numr 0.5.1

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

//! Fused elementwise operation helpers for CPU tensors

use super::super::kernels;
use super::super::{CpuClient, CpuRuntime};
use crate::dispatch_dtype;
use crate::error::{Error, Result};
use crate::runtime::ensure_contiguous;
use crate::tensor::Tensor;

/// Helper for fused_mul_add: out = a * b + c
pub fn fused_mul_add_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    c: &Tensor<CpuRuntime>,
) -> Result<Tensor<CpuRuntime>> {
    let dtype = a.dtype();
    if b.dtype() != dtype || c.dtype() != dtype {
        return Err(Error::DTypeMismatch {
            lhs: dtype,
            rhs: if b.dtype() != dtype {
                b.dtype()
            } else {
                c.dtype()
            },
        });
    }
    if a.shape() != b.shape() || a.shape() != c.shape() {
        return Err(Error::ShapeMismatch {
            expected: a.shape().to_vec(),
            got: if a.shape() != b.shape() {
                b.shape().to_vec()
            } else {
                c.shape().to_vec()
            },
        });
    }

    let a_contig = ensure_contiguous(a);
    let b_contig = ensure_contiguous(b);
    let c_contig = ensure_contiguous(c);
    let out = Tensor::<CpuRuntime>::empty(a.shape(), dtype, &client.device);

    let len = a.numel();
    let a_ptr = a_contig.ptr();
    let b_ptr = b_contig.ptr();
    let c_ptr = c_contig.ptr();
    let out_ptr = out.ptr();

    dispatch_dtype!(dtype, T => {
        unsafe {
            kernels::fused_mul_add_kernel::<T>(
                a_ptr as *const T,
                b_ptr as *const T,
                c_ptr as *const T,
                out_ptr as *mut T,
                len,
            );
        }
    }, "fused_mul_add");

    Ok(out)
}

/// Helper for fused_add_mul: out = (a + b) * c
pub fn fused_add_mul_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    c: &Tensor<CpuRuntime>,
) -> Result<Tensor<CpuRuntime>> {
    let dtype = a.dtype();
    if b.dtype() != dtype || c.dtype() != dtype {
        return Err(Error::DTypeMismatch {
            lhs: dtype,
            rhs: if b.dtype() != dtype {
                b.dtype()
            } else {
                c.dtype()
            },
        });
    }
    if a.shape() != b.shape() || a.shape() != c.shape() {
        return Err(Error::ShapeMismatch {
            expected: a.shape().to_vec(),
            got: if a.shape() != b.shape() {
                b.shape().to_vec()
            } else {
                c.shape().to_vec()
            },
        });
    }

    let a_contig = ensure_contiguous(a);
    let b_contig = ensure_contiguous(b);
    let c_contig = ensure_contiguous(c);
    let out = Tensor::<CpuRuntime>::empty(a.shape(), dtype, &client.device);

    let len = a.numel();
    let a_ptr = a_contig.ptr();
    let b_ptr = b_contig.ptr();
    let c_ptr = c_contig.ptr();
    let out_ptr = out.ptr();

    dispatch_dtype!(dtype, T => {
        unsafe {
            kernels::fused_add_mul_kernel::<T>(
                a_ptr as *const T,
                b_ptr as *const T,
                c_ptr as *const T,
                out_ptr as *mut T,
                len,
            );
        }
    }, "fused_add_mul");

    Ok(out)
}

/// Helper for fused_mul_add_scalar: out = a * scale + bias
pub fn fused_mul_add_scalar_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    scale: f64,
    bias: f64,
) -> Result<Tensor<CpuRuntime>> {
    let dtype = a.dtype();
    let a_contig = ensure_contiguous(a);
    let out = Tensor::<CpuRuntime>::empty(a.shape(), dtype, &client.device);

    let len = a.numel();
    let a_ptr = a_contig.ptr();
    let out_ptr = out.ptr();

    dispatch_dtype!(dtype, T => {
        unsafe {
            kernels::fused_mul_add_scalar_kernel::<T>(
                a_ptr as *const T,
                scale,
                bias,
                out_ptr as *mut T,
                len,
            );
        }
    }, "fused_mul_add_scalar");

    Ok(out)
}