Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_array.py 4.28 KB
Newer Older
root's avatar
clone code from github  
root committed
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 
import warnings

import numpy
import pytest

import cupy
from cupy import testing
import cupyx.distributed._array


@pytest.fixture
def mem_pool():
    try:
        old_pool = cupy.get_default_memory_pool()
        pool = cupy.cuda.memory.MemoryPool()
        cupy.cuda.memory.set_allocator(pool.malloc)
        yield pool
    finally:
        pool.set_limit(size=0)
        pool.free_all_blocks()
        cupy.cuda.memory.set_allocator(old_pool.malloc)


@testing.multi_gpu(2)
class TestDistributedArray:
    def test_array_creation_from_numpy(self, mem_pool):
        array = numpy.arange(64, dtype='q').reshape(8, 8)
        assert mem_pool.used_bytes() == 0
        da = cupyx.distributed._array.array(array, (0, 1), (4, 8))
        assert da.device.id == -1
        # Ensure no memory allocation other than the chunks
        assert da.data.ptr == 0
        assert da.shape == (8, 8)
        assert mem_pool.used_bytes() == array.nbytes
        testing.assert_array_equal(da._chunks[0], array[:4, :])
        testing.assert_array_equal(da._chunks[1], array[4:, :])

    def test_array_creation_from_cupy(self, mem_pool):
        array = cupy.arange(64, dtype='q').reshape(8, 8)
        assert mem_pool.used_bytes() == array.nbytes
        with warnings.catch_warnings():
            warnings.simplefilter('ignore', cupy._util.PerformanceWarning)
            da = cupyx.distributed._array.array(array, (0, 1), (4, 8))
        assert da.device.id == -1
        # Ensure no memory allocation other than chunks & original array
        assert da.data.ptr == 0
        assert da.shape == (8, 8)
        assert mem_pool.used_bytes() == 2 * array.nbytes
        testing.assert_array_equal(da._chunks[0], array[:4, :])
        testing.assert_array_equal(da._chunks[1], array[4:, :])

    def test_invalid_array_creation(self):
        with pytest.raises(TypeError, match=r'numpy or cupy'):
            cupyx.distributed._array.array([1, 2, 3], (0, 1), (2, 8))

    def test_array_creation_invalid_splits(self):
        array = cupy.arange(64).reshape(8, 8)
        with pytest.raises(RuntimeError, match=r'single axis'):
            cupyx.distributed._array.array(array, (0, 1), (4, 4))

    def test_array_creation_more_chunks_than_devices(self):
        array = cupy.arange(64).reshape(8, 8)
        with pytest.raises(RuntimeError, match=r'amount of devices'):
            cupyx.distributed._array.array(array, (0, 1), (2, 8))

    @pytest.mark.parametrize('split_shape', [(8, 4), (4, 8)])
    def test_ufuncs(self, split_shape):
        # TODO parameterize split shape
        np_a = numpy.arange(64).reshape(8, 8)
        np_b = numpy.arange(64).reshape(8, 8) * 2
        np_r = numpy.sin(np_a * np_b)
        d_a = cupyx.distributed._array.array(np_a, (0, 1), split_shape)
        d_b = cupyx.distributed._array.array(np_b, (0, 1), split_shape)
        d_r = cupy.sin(d_a * d_b)
        testing.assert_array_almost_equal(d_r.asnumpy(), np_r)

    @pytest.mark.parametrize('split_shape', [(8, 4), (4, 8)])
    def test_elementwise_kernel(self, split_shape):
        custom_kernel = cupy.ElementwiseKernel(
            'float32 x, float32 y',
            'float32 z',
            'z = (x - y) * (x - y)',
            'custom')
        # TODO parameterize split shape
        np_a = numpy.arange(64).reshape(8, 8).astype(numpy.float32)
        np_b = (numpy.arange(64).reshape(8, 8) * 2.0).astype(numpy.float32)
        np_r = (np_a - np_b) * (np_a - np_b)
        d_a = cupyx.distributed._array.array(np_a, (0, 1), split_shape)
        d_b = cupyx.distributed._array.array(np_b, (0, 1), split_shape)
        d_r = custom_kernel(d_a, d_b)
        testing.assert_array_almost_equal(d_r.asnumpy(), np_r)

    def test_incompatible_chunk_shapes(self):
        np_a = numpy.arange(64).reshape(8, 8)
        np_b = numpy.arange(64).reshape(8, 8) * 2
        d_a = cupyx.distributed._array.array(np_a, (0, 1), (4, 8))
        d_b = cupyx.distributed._array.array(np_b, (0, 1), (8, 4))
        with pytest.raises(RuntimeError, match=r'chunk sizes'):
            cupy.sin(d_a * d_b)

    def test_incompatible_operand(self):
        np_a = numpy.arange(64).reshape(8, 8)
        cp_b = cupy.arange(64).reshape(8, 8)
        d_a = cupyx.distributed._array.array(np_a, (0, 1), (4, 8))
        with pytest.raises(RuntimeError, match=r'Mix `cupy.ndarray'):
            cupy.sin(d_a * cp_b)