luma-cuda 0.3.2

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

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

pub(crate) fn launch_pick<T: DeviceRepr>(
    device: &Cuda,
    val_suffix: &str,
    module: &kernel::Module,
    mask: &CudaSlice<u8>,
    mask_l: &Layout,
    t: &CudaSlice<T>,
    true_l: &Layout,
    f: &CudaSlice<T>,
    false_l: &Layout,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = format!("pick_{}", val_suffix);
    let dims = mask_l.dims();
    let elem_count = mask_l.shape().element_count();
    let num_dims = dims.len();
    let func = device.load_function(&kernel_name, module)?;

    let mut builder = func.builder();

    let dims = device.memcpy_stod(dims)?;
    let mask_strides = device.memcpy_stod(mask_l.stride())?;
    let t_strides = device.memcpy_stod(true_l.stride())?;
    let f_strides = device.memcpy_stod(false_l.stride())?;

    let mask_view = mask.slice(mask_l.start_offset()..);
    let t_view = t.slice(true_l.start_offset()..);
    let f_view = f.slice(false_l.start_offset()..);
    let output = device.alloc::<T>(elem_count)?;

    builder.arg(&elem_count);
    builder.arg(&num_dims);
    builder.arg(&dims);
    builder.arg(&mask_strides);
    builder.arg(&t_strides);
    builder.arg(&f_strides);
    builder.arg(&mask_view);
    builder.arg(&t_view);
    builder.arg(&f_view);
    builder.arg(&output);

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

#[allow(unused_variables)]
pub(crate) fn launch_pick_true<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    val_suffix: &str,
    module: &kernel::Module,
    mask: &CudaSlice<I>,
    mask_l: &Layout,
    val: T,
    f: &CudaSlice<T>,
    false_l: &Layout,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = format!("pick_true_{}", val_suffix);
    let dims = mask_l.dims();
    let elem_count = mask_l.shape().element_count();
    let num_dims = dims.len();
    let func = device.load_function(&kernel_name, module)?;

    let mut builder = func.builder();

    let dims = device.memcpy_stod(dims)?;
    let mask_strides = device.memcpy_stod(mask_l.stride())?;
    let f_strides = device.memcpy_stod(false_l.stride())?;

    let mask_view = mask.slice(mask_l.start_offset()..);
    let f_view = f.slice(false_l.start_offset()..);
    let output = device.alloc::<T>(elem_count)?;

    builder.arg(&elem_count);
    builder.arg(&num_dims);
    builder.arg(&dims);
    builder.arg(&mask_strides);
    builder.arg(&f_strides);
    builder.arg(&mask_view);
    builder.arg(&val);
    builder.arg(&f_view);
    builder.arg(&output);

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

#[allow(unused_variables)]
pub(crate) fn launch_pick_false<T: DeviceRepr, I: DeviceRepr>(
    device: &Cuda,
    val_suffix: &str,
    module: &kernel::Module,
    mask: &CudaSlice<I>,
    mask_l: &Layout,
    t: &CudaSlice<T>,
    true_l: &Layout,
    val: T,
) -> CudaResult<CudaSlice<T>> {
    let kernel_name = format!("pick_false_{}", val_suffix);
    let dims = mask_l.dims();
    let elem_count = mask_l.shape().element_count();
    let num_dims = dims.len();
    let func = device.load_function(&kernel_name, module)?;

    let mut builder = func.builder();

    let dims = device.memcpy_stod(dims)?;
    let mask_strides = device.memcpy_stod(mask_l.stride())?;
    let t_strides = device.memcpy_stod(true_l.stride())?;

    let mask_view = mask.slice(mask_l.start_offset()..);
    let t_view = t.slice(true_l.start_offset()..);
    let output = device.alloc::<T>(elem_count)?;

    builder.arg(&elem_count);
    builder.arg(&num_dims);
    builder.arg(&dims);
    builder.arg(&mask_strides);
    builder.arg(&t_strides);
    builder.arg(&mask_view);
    builder.arg(&t_view);
    builder.arg(&val);
    builder.arg(&output);

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