[FEA] Square Root via __array_ufunc__ protocol

[mrocklin]

As a special case of #256 it would be useful to have a sqrt computation accessible from cudf on the Python side. This would enable the computation of groupby-std (among other things).

Looking at instances of sqrt in the codebase today

MROCKLIN-MLT:cudf mrocklin$ git grep sqrt
cpp/include/cudf/functions.h:gdf_error gdf_sqrt_generic(gdf_column *input, gdf_column *output);
cpp/include/cudf/functions.h:gdf_error gdf_sqrt_f32(gdf_column *input, gdf_column *output);
cpp/include/cudf/functions.h:gdf_error gdf_sqrt_f64(gdf_column *input, gdf_column *output);
cpp/python/libgdf_cffi/tests/test_unaryops.py:def test_sqrt(dtype, ulp):
cpp/python/libgdf_cffi/tests/test_unaryops.py:    math_op_test(dtype, ulp, np.sqrt, libgdf.gdf_sqrt_generic,
cpp/src/unary/unary_ops.cu:        return std::sqrt(data);
cpp/src/unary/unary_ops.cu:DEF_UNARY_OP_REAL(gdf_sqrt)
cpp/src/unary/unary_ops.cu:gdf_error gdf_sqrt_f32(gdf_column *input, gdf_column *output) {
cpp/src/unary/unary_ops.cu:gdf_error gdf_sqrt_f64(gdf_column *input, gdf_column *output) {
python/cudf/bindings/cudf_cpp.pxd:    cdef gdf_error gdf_sqrt_generic(gdf_column *input, gdf_column *output)
python/cudf/bindings/cudf_cpp.pxd:    cdef gdf_error gdf_sqrt_f32(gdf_column *input, gdf_column *output)
python/cudf/bindings/cudf_cpp.pxd:    cdef gdf_error gdf_sqrt_f64(gdf_column *input, gdf_column *output)
python/cudf/dataframe/series.py:        return np.sqrt(self.var(ddof=ddof))

We see references to it in C++ side, a few references to bindings in the Cython side, but no use of it on the Python side as far as I can tell. Oddly we do see np.sqrt used in the Series.std method, which I suspect converts to a numpy array?

[harrism]

This appears to be true of all unary ops. All the unary ops are implemented in C++ and have Cython cdefs. They are just not exposed via the Series interface in python.

[thomcom]

Is the syntax you require for this

import cudf as cd
gdf = cd.DataFrame({'x': [1, 2, 3]})
cd.sqrt(gdf)

?

[mrocklin]

Currently we use np.sqrt(gdf) or np.sqrt(gdf.x)

We could drop that down to gdf ** 0.5 in dask-dataframe if desired.

Or, at some point, we should implement the __array_ufunc__ protocol so that np.sqrt(gdf) calls cudf.sqrt(gdf) appropriately. See #256

[thomcom]

For today, I can give you gd.sqrt

[mrocklin]

If it's ok, I'd like to leave this open until one of np.sqrt(s) or s ** 0.5 work. Those are what are necessary to get operations like groupby-std to work.

[thomcom]

Oh sorry, I thought cudf.sqrt would be adequate. I'll look at array_ufunc now.