I'm trying to reduce the computational time of my MatLab code. For this I use the MatLab profiler.
In some functions I was in the same situation. I loop over a grid (a 2D matrix) and I test if each cell has (or not) some values (two each time). And the condition takes a long time. Most of the time spends in those functions.
Here is an example from my code:
time Calls line
16 function tGrid=some_compute(grid, tGrid, trans)
0.005 125 17 global depth width tWidth
18
< 0.001 125 19 for i=1:depth
0.022 125000 20 for j=1:width
2.270 37500000 21 if (grid(i, j)~=0 && grid(i, j)~=1)
22
23 % Get the transformed coordinates of the current cell
0.864 79997 24 [x, y]=get_trans(trans, i, j);
25
0.006 79997 26 if (x~=-1 && y~=-1)
0.006 79992 27 tGrid(x, y)=grid(i, j);
0.004 79992 28 end % If x~=-1 && y~=-1
0.004 79997 29 end % If grid~=0 && grid~=1
1.933 37500000 30 end % For j=width
0.008 125000 31 end % For i=depth
32
< 0.001 125 33 end
This function takes a total of 5.24s for 125 calls.
As we can see, it's the line 21 who take the more time to execute. While the the grid(i, j) is accessed twice in the condition, and is also used in line 27, I guessed by storing it in a variable before, it will speed-up my code.
So, I did this simple modification:
time Calls line
16 function tGrid=some_compute(grid, tGrid, trans)
0.005 125 17 global depth width tWidth
18
< 0.001 125 19 for i=1:depth
0.021 125000 20 for j=1:width
2.062 37500000 21 val=grid(i, j);
1.681 37500000 22 if (val~=0 && val~=1)
23
24 % Get the transformed coordinates of the current cell
0.793 79997 25 [x, y]=get_trans(trans, i, j);
26
0.005 79997 27 if (x~=-1 && y~=-1)
0.005 79992 28 tGrid(x, y)=val;
0.004 79992 29 end % If x~=-1 && y~=-1
0.004 79997 30 end % If grid~=0 && grid~=1
1.731 37500000 31 end % For j=width
0.007 125000 32 end % For i=depth
33
< 0.001 125 34 end
And now, this function takes a total of 6.32s for 125 calls.
As we can see, getting the value of the cell take as much time as getting it twice and compare it to two values. And the if, line 22, is as much slow as in the first version. I don't understand why.
getting the value of the cell take as much time as getting it twice and compare it to two values. [...] I don't understand why.
This is because your code is run under a JIT (just-in-time compiler). This compiler optimizes your code. Though it's not documented how exactly it works and what kind of optimizations are implemented, from your experiment it seems clear that grid(i, j)~=0 && grid(i, j)~=1 is optimized to only fetch the value grid(i,j) once.