hodu_cuda_kernels 0.2.4

hodu cuda kernels
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

use crate::{
    compat::*,
    cuda::*,
    error::{CudaKernelError, Result},
    kernel::Kernels,
    kernels::macros::ops,
    source::Source,
};

ops!(
    conv1d,
    conv2d,
    conv3d,
    conv_transpose1d,
    conv_transpose2d,
    conv_transpose3d,
    conv1d_grad_weight,
    conv2d_grad_weight,
    conv3d_grad_weight,
    conv_transpose1d_grad_weight,
    conv_transpose2d_grad_weight,
    conv_transpose3d_grad_weight
);

/// Execute a convolution operation (1D, 2D, 3D, or transposed variants)
/// Note: For production use, consider using cuDNN for optimal performance
pub fn call_ops_conv<T>(
    kernel: crate::kernels::macros::Kernel,
    kernels: &Kernels,
    context: &Arc<CudaContext>,
    input: &CudaSlice<T>,
    weight: &CudaSlice<T>,
    output: &mut CudaSlice<T>,
    metadata: &[usize],
) -> Result<()>
where
    T: cudarc::driver::DeviceRepr,
{
    let func = kernels.load_function(context, Source::OpsConv, kernel.0)?;

    let num_els = metadata[0];
    let block_size = 256u32;
    let grid_size = (num_els as u32).div_ceil(block_size).max(1);

    let cfg = LaunchConfig {
        grid_dim: (grid_size, 1, 1),
        block_dim: (block_size, 1, 1),
        shared_mem_bytes: 0,
    };

    let stream = context.default_stream();
    let metadata_dev = stream
        .memcpy_stod(metadata)
        .map_err(|e| CudaKernelError::MemoryError(format!("Failed to copy metadata: {:?}", e)))?;

    unsafe {
        func.launch(&stream, cfg, |args| {
            args.arg(input).arg(weight).arg(output).arg(&metadata_dev);
        })
        .map_err(|e| CudaKernelError::LaunchError(format!("Failed to launch kernel: {:?}", e)))?;
    }

    Ok(())
}

/// Execute a convolution weight gradient operation for backpropagation
pub fn call_ops_conv_grad_weight<T>(
    kernel: crate::kernels::macros::Kernel,
    kernels: &Kernels,
    context: &Arc<CudaContext>,
    input: &CudaSlice<T>,
    grad_output: &CudaSlice<T>,
    grad_weight: &mut CudaSlice<T>,
    metadata: &[usize],
) -> Result<()>
where
    T: cudarc::driver::DeviceRepr,
{
    let func = kernels.load_function(context, Source::OpsConv, kernel.0)?;

    let num_els = metadata[0];
    let block_size = 256u32;
    let grid_size = (num_els as u32).div_ceil(block_size).max(1);

    let cfg = LaunchConfig {
        grid_dim: (grid_size, 1, 1),
        block_dim: (block_size, 1, 1),
        shared_mem_bytes: 0,
    };

    let stream = context.default_stream();
    let metadata_dev = stream
        .memcpy_stod(metadata)
        .map_err(|e| CudaKernelError::MemoryError(format!("Failed to copy metadata: {:?}", e)))?;

    unsafe {
        func.launch(&stream, cfg, |args| {
            args.arg(input).arg(grad_output).arg(grad_weight).arg(&metadata_dev);
        })
        .map_err(|e| CudaKernelError::LaunchError(format!("Failed to launch kernel: {:?}", e)))?;
    }

    Ok(())
}