test it on, I had to assume that it would also make that version faster.Also, Im willing to say that 4716 and 4786 are the same speed anyway. Apart from TomToad and TonyG, both of whoms test speeds can be called "Equivalent", everyone has a slowdown on the present version compared to the older version, rangeing from 5% to 33%, and even after optimising the Variable allocation on the newer version its still slower on the ones its slower on. So 1) What version of Bmax is the original test from? Anyone know how to find out. 2) What Machines are Tonyg and TomToad using? TonyG WinXP SP2, Athlon 2700+, 9800Pro, 1G Ram |
| ||
| Old: 7600 New: 8755 New New: 7729 Looks like Mark's already addressed this (and yes my computer is slow!). |
| ||
| What's the "New New" version? |
| ||
| I was about to ask the same? is it the upcoming update? |
| ||
| Probably the next update. Dabz |
| ||
| And for Mac/Linux someones has old version to compare? Only for curiosity... |
| ||
| New New = latest dev version. |
| ||
| New New = latest dev version. Speaking of which, anyone know when the new update to BlitzMax is coming out? Simon? Mark? |
| ||
| 2346 - c# 2515 - C++ 3063 - OldMax 3435 - NewMax 7578 - Blitz3D 7609 - pb Interesting line up. Glad to hear the next version is getting a re-boost. :) |
| ||
| I still have a copy of Max 1.12 installed and I got 4945 ms with a compile size of 49K, which is about 200ms slower than 1.24 Interesting, I wonder just which version is actually used for their test. System: ASUS A7N8X motherboard AMD Athalon XP 2800+ 2.1Ghz 512 meg Ram NVidia Geforce FX5200 |
| ||
| Hmmm I just noticed the url for that test. I thought I had left a readme in the zip but I guess not :/ I grabbed the c++ and max version from someone else's test. Can't remember who, but I'm sure a search on here will find the original thread and author. No clue on the c++ compiler used( assume its either vs or gcc ). I seem to recall having someone from #blitzcoder compile the blitzmax version for me. I know for sure I ported the c# version( from the c++ one ), and compiled it with the microsoft c# 2.0 compiler. I'm 99% sure I did the purebasic version as well( unless I'm going crazy ). It was compiled with pb 3.93. It would be good to recompile it, since purebasic's gotten some big updates. I'd do it myself, but I've lost my registration details, and my email to get my account sorted was ignored :) Theres a freebasic version, but I guess it didn't make it in that zip. Stonemonkey ported it and compiled it. I don't know if I still have that. From what I remember, it came in about mid-pack. |
| ||
| Speaking of which, anyone know when the new update to BlitzMax is coming out? Simon? Mark? Next week-ish, all going to plan. |
| ||
| Is there anyway to know what has been changed / dropped / added in the new version? |
| ||
| I'm 99% sure I did the purebasic version as well( unless I'm going crazy ). It was compiled with pb 3.93. It would be good to recompile it, since purebasic's gotten some big updates. I'd do it myself, but I've lost my registration details, and my email to get my account sorted was ignored :) Recompiled the PB source under the latest PB. Exactly the same results. |
| ||
Tested like you said: moved locals out of loop, changed (think only visually) "prime = prime + 3" to "prime:+3" and so on.
' BlitzMax sieve test
Framework BRL.StandardIO
Import BRL.System
GCSetMode 2
Const ITERATIONS:Int=50000
Print "SIEVE OF ERATOSTHENES - "+ITERATIONS+" iterations"
Global t:Int=MilliSecs()
Local flags:Int[8191] , i:Int, k:Int, prime:Int
Local Count:Int = 0
Local Iter:Int = 1
For Iter = 1 To ITERATIONS
i = 0
k = 0
Count = 0
For i = 0 To 8190
flags[i] = 1
Next
For i = 0 To 8190
If flags[i]=1
prime=i+i
prime:+3
k=i+prime
While k <= 8190
flags[k] = 0
k:+prime
Wend
Count:+1
EndIf
Next
GCCollect
Next
t=MilliSecs()-t
Print ITERATIONS+" iterations took "+t+" m/secs."
Print "Primes: "+Count
Input "Return to end ..."
End
on my crappy amd 1700xp: old: 6440 new: 5483 new = all update up to august, most recent fasm, normal (old) mingw. bye MB |
| ||
| Well thank you MichealB, but as was pointed out before, this was pointless, unless you have the olderversion of Bmax, and can compile the optimised version on that. |
| ||
| We all have an older version. Its called BM Windows Demo and its 1.12. Much older than stoneage isn't possible. But aboves changes don't make much, even if you replace i+i with (i shl 1) |
| ||
| We all have an older version. Its called BM Windows Demo and its 1.12. And we know that that isnt the one that the old example was compiled on, because Tom already compiled the program on it, and found it slower than the precomiled example in the download.We are not saying that the older the version the faster it is, but rather there was a sweet point. And that it doesnt matter what you do to optimise it now, the sweet point program would also need to be optimised for comparison. |
| ||
| Just wanted to say that the new one finishs faster than the provided application in the first post... That you can't really compare coz of the done changes in sourcecode is known to me. But to help: Exe-File is dated to march 2006. First Version including GC is v1.12 ?! (v1.09 doesn't know GCSetMode 2) Think it should be v1.14 (would fit to the date) - will test it when have the next sparetime ;D. EDIT: argh, its not v1.14 ... filesize 50kb, speed about the same. bye MB |
| ||
| I'm curious what other new features the coming release will have... |
| ||
| Ok... filesize wont help much when it comes to comparison, just had a look into the old exe file ... it got upx-compressed. So stripped and compressed will change a lot. EDIT: ok. uncompressed it ("upx.exe ExeFile -d") and voilą: 50kb. I don't know if my fasm or mingw isn't the same, but 'only' 2-3kb of the old file are different to mine of blitzmax 1.14. May be we narrowed the options down to 1.14-1.16. bye MB |
| ||
| @Michael, it might be a doomed exercise, because all the Mod updates that may or may not have been added to the one in the zip file. Its even possible that the Zipped on was made with a version of Bmax that was sycned in the AM had something wrong with-it that made it faster, and was re-synced in the after-noon with a corrected version without the error but hence slower. All I wanted you to realise (which you have said you did anyway), was that it wasnt a question of opimisation. My personal Vote for the reason, is that the GC was changed/improved which resulted in better and faster auto GC but slower manual GC, as this would explain why ppl with a massive 33% slower wouldnt notice it as much. |
| ||
| Will bmx magui also get an update? |
   |