Speed Up Your Macros

From MapToolDoc
Jump to: navigation, search
ADVANCED
THIS IS AN ADVANCED ARTICLE

Introduction

If you start creating your own framework and if you like the process, then most likely you will get to a point that some of your more advanced macros start to become a drag. MT script isn't the fastest of languages and there are a couple of functions or methods that can really slow things down. Fortunately a couple of users (like Aliasmask) have started testing different methods to speed up their code. Below you can find the results, some tips are based on conjecture others have been throughly tested to be faster. I you find a new faster method, don't hesitate to put it here.

Do it yourself

Recently I've added a new function to the Bag of Tricks called Benchmark. This function can be found in the Setup tab of the virtual menu. On lib:EventMacros there is a group called 'Benchmark Macros' with one macro called 'Default' which should be empty. In this group you can create new macros. When you hit 'Benchmark' then the execution time of these macros will be calculated. The Default macro is used as reference and that execution time is deducted so the actual execution time of your added macros remains. The result is shown in the chat. If add one macro you will be presented with the execution time of that macro on that PC. If you add more then one macros then a comparison report is shown as well so you can get a feeling of which methods is faster.

This way you can run your own tests by simply adding your code to that group and running the Benchmark Macro. In the latest lib:EventMacros you can find two example macros 'if roll option' and 'if function' to show you how it works.

Macro vs UDF vs directly

This is again something to take into consideration. Sometimes you need to split up macro because of the CODE nesting limit, sometimes to prevent stack overflow and sometimes cause its easier. Here's the impact of your choices:

I've tested the macro:

   [r:"Test"]

in three ways. Once directly, once through a User Defined Functions (UDF) Test() and once through the macro call "Test":

   [r:"Test"]
   [r:Test()]
   [r,macro("Test@lib:cifStopwatch"):""]
   [r:eval("Test()")]

The result is (with 10,000 cycles): directly 8.5 seconds, UDF 14.3s, Macro 18.5 and Eval 15.4 seconds. In short, if speed is of the essence, try to keep it into one macro. If you need to split up: use UDF as much as possible. If you feed it one variable (argument) then the total time for both the UDF and the macro increases with 1 second.

Storing and Retrieving Variables

You can store a variable in three ways:

The fastest method to retrieve a simple value is from the identifiers. If the time to retrieve a value from an identifier takes 1 second then the same value takes (on average) 1.2 seconds using getLibProperty() and 1.8 seconds using getProperty(). The same test but using a heavy json object: if we set the identifier again on 1 (still the fastest) we notice that: getLibProperty() is still 1.2 however getProperty() time has increased to 2.8 seconds. The json used was constructed out of 1000 identifiers. And the time average was taken over 10,000 loops.

Now the surprising part: To set a value one would expect similar results but that ain't the case. Using the same heavy json it turns out that token.gm_name was the fastest and token.label the slowest !!! If gm_name is set to 1 second than the rest is: 2 seconds for both setProperty() and setLibProperty() (yes equal speed) and 2.4 seconds for token.label. Again 10,000 loops used to test.

I've ran more test, to see which method is faster to store multiple simple variables onto a libtoken and retrieving them again:

  • Using json
  • Using strProps list
  • Each variable seperately.

The last method is by far the slowest (10x the time for the other methods). Using json or strProps does not really make a lot of difference though strProps are faster. What I tested:

    [testJson = json.set("{}", "test0",test0,"test1",test1,"test2",test2,...,"test9",test9)]
    [testJson = json.fromStrProp(strPropFromVars("test0,test1,test2,...,test9","UNSUFFIXED"))]
    [testStrProp = strPropFromVars("test0,test1,test2,...,test9","UNSUFFIXED")]

I also tried the strFormat trick (though the I could not properly retrieve the json object form the lib with this method:

    [testJson = strformat('{"test0":"%{test0}","test1":"%{test1}","test2":"%{test2}",...,"test9":"%{test9}"}')]
    [testStrProp = strformat('test0=%{test0};test1=%{test1}...;test9=%{test9}')]

Of these 5 methods the strPropFromVars() and the strformat() methods were the fastest: 9.1 seconds (10,000 cycles) and json.set() the slowest 13.1s. The json.fromStrProp() was slightly only slower 9.6s.

Retrieving the data showed roughly the same result, strProps are a bit faster:

    [result = getLibProperty("testJson","lib:OntokenMove")]
    [varsFromStrProp(json.toStrProp(result))]
    [result = getLibProperty("testStrProp","lib:OntokenMove")]
    [varsFromStrProp(result)]

Using another method to retrieve the json vars e.g.

    [foreach(item, result):set(item, eval(item))]

is considerably slower.

Another thing that is interesting is that using the above varsFromStrProp() and strPropFromVars() it hardly matter how many variables you set. I've tested this with setting 2 and 100 variables in one go. It turned out that strPropFromVars() took 4x longer (4ms to set 100 vars vs 1 ms to set 2) and varsFromStrProp() was equally fast for both 2 and 100! (ok a very small difference, 2 takes 0.9ms and 100 take 1.1ms). This was tested again with 10,000 cycles (I divided the results by 1,000 to get to the ms).

Storing and Retrieving Variables II

I've noticed that storing a LOT of properties on a token influences the set get-property method in a negative way. This I wanted to quantify.

Keep in mind, this is quite a bit technical and mainly shows the effect of storing a certain amount of data on a token in an absolute sense. This was tested on an average PC

First test: setProperty() x1,x2,...x10000 to the value 10 and then immediately retrieving it (getProperty())

Result: no time difference between the first and the last set value.

Second test: changing the variable to a 26+ characters and the value to 260 characters. The total get/set time increased a bit, but again not changes whether is was value 1 or 1000.

Third test, same as the second, but now I raised the cap to 10,000. Finally we see something. The initial get/set time is 30,5 ms (averaged over the first 100 samples). The last 100 (average) took 130 ms !!

It turned out that with these variables and values that there was no change for the first 1000 variables and around 1200 a linear increase started of roughly +8ms per 2000 variables. Around 9200 an exponential increase started. Variable 9176 took 96ms and variable 9290 took 118ms. Finally point 10,000 took 130ms (note that the times are averaged over the last 1000). If you look at the individual results then its interesting to note that the last 100 points generally took 100ms but every 14th variable took nearly 5x as long (475 ms). getsetpropertiesTime.jpg

Next test was increasing the length of the variable (so not the value). Starting with 'i', 'i1', 'i12', etc upto 'i1234....1000'. This remained at 21ms per get set 21 variables, after that a pretty steep linear increase of 40 ms extra for the 10000 variant. (20 took 20ms and 1000 took 60ms.)

Fourth test was increasing the length of the value, with the same method as directly above. This resulted in two linear increases with the break at 400. In short the increase from 1 to 400 characters is 22 to 28 ms (1.25ms increase per 100). While from 400 to 800 is 28 to 46 ms (4.5 increase per 100). Here I was still using a new property name for each variable. Which means that the previous variables also stored on the token will also influence the time.

In the final test I simply increased the length of the value with the same method as above, so this gives insight in the actual length of the value. Surprisingly this made hardly any difference. The increase from 1 to 10,000 characters was about 2ms. Which raises the question: when *does* it go wrong. So I bench-marked (in a 1000 test loop) writing and reading 1,10,100,1000,10000 and 100000 characters. For 1 to 1000 its roughly the same, after that it doubles for 10,000 and 100,000 is about 87xItalic text slower!!. Further testing (taking inbetween points) shows that its an exponential of 1.54. If you would assume that a processing time of 100ms (0.1 second) is acceptable, then you're at 50,000 characters.

Token(), GetProperty() and switchToken()

When you need to retrieve a property from a token that is not currentToken() these are the three methods to get it. Now there are quite a lot of ways of how to use these and even combinations of these. I've ran two tests both tests with different approaches. The first test I retrieved 1 value, "Dexterity", which contains a long (150 char.) string value:

  [tmp = getProperty("Dexterity", "Eagle")]              0.9ms
  [token("Eagle"): tmp = getProperty("Dexterity")]       1.3ms
  [switchToken("Eagle")][tmp = getProperty("Dexterity")] 1.3ms
  [token("Eagle"): tmp = Dexterity]                      1.3ms
  [switchToken("Eagle")][tmp = Dexterity]                1.3ms

In the second test I ran the same macros, but now I executed the above lines 10 times, but it turned out they were all equally fast with the exception of [token("Eagle"): tmp = getProperty("Dexterity")] which was about 50% slower then the rest. So I've upped the number to 50x:

  [tmp = getProperty("Dexterity", "Eagle")]                    50x 18ms
  [token("Eagle"): tmp = getProperty("Dexterity")]             50x 29ms
  added:
  [token("Eagle"), CODE:{  [tmp = getProperty("Dexterity")] 50x }] 18ms
  [switchToken("Eagle")] 1x   [tmp = getProperty("Dexterity")] 50x 18ms
  [token("Eagle"), CODE:{ [tmp = Dexterity] 50x }]                 17ms
  [switchToken("Eagle")] 1x [tmp = Dexterity]                  50x 16ms

If you take into account minor fluctuations of my pc then again you can only say something of the second line, which apparently is the slowest, the rest is (surprisingly) still equally fast.

Fortunately I have a humongous json structure lying around (9,000 key value pairs and 175,000 characters) so I tried that one, only realizing after 10 minutes or so that calling such an object say 300,000 times might not be the smartest thing... so after rebooting my pc I tried again, this time first 1 (=300 times retrieving the object which took 14 seconds) and then 10x (137 seconds, its always nice if its scales along a straight line). This only made things even more confusing:

  [tmp = getProperty("Dexterity", "Eagle")]                    50x 2s
  [token("Eagle"): tmp = getProperty("Dexterity")]             50x 2s
  [token("Eagle"), CODE:{  [tmp = getProperty("Dexterity")] 50x }] 2s
  [switchToken("Eagle")] 1x   [tmp = getProperty("Dexterity")] 50x 2s
  [token("Eagle"), CODE:{ [tmp = Dexterity] 50x }]                 1.8s
  [switchToken("Eagle")] 1x [tmp = Dexterity]                  50x 1.6s
  added:
  [switchToken("Eagle")] 1x [tmp = Dexterity] 1x [tmp1 = tmp]  50x 1.6s

I think the conclusions you can derive from this are:

  • Don't do [token(tokID): var = getProperty(property)], either use [var = getProperty(property,tokID)] or [token(tokID): var = property].
  • If you need only a few properties of a token use getPropety(property, tokID).
  • If you need a lot of properties use switchToken() and then access the vars straight away.
  • If you use switchToken() and then assign the property to a local var, or you just keep using the property straightaway, makes no difference (although in the latter you might inadvertently change the property).

jsons

  • Try to avoid nested json objects (so json object within a json object). Objects within a json array is likely better.
  • When storing a json as a property on a token, try to limit the get/setproperty. Do it once, store it in a local variable and pass it along also into submacros. This also accounts if you're changing a property directly (so without get/setproperty) e.g.:
  1. <!-- this (using get/setPropery) -->
  2. [HP = getProperty(tokenName, Hitpoints)]
  3. [HP = HP-1]
  4. [setProperty(tokenNam, HP)]
  5. <!--is the same as this (changing property directly)-->
  6. [Hitpoints = Hitpoints - 1]
  • It might be the case that converting (using encode()) a json to string and then storing it on a token. Retrieving it using decode().
  • If you want to store a huge and complex json variable temporarily on a token, don't use a property but use token.gm_name (or token.label or token.name) to store it (using a lib token for that). It goes without saying that this is a bit a of an extreme method, i.o.w. a hack. If you were to e.g. use the token.name variable on a lib token, interesting (that you don't want) stuff will happen.
  • Overall keep in mind that read/write operations will get slower with the length of the json object. With extremely large json objects the read/write speed can bog down to 1 or more seconds per operation.

jsons object vs json array vs lists

For simple operations:

 slower ------------------------------------------------------> faster 
json object operations --> json array operations --> list operations

The operations were building the structure and retrieving all values. The speed differences are significant!!!
These test were done by comparing getting and setting 1000 x and y coordinates:

  • 1 list with x items, with every item being a list with y items, using different separator: "1,2,3,..; 1,2,3,..; 1,2,3,.."
  • 1 array with x items, where every items contains y items: [[1,2,3,...],[1,2,3,...],[1,2,3,...], etc]
  • 1 json containing x*y keys: {"x1y1":{"x":1,"y":1}, "x1y2":{"x":1,"y":2}, etc}

Obviously there are situations where a json object or array will be faster just because its smarter coding or much easier to use them. So only give value to this test if you want to do something similar as done with this test.

functions

Nested functions

It seems I had it wrong before. I had it from hearsay, now I've benchmarked it myself. Nested is much faster then unnested. First I tried one nested function vs unnested for 10,000 cycles the result was 10s for nested and 15s for unnested. Then next test I ran a really nested function:

    [varsFromStrProp(json.toStrProp(json.fromStrProp(strPropFromVars(theList,"UNSUFFIXED")))]

vs unnested

    [tmp = strPropFromVars(theList,"UNSUFFIXED")]
    [testJson = json.fromStrProp(tmp)]
    [tmp1  = json.toStrProp(testJson)]
    [varsFromStrProp(tmp1)]

Running both 10,000 times resulted in: Nested 14s and Unnested 31s. It might not help the readability of you code, but nesting your functions can be more then twice as fast!!!

Loop speeds

The following loops:
[count():]
[for():]
[foreach():]
take exactly the same amount of time to roll a 1d100 10000 times. In other words, they're equally fast.

  • CIF's stopwatch was used to measure this

This means that you can and should use the right loop function for the right reason. Some examples of good use: Some examples of proper use:

  • use [foreach():] to loop through a list or json array
  • use count(n) if you want to execute a routine n times
  • use for(i, n, 0, -2) if you want to use an a-typical but regular countdown from n to 0, using i in your routine.

macros

When getting arguments within a UDF (user defined function):

  1. <!-- Slow -->
  2. [h: var1 = json.get(macro.args,0)]
  3. [h: var2 = json.get(macro.args,1)]
  4. <!-- Faster -->
  5. [h: var1 = arg(0)]
  6. [h: var2 = arg(1)]

Notes:

  • If you use the [macro():] function you can only make use of the macro.args method (the slow way).
  • This method doesn't work the other way around, if you set macro.return within a UDF you cannot use arg(0) from within the function you called the UDF from. E.g.;
  1. <!--after calling some UDF:-->
  2. [h: doSomething(var)]
  3. <!--this works-->
  4. resultOfDoSomething = macro.return
  5. <!--this won't-->
  6. [resultOfDoSomething = arg(0)]
  7. <!--actually most likely it will 'work' but it won't contain the value you want -->

eval vs evalMacro

Test eval() vs evalMacro(). Used the following two functions:

 [h:eval("table('tbl_Image',2)")]

and

 [h:evalMacro("[table('tbl_Image',2)]")]

not quite surprising result: eval was twice as fast as evalMacro. Actual execution times: eval: 0.8 ms evalMacro: 1.6 ms

Tested other things as well e.g. "2+2", which resulted in roughly the same times.

Basically they're both very fast so it doesn't really matter which you use. However as soon as you start a loop using this then it might get relevant!

Tokens

Though this isn't really about macros, it is about speed. What you put in your tokens will also effect the snappiness of the game play.

  • Having a lot (guesstimation >100) of macrobuttons on a token will influence dragging it on the map (slow it down). Note: this issue has been partially fixed in MT by Rumble around b70-75. It still has impact on speed, but not, by a long shot, as much as it used to be.
  • Having a token with lots of data stored on it, will effect the update of movement of a token on other pc's connected to the server.
  • Large image on a token will also influence speed, try to keep them at 200x200 pixels or lower.
  • Finally having a LOT of data on the token will influence the write speed for ALL properties. here a test result:
 setProperty on empty token      	took 2.4 ms
 setProperty on 'lots of data token'	took 82 ms
 getProperty on empty token      	took 2 ms
 getProperty on 'lots of data token'	took 2 ms

<!-- Comments -->

There are two ways to put comment in MT script:

  <!-- this is open comment -->

Note the space " " after '<!--'. This is essential or it won't be seen as comment. Or:

  [H:'<!-- this is hidden comment -->']

Note the quotes ' ' at the beginning and end, again you get errors if you forget them.

These two methods both have a big pro and a big con. The open comment is processed very fast. On a moderately fast pc it takes about 100 ms to process 10,000 lines (100 ms is the border time you start to notice in macro execution). In short you can use these freely. Do keep in mind, though, that if you put comment in a e.g. count(1000): loop, then this adds 1000 lines of comment to your code! The big con of the open comment however is stack. I've benchmarked this as well and it turns out to be completely system dependent, but I noticed that the text of about half a page of a book, ported straight to the chat, will render a stack overflow with a stack set to 5!!! That is not a lot of text. The best method to omit this issue is by setting the output of the macro standard to 0 in the UDF and use macro.return = result at the end. Another method is by making sure that at least all your loops are hidden [h,foreach(),CODE:{}] so all the comment you put inside can be open.

The hidden comment thus has the big advantage that it does not add to the stack and the chances of a stack overflow are a lot less. However the big drawback is that its relatively slow. Mind you its still pretty fast, on (again a moderately fast pc) it takes 4 ms to execute, which means that it gets noticeable after around 250 lines. If however you keep slower systems in mind as well, this number might easily become half that! Another big advantage for the more experienced coder among use: if you use the console to check the running code [h:'<!-- -->'] shows up, <!-- --> doesnt! So to track which routine is currently active I always start my macros with [h:'<!-- macro name -->'].

What I personally do is use [h:'<!-- -->'] outside any loops and if(),CODE statements and <!-- --> inside these loops and if statements. I obviously make sure that these routines are all hidden.

Here a more in depth discussion on the subject.

To [h: or not to [h:

In a routine like e.g.:

  [h,count(100), CODE:{
    some code
  }]

you can put in [h:some code] or [some code] which to the end result would make no difference. It turns out that no discernible difference between the two.

if(): or if(,,)

There is the if() and [if():]. A simple test:

  [tmp = if(1,1,0)]
  [if(1): tmp = 1, tmp = 0]

Resulted in the [if():] to be twice as slow. Although this is only 0.5 ms. Essentially they are both very fast. The if() takes about 0.5 ms and the [if():] 0.9 ms. Do keep in mind that in the if() evaluates BOTH TRUE AND FALSE and [if():] only true OR false. So if you have a rather complex operation for the true and or false you might be faster of with [if():].

if(x<0,0,x) or max(0,x)

A comparison between if() and max(), where x was randomized (between -5 and 5) Resulted in both functions being equally fast. max() occasionally turned out to be slightly faster, but we're talking 0,08ms (==0,00008 seconds) which really is negligible.

String concatenation, strformat, add, concat and +

The following lines of code were tested 10,000 times:

	[strformat("%{var1}%{var2}%{var3}%{var4}%{var5}%{var6}%{var7}%{var8}%{var9}")]		1.9377ms
	[strformat("%s%s%s%s%s%s%s%s%s", var1, var2, var3, var4, var5, var6, var7, var8, var9)]	2.0979ms
	[concat(var1, var2, var3, var4, var5, var6, var7, var8, var9)]				2.0056ms
	[add(var1, var2, var3, var4, var5, var6, var7, var8, var9)]				2.018ms
	[var1 + var2 + var3 + var4 + var5 + var6 + var7 + var8 + var9]				1.9944ms

The realistic conclusion is that THEY ARE ALL EQUALLY FAST !! Looking at it at a more 'anal retentive' point of view: strformat() is both the fastest and the slowest. This you start to notice when you have more then 600 lines of this type of code in one macro!! A second test made add() the fastest, so there's also some marginal error...boiling down to the realistic fact that there is no difference!

Of course, as is the general belief, you'll notice the real speed difference when using functions... so I tested that as well (different pc so different times, you can't derive anything between the speeds of test1 vs this test.):

	[h:var1	= getProperty("var1")]
	[h:var2	= getProperty("var2")]
	[h:var3	= getProperty("var3")]
	[h:var4	= getProperty("var4")]
	[h:var5	= getProperty("var5")]
	[h:var6	= getProperty("var6")]
	[h:var7	= getProperty("var7")]
	[h:var8	= getProperty("var8")]
	[h:var9	= getProperty("var9")]
	[strformat("%{var1}%{var2}%{var3}%{var4}%{var5}%{var6}%{var7}%{var8}%{var9}")]
 
	[strformat("%s%s%s%s%s%s%s%s%s", getProperty("var1"), getProperty("var2"), getProperty("var3"), getProperty("var4"), getProperty("var5"), getProperty("var6"), getProperty("var7"), getProperty("var8"), getProperty("var9"))]
 
	[concat(getProperty("var1"), getProperty("var2"), getProperty("var3"), getProperty("var4"), getProperty("var5"), getProperty("var6"), getProperty("var7"), getProperty("var8"), getProperty("var9"))]
 
	[add(getProperty("var1"), getProperty("var2"), getProperty("var3"), getProperty("var4"), getProperty("var5"), getProperty("var6"), getProperty("var7"), getProperty("var8"), getProperty("var9"))]
 
	[getProperty("var1") + getProperty("var2") + getProperty("var3") + getProperty("var4") + getProperty("var5") + getProperty("var6") + getProperty("var7") + getProperty("var8") + getProperty("var9")]

The respective times ('nothing' is an empty loop for reference):

 nothing count took			0.2203ms
 strformat("%{var1}%{var2}%{var3}etc")	3.5844ms
 strformat("%s%s etc", var1, var2, etc) 1.0036ms
 concat() 				0.9629ms
 add()					0.9591ms
 var1 + var2				0.9638ms

The first strformat() is obviously slower as the vars need to be defined each loop. So a bit more 'fair' is to move the [var1 = …][var2 = …] etc. outside the first strformat() loop. If you do that and rerun then you get:

 strformat("%{var1}%{var2}%{var3}etc")	0.8833ms
 strformat("%s%s etc", var1, var2, etc) 0.9929ms
 concat()				0.9565ms
 add()					0.9524ms
 var1 + var2 				0.9509ms

And here we see again that strformat() is both the slowest and the fastest, but the speed difference is so insignificant that I would certainly not start rewriting my code. If you happen to have one macro that has 10,000 lines (indeed: ten thousand lines) of the slowest type then rewriting it to fastest type would save you 1 second execution time. In short: not worth the effort.

Assign Function result to Variable first or not (before multiple use)

Sometimes these tests surprises me, like in this case. Lets say you need to retrieve the same x position of a token multiple times, for example in a more complex macro. From a coding perspective its always faster to assign it to a variable first and then use it. However the benchmark told me something completely different:

The test code:

	[h:bot_startTime()]
 
	<!-- empty loop, so you can extract the exact time it takes for the operation alone -->
	[h,count(1000): 0]
	[h:loopTime = bot_subTime()]
 
	[h:Tx		= getTokenX()]
	[h:Ty		= getTokenY()]
	[h,count(1000): y = Tx + Ty]
	variable assignment: [r:bot_formatTime(bot_subTime() - loopTime)]<br>
 
	[h,count(1000): y = getTokenX() + getTokenY()]
	multiple function call: [r:bot_formatTime(bot_subTime() - loopTime)]

which surprisingly resulted in:

variable assignment: 0.112s multiple function call: 0.039s

So using variables takes nearly 3x as long !! Keep in mind though that this is tested for a relatively simple function. Stuff like json.get() might render quite some different results. Which is why I pasted the entire code block here for future use. (The bot_ functions can be found in the Bag of Tricks).


--Wolph42 08:52, 12 August 2010 (UTC)