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
249
250
251
252
253
254
255
256
257
258
259
260
use ::operation::*;
use ::plugin::*;
use ::transpose::*;
use collenchyma::backend::Backend;
use collenchyma::memory::MemoryType;
use collenchyma::frameworks::native::Native;
use collenchyma::plugin::Error;
use rblas::math::mat::Mat;
use rblas::matrix::Matrix;
use rblas;
macro_rules! impl_asum_for {
($t:ident, $b:ty) => (
impl IOperationAsum<$t> for $b {
fn compute(&self, x: &MemoryType, result: &mut MemoryType) -> Result<(), Error> {
let x_slice = try!(x.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_slice::<$t>();
let mut r_slice = try!(result.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `result`."))).as_mut_slice::<$t>();
r_slice[0] = rblas::Asum::asum(x_slice);
Ok(())
}
}
);
}
macro_rules! impl_axpy_for {
($t:ident, $b:ty) => (
impl IOperationAxpy<$t> for $b {
fn compute(&self, a: &MemoryType, x: &MemoryType, y: &mut MemoryType) -> Result<(), Error> {
let a_slice = try!(a.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `a`."))).as_slice::<$t>();
let x_slice = try!(x.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_slice::<$t>();
let y_slice = try!(y.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `y`."))).as_mut_slice::<$t>();
rblas::Axpy::axpy(&a_slice[0], x_slice, y_slice);
Ok(())
}
}
);
}
macro_rules! impl_copy_for {
($t:ident, $b:ty) => (
impl IOperationCopy<$t> for $b {
fn compute(&self, x: &MemoryType, y: &mut MemoryType) -> Result<(), Error> {
let x_slice = try!(x.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_slice::<$t>();
let y_slice = try!(y.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `y`."))).as_mut_slice::<$t>();
rblas::Copy::copy(x_slice, y_slice);
Ok(())
}
}
);
}
macro_rules! impl_dot_for {
($t:ident, $b:ty) => (
impl IOperationDot<$t> for $b {
fn compute(&self, x: &MemoryType, y: &MemoryType, result: &mut MemoryType) -> Result<(), Error> {
let x_slice = try!(x.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_slice::<$t>();
let y_slice = try!(y.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `y`."))).as_slice::<$t>();
let mut r_slice = try!(result.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `result`."))).as_mut_slice::<$t>();
r_slice[0] = rblas::Dot::dot(x_slice, y_slice);
Ok(())
}
}
);
}
macro_rules! impl_nrm2_for {
($t:ident, $b:ty) => (
impl IOperationNrm2<$t> for $b {
fn compute(&self, x: &MemoryType, result: &mut MemoryType) -> Result<(), Error> {
let x_slice = try!(x.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_slice::<$t>();
let mut r_slice = try!(result.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `result`."))).as_mut_slice::<$t>();
r_slice[0] = rblas::Nrm2::nrm2(x_slice);
Ok(())
}
}
);
}
macro_rules! impl_scale_for {
($t:ident, $b:ty) => (
impl IOperationScale<$t> for $b {
fn compute(&self, a: &MemoryType, x: &mut MemoryType) -> Result<(), Error> {
let a_slice = try!(a.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `a`."))).as_slice::<$t>();
let mut x_slice = try!(x.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_mut_slice::<$t>();
rblas::Scal::scal(&a_slice[0], x_slice);
Ok(())
}
}
);
}
macro_rules! impl_swap_for {
($t:ident, $b:ty) => (
impl IOperationSwap<$t> for $b {
fn compute(&self, x: &mut MemoryType, y: &mut MemoryType) -> Result<(), Error> {
let mut x_slice = try!(x.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `x`."))).as_mut_slice::<$t>();
let mut y_slice = try!(y.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `y`."))).as_mut_slice::<$t>();
rblas::Swap::swap(x_slice, y_slice);
Ok(())
}
}
);
}
macro_rules! impl_gemm_for {
($t:ident, $b:ty) => (
impl IOperationGemm<$t> for $b {
fn compute(&self, alpha: &MemoryType, at: Transpose, a_dims: &[usize], a: &MemoryType, bt: Transpose, b_dims: &[usize], b: &MemoryType, beta: &MemoryType, c_dims: &[usize], c: &mut MemoryType) -> Result<(), ::collenchyma::error::Error> {
let alpha_slice = try!(alpha.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `alpha`."))).as_slice::<$t>();
let a_slice = try!(a.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `a`."))).as_slice::<$t>();
let beta_slice = try!(beta.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `beta`."))).as_slice::<$t>();
let b_slice = try!(b.as_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `b`."))).as_slice::<$t>();
let mut c_slice = try!(c.as_mut_native().ok_or(Error::MissingMemoryForDevice("Unable to receive native memory for `c`."))).as_mut_slice::<$t>();
let a_matrix = as_matrix(a_slice, a_dims);
let b_matrix = as_matrix(b_slice, b_dims);
let mut c_matrix = as_matrix(c_slice, c_dims);
rblas::Gemm::gemm(&alpha_slice[0], at.to_rblas(), &a_matrix, bt.to_rblas(), &b_matrix, &beta_slice[0], &mut c_matrix);
read_from_matrix(&c_matrix, c_slice);
Ok(())
}
}
);
}
macro_rules! impl_iblas_for {
($t:ident, $b:ty) => (
impl_asum_for!($t, $b);
impl_axpy_for!($t, $b);
impl_copy_for!($t, $b);
impl_dot_for!($t, $b);
impl_nrm2_for!($t, $b);
impl_scale_for!($t, $b);
impl_swap_for!($t, $b);
impl_gemm_for!($t, $b);
impl IBlas<$t> for $b { }
impl Asum<$t> for $b {
iblas_asum_for!($t, $b);
}
impl Axpy<$t> for $b {
iblas_axpy_for!($t, $b);
}
impl Copy<$t> for $b {
iblas_copy_for!($t, $b);
}
impl Dot<$t> for $b {
iblas_dot_for!($t, $b);
}
impl Nrm2<$t> for $b {
iblas_nrm2_for!($t, $b);
}
impl Scal<$t> for $b {
iblas_scale_for!($t, $b);
}
impl Swap<$t> for $b {
iblas_swap_for!($t, $b);
}
impl Gemm<$t> for $b {
iblas_gemm_for!($t, $b);
}
);
}
impl_iblas_for!(f32, Backend<Native>);
impl_iblas_for!(f64, Backend<Native>);
fn as_matrix<T: Clone + ::std::fmt::Debug>(slice: &[T], dims: &[usize]) -> Mat<T> {
let n = dims[0];
let m = dims.iter().skip(1).fold(1, |prod, i| prod * i);
let mut mat: Mat<T> = Mat::new(n, m);
for i in 0..n {
for j in 0..m {
let index = m * i + j;
unsafe {
*mat.as_mut_ptr().offset(index as isize) = slice[index].clone();
}
}
}
mat
}
fn read_from_matrix<T: Clone>(mat: &Mat<T>, slice: &mut [T]) {
let n = mat.rows();
let m = mat.cols();
for i in 0..n {
for j in 0..m {
let index = m * i + j;
slice[index] = mat[i][j].clone();
}
}
}
#[cfg(test)]
mod test {
use collenchyma::backend::{Backend, BackendConfig};
use collenchyma::framework::IFramework;
use collenchyma::frameworks::Native;
use collenchyma::tensor::SharedTensor;
use collenchyma::memory::MemoryType;
use super::as_matrix;
fn get_native_backend() -> Backend<Native> {
let framework = Native::new();
let hardwares = framework.hardwares().to_vec();
let backend_config = BackendConfig::new(framework, &hardwares);
Backend::new(backend_config).unwrap()
}
pub fn write_to_memory<T: Copy>(mem: &mut MemoryType, data: &[T]) {
match mem {
&mut MemoryType::Native(ref mut mem) => {
let mut mem_buffer = mem.as_mut_slice::<T>();
for (index, datum) in data.iter().enumerate() {
mem_buffer[index] = *datum;
}
},
#[cfg(any(feature = "cuda", feature = "opencl"))]
_ => assert!(false)
}
}
#[test]
fn it_converts_correctly_to_and_from_matrix() {
let backend = get_native_backend();
let mut a = SharedTensor::<f32>::new(backend.device(), &vec![3, 2]).unwrap();
write_to_memory(a.get_mut(backend.device()).unwrap(),
&[2f32, 5f32,
2f32, 5f32,
2f32, 5f32]);
{
let a_slice_in = a.get(backend.device()).unwrap().as_native().unwrap().as_slice::<f32>();
let a_mat = as_matrix(a_slice_in, &[3, 2]);
assert_eq!(a_mat[0][0], 2f32);
assert_eq!(a_mat[0][1], 5f32);
assert_eq!(a_mat[1][0], 2f32);
assert_eq!(a_mat[1][1], 5f32);
assert_eq!(a_mat[2][0], 2f32);
assert_eq!(a_mat[2][1], 5f32);
}
}
}