Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
deform_attention_bwd.py 5.58 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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146 
import hipdnn
import torch


def build_deform_attention_bwd_graph(
    hipdnn_handle,
    torch_tensor_value,
    torch_tensor_spatial_shapes,
    torch_tensor_level_start_index,
    torch_tensor_sampling_locations,
    torch_tensor_attention_weights,
    torch_tensor_grad_output,
    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="deform_attention_bwd",
    )

    # Create hipdnn tensors
    hipdnn_tensor_value = graph.tensor_like(torch_tensor_value)
    hipdnn_tensor_spatial_shapes = graph.tensor_like(torch_tensor_spatial_shapes)
    hipdnn_tensor_level_start_index = graph.tensor_like(torch_tensor_level_start_index)
    hipdnn_tensor_sampling_locations = graph.tensor_like(torch_tensor_sampling_locations)
    hipdnn_tensor_attention_weights = graph.tensor_like(torch_tensor_attention_weights)
    hipdnn_tensor_grad_output = graph.tensor_like(torch_tensor_grad_output)

    # Create deform attn op
    hipdnn_tensor_grad_value, hipdnn_tensor_grad_sampling_loc, hipdnn_tensor_grad_attn_weight = (
        graph.deform_attn_dgrad(
            value=hipdnn_tensor_value,
            spatial_shapes=hipdnn_tensor_spatial_shapes,
            level_start_index=hipdnn_tensor_level_start_index,
            sampling_locations=hipdnn_tensor_sampling_locations,
            attention_weights=hipdnn_tensor_attention_weights,
            grad_output=hipdnn_tensor_grad_output,
            name="deform_attn_dgrad",
        )
    )
    hipdnn_tensor_grad_value.set_output(True)
    hipdnn_tensor_grad_sampling_loc.set_output(True)
    hipdnn_tensor_grad_attn_weight.set_output(True)
    graph.build(hipdnn_handle)

    return (
        graph,
        hipdnn_tensor_value,
        hipdnn_tensor_spatial_shapes,
        hipdnn_tensor_level_start_index,
        hipdnn_tensor_sampling_locations,
        hipdnn_tensor_attention_weights,
        hipdnn_tensor_grad_output,
        hipdnn_tensor_grad_value,
        hipdnn_tensor_grad_sampling_loc,
        hipdnn_tensor_grad_attn_weight,
    )


if __name__ == "__main__":
    # Input dimensions
    n = 2  # batch size
    n_heads = 2
    embed_dims_per_head = 32
    embed_dims = n_heads * embed_dims_per_head
    n_levels = 2
    n_points = 2
    n_queries = 32

    spatial_shapes_cpu = torch.randint(low=1, high=16, size=(n_levels, 2), dtype=torch.int64)
    # calculate n_keys based on spatial_shapes_cpu
    n_keys = spatial_shapes_cpu.prod(dim=1).sum()
    # calculate level_start_index based on spatial_shapes_cpu
    count_per_level = spatial_shapes_cpu.prod(dim=1)
    level_start_index_cpu = torch.zeros_like(count_per_level)
    level_start_index_cpu[1:] = torch.cumsum(count_per_level[:-1], dim=0)

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

    torch_tensor_value = torch.rand(
        n, n_keys, n_heads, embed_dims_per_head, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_spatial_shapes = spatial_shapes_cpu.to("cuda")
    torch_tensor_level_start_index = level_start_index_cpu.to("cuda")
    torch_tensor_sampling_locations = torch.rand(
        n, n_queries, n_heads, n_levels, n_points, 2, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_attention_weights = torch.rand(
        n, n_queries, n_heads, n_levels, n_points, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_grad_output = torch.rand(
        n, n_queries, embed_dims, dtype=torch_data_type, device="cuda"
    )

    hipdnn_handle = hipdnn.create_handle()

    (
        graph,
        hipdnn_tensor_value,
        hipdnn_tensor_spatial_shapes,
        hipdnn_tensor_level_start_index,
        hipdnn_tensor_sampling_locations,
        hipdnn_tensor_attention_weights,
        hipdnn_tensor_grad_output,
        hipdnn_tensor_grad_value,
        hipdnn_tensor_grad_sampling_loc,
        hipdnn_tensor_grad_attn_weight,
    ) = build_deform_attention_bwd_graph(
        hipdnn_handle,
        torch_tensor_value,
        torch_tensor_spatial_shapes,
        torch_tensor_level_start_index,
        torch_tensor_sampling_locations,
        torch_tensor_attention_weights,
        torch_tensor_grad_output,
        hipdnn_data_type,
    )

    torch_tensor_grad_value = torch.empty(
        hipdnn_tensor_grad_value.get_dim(), dtype=torch_data_type, device="cuda"
    )
    torch_tensor_grad_sampling_loc = torch.empty(
        hipdnn_tensor_grad_sampling_loc.get_dim(), dtype=torch_data_type, device="cuda"
    )
    torch_tensor_grad_attn_weight = torch.empty(
        hipdnn_tensor_grad_attn_weight.get_dim(), dtype=torch_data_type, device="cuda"
    )
    variant_pack = {
        hipdnn_tensor_value: torch_tensor_value.data_ptr(),
        hipdnn_tensor_spatial_shapes: torch_tensor_spatial_shapes.data_ptr(),
        hipdnn_tensor_level_start_index: torch_tensor_level_start_index.data_ptr(),
        hipdnn_tensor_sampling_locations: torch_tensor_sampling_locations.data_ptr(),
        hipdnn_tensor_attention_weights: torch_tensor_attention_weights.data_ptr(),
        hipdnn_tensor_grad_output: torch_tensor_grad_output.data_ptr(),
        hipdnn_tensor_grad_value: torch_tensor_grad_value.data_ptr(),
        hipdnn_tensor_grad_sampling_loc: torch_tensor_grad_sampling_loc.data_ptr(),
        hipdnn_tensor_grad_attn_weight: torch_tensor_grad_attn_weight.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("Deform attention bwd graph execution complete.")