cuda-python/cuda_bindings/tests/test_interoperability.py at 883b1a7ad7031f0bef98cdd985944165ff7ede91 · rwgk/cuda-python · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
# SPDX-FileCopyrightText: Copyright (c) 2021-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE

import numpy as np
import pytest

import cuda.bindings.driver as cuda
import cuda.bindings.runtime as cudart
from cuda.bindings._test_helpers.mempool import xfail_if_mempool_oom


def supportsMemoryPool():
    err, isSupported = cudart.cudaDeviceGetAttribute(cudart.cudaDeviceAttr.cudaDevAttrMemoryPoolsSupported, 0)
    return err == cudart.cudaError_t.cudaSuccess and isSupported


def test_interop_stream():
    # DRV to RT
    err_dr, stream = cuda.cuStreamCreate(0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaStreamDestroy(stream)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, stream = cudart.cudaStreamCreate()
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuStreamDestroy(stream)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS


def test_interop_event():
    # DRV to RT
    err_dr, event = cuda.cuEventCreate(0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaEventDestroy(event)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, event = cudart.cudaEventCreate()
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuEventDestroy(event)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS


def test_interop_graph():
    # DRV to RT
    err_dr, graph = cuda.cuGraphCreate(0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaGraphDestroy(graph)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, graph = cudart.cudaGraphCreate(0)
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuGraphDestroy(graph)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS


def test_interop_graphNode():
    err_dr, graph = cuda.cuGraphCreate(0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS

    # DRV to RT
    err_dr, node = cuda.cuGraphAddEmptyNode(graph, [], 0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaGraphDestroyNode(node)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, node = cudart.cudaGraphAddEmptyNode(graph, [], 0)
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuGraphDestroyNode(node)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS

    (err_rt,) = cudart.cudaGraphDestroy(graph)
    assert err_rt == cudart.cudaError_t.cudaSuccess


# cudaUserObject_t
# TODO


# cudaFunction_t
# TODO


@pytest.mark.skipif(not supportsMemoryPool(), reason="Requires mempool operations")
def test_interop_memPool():
    # DRV to RT
    err_dr, pool = cuda.cuDeviceGetDefaultMemPool(0)
    xfail_if_mempool_oom(err_dr, "cuDeviceGetDefaultMemPool", 0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaDeviceSetMemPool(0, pool)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, pool = cudart.cudaDeviceGetDefaultMemPool(0)
    xfail_if_mempool_oom(err_rt, "cudaDeviceGetDefaultMemPool", 0)
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuDeviceSetMemPool(0, pool)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS


def test_interop_graphExec():
    err_dr, graph = cuda.cuGraphCreate(0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    err_dr, node = cuda.cuGraphAddEmptyNode(graph, [], 0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS

    # DRV to RT
    err_dr, graphExec = cuda.cuGraphInstantiate(graph, 0)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS
    (err_rt,) = cudart.cudaGraphExecDestroy(graphExec)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # RT to DRV
    err_rt, graphExec = cudart.cudaGraphInstantiate(graph, 0)
    assert err_rt == cudart.cudaError_t.cudaSuccess
    (err_dr,) = cuda.cuGraphExecDestroy(graphExec)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS

    (err_rt,) = cudart.cudaGraphDestroy(graph)
    assert err_rt == cudart.cudaError_t.cudaSuccess


def test_interop_deviceptr():
    # Allocate dev memory
    size = 1024 * np.uint8().itemsize
    err_dr, dptr = cuda.cuMemAlloc(size)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS

    # Allocate host memory
    h1 = np.full(size, 1).astype(np.uint8)
    h2 = np.full(size, 2).astype(np.uint8)
    assert np.array_equal(h1, h2) is False

    # Initialize device memory
    (err_rt,) = cudart.cudaMemset(dptr, 1, size)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # D to h2
    (err_rt,) = cudart.cudaMemcpy(h2, dptr, size, cudart.cudaMemcpyKind.cudaMemcpyDeviceToHost)
    assert err_rt == cudart.cudaError_t.cudaSuccess

    # Validate h1 == h2
    assert np.array_equal(h1, h2)

    # Cleanup
    (err_dr,) = cuda.cuMemFree(dptr)
    assert err_dr == cuda.CUresult.CUDA_SUCCESS