Optimize manipulation.join.concatenate for len(tup)>3

[LWisteria]

clpy/clpy/core/core.pyx

Lines 2639 to 2674 in 3fd3b23

	# TODO(yoshiki.imaizumi): keep clpy implementation to optimize
	# when len(tup) > 3
	# if len(tup) > 3:
	# all_same_type = True
	# all_one_and_contiguous = True
	# dtype = tup[0].dtype
	# for a in tup:
	# all_same_type = all_same_type and (a.dtype == dtype)
	# all_one_and_contiguous = (
	# all_one_and_contiguous and a._c_contiguous and
	# a._shape[axis] == 1)
	#
	# if all_same_type:
	# ptrs = numpy.ndarray(len(tup), numpy.int64)
	# for i, a in enumerate(tup):
	# ptrs[i] = a.data.ptr
	# x = array(ptrs)
	#
	# raise NotImplementedError("clpy does not support this")
	# if all_one_and_contiguous:
	# base = <int>internal.prod_ssize_t(shape[axis + 1:])
	# _concatenate_kernel_one(x, base, ret)
	# else:
	# ndim = tup[0].ndim
	# x_strides = numpy.ndarray((len(tup), ndim), numpy.int32)
	# cum_sizes = numpy.ndarray(len(tup), numpy.int32)
	# cum = 0
	# for i, a in enumerate(tup):
	# for j in range(ndim):
	# x_strides[i, j] = <int>a._strides[j]
	# cum_sizes[i] = cum
	# cum += <int>a._shape[axis]
	#
	# _concatenate_kernel(
	# x, axis, array(cum_sizes), array(x_strides), ret)
	# return ret

They are commented out in current ClPy. However, should be implement same as CuPy's due to performance problem.