10 Minutes to cuDF and CuPy

This notebook provides introductory examples of how you can use cuDF and CuPy together to take advantage of CuPy array functionality (such as advanced linear algebra operations).

[1]:
import timeit

import cupy as cp
import cudf

Converting a cuDF DataFrame to a CuPy Array

If we want to convert a cuDF DataFrame to a CuPy ndarray, There are multiple ways to do it:

  1. We can use the dlpack interface.

  2. We can also use DataFrame.values.

  3. We can also convert via the CUDA array interface by using cuDF’s as_gpu_matrix and CuPy’s asarray functionality.

[2]:
nelem = 10000
df = cudf.DataFrame({'a':range(nelem),
                     'b':range(500, nelem + 500),
                     'c':range(1000, nelem + 1000)}
                   )

%timeit arr_cupy = cp.fromDlpack(df.to_dlpack())
%timeit arr_cupy = df.values
%timeit arr_cupy = cp.asarray(df.as_gpu_matrix())
44.1 µs ± 689 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
209 µs ± 2.77 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
208 µs ± 3.14 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
[3]:
arr_cupy = cp.fromDlpack(df.to_dlpack())
arr_cupy
[3]:
array([[    0,   500,  1000],
       [    1,   501,  1001],
       [    2,   502,  1002],
       ...,
       [ 9997, 10497, 10997],
       [ 9998, 10498, 10998],
       [ 9999, 10499, 10999]])

Converting a cuDF Series to a CuPy Array

There are also multiple ways to convert a cuDF Series to a CuPy array:

  1. We can pass the Series to cupy.asarray as cuDF Series exposes `__cuda_array_interface__ <https://docs-cupy.chainer.org/en/stable/reference/interoperability.html>`__.

  2. We can leverage the dlpack interface to_dlpack().

  3. We can also use Series.values

[4]:
col = 'a'

%timeit cola_cupy = cp.asarray(df[col])
%timeit cola_cupy = cp.fromDlpack(df[col].to_dlpack())
%timeit cola_cupy = df[col].values
22.1 µs ± 518 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
58.3 µs ± 647 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
80.2 µs ± 647 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
[5]:
cola_cupy = cp.asarray(df[col])
cola_cupy
[5]:
array([   0,    1,    2, ..., 9997, 9998, 9999])

From here, we can proceed with normal CuPy workflows, such as reshaping the array, getting the diagonal, or calculating the norm.

[6]:
reshaped_arr = cola_cupy.reshape(50, 200)
reshaped_arr
[6]:
array([[   0,    1,    2, ...,  197,  198,  199],
       [ 200,  201,  202, ...,  397,  398,  399],
       [ 400,  401,  402, ...,  597,  598,  599],
       ...,
       [9400, 9401, 9402, ..., 9597, 9598, 9599],
       [9600, 9601, 9602, ..., 9797, 9798, 9799],
       [9800, 9801, 9802, ..., 9997, 9998, 9999]])
[7]:
reshaped_arr.diagonal()
[7]:
array([   0,  201,  402,  603,  804, 1005, 1206, 1407, 1608, 1809, 2010,
       2211, 2412, 2613, 2814, 3015, 3216, 3417, 3618, 3819, 4020, 4221,
       4422, 4623, 4824, 5025, 5226, 5427, 5628, 5829, 6030, 6231, 6432,
       6633, 6834, 7035, 7236, 7437, 7638, 7839, 8040, 8241, 8442, 8643,
       8844, 9045, 9246, 9447, 9648, 9849])
[8]:
cp.linalg.norm(reshaped_arr)
[8]:
array(577306.967739)

Converting a CuPy Array to a cuDF DataFrame

We can also convert a CuPy ndarray to a cuDF DataFrame. Like before, there are multiple ways to do it:

  1. Easiest; We can directly use the DataFrame constructor.

  2. We can use CUDA array interface with the DataFrame constructor.

  3. We can also use the dlpack interface.

For the latter two cases, we’ll need to make sure that our CuPy array is Fortran contiguous in memory (if it’s not already). We can either transpose the array or simply coerce it to be Fortran contiguous beforehand.

[9]:
%timeit reshaped_df = cudf.DataFrame(reshaped_arr)
13.1 ms ± 193 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
[10]:
reshaped_df = cudf.DataFrame(reshaped_arr)
reshaped_df.head()
[10]:
0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
0 0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
1 200 201 202 203 204 205 206 207 208 209 ... 390 391 392 393 394 395 396 397 398 399
2 400 401 402 403 404 405 406 407 408 409 ... 590 591 592 593 594 595 596 597 598 599
3 600 601 602 603 604 605 606 607 608 609 ... 790 791 792 793 794 795 796 797 798 799
4 800 801 802 803 804 805 806 807 808 809 ... 990 991 992 993 994 995 996 997 998 999

5 rows × 200 columns

We can check whether our array is Fortran contiguous by using cupy.isfortran or looking at the flags of the array.

[11]:
cp.isfortran(reshaped_arr)
[11]:
False

In this case, we’ll need to convert it before going to a cuDF DataFrame. In the next two cells, we create the DataFrame by leveraging dlpack and the CUDA array interface, respectively.

[12]:
%%timeit

fortran_arr = cp.asfortranarray(reshaped_arr)
reshaped_df = cudf.DataFrame(fortran_arr)
4.9 ms ± 26.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
[13]:
%%timeit

fortran_arr = cp.asfortranarray(reshaped_arr)
reshaped_df = cudf.from_dlpack(fortran_arr.toDlpack())
5.1 ms ± 23.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
[14]:
fortran_arr = cp.asfortranarray(reshaped_arr)
reshaped_df = cudf.DataFrame(fortran_arr)
reshaped_df.head()
[14]:
0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
0 0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
1 200 201 202 203 204 205 206 207 208 209 ... 390 391 392 393 394 395 396 397 398 399
2 400 401 402 403 404 405 406 407 408 409 ... 590 591 592 593 594 595 596 597 598 599
3 600 601 602 603 604 605 606 607 608 609 ... 790 791 792 793 794 795 796 797 798 799
4 800 801 802 803 804 805 806 807 808 809 ... 990 991 992 993 994 995 996 997 998 999

5 rows × 200 columns

Converting a CuPy Array to a cuDF Series

To convert an array to a Series, we can directly pass the array to the Series constructor.

[15]:
cudf.Series(reshaped_arr.diagonal()).head()
[15]:
0      0
1    201
2    402
3    603
4    804
dtype: int64

Interweaving CuDF and CuPy for Smooth PyData Workflows

RAPIDS libraries and the entire GPU PyData ecosystem are developing quickly, but sometimes a one library may not have the functionality you need. One example of this might be taking the row-wise sum (or mean) of a Pandas DataFrame. cuDF’s support for row-wise operations isn’t mature, so you’d need to either transpose the DataFrame or write a UDF and explicitly calculate the sum across each row. Transposing could lead to hundreds of thousands of columns (which cuDF wouldn’t perform well with) depending on your data’s shape, and writing a UDF can be time intensive.

By leveraging the interoperability of the GPU PyData ecosystem, this operation becomes very easy. Let’s take the row-wise sum of our previously reshaped cuDF DataFrame.

[16]:
reshaped_df.head()
[16]:
0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
0 0 1 2 3 4 5 6 7 8 9 ... 190 191 192 193 194 195 196 197 198 199
1 200 201 202 203 204 205 206 207 208 209 ... 390 391 392 393 394 395 396 397 398 399
2 400 401 402 403 404 405 406 407 408 409 ... 590 591 592 593 594 595 596 597 598 599
3 600 601 602 603 604 605 606 607 608 609 ... 790 791 792 793 794 795 796 797 798 799
4 800 801 802 803 804 805 806 807 808 809 ... 990 991 992 993 994 995 996 997 998 999

5 rows × 200 columns

We can just transform it into a CuPy array and use the axis argument of sum.

[17]:
new_arr = cp.fromDlpack(reshaped_df.to_dlpack())
new_arr.sum(axis=1)
[17]:
array([  19900,   59900,   99900,  139900,  179900,  219900,  259900,
        299900,  339900,  379900,  419900,  459900,  499900,  539900,
        579900,  619900,  659900,  699900,  739900,  779900,  819900,
        859900,  899900,  939900,  979900, 1019900, 1059900, 1099900,
       1139900, 1179900, 1219900, 1259900, 1299900, 1339900, 1379900,
       1419900, 1459900, 1499900, 1539900, 1579900, 1619900, 1659900,
       1699900, 1739900, 1779900, 1819900, 1859900, 1899900, 1939900,
       1979900])

With just that single line, we’re able to seamlessly move between data structures in this ecosystem, giving us enormous flexibility without sacrificing speed.

Converting a cuDF DataFrame to a CuPy Sparse Matrix

We can also convert a DataFrame or Series to a CuPy sparse matrix. We might want to do this if downstream processes expect CuPy sparse matrices as an input.

The sparse matrix data structure is defined by three dense arrays. We’ll define a small helper function for cleanliness.

[18]:
def cudf_to_cupy_sparse_matrix(data, sparseformat='column'):
    """Converts a cuDF object to a CuPy Sparse Column matrix.
    """
    if sparseformat not in ('row', 'column',):
        raise ValueError("Let's focus on column and row formats for now.")

    _sparse_constructor = cp.sparse.csc_matrix
    if sparseformat == 'row':
        _sparse_constructor = cp.sparse.csr_matrix

    return _sparse_constructor(cp.fromDlpack(data.to_dlpack()))

We can define a sparsely populated DataFrame to illustrate this conversion to either sparse matrix format.

[19]:
df = cudf.DataFrame()
nelem = 10000
nonzero = 1000
for i in range(20):
    arr = cp.random.normal(5, 5, nelem)
    arr[cp.random.choice(arr.shape[0], nelem-nonzero, replace=False)] = 0
    df['a' + str(i)] = arr
[20]:
df.head()
[20]:
a0 a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 a13 a14 a15 a16 a17 a18 a19
0 0.0 0.0 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.000000 0.00000 0.0 16.822959 0.0 0.000000 0.0 0.0 0.0 0.000000
1 0.0 0.0 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.000000 0.00000 0.0 0.000000 0.0 0.000000 0.0 0.0 0.0 0.000000
2 0.0 0.0 6.618972 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.000000 2.25678 0.0 0.000000 0.0 0.000000 0.0 0.0 0.0 0.000000
3 0.0 0.0 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.000000 0.00000 0.0 0.000000 0.0 2.715802 0.0 0.0 0.0 0.000000
4 0.0 0.0 0.000000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 4.296568 0.00000 0.0 0.000000 0.0 0.000000 0.0 0.0 0.0 4.865495
[21]:
sparse_data = cudf_to_cupy_sparse_matrix(df)
sparse_data
[21]:
<cupyx.scipy.sparse.csc.csc_matrix at 0x7f25e49466a0>

From here, we could continue our workflow with a CuPy sparse matrix.

For a full list of the functionality built into these libraries, we encourage you to check out the API docs for cuDF and CuPy.