G3P - The GPU General Graphing Project

[Magos Mechanicus]

G3P
The GPU General Graphing Project

The GPU General Graphing Project aims to create a system for producing, on the GPU, graphs of arbitrary input functions.
This thread will be used as a project hub, so will include everything from status reports to help requests to (hopefully) 1 or more releases.
Any help is very welcome!

Current status (20/04/2010):
Function return values work! Time to start writing the GPU.

Jobs to do/Roadmap:

1. Write the GPU framework - displaying (graduated?) coordinate axes, function outputs as dlines etc.
2. Write the GPU components to compute the required input functions and give them to the GPU display portion.
3. Parsing of infix notation input?
4. Go 3D?
5. Port parser+compiler from E2 to CPU?

[Magos Mechanicus]

The reason I'm coming to you now is that I'm in a bit of a pickle in terms of getting my E2's machine code to execute and then hand control back to the main program.
First up, of course: Show the compiler in case that's screwing up.

@name G3P Compiler
@inputs RPN:string Device:wirelink Offset
@persist Bytecodes:table Index Compiling:array
@outputs Compiled:array
if(~RPN & RPN:length()){Compiled = array(30,4)
    #All functions produced start by popping (30) 
    #the return value off the stack into edx (4)
    #Make sure to push it back on before the ret! (9,4,40)
    Compiling = RPN:explode(" ")
    }
    
while(Compiling:count()){
    if(minquota()<300){interval(20) break} #Don't run out of quota
    Token = Compiling:shiftString()
    if(Token:toNumber()|Token == "0"){Compiled:pushNumber(9)
    Compiled:pushNumber(0)
    Compiled:pushNumber(Token:toNumber())}
    elseif(Bytecodes[Token,string]){
    Temp = Bytecodes[Token,string]:explode(",")
    while(Temp:count()){Compiled:pushNumber(Temp:shiftString():toNumber())}
    }
    else{print("Compiler error: \"" + Token + "\" is not a recognized token.")}
    #Todo: Finish clause, including pushing the return pointer onto stack
    #Todo: Check for stack safety. Balance between pushes(9) and pops(30)
        #How to exclude values which are not code but the numbers?
        #Check for next value being 1-4?
    #Output to hispeed memory
}
if(!Compiling:count()){Compiled:pushNumber(9)
    Compiled:pushNumber(4)
    Compiled:pushNumber(40)
    printTable(Compiled)}

if(first()){
    #Initialize bytecodes for operations
    #General structure of operations: Pop any arguments off stack, do operation, push result back on
    #Numbers: handled at compiletime, except for
    Bytecodes["par",string] = "9,3" #The parameter, which is kept in ecx and can be pushed to stack as needed
    #Operators:
    Bytecodes["+",string] = "30,1,30,2,10,20001,9,1" 
    Bytecodes["-",string] = "30,1,30,2,11,20001,9,1"
    Bytecodes["*",string] = "30,1,30,2,12,20001,9,1"
    Bytecodes["/",string] = "30,1,30,2,14,20001,9,1"
    Bytecodes["^",string] = "30,1,30,2,80,20001,9,1"
    #Functions:
    Bytecodes["sin",string] = "30,1,54,10001,9,1"
    Bytecodes["cos",string] = "30,1,55,10001,9,1"
    Bytecodes["tan",string] = "30,1,57,10001,9,1"
    Bytecodes["asin",string] = "30,1,58,10001,9,1"
    Bytecodes["acos",string] = "30,1,59,10001,9,1"
    Bytecodes["asin",string] = "30,1,58,10001,9,1"
    Bytecodes["log",string] = "30,1,82,10001,9,1" #Natural logarithm
    Bytecodes["log10",string] = "30,1,83,10001,9,1" #Briggs logarithm
    }

A sample of its output is perhaps also in order. For the input "3 8 * sin 20 *", or in other words the calculation 20*sin(3*8) ~ -18.1, it outputs this machine code:

db 30,4,9,0,3,9,0,8,30,1,30,2,12,20001,9,1,30,1,54,10001,9,1,9,0,20,30,1,30,2,12,20001,9,1,9,4,40

or, in zASM

pop edx; //Pop the return value for safekeeping
push 3; //Numbers are simply pushed on
push 8;
pop eax;//Operators pop arguments off, then operate on them
pop ebx;
mul eax,ebx;
push eax;//then push results
pop eax;
fsin eax,eax;//The same applies for functions
push eax;
push 20;
pop eax;
pop ebx;
mul eax,ebx;
push eax;
push edx; // And lastly we put the return value back on top
ret;

To me, it looks like it should work. then there's trying to get the CPU to execute this, which is much harder.
My test CPU is this:

// G3P Evaluation test
main:
in ecx,0;
cmp #20000,0;
je main;
call 32000;
pop eax;
out 0,eax;
out 1,1;
jmp main;

Which steadfastly outputs 0 to its port 0 and 1 no matter how I try to first write the machine code to cell 32000, then a 1 to cell 20000.
HALP PLOX!

I'm also very interested in opinions/experience with job 2: Can I actually do the calculation on the GPU, or will it have to be done on a CPU which provides ready made coordinates to dline?

[-HP-]

Expression 2 RPN compiler... damn you. I was doing that too :P

[Matte]

If I understood your question right:
I think, for this project, the GPU would suffice in computing the functions. From my understanding, nothing you're trying to do is too heavy.

[Magos Mechanicus]

The main thing worrying me was the client/server divide and lagging people's computers since the GPU executes every frame. I guess there is much that can be done by staggering the computation of the points, though.

[Hitman271]

haha, I'm actually also working on something like this.

My parser turns a chat string like this " -27x^3 + 6x - 10" and turns it into an array: ' -27x3, 0x2, 6x1, -10x0" (those zeroes are for the synthetic division)

I'm using mine to determine the roots using Newton's approximation and synthetic division aswell

[Black Phoenix]

You can draw function of any complexity (including 3D) if you draw it in blocks - spreading the whole rendering over several frames. You can draw graph with 1000 points if you wish, it'll just take some time. Or surface with 1000 polygons.

[Magos Mechanicus]

Well, a little progress - I've successfully demonstrated that the functions can be called and then return properly. Their return value is not working so well, however. Even

30,4,9,0,3,9,4,40
i.e.
pop edx; push 3; push edx; ret;

somehow manages to return 8.48 instead of 3. Every function regardless of the input ends up doing so. I'll keep plugging away at it.

Interesting that so many of us started doing this kind of thing simultaneously. Wonder if there was anything that set us all off?

[-HP-]

Well, a little progress - I've successfully demonstrated that the functions can be called and then return properly. Their return value is not working so well, however. Even

30,4,9,0,3,9,4,40
i.e.
pop edx; push 3; push edx; ret;

somehow manages to return 8.48 instead of 3. Every function regardless of the input ends up doing so. I'll keep plugging away at it.

Interesting that so many of us started doing this kind of thing simultaneously. Wonder if there was anything that set us all off?

I thought of it when you posted somewhere something about compiling CPU and how the code works etc. - Even if I already knew it, that put me on the idea. And since I already had an RPN E2 thing, I thought I might give it a shot. But I never really got started, I don't play GMod very much any more.

[Fizyk]

I thought once of making a program that would graph functions on a digital screen My approach would be a bit different though - I would make a parser function, that would parse the whole expression for each point to generate table of values, and then graph it. This wouldn't require compiling anything (I made something like that a long time ago in Pascal).

Well, a little progress - I've successfully demonstrated that the functions can be called and then return properly. Their return value is not working so well, however. Even

30,4,9,0,3,9,4,40
i.e.
pop edx; push 3; push edx; ret;

somehow manages to return 8.48 instead of 3. Every function regardless of the input ends up doing so. I'll keep plugging away at it.

Try returning values through eax instead of stack, usually it's a better idea.