cutile-python/test/test_copy.py at main · NVIDIA/cutile-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
151
152
# SPDX-FileCopyrightText: Copyright (c) <2025> NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0

from cuda.tile._bytecode.version import BytecodeVersion
import pytest
import torch

from math import ceil
import cuda.tile as ct
from util import assert_equal, make_test_tensor, require_hopper_or_newer, require_blackwell_or_newer
from conftest import (
    float_dtypes, bool_dtypes, int_dtypes, dtype_id, requires_tileiras, uint_dtypes
)
from torch.testing import make_tensor
from cuda.tile import PaddingMode
from typing import Optional

float8_dtypes = [
    pytest.param(torch.float8_e5m2, marks=require_hopper_or_newer()),
    pytest.param(torch.float8_e4m3fn, marks=require_hopper_or_newer()),
    pytest.param(torch.float8_e8m0fnu, marks=(require_blackwell_or_newer(),
                                              requires_tileiras(BytecodeVersion.V_13_2))),
]

test_dtypes = (float_dtypes + bool_dtypes + int_dtypes + uint_dtypes +
               [torch.float64] + float8_dtypes)


def assert_tensors_contiguity(tensors, predicate):
    is_contiguous = all(t.is_contiguous() for t in tensors)
    assert is_contiguous if predicate else not is_contiguous


@ct.kernel
def array_copy_1d(x, y, TILE: ct.Constant[int]):
    bid = ct.bid(0)
    tx = ct.load(x, index=(bid,), shape=(TILE,))
    ct.store(y, index=(bid,), tile=tx)


@pytest.mark.parametrize("shape", [(128,), (225,)])
@pytest.mark.parametrize("tile", [64, 128])
@pytest.mark.parametrize("stride_step", [1, 2])
@pytest.mark.parametrize("dtype", test_dtypes, ids=dtype_id)
def test_array_copy_1d(shape, stride_step, dtype, tile):
    x = make_test_tensor(shape, dtype=dtype, device='cuda')
    y = torch.zeros_like(x, device='cuda')
    xx = x[::stride_step]
    assert xx.stride() == (stride_step,)
    yy = y[::stride_step]
    assert yy.stride() == (stride_step,)
    assert_tensors_contiguity((xx, yy), stride_step == 1)
    grid = (ceil(len(xx) / tile), 1, 1)
    ct.launch(torch.cuda.current_stream(), grid, array_copy_1d, (xx, yy, tile))
    y_expected = torch.zeros_like(x)
    y_expected[::stride_step] = x[::stride_step]
    assert_equal(y, y_expected)


@ct.kernel
def array_copy_2d(x, y, TILE_X: ct.Constant[int], TILE_Y: ct.Constant[int]):
    bidx = ct.bid(0)
    bidy = ct.bid(1)
    tx = ct.load(x, index=(bidx, bidy), shape=(TILE_X, TILE_Y))
    ct.store(y, index=(bidx, bidy), tile=tx)


@pytest.mark.parametrize("shape", [(128, 128), (192, 192)])
@pytest.mark.parametrize("tile", [(64, 64), (128, 128)])
@pytest.mark.parametrize("permute", [(0, 1), (1, 0)])
@pytest.mark.parametrize("stride_step", [1, 2])
@pytest.mark.parametrize("dtype", test_dtypes, ids=dtype_id)
def test_array_copy_2d(shape, stride_step, permute, dtype, tile):
    x = make_test_tensor(shape, dtype=dtype, device='cuda')
    y = torch.zeros_like(x)
    xx = x[::stride_step, ::stride_step].permute(permute)
    yy = y[::stride_step, ::stride_step].permute(permute)
    assert_tensors_contiguity((xx, yy), stride_step == 1 and permute == (0, 1))
    grid = (*(ceil(i / j) for i, j in zip(xx.shape, tile)), 1)
    ct.launch(torch.cuda.current_stream(), grid, array_copy_2d, (xx, yy, tile[0], tile[1]))
    y_expected = torch.zeros_like(x)
    y_expected[::stride_step, ::stride_step] = x[::stride_step, ::stride_step]
    assert_equal(y, y_expected)


@ct.kernel
def array_copy_3d(x, y,
                  TILE_BATCH: ct.Constant[int],
                  TILE_X: ct.Constant[int],
                  TILE_Y: ct.Constant[int]):
    bidb = ct.bid(0)
    bidx = ct.bid(1)
    bidy = ct.bid(2)
    tx = ct.load(x, index=(bidb, bidx, bidy), shape=(TILE_BATCH, TILE_X, TILE_Y))
    ct.store(y, index=(bidb, bidx, bidy), tile=tx)


@pytest.mark.parametrize("shape", [(4, 128, 128), (2, 192, 192)])
@pytest.mark.parametrize("tile", [(2, 64, 64), (1, 128, 128)])
@pytest.mark.parametrize("permute", [(0, 1, 2), (2, 1, 0), (2, 0, 1)])
@pytest.mark.parametrize("stride_step", [1, 2])
@pytest.mark.parametrize("dtype", test_dtypes, ids=dtype_id)
def test_array_copy_3d(shape, stride_step, permute, dtype, tile):
    x = make_test_tensor(shape, dtype=dtype, device='cuda')
    y = torch.zeros_like(x, device='cuda')
    xx = x[:, :, ::stride_step].permute(permute)
    yy = y[:, :, ::stride_step].permute(permute)
    assert_tensors_contiguity((xx, yy), stride_step == 1 and permute == (0, 1, 2))
    permuted_tile = tile[permute[0]], tile[permute[1]], tile[permute[2]]
    grid = tuple(ceil(i / j) for i, j in zip(xx.shape, permuted_tile))
    ct.launch(torch.cuda.current_stream(), grid, array_copy_3d,
              (xx, yy, permuted_tile[0], permuted_tile[1], permuted_tile[2]))
    y_expected = torch.zeros_like(x)
    y_expected[:, :, ::stride_step] = x[:, :, ::stride_step]
    assert_equal(y, y_expected)


def make_array_copy_2d_with_padding_kernel(padding_mode: Optional[PaddingMode]):

    @ct.kernel
    def kernel(x, y, TILE_X: ct.Constant[int], TILE_Y: ct.Constant[int]):
        bidx = ct.bid(0)
        bidy = ct.bid(1)
        tx = ct.load(x, index=(bidx, bidy), shape=(TILE_X, TILE_Y), padding_mode=padding_mode)
        ct.store(y, index=(bidx, bidy), tile=tx)
    return kernel


@pytest.mark.parametrize("padding_value, float_padding_value", [
    (PaddingMode.UNDETERMINED, None),
    (PaddingMode.ZERO, 0.0),
    (PaddingMode.NEG_ZERO, -0.0),
    (PaddingMode.NAN, float('nan')),
    (PaddingMode.POS_INF, float('inf')),
    (PaddingMode.NEG_INF, float('-inf'))
])
def test_array_copy_2d_with_padding(padding_value, float_padding_value):
    shape = (63, 63)
    tile = (64, 64)
    x = make_tensor(shape, dtype=torch.float32, device='cuda')
    y = make_tensor(tile, dtype=torch.float32, device='cuda')
    grid = (ceil(shape[0] / tile[0]), ceil(shape[1] / tile[1]), 1)
    ct.launch(torch.cuda.current_stream(), grid,
              make_array_copy_2d_with_padding_kernel(padding_value),
              (x, y, tile[0], tile[1]))
    if float_padding_value is None:
        assert_equal(y[:shape[0], :shape[1]], x)
    else:
        y_expected = torch.ones(tile, dtype=torch.float32, device='cuda') * float_padding_value
        y_expected[:shape[0], :shape[1]] = x
        assert_equal(y, y_expected)