luma-cuda 0.3.0

luma cuda implement
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
155
156
157
158
159
160
161
162
163
164
165

use luma_tensor::Layout;
use crate::builder_arg;
use crate::kernel;
use crate::{Cuda, CudaError, CudaResult};
use cudarc::driver::{CudaSlice, DeviceRepr, LaunchConfig, PushKernelArg};

pub(crate) fn indexing_kernel_name(op: &str, idx_suffix: &str, val_suffix: &str) -> String {
    format!("{}_{}_{}", op, idx_suffix, val_suffix)
}

pub(crate) fn launch_index_select<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    idx_suffix: &str,
    val_suffix: &str,
    module: &kernel::Module,
    src: &CudaSlice<T>,
    src_l: &Layout,
    ids: &CudaSlice<I>,
    ids_l: &Layout,
    dim: usize,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = indexing_kernel_name("is", idx_suffix, val_suffix);
    let src_dims = src_l.dims();
    let num_dims = src_dims.len();

    let left_size: usize = src_dims[..dim].iter().product();
    let right_size: usize = src_dims[dim + 1..].iter().product();
    let src_dim_size = src_dims[dim];
    let ids_dim_size = ids_l.shape().element_count();
    let numel = ids_dim_size * left_size * right_size;

    let func = device.load_function(&kernel_name, module)?;

    let info: Vec<usize> = [src_dims, src_l.stride()].concat();
    let info_dev = device.memcpy_stod(&info)?;
    let out = device.alloc::<T>(numel)?;

    let mut builder = func.builder();
    builder_arg!(builder, numel);
    builder_arg!(builder, num_dims);
    builder.arg(&info_dev);
    builder.arg(ids);
    builder.arg(src);
    builder.arg(&out);
    builder_arg!(builder, left_size, src_dim_size, ids_dim_size, right_size);

    let config = LaunchConfig::for_num_elems(numel as u32);
    unsafe { builder.launch(config) }.map_err(CudaError::CudaDriver)?;
    Ok(out)
}

pub(crate) fn launch_gather<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    idx_suffix: &str,
    val_suffix: &str,
    module: &kernel::Module,
    src: &CudaSlice<T>,
    src_l: &Layout,
    ids: &CudaSlice<I>,
    ids_l: &Layout,
    dim: usize,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = indexing_kernel_name("gather", idx_suffix, val_suffix);
    let src_dims = src_l.dims();

    let left_size: usize = src_dims[..dim].iter().product();
    let right_size: usize = src_dims[dim + 1..].iter().product();
    let src_dim_size = src_dims[dim];
    let ids_dim_size = ids_l.dims()[dim];
    let numel = ids_l.shape().element_count();

    let func = device.load_function(&kernel_name, module)?;
    let out = device.alloc::<T>(numel)?;

    let mut builder = func.builder();
    builder_arg!(builder, numel);
    builder.arg(ids);
    builder.arg(src);
    builder.arg(&out);
    builder_arg!(builder, left_size, src_dim_size, ids_dim_size, right_size);

    let config = LaunchConfig::for_num_elems(numel as u32);
    unsafe { builder.launch(config) }.map_err(CudaError::CudaDriver)?;
    Ok(out)
}

pub(crate) fn launch_index_add<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    idx_suffix: &str,
    val_suffix: &str,
    module: &kernel::Module,
    init: &CudaSlice<T>,
    init_l: &Layout,
    ids: &CudaSlice<I>,
    ids_l: &Layout,
    src: &CudaSlice<T>,
    src_l: &Layout,
    dim: usize,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = indexing_kernel_name("ia", idx_suffix, val_suffix);
    let dst_dims = init_l.dims();
    let src_dims = src_l.dims();

    let left_size: usize = src_dims[..dim].iter().product();
    let right_size: usize = src_dims[dim + 1..].iter().product();
    let src_dim_size = src_dims[dim];
    let dst_dim_size = dst_dims[dim];
    let ids_dim_size = ids_l.dims()[0];
    let numel = left_size * right_size;

    let func = device.load_function(&kernel_name, module)?;

    let init_data = device.memcpy_dtov(init)?;
    let out = device.memcpy_stod(&init_data)?;

    let mut builder = func.builder();
    builder.arg(ids);
    builder_arg!(builder, ids_dim_size);
    builder.arg(src);
    builder.arg(&out);
    builder_arg!(builder, left_size, src_dim_size, dst_dim_size, right_size);

    let config = LaunchConfig::for_num_elems(numel as u32);
    unsafe { builder.launch(config) }.map_err(CudaError::CudaDriver)?;
    Ok(out)
}

pub(crate) fn launch_scatter_add<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    idx_suffix: &str,
    val_suffix: &str,
    module: &kernel::Module,
    init: &CudaSlice<T>,
    init_l: &Layout,
    ids: &CudaSlice<I>,
    _ids_l: &Layout,
    src: &CudaSlice<T>,
    src_l: &Layout,
    dim: usize,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = indexing_kernel_name("sa", idx_suffix, val_suffix);
    let dst_dims = init_l.dims();
    let src_dims = src_l.dims();

    let left_size: usize = src_dims[..dim].iter().product();
    let right_size: usize = src_dims[dim + 1..].iter().product();
    let src_dim_size = src_dims[dim];
    let dst_dim_size = dst_dims[dim];
    let numel = left_size * right_size;

    let func = device.load_function(&kernel_name, module)?;

    let init_data = device.memcpy_dtov(init)?;
    let out = device.memcpy_stod(&init_data)?;

    let mut builder = func.builder();
    builder.arg(ids);
    builder.arg(src);
    builder.arg(&out);
    builder_arg!(builder, left_size, src_dim_size, dst_dim_size, right_size);

    let config = LaunchConfig::for_num_elems(numel as u32);
    unsafe { builder.launch(config) }.map_err(CudaError::CudaDriver)?;
    Ok(out)
}