cumath 0.2.7

Cuda-based matrix/vector computations
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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345



use std::ptr;

use super::{cuda::*, cublas_ffi::*};
use ::cuvector::*;
use ::cumatrix::*;
use ::meta::result::*;
#[cfg(not(feature = "disable_checks"))]
use ::meta::assert::*;



pub struct Cublas {
    handle: *mut StructCublasContext,
}
impl Drop for Cublas {
    fn drop(&mut self) {
        cublas_destroy(self.handle)
    }
}
impl Cublas {

    /// Returns a new instance of Cublas.
    pub fn new() -> CumathResult<Cublas> {
        let mut handle = ptr::null_mut();
        match cublas_create(&mut handle) {
            Some(err) => Err(CumathError::new(err)),
            None => Ok(Cublas { handle }),
        }
    }

    /// Sets the cuda stream used by this instance of Cublas.
    /// Uses cuda's default stream by default.
    pub fn set_stream(&mut self, stream: &CudaStream) {
        cublas_set_stream(self.handle, stream.stream)
    }

    /// Returns the smallest i where abs(vector[i-1]) = max(abs(value) for value in vector)
    pub fn amax_idx(&self, vector: &CuVectorDeref<f32>) -> i32 {
        let mut output = 0;
        cublas_isamax(self.handle, vector.len() as i32, vector.as_ptr(), 1, &mut output);
        output
    }

    /// Returns the smallest i where abs(vector[i-1]) = min(abs(value) for value in vector)
    pub fn amin_idx(&self, vector: &CuVectorDeref<f32>) -> i32 {
        let mut output = 0;
        cublas_isamin(self.handle, vector.len() as i32, vector.as_ptr(), 1, &mut output);
        output
    }

    /// Returns sum(abs(value) for value in vector)
    pub fn asum(&self, vector: &CuVectorDeref<f32>) -> f32 {
        let mut output = 0.0;
        cublas_sasum(self.handle, vector.len() as i32, vector.as_ptr(), 1, &mut output);
        output
    }

    /// y[i] = y[i] + alpha*x[i]
    pub fn axpy(&self, alpha: f32, x: &CuVectorDeref<f32>, y: &mut CuVectorDeref<f32>) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(x.len(), "x.len()", y.len(), "y.len()");
        }
        cublas_saxpy(self.handle, x.len() as i32, &alpha, x.as_ptr(), 1, y.as_mut_ptr(), 1)
    }

    /// y[i] = y[i] * alpha
    pub fn scal(&self, vector: &mut CuVectorDeref<f32>, alpha: f32) {
        cublas_sscal(self.handle, vector.len() as i32, &alpha, vector.as_mut_ptr(), 1);
    }

    /// output = matrix_mult(left_op, right_op)
    pub fn mult_m_m(&self, left_op: &CuMatrixDeref<f32>, right_op: &CuMatrixDeref<f32>, output: &mut CuMatrixDeref<f32>) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(left_op.cols(), "left_op.cols()", right_op.rows(), "right_op.rows()");
            assert_eq_usize(left_op.rows(), "left_op.rows()", output.rows(), "output.rows()");
            assert_eq_usize(right_op.cols(), "right_op.cols()", output.cols(), "output.cols()");
        }
        cublas_sgemm(self.handle,
                     CublasOperation::None, CublasOperation::None,
                     left_op.rows() as i32, right_op.cols() as i32, left_op.cols() as i32, &1.0,
                     left_op.as_ptr(), left_op.rows() as i32,
                     right_op.as_ptr(), right_op.rows() as i32,
                     &0.0, output.as_mut_ptr(), output.rows() as i32)
    }

    /// output = matrix_mult(left_op as RowMatrix, right_op)
    pub fn mult_row_m(&self, left_op: &CuVectorDeref<f32>, right_op: &CuMatrixDeref<f32>, output: &mut CuVectorDeref<f32>) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(left_op.len(), "left_op.len()", right_op.rows(), "right_op.rows()");
            assert_eq_usize(right_op.cols(), "right_op.cols()", output.len(), "output.len()");
        }
        cublas_sgemv(self.handle,
                     CublasOperation::Transpose,
                     right_op.rows() as i32, right_op.cols() as i32, &1.0,
                     right_op.as_ptr(), right_op.rows() as i32,
                     left_op.as_ptr(), 1,
                     &0.0, output.as_mut_ptr(), 1)
    }

    /// output = matrix_mult(left_op, right_op as ColMatrix)
    pub fn mult_m_col(&self, left_op: &CuMatrixDeref<f32>, right_op: &CuVectorDeref<f32>, output: &mut CuVectorDeref<f32>) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(left_op.cols(), "left_op.cols()", right_op.len(), "right_op.len()");
            assert_eq_usize(left_op.rows(), "left_op.rows()", output.len(), "output.len()");
        }
        cublas_sgemv(self.handle,
                     CublasOperation::None,
                     left_op.rows() as i32, left_op.cols() as i32, &1.0,
                     left_op.as_ptr(), left_op.rows() as i32,
                     right_op.as_ptr(), 1,
                     &0.0, output.as_mut_ptr(), 1)
    }

    /// output = matrix_mult(left_op as RowMatrix, right_op as ColMatrix)
    pub fn mult_col_row(&self, left_op: &CuVectorDeref<f32>, right_op: &CuVectorDeref<f32>, output: &mut CuMatrixDeref<f32>) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(left_op.len(), "left_op.len()", output.rows(), "output.rows()");
            assert_eq_usize(right_op.len(), "right_op.len()", output.cols(), "output.cols()");
        }
        cublas_sgemm(self.handle,
                     CublasOperation::None, CublasOperation::None,
                     left_op.len() as i32, right_op.len() as i32, 1, &1.0,
                     left_op.as_ptr(), left_op.len() as i32,
                     right_op.as_ptr(), 1,
                     &0.0, output.as_mut_ptr(), output.rows() as i32)
    }

    /// output = out_scl * output + in_scl * matrix_mult(left_op, right_op)
    pub fn mult_col_row_rescaled(&self, left_op: &CuVectorDeref<f32>, right_op: &CuVectorDeref<f32>, output: &mut CuMatrixDeref<f32>, in_scl: f32, out_scl: f32) {
        #[cfg(not(feature = "disable_checks"))] {
            assert_eq_usize(left_op.len(), "left_op.len()", output.rows(), "output.rows()");
            assert_eq_usize(right_op.len(), "right_op.len()", output.cols(), "output.cols()");
        }
        cublas_sgemm(self.handle,
                     CublasOperation::None, CublasOperation::None,
                     left_op.len() as i32, right_op.len() as i32, 1, &in_scl,
                     left_op.as_ptr(), left_op.len() as i32,
                     right_op.as_ptr(), 1,
                     &out_scl, output.as_mut_ptr(), output.rows() as i32)
    }

}


#[cfg(test)]
mod tests {

    use super::*;

    #[test]
    fn amax_idx() {
        let input_data = vec![-1.0, 2.0, 1.0, -2.0, 7.0, 5.5, -3.7, 1.1, 0.7];

        let cublas = Cublas::new().unwrap();
        let vector = CuVector::<f32>::from_host_data(input_data.as_slice());
        let idx = cublas.amax_idx(&vector);

        assert_eq!(4, idx -1);
    }

    #[test]
    fn amin_idx() {
        let input_data = vec![-1.0, 2.0, 1.0, -2.0, 7.0, 5.5, -3.7, 1.1, -0.7];

        let cublas = Cublas::new().unwrap();
        let vector = CuVector::<f32>::from_host_data(input_data.as_slice());
        let idx = cublas.amin_idx(&vector);

        assert_eq!(8, idx -1);
    }

    #[test]
    fn asum() {
        let input_data = vec![-1.0, 2.0, 1.0, -2.0, 7.0, 5.5, -3.7, 1.1, 0.7];

        let cublas = Cublas::new().unwrap();
        let vector = CuVector::<f32>::from_host_data(input_data.as_slice());
        let asum = cublas.asum(&vector);

        assert_eq!(24.0, asum);
    }

    #[test]
    fn axpy() {
        let x_data = vec![-1.0, 1.5, 0.2, -2.0];
        let y_data = vec![1.0, 0.0, -0.15, 5.0];

        let cublas = Cublas::new().unwrap();
        let x = CuVector::<f32>::from_host_data(x_data.as_slice());
        let mut y = CuVector::<f32>::from_host_data(y_data.as_slice());
        cublas.axpy(2.0, &x, &mut y);

        x.dev_assert_equals(&[-1.0, 1.5, 0.2, -2.0]);
        y.dev_assert_equals(&[-1.0, 3.0, 0.25, 1.0]);
    }

    #[test]
    fn scal() {
        let x_data = vec![-1.0, 1.5, 0.2, -2.0];

        let cublas = Cublas::new().unwrap();
        let mut x = CuVector::<f32>::from_host_data(x_data.as_slice());
        cublas.scal(&mut x, 2.0);

        x.dev_assert_equals(&[-2.0, 3.0, 0.4, -4.0]);
    }

    #[test]
    fn mult_m_m() {
        let input_data = vec![-1.0, 2.0, 1.0, -2.0, 7.0, 5.5];

        let cublas = Cublas::new().unwrap();
        let matrix1 = CuMatrix::<f32>::from_host_data(2, 3, input_data.as_slice());
        let matrix2 = CuMatrix::<f32>::from_host_data(3, 2, input_data.as_slice());
        let mut output = CuMatrix::<f32>::zero(2, 2);

        cublas.mult_m_m(&matrix1, &matrix2, &mut output);

        let mut output_buffer = vec![0.0; 4];
        output.clone_to_host(output_buffer.as_mut_slice());

        assert_equals_float(output_buffer[0], 10.0);
        assert_equals_float(output_buffer[1], -0.5);
        assert_equals_float(output_buffer[2], 47.5);
        assert_equals_float(output_buffer[3], 12.25);
    }

    #[test]
    fn mult_m_col() {
        let col_vector_data = vec![1.0, -2.0, 3.0];
        let matrix_data = vec![-1.5, 2.0, 1.5, -0.5, 1.0, 3.5];

        let cublas = Cublas::new().unwrap();
        let col_vector = CuVector::<f32>::from_host_data(col_vector_data.as_slice());
        let matrix = CuMatrix::<f32>::from_host_data(2, 3, matrix_data.as_slice());
        let mut output = CuVector::<f32>::zero(2);

        cublas.mult_m_col(&matrix, &col_vector, &mut output);

        let mut output_buffer = vec![0.0; 2];
        output.clone_to_host(output_buffer.as_mut_slice());

        assert_equals_float(output_buffer[0], -1.5);
        assert_equals_float(output_buffer[1], 13.5);
    }

    #[test]
    fn mult_col_row() {
        let col_vector_data = vec![2.2, -3.2, 1.1];
        let row_vector_data = vec![-1.0, 2.0];

        let cublas = Cublas::new().unwrap();
        let col_vector = CuVector::<f32>::from_host_data(col_vector_data.as_slice());
        let row_vector = CuVector::<f32>::from_host_data(row_vector_data.as_slice());
        let mut output = CuMatrix::<f32>::zero(3, 2);

        cublas.mult_col_row(&col_vector, &row_vector, &mut output);

        let mut output_buffer = vec![0.0; 6];
        output.clone_to_host(output_buffer.as_mut_slice());

        assert_equals_float(output_buffer[0], -2.2);
        assert_equals_float(output_buffer[1], 3.2);
        assert_equals_float(output_buffer[2], -1.1);
        assert_equals_float(output_buffer[3], 4.4);
        assert_equals_float(output_buffer[4], -6.4);
        assert_equals_float(output_buffer[5], 2.2);
    }

    #[test]
    fn mult_v_m() {
        let vector_data = vec![2.2, -3.2, 1.1];
        let matrix_data = vec![-1.0, 2.0, 1.0, -2.0, 7.0, 5.5];

        let cublas = Cublas::new().unwrap();
        let vector = CuVector::<f32>::from_host_data(vector_data.as_slice());
        let matrix = CuMatrix::<f32>::from_host_data(3, 2, matrix_data.as_slice());
        let mut output = CuVector::<f32>::zero(2);

        cublas.mult_row_m(&vector, &matrix, &mut output);

        let mut output_buffer = vec![0.0; 2];
        output.clone_to_host(output_buffer.as_mut_slice());

        assert_equals_float(output_buffer[0], -7.5);
        assert_equals_float(output_buffer[1], -20.75);
    }

    /*
    #[test]
    fn benchmark_stream() {
        use std::time::Instant;

        let input_data = vec![-1.0; 900];

        let mut cublas0 = Cublas::new();
        let mut cublas1 = Cublas::new();

        let stream0 = CudaStream::new();
        let stream1 = CudaStream::new();

        let matrix1 = CuMatrix::from_data(30, 30, input_data.as_slice());
        let matrix2 = CuMatrix::from_data(30, 30, input_data.as_slice());
        let mut output0 = CuMatrix::new(30, 30, 0.0);
        let mut output1 = CuMatrix::new(30, 30, 0.0);


        let t0 = Instant::now();
        cublas0.set_stream(&stream0);
        for _ in 0..100000 {
            cublas0.mult_m_m(&matrix1, &matrix2, &mut output0);
        }
        stream0.synchronize();
        let dt = t0.elapsed();
        println!("1 finished in {}.{}", dt.as_secs(), dt.subsec_nanos());

        let t0 = Instant::now();
        for _ in 0..50000 {
            cublas0.set_stream(&stream0);
            cublas0.mult_m_m(&matrix1, &matrix2, &mut output0);
            cublas0.set_stream(&stream1);
            cublas0.mult_m_m(&matrix1, &matrix2, &mut output1);
        }
        stream0.synchronize();
        stream1.synchronize();
        let dt = t0.elapsed();
        println!("2 finished in {}.{}", dt.as_secs(), dt.subsec_nanos());

        let t0 = Instant::now();
        cublas0.set_stream(&stream0);
        cublas1.set_stream(&stream1);
        for _ in 0..50000 {
            cublas0.mult_m_m(&matrix1, &matrix2, &mut output0);
            cublas1.mult_m_m(&matrix1, &matrix2, &mut output1);
        }
        stream0.synchronize();
        stream1.synchronize();
        let dt = t0.elapsed();
        println!("3 finished in {}.{}", dt.as_secs(), dt.subsec_nanos());
    }
    */

}