numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
Documentation 
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

//! CUDA kernel launchers for 2:4 structured sparsity
//!
//! Kernel source: sparse_24.cu

use cudarc::driver::PushKernelArg;
use cudarc::driver::safe::{CudaContext, CudaStream};
use std::sync::Arc;

use crate::dtype::DType;
use crate::error::{Error, Result};
use crate::runtime::cuda::kernels::loader::{
    BLOCK_SIZE, elementwise_launch_config, get_kernel_function, get_or_load_module, kernel_name,
    launch_config,
};

const MODULE_NAME: &str = "sparse_24";

/// Launch prune-to-2:4 kernel.
///
/// # Safety
/// All pointers must be valid device memory of correct size.
pub unsafe fn launch_sparse_24_prune(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    dense_ptr: u64,
    compressed_ptr: u64,
    metadata_ptr: u64,
    m: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, MODULE_NAME)?;
    let func_name = kernel_name("sparse_24_prune", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let total_groups = (m * (k / 4)) as u32;
    let grid = elementwise_launch_config(total_groups as usize);
    let block = (BLOCK_SIZE, 1, 1);
    let cfg = launch_config(grid, block, 0);

    let m_u32 = m as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&dense_ptr);
        builder.arg(&compressed_ptr);
        builder.arg(&metadata_ptr);
        builder.arg(&m_u32);
        builder.arg(&k_u32);
        builder
            .launch(cfg)
            .map_err(|e| Error::Internal(format!("CUDA sparse_24_prune launch failed: {e:?}")))?;
    }

    Ok(())
}

/// Launch decompress-from-2:4 kernel.
///
/// # Safety
/// All pointers must be valid device memory of correct size.
pub unsafe fn launch_sparse_24_decompress(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    compressed_ptr: u64,
    metadata_ptr: u64,
    dense_ptr: u64,
    m: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, MODULE_NAME)?;
    let func_name = kernel_name("sparse_24_decompress", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let total_groups = (m * (k / 4)) as u32;
    let grid = elementwise_launch_config(total_groups as usize);
    let block = (BLOCK_SIZE, 1, 1);
    let cfg = launch_config(grid, block, 0);

    let m_u32 = m as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&compressed_ptr);
        builder.arg(&metadata_ptr);
        builder.arg(&dense_ptr);
        builder.arg(&m_u32);
        builder.arg(&k_u32);
        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!("CUDA sparse_24_decompress launch failed: {e:?}"))
        })?;
    }

    Ok(())
}

/// Launch 2:4 sparse matmul kernel: C = A @ B^T where B is 2:4 compressed.
///
/// # Safety
/// All pointers must be valid device memory of correct size.
pub unsafe fn launch_sparse_24_matmul(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    dtype: DType,
    a_ptr: u64,            // [N, K]
    b_compressed_ptr: u64, // [M, K/2]
    b_metadata_ptr: u64,   // [M, meta_cols]
    c_ptr: u64,            // [N, M]
    n: usize,
    m: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, MODULE_NAME)?;
    let func_name = kernel_name("sparse_24_matmul", dtype);
    let func = get_kernel_function(&module, &func_name)?;

    let tile_size = 16u32;
    let grid_x = (m as u32 + tile_size - 1) / tile_size;
    let grid_y = (n as u32 + tile_size - 1) / tile_size;
    let grid = (grid_x, grid_y, 1);
    let block = (tile_size, tile_size, 1);
    let cfg = launch_config(grid, block, 0);

    let n_u32 = n as u32;
    let m_u32 = m as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_compressed_ptr);
        builder.arg(&b_metadata_ptr);
        builder.arg(&c_ptr);
        builder.arg(&n_u32);
        builder.arg(&m_u32);
        builder.arg(&k_u32);
        builder
            .launch(cfg)
            .map_err(|e| Error::Internal(format!("CUDA sparse_24_matmul launch failed: {e:?}")))?;
    }

    Ok(())
}