tenferro-cpu 0.4.0

CPU backend, kernels, provider selection, and CPU resource pools for tenferro.
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

use num_complex::{Complex32, Complex64};

use crate::provider::CpuExecutionContext;

pub(crate) trait FaerGemm: Sized {
    #[allow(clippy::too_many_arguments, dead_code)]
    unsafe fn strided_gemm(
        ctx: &CpuExecutionContext<'_>,
        alpha: Self,
        a_ptr: *const Self,
        m: usize,
        k: usize,
        a_rs: isize,
        a_cs: isize,
        b_ptr: *const Self,
        n: usize,
        b_rs: isize,
        b_cs: isize,
        beta: Self,
        c_ptr: *mut Self,
        c_rs: isize,
        c_cs: isize,
    ) {
        unsafe {
            Self::strided_gemm_with_conj(
                ctx, alpha, a_ptr, m, k, a_rs, a_cs, false, b_ptr, n, b_rs, b_cs, false, beta,
                c_ptr, c_rs, c_cs,
            )
        }
    }

    #[allow(clippy::too_many_arguments, dead_code)]
    unsafe fn strided_gemm_with_conj(
        ctx: &CpuExecutionContext<'_>,
        alpha: Self,
        a_ptr: *const Self,
        m: usize,
        k: usize,
        a_rs: isize,
        a_cs: isize,
        conj_a: bool,
        b_ptr: *const Self,
        n: usize,
        b_rs: isize,
        b_cs: isize,
        conj_b: bool,
        beta: Self,
        c_ptr: *mut Self,
        c_rs: isize,
        c_cs: isize,
    ) {
        unsafe {
            Self::strided_gemm_with_conj_par(
                ctx,
                ctx.faer_parallelism(),
                alpha,
                a_ptr,
                m,
                k,
                a_rs,
                a_cs,
                conj_a,
                b_ptr,
                n,
                b_rs,
                b_cs,
                conj_b,
                beta,
                c_ptr,
                c_rs,
                c_cs,
            )
        }
    }

    #[allow(clippy::too_many_arguments)]
    unsafe fn strided_gemm_with_conj_par(
        ctx: &CpuExecutionContext<'_>,
        par: faer::Par,
        alpha: Self,
        a_ptr: *const Self,
        m: usize,
        k: usize,
        a_rs: isize,
        a_cs: isize,
        conj_a: bool,
        b_ptr: *const Self,
        n: usize,
        b_rs: isize,
        b_cs: isize,
        conj_b: bool,
        beta: Self,
        c_ptr: *mut Self,
        c_rs: isize,
        c_cs: isize,
    );
}

macro_rules! impl_faer_gemm {
    ($ty:ty) => {
        impl FaerGemm for $ty {
            unsafe fn strided_gemm_with_conj_par(
                ctx: &CpuExecutionContext<'_>,
                par: faer::Par,
                alpha: $ty,
                a_ptr: *const $ty,
                m: usize,
                k: usize,
                a_rs: isize,
                a_cs: isize,
                conj_a: bool,
                b_ptr: *const $ty,
                n: usize,
                b_rs: isize,
                b_cs: isize,
                conj_b: bool,
                beta: $ty,
                c_ptr: *mut $ty,
                c_rs: isize,
                c_cs: isize,
            ) {
                let _ = ctx;
                use faer::{Accum, Conj, MatMut, MatRef};
                let a_rs = super::normalize_singleton_stride(a_rs, m, k);
                let a_cs = super::normalize_singleton_stride(a_cs, k, m);
                let b_rs = super::normalize_singleton_stride(b_rs, k, n);
                let b_cs = super::normalize_singleton_stride(b_cs, n, k);
                let c_rs = super::normalize_singleton_stride(c_rs, m, 1);
                let c_cs = super::normalize_singleton_stride(c_cs, n, m);
                // SAFETY: callers pass dense tensor storage plus validated GEMM
                // dimensions/strides; singleton strides are normalized above.
                let a_mat = MatRef::<$ty>::from_raw_parts(a_ptr, m, k, a_rs, a_cs);
                let b_mat = MatRef::<$ty>::from_raw_parts(b_ptr, k, n, b_rs, b_cs);
                let zero = <$ty as num_traits::Zero>::zero();
                let one = <$ty as num_traits::One>::one();
                let accum = if beta == zero {
                    Accum::Replace
                } else {
                    if beta != one {
                        // SAFETY: beta != 0 callers must pass initialized C; pooled dot_general uses beta = 0.
                        let mut col_off = 0isize;
                        for _ in 0..n {
                            let mut off = col_off;
                            for _ in 0..m {
                                *c_ptr.offset(off) *= beta;
                                off += c_rs;
                            }
                            col_off += c_cs;
                        }
                    }
                    Accum::Add
                };
                // SAFETY: `c_ptr` points at the caller-owned output buffer for
                // an m x n dense result; beta=0 paths may receive uninitialized
                // storage because faer uses Accum::Replace.
                let mut c_mat = MatMut::<$ty>::from_raw_parts_mut(c_ptr, m, n, c_rs, c_cs);
                let conj_a = if conj_a { Conj::Yes } else { Conj::No };
                let conj_b = if conj_b { Conj::Yes } else { Conj::No };
                faer::linalg::matmul::matmul_with_conj(
                    &mut c_mat, accum, &a_mat, conj_a, &b_mat, conj_b, alpha, par,
                );
            }
        }
    };
}

impl_faer_gemm!(f64);
impl_faer_gemm!(f32);
impl_faer_gemm!(Complex64);
impl_faer_gemm!(Complex32);