5dc3e70446c09312e2bd6dbe263b9af6f8405e62,scikits/cuda/special.py,,exp1,#Any#,145

Before Change

    // Get block/grid sizes; the number of threads per block is limited
    // to 256 because the kernel defined above uses too many
    // registers to be invoked more threads per block:
    dev = misc.get_current_device()
    max_threads_per_block = 256
    block_dim, grid_dim = \
               misc.select_block_grid_sizes(dev, z_gpu.shape, max_threads_per_block)

    // Set this to False when debugging to make sure the compiled kernel is
    // not cached:
    cache_dir=None
    expi_mod = \
             SourceModule(expi_template.substitute(use_double=use_double),
                          cache_dir=cache_dir,
                          options=["-I", install_headers])
    exp1_func = expi_mod.get_function("exp1_array")

    e_gpu = gpuarray.empty_like(z_gpu)
    exp1_func(z_gpu, e_gpu,
              np.uint32(z_gpu.size),
              block=block_dim,
              grid=grid_dim)
    return e_gpu

After Change

        raise ValueError("unsupported type")
    op = "e[i] = exp1(z[i])"
    
    try:
        func = exp1.cache[z_gpu.dtype]
    except KeyError:
        func = elementwise.ElementwiseKernel(args, op,
                                 options=["-I", install_headers],
                                 preamble="//include "cuSpecialFuncs.h"")
        exp1.cache[z_gpu.dtype] = func

    e_gpu = gpuarray.empty_like(z_gpu)
    func(z_gpu, e_gpu)

    return e_gpu
exp1.cache = {}

Instances