Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
groupnorm_bwd.py 4.12 KB
Newer Older
yanjl1's avatar
Initial  
yanjl1 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
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129 
import hipdnn
import torch


def build_groupnorm_bwd_graph(
    hipdnn_handle,
    torch_tensor_x,
    torch_tensor_dy,
    torch_tensor_scale,
    torch_tensor_epsilon,
    torch_tensor_mean,
    torch_tensor_inv_var,
    groups,
    hipdnn_data_type,
):
    # Create graph
    graph = hipdnn.pygraph(
        handle=hipdnn_handle,
        io_data_type=hipdnn_data_type,
        intermediate_data_type=hipdnn.data_type.FLOAT,
        compute_data_type=hipdnn.data_type.FLOAT,
        name="groupnorm_bwd",
    )

    # Create hipdnn tensors
    hipdnn_tensor_x = graph.tensor_like(torch_tensor_x)
    hipdnn_tensor_dy = graph.tensor_like(torch_tensor_dy)
    hipdnn_tensor_scale = graph.tensor_like(torch_tensor_scale)
    hipdnn_tensor_epsilon = graph.tensor_like(torch_tensor_epsilon)
    hipdnn_tensor_mean = graph.tensor_like(torch_tensor_mean)
    hipdnn_tensor_inv_var = graph.tensor_like(torch_tensor_inv_var)

    # Create groupnorm op
    hipdnn_tensor_dx, hipdnn_tensor_dbias, hipdnn_tensor_dscale = graph.groupnorm_backward(
        x=hipdnn_tensor_x,
        dy=hipdnn_tensor_dy,
        scale=hipdnn_tensor_scale,
        epsilon=hipdnn_tensor_epsilon,
        mean=hipdnn_tensor_mean,
        inv_variance=hipdnn_tensor_inv_var,
        groups=groups,
        name="groupnorm_backward",
    )
    hipdnn_tensor_dx.set_output(True)
    hipdnn_tensor_dbias.set_output(True)
    hipdnn_tensor_dscale.set_output(True)
    graph.build(hipdnn_handle)

    return (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_dy,
        hipdnn_tensor_scale,
        hipdnn_tensor_epsilon,
        hipdnn_tensor_mean,
        hipdnn_tensor_inv_var,
        hipdnn_tensor_dx,
        hipdnn_tensor_dbias,
        hipdnn_tensor_dscale,
    )


if __name__ == "__main__":
    # Input dimensions
    batch, channels, height, width = 2, 16, 512, 512

    hipdnn_data_type = hipdnn.data_type.FLOAT
    torch_data_type = torch.float32

    torch_tensor_x = torch.rand(
        batch, channels, height, width, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_dy = torch.rand(
        batch, channels, height, width, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_scale = torch.rand(channels, dtype=torch_data_type, device="cuda")
    torch_tensor_epsilon = torch.full(
        (1, 1, 1, 1), 1e-5, dtype=torch.float32, requires_grad=False, device="cpu"
    )
    groups = 2
    torch_tensor_mean = torch.rand(groups * batch, dtype=torch.float32, device="cuda")
    torch_tensor_inv_var = torch.rand(groups * batch, dtype=torch.float32, device="cuda")

    hipdnn_handle = hipdnn.create_handle()

    (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_dy,
        hipdnn_tensor_scale,
        hipdnn_tensor_epsilon,
        hipdnn_tensor_mean,
        hipdnn_tensor_inv_var,
        hipdnn_tensor_dx,
        hipdnn_tensor_dbias,
        hipdnn_tensor_dscale,
    ) = build_groupnorm_bwd_graph(
        hipdnn_handle,
        torch_tensor_x,
        torch_tensor_dy,
        torch_tensor_scale,
        torch_tensor_epsilon,
        torch_tensor_mean,
        torch_tensor_inv_var,
        groups,
        hipdnn_data_type,
    )

    torch_tensor_dx = torch.empty(hipdnn_tensor_dx.get_dim(), dtype=torch_data_type, device="cuda")
    torch_tensor_dscale = torch.empty(
        hipdnn_tensor_dscale.get_dim(), dtype=torch_data_type, device="cuda"
    )
    torch_tensor_dbias = torch.empty(
        hipdnn_tensor_dbias.get_dim(), dtype=torch_data_type, device="cuda"
    )
    variant_pack = {
        hipdnn_tensor_x: torch_tensor_x.data_ptr(),
        hipdnn_tensor_dy: torch_tensor_dy.data_ptr(),
        hipdnn_tensor_scale: torch_tensor_scale.data_ptr(),
        hipdnn_tensor_mean: torch_tensor_mean.data_ptr(),
        hipdnn_tensor_inv_var: torch_tensor_inv_var.data_ptr(),
        hipdnn_tensor_dx: torch_tensor_dx.data_ptr(),
        hipdnn_tensor_dbias: torch_tensor_dbias.data_ptr(),
        hipdnn_tensor_dscale: torch_tensor_dscale.data_ptr(),
    }
    workspace = torch.empty(graph.get_workspace_size(), dtype=torch.uint8, device="cuda")

    graph.exec(variant_pack=variant_pack, workspace=workspace.data_ptr())
    print("groupnorm bwd graph execution complete.")