Add fast paths for DataFrame.to_cupy

python/cudf/benchmarks/API/bench_dataframe.py

@@ -351,7 +351,9 @@ def bench_where(benchmark, dataframe, cond, other):
     benchmark(dataframe.where, cond, other)
 
 
-@benchmark_with_object(cls="dataframe", dtype="float", nulls=False, cols=20)
+@benchmark_with_object(
+    cls="dataframe", dtype="float", nulls=False, cols=20, rows=20
+)
 @pytest.mark.pandas_incompatible
 def bench_to_cupy(benchmark, dataframe):
     benchmark(dataframe.to_cupy)

python/cudf/benchmarks/API/bench_series.py

@@ -23,13 +23,17 @@ def bench_series_nsmallest(benchmark, series, n):
     benchmark(series.nsmallest, n)
 
 
-@benchmark_with_object(cls="series", dtype="int")
+@benchmark_with_object(cls="series", dtype="int", nulls=False)
 def bench_series_cp_asarray(benchmark, series):
-    series = series.dropna()
     benchmark(cupy.asarray, series)
 
 
-@benchmark_with_object(cls="series", dtype="int")
+@benchmark_with_object(cls="series", dtype="int", nulls=False)
+@pytest.mark.pandas_incompatible
+def bench_to_cupy(benchmark, series):
+    benchmark(lambda: series.values)
+
+
+@benchmark_with_object(cls="series", dtype="int", nulls=False)
 def bench_series_values(benchmark, series):
-    series = series.dropna()
     benchmark(lambda: series.values)

python/cudf/benchmarks/common/config.py

@@ -66,5 +66,5 @@ def pytest_sessionfinish(session, exitstatus):
     NUM_ROWS = [10, 20]
     NUM_COLS = [1, 6, 20]
 else:
-    NUM_ROWS = [100, 10_000, 1_000_000]
-    NUM_COLS = [1, 6, 20]
+    NUM_ROWS = [100, 1_000, 10_000, 50_000, 1_000_000]
+    NUM_COLS = [1, 6, 20, 1_000, 10_000, 50_000]

python/cudf/benchmarks/conftest.py

@@ -45,7 +45,6 @@
 """
 
 import os
-import string
 import sys
 
 import pytest_cases
@@ -83,14 +82,8 @@ def axis(request):
 for dtype, column_generator in column_generators.items():
 
     def make_dataframe(nr, nc, column_generator=column_generator):
-        assert nc <= len(string.ascii_lowercase), (
-            "make_dataframe only supports a maximum of 26 columns"
-        )
         return cudf.DataFrame(
-            {
-                f"{string.ascii_lowercase[i]}": column_generator(nr)
-                for i in range(nc)
-            }
+            {f"col{i}": column_generator(nr) for i in range(nc)}
         )
 
     for nr in NUM_ROWS:

python/cudf/cudf/core/frame.py

@@ -8,7 +8,6 @@
 from typing import TYPE_CHECKING, Any, Literal
 
 import cupy
-import cupy as cp
 import numpy
 import numpy as np
 import pyarrow as pa
@@ -570,14 +569,14 @@ def to_cupy(
             )
         ):
             if dtype is None:
-                dtype = np.dtype(self._columns[0].dtype)
+                dtype = self._columns[0].dtype
 
             shape = (len(self), self._num_columns)
             out = cupy.empty(shape, dtype=dtype, order="F")
 
-            table = self.to_pylibcudf()[0]
-            if isinstance(table, plc.Column):
-                table = plc.Table([table])
+            table = plc.Table(
+                [col.to_pylibcudf(mode="read") for col in self._columns]
+            )
             plc.reshape.table_to_array(
                 table,
                 out.data.ptr,
@@ -591,7 +590,7 @@ def to_cupy(
             if (
                 not copy
                 and col.dtype.kind in {"i", "u", "f", "b"}
-                and cp.can_cast(col.dtype, final_dtype)
+                and cupy.can_cast(col.dtype, final_dtype)
             ):
                 if col.has_nulls():
                     if na_value is not None:
@@ -604,7 +603,7 @@ def to_cupy(
                             dtype,
                             na_value,
                         )
-                return cp.asarray(col, dtype=final_dtype).reshape((-1, 1))
+                return cupy.asarray(col, dtype=final_dtype).reshape((-1, 1))
         return self._to_array(
             lambda col: col.values,
             cupy,

python/cudf/cudf/core/single_column_frame.py

@@ -139,25 +139,11 @@ def to_cupy(
         -------
         cupy.ndarray
         """
-        col = self._column
-        final_dtype = (
-            col.dtype if dtype is None else dtype
-        )  # some types do not support | operator
-        if (
-            not copy
-            and col.dtype.kind in {"i", "u", "f", "b"}
-            and cp.can_cast(col.dtype, final_dtype)
-        ):
-            if col.has_nulls():
-                if na_value is not None:
-                    col = col.fillna(na_value)
-                else:
-                    return super().to_cupy(
-                        dtype=dtype, copy=copy, na_value=na_value
-                    )
-            return cp.asarray(col, dtype=final_dtype)
-
-        return super().to_cupy(dtype=dtype, copy=copy, na_value=na_value)
+        return (
+            super()
+            .to_cupy(dtype=dtype, copy=copy, na_value=na_value)
+            .reshape(len(self), order="F")
+        )
 
     @property  # type: ignore
     @_performance_tracking

python/pylibcudf/pylibcudf/reshape.pxd

@@ -1,5 +1,6 @@
 # Copyright (c) 2024-2025, NVIDIA CORPORATION.
 
+from libc.stddef cimport size_t
 from libc.stdint cimport uintptr_t
 
 from pylibcudf.libcudf.types cimport size_type
@@ -18,6 +19,6 @@ cpdef Table tile(Table source_table, size_type count)
 cpdef void table_to_array(
     Table input_table,
     uintptr_t ptr,
-    size_type size,
+    size_t size,
     Stream stream=*
 )

python/pylibcudf/pylibcudf/reshape.pyx

@@ -1,5 +1,6 @@
 # Copyright (c) 2024-2025, NVIDIA CORPORATION.
 
+from libc.stddef cimport size_t
 from libc.stdint cimport uintptr_t
 from libcpp.memory cimport unique_ptr
 from libcpp.utility cimport move
@@ -81,11 +82,12 @@ cpdef Table tile(Table source_table, size_type count):
 cpdef void table_to_array(
     Table input_table,
     uintptr_t ptr,
-    size_type size,
+    size_t size,
     Stream stream=None
 ):
     """
     Copy a table into a preallocated column-major device array.
+
     Parameters
     ----------
     input_table : Table
@@ -98,7 +100,7 @@ cpdef void table_to_array(
     stream : Stream | None
         CUDA stream on which to perform the operation.
     """
-    if size > numeric_limits[size_type].max():
+    if size > numeric_limits[size_t].max():
         raise ValueError(
             "Size exceeds the int32_t limit."
         )

python/pylibcudf/pylibcudf/tests/test_reshape.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2024-2025, NVIDIA CORPORATION.
 
+import cupy as cp
 import pyarrow as pa
 import pytest
 from utils import assert_column_eq, assert_table_eq
@@ -8,16 +9,6 @@
 from pylibcudf.types import TypeId
 
 
-@pytest.fixture(scope="module")
-def np():
-    return pytest.importorskip("cupy")
-
-
-@pytest.fixture(scope="module")
-def cp():
-    return pytest.importorskip("cupy")
-
-
 @pytest.fixture(scope="module")
 def reshape_data():
     data = [[1, 2, 3], [4, 5, 6]]
@@ -59,8 +50,8 @@ def test_tile(reshape_data, cnt):
         ("float64", TypeId.FLOAT64),
     ],
 )
-def test_table_to_array(dtype, type_id, np, cp):
-    arrow_type = pa.from_numpy_dtype(getattr(np, dtype))
+def test_table_to_array(dtype, type_id):
+    arrow_type = pa.from_numpy_dtype(getattr(cp, dtype))
     arrs = [
         pa.array([1, 2, 3], type=arrow_type),
         pa.array([4, 5, 6], type=arrow_type),