Python Multiprocessing Data Output Wrong

I am trying Multiprocessing in Python. I have written some code which does vector add, but couldn't get the output out of the function. Which mean, the output Z prints out 0 rather

Solution 1:

The reason your last line print Z[0] returns [0] instead of [2] is that each of the processes makes an independent copy of Z (or may be Z[j] - not completely sure about that) before modifying it. Either way, a separate process run will guarantee that your original version will be unchanged.

If you were to use the threading module instead the last line would indeed return [2] as expected, but that is not multi-processing.

So, you probably want to use multiprocessing.Pool instead. Going along your experiment purely for illustration, one could do the following:

In [40]: pool = multiprocessing.Pool()
In [41]: def add_func(j):
   ....:     return X[j] + Y[j]
In [42]: pool = multiprocessing.Pool(numThreads)
In [43]: pool.map(add_func, range(numThreads))
Out[43]: 
[array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.]),
 array([ 2.])]

Have fun!

For your second part to your question, the problem is that the conv() function does not return any value. While the process pool gets a copy of X, B and W for pulling values from, the Y inside conv() is local to each process that is launched. To get the new computed value of Y, you would use something like this:

defconv(idx):
    Ylocal_section = X[idx*stride:idx*stride+stride].dot(W) +  B[idx*stride:idx*stride+stride]
    return Ylocal_section 

results = pool.map(conv, range(numThreads)) # then apply each result to Yfor idx inrange(numThreads):
    Y[idx*stride:idx*stride+stride] = results[idx] 

Parallelism can get complicated really fast, and at this point I would evaluate existing libraries that can perform fast 2D convolution. numpy and scipy libraries can be super efficient and might serve your needs better.