def c_code(self, node, name, inp, out, sub):
        ndim = len(inp)
        zz = out[0]
        fail = sub["fail"]

        code = [
            """
int i;
size_t shape[%(ndim)s];
"""
            % dict(ndim=ndim)
        ]

        for i, shp_i in enumerate(inp):
            code.append(
                """
shape[%(i)s] = ((dtype_%(shp_i)s *)PyArray_DATA(%(shp_i)s))[0];
"""
                % dict(i=i, shp_i=shp_i)
            )

        code.append(
            """
if (theano_prep_output(&%(zz)s, %(ndim)s, shape, %(type)s, GA_C_ORDER,
                       %(ctx)s)) {
  %(fail)s
}
"""
            % dict(zz=zz, ndim=ndim, type=gpuarray.dtype_to_typecode(self.dtype), fail=fail, ctx=sub["params"])
        )

        return "".join(code)

 def __init__(self, dtype, broadcastable, name=None):
     # In case this was not provided and no global value is available
     self.dtype = str(dtype)
     self.broadcastable = tuple(bool(b) for b in broadcastable)
     self.ndim = len(self.broadcastable)
     self.name = name
     try:
         self.typecode = gpuarray.dtype_to_typecode(self.dtype)
     except gpuarray.GpuArrayException:
         raise TypeError("Unsupported dtype for %s: %s" %
                         (self.__class__.__name__, self.dtype))

 def __init__(self, dtype, broadcastable, context_name=None, name=None):
     # In case this was not provided and no global value is available
     self.dtype = str(dtype)
     self.broadcastable = tuple(bool(b) for b in broadcastable)
     self.ndim = len(self.broadcastable)
     self.name = name
     self.context_name = context_name
     # This will check that the passed context name is valid and registered.
     get_context(self.context_name)
     try:
         self.typecode = gpuarray.dtype_to_typecode(self.dtype)
     except gpuarray.GpuArrayException:
         raise TypeError("Unsupported dtype for %s: %s" % (self.__class__.__name__, self.dtype))

 def get_params(self, node):
     from pygpu.gpuarray import dtype_to_typecode
     return self.params_type.get_params(typecode=dtype_to_typecode(self.dtype),
                                        context=get_context(self.context_name))

 def m(t):
     if t == gpuarray.GpuArray:
         return "GA_BUFFER"
     else:
         return str(gpuarray.dtype_to_typecode(t))

    def c_code(self, node, nodename, inps, outs, sub):
        context = node.inputs[0].type.context
        if context.kind != b'cuda':
            raise NotImplementedError(
                '%s: We only have CUDA '
                'implementation so far.' % self.__class__.__name__)
        x, k = inps
        inp_dtc = ga.dtype_to_typecode(node.inputs[0].dtype)
        if not self.return_indices:
            yv, = outs
        elif self.return_values:
            yv, yi = outs
        else:
            yi, = outs
        out_dtype_s = self.idx_dtype
        out_dtc = ga.dtype_to_typecode(out_dtype_s)
        fail = sub['fail']
        ctx = sub['params']
        k_dtype = node.inputs[1].type.dtype_specs()[1]
        # max threads per block
        MAX_TPB = context.maxlsize
        # max blocks per grid
        MAX_BPG = context.maxgsize0
        WARP_SIZE = 32

        ndim = node.inputs[0].ndim
        reordered_axes = list(range(ndim))
        axis = self.axis % ndim
        del(reordered_axes[axis])
        reordered_axes = [axis] + reordered_axes
        dims = ''.join('(void*)(dims+%d), ' % i for i in reordered_axes[1:])
        prep_output = ''
        if self.return_values:
            def_dvstrides = 'const ssize_t *dvstrides = PyGpuArray_STRIDES(%s)' % yv
            params_dv = '(void*)((char*)(%s->ga.data) + (%s->ga.offset)),\n' % (yv, yv)
            params_dv += ''.join('(void*)(dvstrides+%d), ' % i for i in reordered_axes)
            prep_output += '''
    if (0 != theano_prep_output(
        &%(yv)s, %(ndim)d, odims,
        %(inp_dtc)s, GA_C_ORDER, %(ctx)s)) {
        %(fail)s;
    }\n''' % locals()
        else:
            def_dvstrides = params_dv = ''

        if self.return_indices:
            def_distrides = 'const ssize_t *distrides = PyGpuArray_STRIDES(%s)' % yi
            params_di = '(void*)((char*)(%s->ga.data) + (%s->ga.offset)),\n' % yi
            params_di += ''.join('(void*)(distrides+%d), ' % i for i in reordered_axes)
            prep_output += '''
    if (0 != theano_prep_output(
        &%(yi)s, %(ndim)d, odims,
        %(out_dtc)s, GA_C_ORDER, %(ctx)s)) {
        %(fail)s;
    }\n''' % locals()
        else:
            def_distrides = params_di = ''
        sstrides = ', '.join('(void*)(sstrides+%d)' % i for i in reordered_axes)
        code = '''
{
    const ssize_t k_ = ((%(k_dtype)s*)(PyArray_DATA(%(k)s)))[0];
    const size_t *dims = PyGpuArray_DIMS(%(x)s);
    size_t odims[%(ndim)d];
    for (int i=0; i<%(ndim)d; i++)
        odims[i] = dims[i];

    odims[%(axis)d] = k_>=0 ? k_ : -k_;

    if (0 == odims[%(axis)d]) {
        PyErr_SetString(
            PyExc_ValueError,
            "topk: kth must not be zero");
        %(fail)s;
    } else if (dims[%(axis)d] < odims[%(axis)d]) {
        PyErr_SetString(
            PyExc_ValueError,
            "topk: kth cannot be larger than the size of specified axis %(axis)d");
        %(fail)s;
    }
    %(prep_output)s

    size_t grid_size=1, block_size=1;
    for (int i=0; i<%(ndim)d; ++i) {
        if (i!=%(axis)d)
            grid_size *= dims[i];
        else
            block_size = dims[i];
    }
    // round up to multiples of warp size
    block_size = ((block_size + %(WARP_SIZE)d - 1) / %(WARP_SIZE)d) * %(WARP_SIZE)d;

    if (grid_size > %(MAX_BPG)d) {
        PyErr_SetString(
            PyExc_ValueError,
            "topk: too many slices to work with, expected <= %(MAX_BPG)d");
        %(fail)s;
    }

    %(def_dvstrides)s;
    %(def_distrides)s;
#.........这里部分代码省略.........

 def get_op_params(self):
     return [('TYPECODE', str(dtype_to_typecode(self.dtype)))]

    def get_op_params(self):
        from pygpu.gpuarray import dtype_to_typecode

        return [('TYPECODE', str(dtype_to_typecode(self.dtype)))]

    def c_code(self, node, name, inp, out, sub):
        if not any(getattr(self, 'redux', [node.inputs[0].ndim != 0])):
            # We special case the no-reduction case since the gpu
            # kernel has trouble handling it.
            return """
        Py_XDECREF(%(out)s);
        %(out)s = pygpu_copy(%(inp)s, GA_ANY_ORDER);
        if (!%(out)s) {
            %(fail)s
        }

        if (%(sync)d)
            GpuArray_sync(&%(out)s->ga);
""" % dict(out=out[0], inp=inp[0], fail=sub['fail'],
           sync=bool(config.gpuarray.sync))
        k = self.get_kernel_cache(node)
        _, src, _, ls = k._get_basic_kernel(k.init_local_size,
                                           node.inputs[0].ndim)
        if self.axis is None:
            redux = [True] * node.inputs[0].ndim
        else:
            redux = self.redux
        acc_dtype = getattr(self, 'acc_dtype', None)
        if acc_dtype is None:
            acc_dtype = node.outputs[0].type.dtype
        input = inp[0]
        output = out[0]
        nd_out = node.outputs[0].ndim
        code = """
        size_t gs = 1;
        unsigned int n = 1;
        unsigned int proxy_dim[%(nd_in)s];
        unsigned int proxy_off;
        int proxy_str[%(nd_in)s];
        void *args[%(n_args)s];
        PyGpuArrayObject *tmp;
        int err;
""" % dict(n_args=4 + (node.inputs[0].ndim * 2), nd_in=node.inputs[0].ndim)

        if nd_out != 0:
            code += """
        size_t out_dims[%(nd_out)s];
        int need_out = %(output)s == NULL || %(output)s->ga.nd != %(nd_out)s;
""" % dict(nd_out=nd_out, output=output)
            j = 0
            for i in range(node.inputs[0].ndim):
                if not self.redux[i]:
                    code += """
         out_dims[%(j)s] = %(input)s->ga.dimensions[%(i)s];
         if (!need_out)
             need_out |= %(output)s->ga.dimensions[%(j)s] != out_dims[%(j)s];
""" % dict(j=j, i=i, input=input, output=output)
                    j += 1
            code += """
         if (need_out) {
             %(output)s = pygpu_empty(%(nd_out)s, out_dims, %(out_type)s, GA_C_ORDER, pygpu_default_context(), Py_None);
             if (!%(output)s) {
                 %(fail)s
             }
         }
""" % dict(output=output, nd_out=nd_out, fail=sub['fail'],
           out_type=dtype_to_typecode(node.outputs[0].type.dtype))
        else:
            code += """
        if (%(output)s == NULL || %(output)s->ga.nd != 0) {
            Py_XDECREF(%(output)s);
            %(output)s = pygpu_empty(0, NULL, %(out_type)s, GA_C_ORDER,
                                     pygpu_default_context(), Py_None);
            if (!%(output)s) {
                %(fail)s
            }
        }
""" % dict(output=output, fail=sub['fail'],
           out_type=dtype_to_typecode(node.outputs[0].type.dtype))

        if acc_dtype != node.outputs[0].type.dtype:
            code += """
        tmp = pygpu_empty(%(output)s->ga.nd, %(output)s->ga.dimensions,
                          %(acc_type)s, GA_C_ORDER, pygpu_default_context(),
                          Py_None);
        if (!tmp) %(fail)s
""" % dict(output=output, fail=sub['fail'], acc_type=dtype_to_typecode(acc_dtype))
        else:
            code += """
        tmp = %(output)s;
        Py_INCREF(tmp);
""" % dict(output=output)

        # We need the proxies since we are passing a pointer to the
        # data into the call and therefore we need a real copy of the
        # data in the proper type.
        code += """
        args[0] = &n;
        args[1] = &tmp->ga;
""" % dict(output=output)

        p = 2
        for i in range(node.inputs[0].ndim):
            code += """
        proxy_dim[%(i)s] = %(input)s->ga.dimensions[%(i)s];
#.........这里部分代码省略.........