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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

//! Masked operations for CUDA runtime

use super::helpers::{BroadcastContext, validate_mask_dtype};
use crate::error::Result;
use crate::runtime::cuda::kernels::{
    launch_masked_count, launch_masked_count_broadcast, launch_masked_fill,
    launch_masked_fill_broadcast, launch_masked_prefix_sum, launch_masked_prefix_sum_broadcast,
    launch_masked_select, launch_masked_select_broadcast,
};
use crate::runtime::cuda::{CudaClient, CudaRuntime};
use crate::runtime::{Runtime, ensure_contiguous};
use crate::tensor::Tensor;

/// Execute masked_select operation.
pub fn masked_select(
    client: &CudaClient,
    a: &Tensor<CudaRuntime>,
    mask: &Tensor<CudaRuntime>,
) -> Result<Tensor<CudaRuntime>> {
    validate_mask_dtype(mask)?;

    let dtype = a.dtype();
    let a_contig = ensure_contiguous(a);
    let mask_contig = ensure_contiguous(mask);
    let numel = a.numel();

    let bcast = BroadcastContext::prepare(a, mask, &client.device)?;

    // Phase 1: Count true elements in mask
    let count_bytes = std::mem::size_of::<u32>();
    let count_ptr = CudaRuntime::allocate(count_bytes, &client.device)?;

    let zero: u32 = 0;
    CudaRuntime::copy_to_device(bytemuck::bytes_of(&zero), count_ptr, &client.device)?;

    if bcast.needs_broadcast {
        unsafe {
            launch_masked_count_broadcast(
                &client.context,
                &client.stream,
                client.device.index,
                mask_contig.ptr(),
                count_ptr,
                bcast.strides_ptr(),
                bcast.shape_ptr(),
                bcast.ndim,
                numel,
            )?;
        }
    } else {
        unsafe {
            launch_masked_count(
                &client.context,
                &client.stream,
                client.device.index,
                mask_contig.ptr(),
                count_ptr,
                numel,
            )?;
        }
    }

    let mut count_buf = [0u32; 1];
    CudaRuntime::copy_from_device(
        count_ptr,
        bytemuck::bytes_of_mut(&mut count_buf),
        &client.device,
    )?;
    let count = count_buf[0] as usize;
    CudaRuntime::deallocate(count_ptr, count_bytes, &client.device);

    let out = Tensor::<CudaRuntime>::empty(&[count], dtype, &client.device);
    if count == 0 {
        return Ok(out);
    }

    // Phase 2: Compute prefix sum
    let prefix_sum_bytes = numel * std::mem::size_of::<u32>();
    let prefix_sum_ptr = CudaRuntime::allocate(prefix_sum_bytes, &client.device)?;

    if bcast.needs_broadcast {
        unsafe {
            launch_masked_prefix_sum_broadcast(
                &client.context,
                &client.stream,
                client.device.index,
                mask_contig.ptr(),
                prefix_sum_ptr,
                bcast.strides_ptr(),
                bcast.shape_ptr(),
                bcast.ndim,
                numel,
            )?;
        }
    } else {
        unsafe {
            launch_masked_prefix_sum(
                &client.context,
                &client.stream,
                client.device.index,
                mask_contig.ptr(),
                prefix_sum_ptr,
                numel,
            )?;
        }
    }

    // Phase 3: Gather selected elements
    if bcast.needs_broadcast {
        unsafe {
            launch_masked_select_broadcast(
                &client.context,
                &client.stream,
                client.device.index,
                dtype,
                a_contig.ptr(),
                mask_contig.ptr(),
                out.ptr(),
                prefix_sum_ptr,
                bcast.strides_ptr(),
                bcast.shape_ptr(),
                bcast.ndim,
                numel,
            )?;
        }
    } else {
        unsafe {
            launch_masked_select(
                &client.context,
                &client.stream,
                client.device.index,
                dtype,
                a_contig.ptr(),
                mask_contig.ptr(),
                out.ptr(),
                prefix_sum_ptr,
                numel,
            )?;
        }
    }

    CudaRuntime::deallocate(prefix_sum_ptr, prefix_sum_bytes, &client.device);
    Ok(out)
}

/// Execute masked_fill operation.
pub fn masked_fill(
    client: &CudaClient,
    a: &Tensor<CudaRuntime>,
    mask: &Tensor<CudaRuntime>,
    value: f64,
) -> Result<Tensor<CudaRuntime>> {
    validate_mask_dtype(mask)?;

    let dtype = a.dtype();
    let a_contig = ensure_contiguous(a);
    let mask_contig = ensure_contiguous(mask);
    let numel = a.numel();

    let bcast = BroadcastContext::prepare(a, mask, &client.device)?;
    let out = Tensor::<CudaRuntime>::empty(a.shape(), dtype, &client.device);

    if bcast.needs_broadcast {
        unsafe {
            launch_masked_fill_broadcast(
                &client.context,
                &client.stream,
                client.device.index,
                dtype,
                a_contig.ptr(),
                mask_contig.ptr(),
                out.ptr(),
                value,
                bcast.strides_ptr(),
                bcast.shape_ptr(),
                bcast.ndim,
                numel,
            )?;
        }
    } else {
        unsafe {
            launch_masked_fill(
                &client.context,
                &client.stream,
                client.device.index,
                dtype,
                a_contig.ptr(),
                mask_contig.ptr(),
                out.ptr(),
                value,
                numel,
            )?;
        }
    }

    Ok(out)
}