Jérôme Parent <jerome.par...@lynceetec.com> writes: > Hello, > > I am new to Opencl, I use to code using numpy and I am dealing with stack > of images. At the end I have 3D float array of the size (m,n,n). > I made my first customized kernel and it works like a charm and it's so > fast, I love it! > > Now I am converting other algorithm and many of them has operation over a > given axis. I also realized that I need to use scan. > > Now, I am stuck with cumsum along a given axis : data is a float array > (m,n,n). I would like to compute like in numpy : numpy.cumsum(data, axis=0) > to perform for every pixel of my 3D stack a cumsum along m direction > In the Pyopencl documentation : > https://documen.tician.de/pyopencl/a...edefined-scans > <https://documen.tician.de/pyopencl/algorithm.html#predefined-scans>, there > is an example : > > Code : > > knl = InclusiveScanKernel(context, np.int32, "a+b") > > n = 2**20-2**18+5 > host_data = np.random.randint(0, 10, n).astype(np.int32) > dev_data = cl_array.to_device(queue, host_data) > > knl(dev_data) > assert (dev_data.get() == np.cumsum(host_data, axis=0)).all() > > > This code works for 1D input data, but I do not know how ot perform scan > along a given direction of n dim array. > I hope someone could help me and teach the right "opencl" approach for such > operation
First of all, if there is enough parallelism available in the rest (i.e. the remaining axes) of your array, you'd be crazy to use parallel scan! "Embarrassing" parallelization of sequential for loops for scan will beat a parallel scan any day of the week. More generally, (i.e. if things aren't quite that simple, or if you can't get the data layout to play ball), the best solution I can offer is loopy [1], a code generator for computational code with multi-dimensional arrays. It is capable of emitting and transforming scans (grep the tests). (This will be better documented once the related paper is submitted, hopefully this summer.) [1] https://github.com/inducer/loopy HTH, Andreas _______________________________________________ PyOpenCL mailing list PyOpenCL@tiker.net https://lists.tiker.net/listinfo/pyopencl
