PyOpenCL Matrix multiplication

Question

I have this code for matrix multiplication using pyopenCL. My problem is that the result is wrong in some matrices, and I dont understand why. After some research i think its related with global size of something like that but i dont understand how to set that values.

For example:

matrices using numpy dtype = float32

matrix 1:

[[ 0.99114645  0.09327769  0.90075564  0.8913309 ]
[ 0.59739089  0.13906649  0.94246316  0.65673178]
[ 0.24535166  0.68942326  0.41361505  0.5789603 ]
[ 0.31962237  0.17714553  0.49025267  0.21861202]]

matrix2:

[[ 0.41509482  0.82779616  0.74143827  0.37681136]
[ 0.88058949  0.01039944  0.4342753   0.45752665]
[ 0.60375261  0.21243185  0.88312167  0.97394323]
[ 0.60855824  0.69482827  0.61627114  0.57155776]]

expected result:

[[ 1.57981943  1.63210835  2.12016045  1.80288424]
[ 1.3391085   1.15248911  1.7403561   1.58199609]
[ 1.31099532  0.70041376  1.20338154  1.14162762]
[ 0.71769556  0.52246746  0.88158722  0.8039138 ]]

script result:

[[ 1.20828819  0.73175305  1.64546931  1.42526579]
[ 1.13179159  0.46403384  1.20692348  1.14317513]
[ 1.25328159  0.86723316  1.58679342  1.40186214]
[ 1.35214019  0.6795128   1.73811913  1.48048854]]

script:

def openCL_multiplication(matrix1, matrix2, res):

import pyopencl as cl
import numpy as np
import numpy.linalg as la

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)

mf = cl.mem_flags
a_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=matrix1)
b_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=matrix2)
dest_buf = cl.Buffer(ctx, mf.WRITE_ONLY, matrix1.nbytes )


prg = cl.Program(ctx, """
    __kernel void multiplymatrices(const unsigned int size, __global float * matrix1, __global float * matrix2, __global float * res) {

    int i = get_global_id(1); 
    int j = get_global_id(0);

    res[i + size * j] = 0;

    for (int k = 0; k < size; k++)
    {
        res[i + size * j] += matrix1[i + size * k] * matrix2[k + size * j];
    }

    }
    """).build()

t0 = datetime.datetime.now()

prg.multiplymatrices(queue, matrix1.shape, None,np.int32(len(matrix1)) ,a_buf, b_buf, dest_buf)

final_matrix = np.empty_like(matrix1)
cl.enqueue_copy(queue, final_matrix , dest_buf)

print  final_matrix


delta_t = datetime.datetime.now() - t0
print 'OpenCL Multiplication: ' + str(delta_t)

return final_matrix

Thank you!

Answer 1

Well, I think the kernel does all right. I can even call script result correct. It all depends on how you treat your matrices :-) If you want your expected result. I'd change this:

res[i + size * j] += matrix1[i + size * k] * matrix2[k + size * j];

to this:

res[i + size * j] += matrix1[k + size * i] * matrix2[j + size * k];

Hope this helps.