Hi,
The attached script shows highly inconsistent results (> 10% error at
times) between the numpy and gpuarray dot products. The inconsistent
results seem to only appear for large gpuarrays of data type complex64 or
complex128. Any ideas on what's going on? Thanks!
Jesse
import pycuda.autoinit
from pycuda import gpuarray
import numpy
shapes = [(10000,), (100000,), (1000000,), (10000000,), \
(100,100), (1000,1000), \
(10,20,30), (40,50,60), (200,200,200)]
dtypes = [numpy.float32, numpy.float64, numpy.complex64, numpy.complex128]
for dtype in dtypes:
for shape in shapes:
x = gpuarray.to_gpu(numpy.random.randn(*shape).astype(dtype))
y = gpuarray.to_gpu(numpy.random.randn(*shape).astype(dtype))
dot_cpu = numpy.dot(x.get().flatten(), y.get().flatten())
dot_gpu = gpuarray.dot(x, y).get()
print 'shape:', shape
print 'data type:', dtype
print 'numpy computed dot product:', dot_cpu
print 'gpuarray computed dot product:', dot_gpu
print 'percent error:', abs(dot_cpu-dot_gpu)/abs(dot_cpu)*100, '%'
print '\n'
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
