redstone-ml 0.0.0

High-performance Machine Learning, Auto-Differentiation and Tensor Algebra crate for Rust
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
198
199
200
201
202
203
204
205
206
207
208

use crate::acceleration::simd::Simd;
#[cfg(neon_simd)]
use crate::util::partial_ord::{partial_max, partial_min};

pub(crate) trait SimdReduceOps: Simd {
    #[cfg(neon_simd)]
    unsafe fn simd_sum_uniform(mut ptr: *const Self, mut count: usize, stride: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::zero());

        while count >= Self::LANES {
            let vec = Self::simd_vec_from_stride(ptr, stride);
            acc = Self::simd_add(acc, vec);

            count -= Self::LANES;
            ptr = ptr.add(stride * Self::LANES);
        }

        let mut output = Self::simd_horizontal_sum(acc);

        for _ in 0..count {
            output += *ptr;
            ptr = ptr.add(stride);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_sum_contiguous(mut ptr: *const Self, mut count: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::zero());

        while count >= 4 * Self::LANES {
            let a = Self::simd_load(ptr.add(0 * Self::LANES));
            let b = Self::simd_load(ptr.add(1 * Self::LANES));
            let c = Self::simd_load(ptr.add(2 * Self::LANES));
            let d = Self::simd_load(ptr.add(3 * Self::LANES));

            let ab = Self::simd_add(a, b);
            let cd = Self::simd_add(c, d);
            let abcd = Self::simd_add(ab, cd);

            count -= 4 * Self::LANES;
            ptr = ptr.add(4 * Self::LANES);

            acc = Self::simd_add(acc, abcd);
        }

        let mut output = Self::simd_horizontal_sum(acc);
        for _ in 0..count {
            output += *ptr;
            ptr = ptr.add(1);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_product_uniform(mut ptr: *const Self, mut count: usize, stride: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::one());

        while count >= Self::LANES {
            let vec = Self::simd_vec_from_stride(ptr, stride);
            acc = Self::simd_mul(acc, vec);

            count -= Self::LANES;
            ptr = ptr.add(stride * Self::LANES);
        }

        let mut output = Self::simd_horizontal_mul(acc);

        for _ in 0..count {
            output *= *ptr;
            ptr = ptr.add(stride);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_product_contiguous(mut ptr: *const Self, mut count: usize) -> Self {
        let mut output = Self::one();

        while count >= 4 * Self::LANES {
            let a = Self::simd_load(ptr.add(0 * Self::LANES));
            let b = Self::simd_load(ptr.add(1 * Self::LANES));
            let c = Self::simd_load(ptr.add(2 * Self::LANES));
            let d = Self::simd_load(ptr.add(3 * Self::LANES));

            let ab = Self::simd_mul(a, b);
            let cd = Self::simd_mul(c, d);
            let abcd = Self::simd_mul(ab, cd);

            count -= 4 * Self::LANES;
            ptr = ptr.add(4 * Self::LANES);

            output *= Self::simd_horizontal_mul(abcd);
        }

        for _ in 0..count {
            output *= *ptr;
            ptr = ptr.add(1);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_min_uniform(mut ptr: *const Self, mut count: usize, stride: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::max_value());

        while count >= Self::LANES {
            let vec = Self::simd_vec_from_stride(ptr, stride);
            acc = Self::simd_min(acc, vec);

            count -= Self::LANES;
            ptr = ptr.add(stride * Self::LANES);
        }

        let mut output = Self::simd_horizontal_min(acc);
        for _ in 0..count {
            output = partial_min(*ptr, output);
            ptr = ptr.add(stride);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_min_contiguous(mut ptr: *const Self, mut count: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::max_value());

        while count >= 4 * Self::LANES {
            let a = Self::simd_load(ptr.add(0 * Self::LANES));
            let b = Self::simd_load(ptr.add(1 * Self::LANES));
            let c = Self::simd_load(ptr.add(2 * Self::LANES));
            let d = Self::simd_load(ptr.add(3 * Self::LANES));

            let ab = Self::simd_min(a, b);
            let cd = Self::simd_min(c, d);
            let abcd = Self::simd_min(ab, cd);

            count -= 4 * Self::LANES;
            ptr = ptr.add(4 * Self::LANES);

            acc = Self::simd_min(acc, abcd);
        }

        let mut output = Self::simd_horizontal_min(acc);
        for _ in 0..count {
            output = partial_min(*ptr, output);
            ptr = ptr.add(1);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_max_uniform(mut ptr: *const Self, mut count: usize, stride: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::min_value());

        while count >= Self::LANES {
            let vec = Self::simd_vec_from_stride(ptr, stride);
            acc = Self::simd_max(acc, vec);

            count -= Self::LANES;
            ptr = ptr.add(stride * Self::LANES);
        }

        let mut output = Self::simd_horizontal_max(acc);
        for _ in 0..count {
            output = partial_max(*ptr, output);
            ptr = ptr.add(stride);
        }

        output
    }

    #[cfg(neon_simd)]
    unsafe fn simd_max_contiguous(mut ptr: *const Self, mut count: usize) -> Self {
        let mut acc = Self::simd_from_constant(Self::min_value());

        while count >= 4 * Self::LANES {
            let a = Self::simd_load(ptr.add(0 * Self::LANES));
            let b = Self::simd_load(ptr.add(1 * Self::LANES));
            let c = Self::simd_load(ptr.add(2 * Self::LANES));
            let d = Self::simd_load(ptr.add(3 * Self::LANES));

            let ab = Self::simd_max(a, b);
            let cd = Self::simd_max(c, d);
            let abcd = Self::simd_max(ab, cd);

            count -= 4 * Self::LANES;
            ptr = ptr.add(4 * Self::LANES);

            acc = Self::simd_max(acc, abcd);
        }

        let mut output = Self::simd_horizontal_max(acc);
        for _ in 0..count {
            output = partial_max(*ptr, output);
            ptr = ptr.add(1);
        }

        output
    }
}

impl<T: Simd> SimdReduceOps for T {}