How to measure FLOPS of a MatLab function?

I'm running MS-OS XP, service pack 3, Pentium 4, CPU 3.4GHz.

Matlab 7.11.0.584, 32 BIT.

BACKGROUND:

I recently installed a FLOPS toolbox called Lightspeed by Tom Minka. http://research.microsoft.com/en-us/um/people/minka/software/lightspeed/

The install went well, and I'm able to access the commands.

I'm running a function I've created that involves addition, subtraction, index searching, if-statements, for-loops, conditional statements, large database structures, etc...

QUESTION:

How do I use the FLOPS commands from the Lightspeed Toolbox to measure the number of flops performed by the function? I only want the total number of flops performed by the function.

And if that doesn't work, what other method exists to measure flops of a function in Matlab and how do I use those?

Answer 1

[Walter Roberson]

You cannot use the flops() command from that toolbox to measure the number of flops performed by one function: you have to rewrite the code so that every operation you want to be counted is specifically measured, such as by changing multiplication calls to flops_mult() calls.

What are you trying to measure, exactly? "Idealized" floating point operation counts sometimes have only a rough resemblance to actual code execution times.

[Steven Bratt]

I was trying to measure the overall number of flops to run the function. I wanted to measure flops for multiplication, for subtraction, for array index searching, for-loops, things of that nature.

Now by using flops_mult(), do I need to make a separate line of code for that? Or can it be part of my equations? What I mean is, do I choose 1) or 2)?

1)

x = A*b; %A is a matrix, b is an array

addflops(A*b,flops_multi)

2)

x = flops_multi(A*b);

Honestly, I plead ignorance when it comes to this toolbox. I don't know how to use the commands properly.

[Alexey]

You probably already found the answer, but for the sake of others who might come across this, the answer is more like 2 - you need to add to your code a counting line for every operation you want counted. For example, if you want to do x=A*b:

flops(0)           %start global flop count at 0
A=[1 2 3; 4 5 6];  
b=[7 8 9]';
x=A*b;             %do the operation
addflops(flops_mul(A, b)) %do the counting
flops              %display count so far

you will get 10, which makes sense, there are 6 multiplications and 4 additions each costing 1 flop: [1*7+2*8+3*9; 4*7+5*8+6*9]. Note that the arguments to flops_mul should be the same as what you did. I say that for 2 reasons: 1) is obviously you want to count what was computed exactly, but 2) is that flops_mul doesn't seem to do any matrix multiplication checking - eg flops_mul([2 2; 2 2], [5 5 5 5 5]') happily returns 6 even though you can't multiply a 2x2 matrix with a 5x1 vector... Hope this helps

Answer 2

[wided hechkel]

please can you give me the matlab code for calculating the FLOPs of this algorithm:

zf=10 
for zj=0:0.01:20
    for xf=1:1:64
        for xi=0:0.01:10
            for xr=1:1:128  
            ZT=sign(zj-zf)*sqrt((zj-zf)^2+(xf-xi)^2)-abs(zj-zf);
            ZD=sqrt(zf^2+(xr-xi)^2)-zf;
            somme=sum(zj-ZD-2*ZT);
            end
        end
    end
end

i want to know the method of utilizing the lightspeed matlab toolbox, please can you give me the code not only the results of FLOPs number

[Walter Roberson]

The operation sqrt((zj-zf)^2+(xf-xi)^2) can be implemented in some architectures as hypot(zj-sf, xf-xi) where hypot is a built-in instruction. Such instructions can be more accurate than the naive computation done here for the case where the two values are between 0 and sqrt(realmin) as squaring such values would underflow to 0. How do you count FLOPS for a hybrid hardware instruction, one which might take fewer or more cycles than the naive equivalent (due to accuracy issues)? How do you count FLOPS for sqrt(), which is neither a multiply nor an addition?