Add test for reverse slicing

[loikki]

Hi,

I am currently trying to use reverse slicing with a gpuarray (e.g. [end:start:-step]), but I am getting an error. I have written quickly a new test in order to reproduce the bug (feel free to discard my merge request if you are able to work on it).

>       copy.src_pitch = src_strides[1]
E       OverflowError: can't convert negative value to unsigned int

/usr/local/lib/python3.5/dist-packages/pycuda-2017.1.1-py3.5-linux-x86_64.egg/pycuda/gpuarray.py:1300: OverflowError

I have been through the code looking for the definition of src_pitch and I suppose that it comes from cudaMemcpy defined by Nvidia, right? Therefore, it would not be possible to change the unsigned int to a signed one.

[inducer]

a_gpu_slice = a_gpu[end:start:-2]
a_slice = a[start:end]

assert la.norm(a_gpu_slice.get()-a_slice) == 0

Those are some weird semantics. I'd call that a bug. I'd prefer this to be true:

a_gpu_slice = a_gpu[end:start:-2]
a_slice = a[end:start:-2]

assert la.norm(a_gpu_slice.get()-a_slice) == 0

[loikki]

Sorry, I was a little bit too fast in writing the test.

The error was not on the assert, but on the slicing a_gpu_slice = a_gpu[end:start:-2]

I have just fixed the test

[inducer]

That looks OK. I'd be happy to take a patch that makes this pass.

[loikki]

I suppose that we need to trick cuda in order avoid the negative value in an unsigned int.
What do you think about selecting in the normal direction and then reversing it?

[inducer]

I'd be OK with that, as long as it gives rise to correct behavior in all cases.

[loikki]

Hi,

I have been able to implement it.
To do it, I add to modify the reverse kernel and modify getitem in order to switch start and stop

[inducer]

CI at https://gitlab.tiker.net/inducer/pycuda/merge_requests/7

[inducer]

I actually don't see why a call to reverse should be necessary. (And I also view that as a negative since it adds a cost that a user might not expect.) I.e. why not copy the array in the "forward" direction and then adjust the strides? Could you explain?

[loikki]

I am not sure if it is possible. When I tried to use a negative step, the CUDA api was raising an error due to the use of unsigned int.

Anyway, I think that my work on the reverse function will be useful as now it is able to deal with discontinuous data

[inducer]

I am not sure if it is possible. When I tried to use a negative step, the CUDA api was raising an error due to the use of unsigned int.

I think it should be possible. Any data "footprint" that can be reached with negative strides can also be reached with "positive" strides. So, suppose we call that operation an "axis flip". Then you would pick out all the axes with negative strides, flip them, do the copy, and then flip them again. No additional code execution/second-stage data copying needed.