Issue with Using Image Interpolation about pyopencl HOT 10 CLOSED

Which implementation are you using? Have you tried others?

from pyopencl.

I tried on Apple with Iris Pro and Nvidia and on Windows with HD Graphics. I get the same results so I assume that It is no drivers or hardware.

from pyopencl.

These seem relevant but they don't provide answers:

• http://stackoverflow.com/questions/20017990/opencl-image3d-linear-sampling
• http://stackoverflow.com/questions/20027292/precision-when-reading-image-with-clk-filter-linear-in-opencl

from pyopencl.

Have you tried a CPU implemnetation? How is the accuracy there? Alternatively, post a full working script, and i'd be happy to try on a few machines.

from pyopencl.

http://nbviewer.ipython.org/gist/cancan101/654c070ef7b4bf34aac1

I changed the offset so it goes 0,1,2 rather than 10,11,12 but the issue is the same,

from pyopencl.

Not a perfect test of CPU, but:

In [57]:

a = (0.5+0.1*np.arange(10, dtype=np.float32))-0.5
In [52]:

a[1] * test[1]
Out[52]:
1.0000002

from pyopencl.

@inducer Did you have a chance to test this out?

from pyopencl.

Appears to work as documented for me, with a few fixes to your code.

https://gist.github.com/inducer/c83d58aa1ea3640b16ee

from pyopencl.

Interesting. When I run the same script, I get the following output in the last cell:

[ 0.          0.1         0.2         0.30000001  0.40000001  0.5
  0.60000002  0.69999999  0.80000001  0.90000004]
[ 0.          0.1015625   0.19921875  0.30078125  0.3984375   0.5
  0.6015625   0.69921875  0.80078125  0.8984375 ]

from pyopencl.

As far as I know, image sampling does not come with any particular accuracy guarantees. ("Visual accuracy"?) In CUDA, it appears you get about two digits, which is about what you're seeing: https://stackoverflow.com/questions/14460514/accuracy-of-1d-linear-interpolation-by-cuda-texture-interpolation?rq=1

from pyopencl.