pllm 0.4.1

Portable LLM
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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248

#![feature(portable_simd)]
#![feature(array_chunks)]

// extern crate blas_src;
use std::simd::num::SimdInt;
use std::simd::{f32x16, f32x32, f32x8, i16x16, i16x8, i32x16, i32x4, i32x8};
use std::simd::{f32x4, num::SimdFloat, StdFloat};
use std::time::Instant;

// use ndarray::linalg::Dot;
// use ndarray::{ArcArray, CowArray};

// pub fn f32_dot_prod_ndarray(a: &[f32], b: &[f32]) -> f32 {
//     let ar = CowArray::from(a);
//     let br = CowArray::from(b);
//     ar.dot(&br)
// }

// pub fn i32_dot_prod_ndarray(a: &[i32], b: &[i32]) -> i32 {
//     let ar = CowArray::from(a);
//     let br = CowArray::from(b);
//     ar.dot(&br)
// }

pub fn f32_dot_prod_test() {
    let N = 2048;
    let K = 256128;
    let a: Vec<f32> = (0..N).map(|i| ((i as f32) / 100.0).sin()).collect();
    let b: Vec<f32> = (0..N).map(|i| ((i as f32) / 100.0).cos()).collect();
    let mut result = 0.0;
    let mut before = Instant::now();
    for _ in 0..K {
        result = f32_dot_prod_simd_4(&a, &b);
    }
    println!("f32x4 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = f32_dot_prod_simd_8(&a, &b);
    }
    println!("f32x8 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = f32_dot_prod_simd_16(&a, &b);
    }
    println!("f32x16 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = f32_dot_prod_simd_32(&a, &b);
    }
    println!("f32x32 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = f32_dot_prod(&a, &b);
    }
    println!("f32 plain time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    // before = Instant::now();
    // for _ in 0..K {
    //     result = f32_dot_prod_ndarray(&a, &b);
    // }
    // println!("f32 ndarray time: {:.2?}", before.elapsed());
    // println!("result: {}", result);
}

pub fn i16_dot_prod_test() {
    let N = 2048;
    let K = 256128;
    let a: Vec<i16> = (0..N).collect();
    let b: Vec<i16> = (0..N).collect();
    let mut result: i32 = 0;
    let mut before = Instant::now();
    for _ in 0..K {
        result = i16_dot_prod_simd_8(&a, &b);
    }
    println!("i16x8 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = i16_dot_prod_simd_16(&a, &b);
    }
    println!("i16x16 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = i16_dot_prod(&a, &b);
    }
    println!("i16 plain time: {:.2?}", before.elapsed());
    println!("result: {}", result);
}

pub fn i32_dot_prod_test() {
    let N = 2048;
    let K = 256128;
    let a: Vec<i32> = (0..N).collect();
    let b: Vec<i32> = (0..N).collect();
    let mut result = 0;
    let mut before = Instant::now();
    for _ in 0..K {
        result = i32_dot_prod_simd_4(&a, &b);
    }
    println!("i32x4 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = i32_dot_prod_simd_8(&a, &b);
    }
    println!("i32x8 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = i32_dot_prod_simd_16(&a, &b);
    }
    println!("i32x16 time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    before = Instant::now();
    for _ in 0..K {
        result = i32_dot_prod(&a, &b);
    }
    println!("i32 plain time: {:.2?}", before.elapsed());
    println!("result: {}", result);

    // before = Instant::now();
    // for _ in 0..K {
    //     result = i32_dot_prod_ndarray(&a, &b);
    // }
    // println!("i32 ndarray time: {:.2?}", before.elapsed());

    // println!("result: {}", result);
}

fn f32_dot_prod_simd_4(a: &[f32], b: &[f32]) -> f32 {
    a.array_chunks::<4>()
        .map(|&a| f32x4::from_array(a))
        .zip(b.array_chunks::<4>().map(|&b| f32x4::from_array(b)))
        .fold(f32x4::splat(0.), |acc, (a, b)| a.mul_add(b, acc))
        .reduce_sum()
}

fn i32_dot_prod_simd_4(a: &[i32], b: &[i32]) -> i32 {
    a.array_chunks::<4>()
        .map(|&a| i32x4::from_array(a))
        .zip(b.array_chunks::<4>().map(|&b| i32x4::from_array(b)))
        .fold(i32x4::splat(0), |acc, (a, b)| a * b + acc)
        .reduce_sum()
}

fn f32_dot_prod_simd_8(a: &[f32], b: &[f32]) -> f32 {
    a.array_chunks::<8>()
        .map(|&a| f32x8::from_array(a))
        .zip(b.array_chunks::<8>().map(|&b| f32x8::from_array(b)))
        .fold(f32x8::splat(0.), |acc, (a, b)| a.mul_add(b, acc))
        .reduce_sum()
}

fn i32_dot_prod_simd_8(a: &[i32], b: &[i32]) -> i32 {
    a.array_chunks::<8>()
        .map(|&a| i32x8::from_array(a))
        .zip(b.array_chunks::<8>().map(|&b| i32x8::from_array(b)))
        .fold(i32x8::splat(0), |acc, (a, b)| a * b + acc)
        .reduce_sum()
}

fn i16_dot_prod_simd_8(a: &[i16], b: &[i16]) -> i32 {
    a.array_chunks::<8>()
        .map(|&a| i16x8::from_array(a))
        .zip(b.array_chunks::<8>().map(|&b| i16x8::from_array(b)))
        .fold(i16x8::splat(0), |acc, (a, b)| a * b + acc)
        .reduce_sum() as i32
}

fn f32_dot_prod_simd_16(a: &[f32], b: &[f32]) -> f32 {
    a.array_chunks::<16>()
        .map(|&a| f32x16::from_array(a))
        .zip(b.array_chunks::<16>().map(|&b| f32x16::from_array(b)))
        .fold(f32x16::splat(0.), |acc, (a, b)| a.mul_add(b, acc))
        .reduce_sum()
}

fn i32_dot_prod_simd_16(a: &[i32], b: &[i32]) -> i32 {
    a.array_chunks::<16>()
        .map(|&a| i32x16::from_array(a))
        .zip(b.array_chunks::<16>().map(|&b| i32x16::from_array(b)))
        .fold(i32x16::splat(0), |acc, (a, b)| a * b + acc)
        .reduce_sum()
}

fn i16_dot_prod_simd_16(a: &[i16], b: &[i16]) -> i32 {
    a.array_chunks::<16>()
        .map(|&a| i16x16::from_array(a))
        .zip(b.array_chunks::<16>().map(|&b| i16x16::from_array(b)))
        .fold(i16x16::splat(0), |acc, (a, b)| a * b + acc)
        .reduce_sum() as i32
}

fn f32_dot_prod_simd_32(a: &[f32], b: &[f32]) -> f32 {
    a.array_chunks::<32>()
        .map(|&a| f32x32::from_array(a))
        .zip(b.array_chunks::<32>().map(|&b| f32x32::from_array(b)))
        .fold(f32x32::splat(0.), |acc, (a, b)| a.mul_add(b, acc))
        .reduce_sum()
}

fn f32_dot_prod(a: &[f32], b: &[f32]) -> f32 {
    assert_eq!(a.len(), b.len());
    let mut product = 0.0;
    for i in 0..a.len() {
        product += a[i] * b[i];
    }
    product
}

fn i32_dot_prod(a: &[i32], b: &[i32]) -> i32 {
    assert_eq!(a.len(), b.len());
    let mut product = 0;
    for i in 0..a.len() {
        product += a[i] * b[i];
    }
    product
}

fn i16_dot_prod(a: &[i16], b: &[i16]) -> i32 {
    assert_eq!(a.len(), b.len());
    let mut product = 0;
    for i in 0..a.len() {
        product += a[i] * b[i];
    }
    product as i32
}

fn main() {
    f32_dot_prod_test();
    i32_dot_prod_test();
    i16_dot_prod_test();
}