Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
pw_conv_bn.py 4.65 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
147
148
149
150
151
152
153
154 
import hipdnn
import torch


def build_pointwise_convolution_batchnorm_graph(
    hipdnn_handle,
    torch_tensor_x,
    torch_tensor_w,
    torch_tensor_scale,
    torch_tensor_bias,
    padding,
    stride,
    dilation,
    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="pointwise_convolution_batchnorm",
    )

    # Create hipdnn tensors
    hipdnn_tensor_x = graph.tensor_like(torch_tensor_x)
    hipdnn_tensor_w = graph.tensor_like(torch_tensor_w)
    hipdnn_tensor_scale = graph.tensor_like(torch_tensor_scale)
    hipdnn_tensor_bias = graph.tensor_like(torch_tensor_bias)

    # Create op
    hipdnn_tensor_mul_out = graph.mul(a=hipdnn_tensor_x, b=hipdnn_tensor_scale, name="mul")
    hipdnn_tensor_add_out = graph.add(a=hipdnn_tensor_mul_out, b=hipdnn_tensor_bias, name="add")
    hipdnn_tensor_relu_out = graph.relu(input=hipdnn_tensor_add_out, name="relu")
    hipdnn_tensor_conv_out = graph.conv_fprop(
        image=hipdnn_tensor_relu_out,
        weight=hipdnn_tensor_w,
        padding=padding,
        stride=stride,
        dilation=dilation,
        name="conv2d",
    )
    hipdnn_tensor_conv_out.set_output(True)

    [hipdnn_tensor_sum_out, hipdnn_tensor_sq_sum_out] = graph.genstats(
        hipdnn_tensor_conv_out, name="genstats"
    )
    hipdnn_tensor_sum_out.set_output(True)
    hipdnn_tensor_sq_sum_out.set_output(True)

    graph.build(hipdnn_handle)

    return (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_w,
        hipdnn_tensor_scale,
        hipdnn_tensor_bias,
        hipdnn_tensor_conv_out,
        hipdnn_tensor_sum_out,
        hipdnn_tensor_sq_sum_out,
    )


if __name__ == "__main__":
    # Input dimensions
    n = 4  # Batch size
    c = 64  # Number of input channels
    h = 16  # Height
    w = 16  # Width

    # Filter dimensions
    k = 32  # Number of output channels
    r = 3  # Filter height
    s = 3  # Filter width

    # Convolution parameters
    stride_h = 1  # Height stride
    stride_w = 1  # Width stride
    pad_h = 1  # Height padding
    pad_w = 1  # Width padding
    dil_h = 1  # Height dilation
    dil_w = 1  # Width dilation

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

    torch_tensor_x = torch.rand(n, c, h, w, dtype=torch_data_type, device="cuda").to(
        memory_format=torch.channels_last
    )
    torch_tensor_w = torch.rand(k, c, r, s, dtype=torch_data_type, device="cuda").to(
        memory_format=torch.channels_last
    )
    torch_tensor_scale = torch.rand(1, c, 1, 1, dtype=torch_data_type, device="cuda").to(
        memory_format=torch.channels_last
    )
    torch_tensor_bias = torch.rand(1, c, 1, 1, dtype=torch_data_type, device="cuda").to(
        memory_format=torch.channels_last
    )

    hipdnn_handle = hipdnn.create_handle()

    (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_w,
        hipdnn_tensor_scale,
        hipdnn_tensor_bias,
        hipdnn_tensor_conv_out,
        hipdnn_tensor_sum_out,
        hipdnn_tensor_sq_sum_out,
    ) = build_pointwise_convolution_batchnorm_graph(
        hipdnn_handle,
        torch_tensor_x,
        torch_tensor_w,
        torch_tensor_scale,
        torch_tensor_bias,
        [pad_h, pad_w],
        [stride_h, stride_w],
        [dil_h, dil_w],
        hipdnn_data_type,
    )

    torch_tensor_conv_out = torch.empty(
        hipdnn_tensor_conv_out.get_dim(),
        dtype=torch_data_type,
        memory_format=torch.channels_last,
        device="cuda",
    )
    torch_tensor_sum_out = torch.empty(
        hipdnn_tensor_sum_out.get_dim(),
        dtype=torch_data_type,
        memory_format=torch.channels_last,
        device="cuda",
    )
    torch_tensor_sq_sum_out = torch.empty(
        hipdnn_tensor_sq_sum_out.get_dim(),
        dtype=torch_data_type,
        memory_format=torch.channels_last,
        device="cuda",
    )
    variant_pack = {
        hipdnn_tensor_x: torch_tensor_x.data_ptr(),
        hipdnn_tensor_w: torch_tensor_w.data_ptr(),
        hipdnn_tensor_scale: torch_tensor_scale.data_ptr(),
        hipdnn_tensor_bias: torch_tensor_bias.data_ptr(),
        hipdnn_tensor_conv_out: torch_tensor_conv_out.data_ptr(),
        hipdnn_tensor_sum_out: torch_tensor_sum_out.data_ptr(),
        hipdnn_tensor_sq_sum_out: torch_tensor_sq_sum_out.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("pointwise_convolution_batchnorm graph execution complete.")