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
use crate::{
core::{Array, MajorOrder},
math::{MatmulPolicy, matmul::Matmul},
};
#[cfg(target_arch = "aarch64")]
use std::arch::aarch64::{
vfmaq_f32, vfmaq_f64, vld1q_dup_f32, vld1q_dup_f64, vld1q_f32, vld1q_f64, vst1q_f32, vst1q_f64,
};
impl Matmul for Array<2, f32> {
fn matmul_with_policy(&self, rhs: &Self, policy: MatmulPolicy) -> Self {
assert_eq!(self.shape()[1], rhs.shape()[0]);
match policy {
MatmulPolicy::Faer => {
let l = self.as_faer();
let r = rhs.as_faer();
(l * r).into()
}
_ => {
let m = self.shape[0];
let k = self.shape[1];
let n = rhs.shape[1];
let result_shape = [m, n];
let mut result_data = vec![0.0; result_shape.iter().product()];
match policy {
#[cfg(target_arch = "aarch64")]
MatmulPolicy::LoopRecorderSimd => {
let l_s_r = self.strides[0];
let l_s_c = self.strides[1];
let r_s_r = rhs.strides[0];
let r_s_c = rhs.strides[1];
for i in 0..m {
for l in 0..k {
// In practice, using get_unchecked directly only reduces runtime by 1%
// let a_v = unsafe { self.data.get_unchecked(i * k + l) };
// let a_v = self.data[i * l_s_c + l];
let a_v =
unsafe { vld1q_dup_f32(&self.data[i * l_s_r + l * l_s_c]) };
let mut j = 0;
while j + 3 < n {
unsafe {
let b_vec = vld1q_f32(&rhs.data[l * r_s_r + j * r_s_c]);
let c_vec = vld1q_f32(&result_data[i * n + j]);
let res_vec = vfmaq_f32(c_vec, a_v, b_vec);
vst1q_f32(&mut result_data[i * n + j], res_vec);
}
j += 4;
}
for j in j..n {
result_data[i * n + j] += self.data[i * l_s_r + l * l_s_c]
* rhs.data[l * r_s_r + j * r_s_c].clone();
}
}
}
}
MatmulPolicy::Blas => {
let m = m as i32;
let k = k as i32;
let n = n as i32;
let (transa, transb, lda, ldb) = match (self.major_order, rhs.major_order) {
(MajorOrder::RowMajor, MajorOrder::RowMajor) => (b'N', b'N', n, k),
(MajorOrder::RowMajor, MajorOrder::ColumnMajor) => (b'T', b'N', k, k),
(MajorOrder::ColumnMajor, MajorOrder::RowMajor) => (b'N', b'T', n, m),
(MajorOrder::ColumnMajor, MajorOrder::ColumnMajor) => {
(b'T', b'T', k, m)
}
};
let c = result_data.as_mut_ptr();
let alpha = 1.0;
let beta = 0.0;
unsafe {
blas_sys::sgemm_(
&(transa as std::ffi::c_char),
&(transb as std::ffi::c_char),
&n,
&m,
&k,
&alpha,
rhs.data().as_ptr().cast(),
&lda,
self.data.as_ptr().cast(),
&ldb,
&beta,
c.cast(),
&n,
);
}
}
_ => return super::matmul_general(self, rhs, policy),
}
Array::from_vec_major(result_data, result_shape, MajorOrder::RowMajor)
}
}
}
}
impl Matmul for Array<2, f64> {
fn matmul_with_policy(&self, rhs: &Self, policy: MatmulPolicy) -> Self {
assert_eq!(self.shape()[1], rhs.shape()[0]);
match policy {
MatmulPolicy::Faer => {
let l = self.as_faer();
let r = rhs.as_faer();
(l * r).into()
}
_ => {
let m = self.shape[0];
let k = self.shape[1];
let n = rhs.shape[1];
let result_shape = [m, n];
let mut result_data = vec![0.0; result_shape.iter().product()];
match policy {
#[cfg(target_arch = "aarch64")]
MatmulPolicy::LoopRecorderSimd => {
let l_s_r = self.strides[0];
let l_s_c = self.strides[1];
let r_s_r = rhs.strides[0];
let r_s_c = rhs.strides[1];
for i in 0..m {
for l in 0..k {
// In practice, using get_unchecked directly only reduces runtime by 1%
// let a_v = unsafe { self.data.get_unchecked(i * k + l) };
// let a_v = self.data[i * l_s_c + l];
// 因为一次读取2个位置数据，所以每行的结尾需要避免越界
let mut j = 0;
let a_v =
unsafe { vld1q_dup_f64(&self.data[i * l_s_r + l * l_s_c]) };
while j + 1 < n {
unsafe {
let b_vec = vld1q_f64(&rhs.data[l * r_s_r + j * r_s_c]);
let c_vec = vld1q_f64(&result_data[i * n + j]);
let res_vec = vfmaq_f64(c_vec, a_v, b_vec);
vst1q_f64(&mut result_data[i * n + j], res_vec);
}
j += 2;
}
for j in j..n {
result_data[i * n + j] += self.data[i * l_s_r + l * l_s_c]
* rhs.data[l * r_s_r + j * r_s_c].clone();
}
}
}
}
MatmulPolicy::Blas => {
let m = m as i32;
let n = n as i32;
let k = k as i32;
let a = self.data.as_ptr();
let b = rhs.data.as_ptr();
let (transa, transb, lda, ldb) = match (self.major_order, rhs.major_order) {
(MajorOrder::RowMajor, MajorOrder::RowMajor) => (
cblas_sys::CBLAS_TRANSPOSE::CblasNoTrans,
cblas_sys::CBLAS_TRANSPOSE::CblasNoTrans,
k as i32,
n as i32,
),
(MajorOrder::RowMajor, MajorOrder::ColumnMajor) => (
cblas_sys::CBLAS_TRANSPOSE::CblasNoTrans,
cblas_sys::CBLAS_TRANSPOSE::CblasTrans,
k as i32,
k as i32,
),
(MajorOrder::ColumnMajor, MajorOrder::RowMajor) => (
cblas_sys::CBLAS_TRANSPOSE::CblasTrans,
cblas_sys::CBLAS_TRANSPOSE::CblasNoTrans,
m as i32,
n as i32,
),
(MajorOrder::ColumnMajor, MajorOrder::ColumnMajor) => (
cblas_sys::CBLAS_TRANSPOSE::CblasTrans,
cblas_sys::CBLAS_TRANSPOSE::CblasTrans,
m as i32,
k as i32,
),
};
let c = result_data.as_mut_ptr();
let alpha = 1.0;
let beta = 0.0;
unsafe {
cblas_sys::cblas_dgemm(
cblas_sys::CBLAS_LAYOUT::CblasRowMajor,
transa,
transb,
m,
n,
k,
alpha,
a.cast(),
lda,
b.cast(),
ldb,
beta,
c.cast(),
n,
);
}
}
_ => return super::matmul_general(self, rhs, policy),
}
Array::from_vec_major(result_data, result_shape, MajorOrder::RowMajor)
}
}
}
}