lak-kernels 0.1.0

BLAS-like linear algebra kernels in fully-safe Rust.
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

// nt_direct_microkernel.rs 

use std::simd::Simd; 
use crate::l1::axpy;
use crate::traits::Fma; 
use crate::types::{MatMut, MatRef, VecMut, VecRef}; 

// tunable 
pub(crate) const MR_F32: usize = 16; 
pub(crate) const MR_F64: usize = 8; 

// not tunable 
pub(crate) const NR_F32: usize = 4; 
pub(crate) const NR_F64: usize = 4; 

#[inline(always)]
pub(crate) fn sgemm_nt_micro( 
    alpha: f32, 
    a_panel: MatRef<'_, f32>, 
    b: MatRef<'_, f32>, 
    mut c_panel: MatMut<'_, f32>, 
    kc_beg: usize, 
    j_beg: usize, 
) { 
    let (m, kc) = a_panel.dimension(); 
    let (n, k) = b.dimension(); 
    let nc = c_panel.n_cols(); 
    debug_assert_eq!(m,  c_panel.n_rows()); 
    debug_assert!(j_beg + nc <= n); 
    debug_assert!(kc_beg + kc <= k);

    for j in (0..nc).step_by(NR_F32) {
        let nr = (nc - j).min(NR_F32); 

        // full NR wide panel 
        if nr == NR_F32 { 

            skernel_mrnr(
                alpha, 
                a_panel, 
                b, 
                c_panel.reborrow(), 
                kc_beg, 
                j_beg,
                kc, 
                j,
                m, 
                n,
            );

        } else {
            // leftover cols 
            let j0 = j; 
            let j1 = nc; 
            let b_slice = b.as_slice(); 

            for jj in j0..j1 { 
                let mut ccol = c_panel.col_mut(jj); 
                let jeff = j_beg + jj; 

                for kk in 0..kc { 
                    let keff = kc_beg + kk; 
                    let bval = b_slice[keff * n + jeff]; 
                    axpy(
                        alpha * bval, 
                        a_panel.col(kk), 
                        ccol.reborrow(),
                    );
                }
            }
        }
    }
} 



#[inline(always)]
pub(crate) fn dgemm_nt_micro( 
    alpha: f64, 
    a_panel: MatRef<'_, f64>, 
    b: MatRef<'_, f64>, 
    mut c_panel: MatMut<'_, f64>, 
    kc_beg: usize, 
    j_beg: usize, 
) { 
    let (m, kc) = a_panel.dimension(); 
    let (n, k) = b.dimension(); 
    let nc = c_panel.n_cols(); 
    debug_assert_eq!(m,  c_panel.n_rows()); 
    debug_assert!(j_beg + nc <= n); 
    debug_assert!(kc_beg + kc <= k);

    for j in (0..nc).step_by(NR_F64) {
        let nr = (nc - j).min(NR_F64); 

        // full NR wide panel 
        if nr == NR_F64 { 

            dkernel_mrnr(alpha, 
                a_panel, 
                b, 
                c_panel.reborrow(), 
                kc_beg,
                j_beg,
                kc, 
                j,
                m,
                n,
            );

        } else {
            // leftover cols 
            let j0 = j; 
            let j1 = nc; 
            let b_slice = b.as_slice(); 

            for jj in j0..j1 { 
                let mut ccol = c_panel.col_mut(jj); 
                let jeff = j_beg + jj; 

                for kk in 0..kc { 
                    let keff = kc_beg + kk; 
                    let bval = b_slice[keff * n + jeff]; 
                    axpy(
                        alpha * bval, 
                        a_panel.col(kk), 
                        ccol.reborrow(),
                    );
                }
            }
        }
    }
}


fn skernel_mrnr( 
    alpha: f32, 
    a_panel: MatRef<'_, f32>, 
    b: MatRef<'_, f32>, 
    mut c_panel: MatMut<'_, f32>, 
    kc_beg: usize, 
    j_beg: usize, 
    kc: usize, 
    j: usize, 
    m: usize, 
    n: usize, 
) { 
    debug_assert!(NR_F32 == 4); 

    let c_slice  = c_panel.as_slice_mut();
    let c_base   = j * m; 
    let c_block  = &mut c_slice[c_base..c_base + NR_F32 * m]; 
    let (cl, cr) = c_block.split_at_mut(m * 2); 
    let (c0, c1) = cl.split_at_mut(m); 
    let (c2, c3) = cr.split_at_mut(m);  

    let (c0_chunks, c0_tail) = c0.as_chunks_mut::<MR_F32>();
    let (c1_chunks, c1_tail) = c1.as_chunks_mut::<MR_F32>(); 
    let (c2_chunks, c2_tail) = c2.as_chunks_mut::<MR_F32>();
    let (c3_chunks, c3_tail) = c3.as_chunks_mut::<MR_F32>(); 
    let n_chunks = c0_chunks.len(); 

    let a_slice = a_panel.as_slice(); 
    let b_slice = b.as_slice(); 

    for chunk_idx in 0..n_chunks { 
        let mut c0vec = Simd::from_array(c0_chunks[chunk_idx]);
        let mut c1vec = Simd::from_array(c1_chunks[chunk_idx]); 
        let mut c2vec = Simd::from_array(c2_chunks[chunk_idx]); 
        let mut c3vec = Simd::from_array(c3_chunks[chunk_idx]); 

        for kk in 0..kc { 
            let row_beg = chunk_idx * MR_F32; 
            let col_beg = kk * m; 

            let acol = &a_slice[col_beg + row_beg..col_beg + row_beg + MR_F32]; 
            let avec = Simd::<f32, MR_F32>::from_slice(acol);

            let keff = kc_beg + kk; 
            let jeff = j_beg + j; 
            let b0 = b_slice[keff * n + jeff]; 
            let b1 = b_slice[keff * n + jeff + 1]; 
            let b2 = b_slice[keff * n + jeff + 2]; 
            let b3 = b_slice[keff * n + jeff + 3]; 
            let b0vec = Simd::<f32, MR_F32>::splat(alpha * b0); 
            let b1vec = Simd::<f32, MR_F32>::splat(alpha * b1); 
            let b2vec = Simd::<f32, MR_F32>::splat(alpha * b2); 
            let b3vec = Simd::<f32, MR_F32>::splat(alpha * b3); 

            c0vec = b0vec.fma(avec, c0vec); 
            c1vec = b1vec.fma(avec, c1vec); 
            c2vec = b2vec.fma(avec, c2vec); 
            c3vec = b3vec.fma(avec, c3vec); 
        }

        c0vec.copy_to_slice(&mut c0_chunks[chunk_idx]);
        c1vec.copy_to_slice(&mut c1_chunks[chunk_idx]); 
        c2vec.copy_to_slice(&mut c2_chunks[chunk_idx]); 
        c3vec.copy_to_slice(&mut c3_chunks[chunk_idx]); 
    }

    // leftover tail
    let tail_len = c0_tail.len(); 
    if tail_len != 0 { 
        let mr_tail_idx = n_chunks * MR_F32; 

        let mut c0 = VecMut::new(c0_tail); 
        let mut c1 = VecMut::new(c1_tail); 
        let mut c2 = VecMut::new(c2_tail); 
        let mut c3 = VecMut::new(c3_tail); 

        for kk in 0..kc { 
            let row_beg = mr_tail_idx; 
            let col_beg = kk * m; 

            let acol = VecRef::new(
                &a_slice[col_beg + row_beg..col_beg + row_beg + tail_len]
            ); 

            let keff = kc_beg + kk;
            let jeff = j_beg + j; 
            let b0 = b_slice[keff * n + jeff]; 
            let b1 = b_slice[keff * n + jeff + 1]; 
            let b2 = b_slice[keff * n + jeff + 2]; 
            let b3 = b_slice[keff * n + jeff + 3]; 

            axpy(alpha * b0, acol, c0.reborrow());
            axpy(alpha * b1, acol, c1.reborrow()); 
            axpy(alpha * b2, acol, c2.reborrow()); 
            axpy(alpha * b3, acol, c3.reborrow()); 
        }   
    }
}

fn dkernel_mrnr( 
    alpha: f64, 
    a_panel: MatRef<'_, f64>, 
    b: MatRef<'_, f64>, 
    mut c_panel: MatMut<'_, f64>, 
    kc_beg: usize, 
    j_beg: usize, 
    kc: usize, 
    j: usize, 
    m: usize, 
    n: usize, 
) { 
    debug_assert!(NR_F64 == 4); 

    let c_slice  = c_panel.as_slice_mut();
    let c_base   = j * m; 
    let c_block  = &mut c_slice[c_base..c_base + NR_F64 * m]; 
    let (cl, cr) = c_block.split_at_mut(m * 2); 
    let (c0, c1) = cl.split_at_mut(m); 
    let (c2, c3) = cr.split_at_mut(m);  

    let (c0_chunks, c0_tail) = c0.as_chunks_mut::<MR_F64>();
    let (c1_chunks, c1_tail) = c1.as_chunks_mut::<MR_F64>(); 
    let (c2_chunks, c2_tail) = c2.as_chunks_mut::<MR_F64>();
    let (c3_chunks, c3_tail) = c3.as_chunks_mut::<MR_F64>(); 
    let n_chunks = c0_chunks.len(); 

    let a_slice = a_panel.as_slice(); 
    let b_slice = b.as_slice(); 

    for chunk_idx in 0..n_chunks { 
        let mut c0vec = Simd::from_array(c0_chunks[chunk_idx]);
        let mut c1vec = Simd::from_array(c1_chunks[chunk_idx]); 
        let mut c2vec = Simd::from_array(c2_chunks[chunk_idx]); 
        let mut c3vec = Simd::from_array(c3_chunks[chunk_idx]); 

        for kk in 0..kc { 
            let row_beg = chunk_idx * MR_F64; 
            let col_beg = kk * m; 

            let acol = &a_slice[col_beg + row_beg..col_beg + row_beg + MR_F64]; 
            let avec = Simd::<f64, MR_F64>::from_slice(acol);

            let keff = kc_beg + kk; 
            let jeff = j_beg + j; 
            let b0 = b_slice[keff * n + jeff]; 
            let b1 = b_slice[keff * n + jeff + 1]; 
            let b2 = b_slice[keff * n + jeff + 2]; 
            let b3 = b_slice[keff * n + jeff + 3]; 
            let b0vec = Simd::<f64, MR_F64>::splat(alpha * b0); 
            let b1vec = Simd::<f64, MR_F64>::splat(alpha * b1); 
            let b2vec = Simd::<f64, MR_F64>::splat(alpha * b2); 
            let b3vec = Simd::<f64, MR_F64>::splat(alpha * b3); 

            c0vec = b0vec.fma(avec, c0vec); 
            c1vec = b1vec.fma(avec, c1vec); 
            c2vec = b2vec.fma(avec, c2vec); 
            c3vec = b3vec.fma(avec, c3vec); 
        }

        c0vec.copy_to_slice(&mut c0_chunks[chunk_idx]);
        c1vec.copy_to_slice(&mut c1_chunks[chunk_idx]); 
        c2vec.copy_to_slice(&mut c2_chunks[chunk_idx]); 
        c3vec.copy_to_slice(&mut c3_chunks[chunk_idx]); 
    }

    // leftover tail
    let tail_len = c0_tail.len(); 
    if tail_len != 0 { 
        let mr_tail_idx = n_chunks * MR_F64; 

        let mut c0 = VecMut::new(c0_tail); 
        let mut c1 = VecMut::new(c1_tail); 
        let mut c2 = VecMut::new(c2_tail); 
        let mut c3 = VecMut::new(c3_tail); 

        for kk in 0..kc { 
            let row_beg = mr_tail_idx; 
            let col_beg = kk * m; 

            let acol = VecRef::new(
                &a_slice[col_beg + row_beg..col_beg + row_beg + tail_len]
            ); 

            let keff = kc_beg + kk;
            let jeff = j_beg + j; 
            let b0 = b_slice[keff * n + jeff]; 
            let b1 = b_slice[keff * n + jeff + 1]; 
            let b2 = b_slice[keff * n + jeff + 2]; 
            let b3 = b_slice[keff * n + jeff + 3]; 

            axpy(alpha * b0, acol, c0.reborrow());
            axpy(alpha * b1, acol, c1.reborrow()); 
            axpy(alpha * b2, acol, c2.reborrow()); 
            axpy(alpha * b3, acol, c3.reborrow()); 
        }   
    }
}