I've written this test code:
Where I've inserted the call to the opencl prg in a loop.
But after about 150 iterations I experiencing a dramatic loss
of performance, and the velocity became too slow.
What's the better way for calling an opencl program in a python for, like
the example bellow, without any loss of performance.
That's the output:
the two loop do exactly the same operation.
start ....
Prg : 0.256917
start b ....
Prg b: 1.663486
Tnx.
The code
----------------------------------------------------------------------
import pyopencl as cl
import pyopencl.array as cla
import numpy
import numpy.linalg as la
import time
lnn = 100000
szz = lnn*32
a = numpy.random.rand(szz,3).astype(numpy.float32)
b = numpy.random.rand(szz,3).astype(numpy.float32)
c = numpy.random.rand(szz,3).astype(numpy.float32)
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
queue2 = cl.CommandQueue(ctx)
mf = cl.mem_flags
a_array = cla.to_device( queue , a )
b_array = cla.to_device( queue , b )
dest_array = cla.Array( queue , (szz,3) , numpy.float32 )
dest_array_b = cla.Array( queue , (szz,3) , numpy.float32 )
prg_b = cl.Program(ctx, """
__kernel void sum_b(__global const float *a,
__global const float *b, __global float *c)
{
int i = get_global_id(0);
float m = sqrt( pown( a[3*i] , 2 ) + pown( a[3*i+1] , 2 ) + pown(
a[3*i+2] , 2 ) ) ;
c[3*i] = i*10.0f + m ;
c[3*i+1] = i*10.0f + 1 ;
c[3*i+2] = i*10.0f + 2 ;
}
""").build()
rep = 400
print("\nstart ....")
ta = time.time()
for fooo in range(rep):
prg_b.sum_b(queue, (szz,), None, a_array.data , b_array.data ,
dest_array.data )
tb = time.time()
print( "Prg : %f" % (tb - ta) )
#dest_array.get( queue , c )
#print dest_array
print("\nstart b ....")
taa = time.time()
for foo in range(rep):
prg_b.sum_b(queue, (szz,), None, a_array.data , b_array.data ,
dest_array_b.data )
tbb = time.time()
print( "Prg b: %f" % (tbb - taa) )
#dest_array_b.get( queue , c )
#print ( dest_array_b - dest_array )
_______________________________________________
PyOpenCL mailing list
[email protected]
http://lists.tiker.net/listinfo/pyopencl
