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
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
// Copyright 2014-2016 bluss and ndarray developers.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

use crate::imp_prelude::*;
use crate::numeric_util;

use crate::{LinalgScalar, Zip};

use std::any::TypeId;

#[cfg(feature = "blas")]
use std::cmp;
#[cfg(feature = "blas")]
use std::mem::swap;
#[cfg(feature = "blas")]
use std::os::raw::c_int;

#[cfg(feature = "blas")]
use cblas_sys as blas_sys;
#[cfg(feature = "blas")]
use cblas_sys::{CblasNoTrans, CblasRowMajor, CblasTrans, CBLAS_LAYOUT};

/// len of vector before we use blas
#[cfg(feature = "blas")]
const DOT_BLAS_CUTOFF: usize = 32;
/// side of matrix before we use blas
#[cfg(feature = "blas")]
const GEMM_BLAS_CUTOFF: usize = 7;
#[cfg(feature = "blas")]
#[allow(non_camel_case_types)]
type blas_index = c_int; // blas index type

impl<A, S> ArrayBase<S, Ix1>
where
    S: Data<Elem = A>,
{
    /// Perform dot product or matrix multiplication of arrays `self` and `rhs`.
    ///
    /// `Rhs` may be either a one-dimensional or a two-dimensional array.
    ///
    /// If `Rhs` is one-dimensional, then the operation is a vector dot
    /// product, which is the sum of the elementwise products (no conjugation
    /// of complex operands, and thus not their inner product). In this case,
    /// `self` and `rhs` must be the same length.
    ///
    /// If `Rhs` is two-dimensional, then the operation is matrix
    /// multiplication, where `self` is treated as a row vector. In this case,
    /// if `self` is shape *M*, then `rhs` is shape *M* × *N* and the result is
    /// shape *N*.
    ///
    /// **Panics** if the array shapes are incompatible.<br>
    /// *Note:* If enabled, uses blas `dot` for elements of `f32, f64` when memory
    /// layout allows.
    pub fn dot<Rhs>(&self, rhs: &Rhs) -> <Self as Dot<Rhs>>::Output
    where
        Self: Dot<Rhs>,
    {
        Dot::dot(self, rhs)
    }

    fn dot_generic<S2>(&self, rhs: &ArrayBase<S2, Ix1>) -> A
    where
        S2: Data<Elem = A>,
        A: LinalgScalar,
    {
        debug_assert_eq!(self.len(), rhs.len());
        assert!(self.len() == rhs.len());
        if let Some(self_s) = self.as_slice() {
            if let Some(rhs_s) = rhs.as_slice() {
                return numeric_util::unrolled_dot(self_s, rhs_s);
            }
        }
        let mut sum = A::zero();
        for i in 0..self.len() {
            unsafe {
                sum = sum + *self.uget(i) * *rhs.uget(i);
            }
        }
        sum
    }

    #[cfg(not(feature = "blas"))]
    fn dot_impl<S2>(&self, rhs: &ArrayBase<S2, Ix1>) -> A
    where
        S2: Data<Elem = A>,
        A: LinalgScalar,
    {
        self.dot_generic(rhs)
    }

    #[cfg(feature = "blas")]
    fn dot_impl<S2>(&self, rhs: &ArrayBase<S2, Ix1>) -> A
    where
        S2: Data<Elem = A>,
        A: LinalgScalar,
    {
        // Use only if the vector is large enough to be worth it
        if self.len() >= DOT_BLAS_CUTOFF {
            debug_assert_eq!(self.len(), rhs.len());
            assert!(self.len() == rhs.len());
            macro_rules! dot {
                ($ty:ty, $func:ident) => {{
                    if blas_compat_1d::<$ty, _>(self) && blas_compat_1d::<$ty, _>(rhs) {
                        unsafe {
                            let (lhs_ptr, n, incx) =
                                blas_1d_params(self.ptr.as_ptr(), self.len(), self.strides()[0]);
                            let (rhs_ptr, _, incy) =
                                blas_1d_params(rhs.ptr.as_ptr(), rhs.len(), rhs.strides()[0]);
                            let ret = blas_sys::$func(
                                n,
                                lhs_ptr as *const $ty,
                                incx,
                                rhs_ptr as *const $ty,
                                incy,
                            );
                            return cast_as::<$ty, A>(&ret);
                        }
                    }
                }};
            }

            dot! {f32, cblas_sdot};
            dot! {f64, cblas_ddot};
        }
        self.dot_generic(rhs)
    }
}

/// Return a pointer to the starting element in BLAS's view.
///
/// BLAS wants a pointer to the element with lowest address,
/// which agrees with our pointer for non-negative strides, but
/// is at the opposite end for negative strides.
#[cfg(feature = "blas")]
unsafe fn blas_1d_params<A>(
    ptr: *const A,
    len: usize,
    stride: isize,
) -> (*const A, blas_index, blas_index) {
    // [x x x x]
    //        ^--ptr
    //        stride = -1
    //  ^--blas_ptr = ptr + (len - 1) * stride
    if stride >= 0 || len == 0 {
        (ptr, len as blas_index, stride as blas_index)
    } else {
        let ptr = ptr.offset((len - 1) as isize * stride);
        (ptr, len as blas_index, stride as blas_index)
    }
}

/// Matrix Multiplication
///
/// For two-dimensional arrays, the dot method computes the matrix
/// multiplication.
pub trait Dot<Rhs> {
    /// The result of the operation.
    ///
    /// For two-dimensional arrays: a rectangular array.
    type Output;
    fn dot(&self, rhs: &Rhs) -> Self::Output;
}

impl<A, S, S2> Dot<ArrayBase<S2, Ix1>> for ArrayBase<S, Ix1>
where
    S: Data<Elem = A>,
    S2: Data<Elem = A>,
    A: LinalgScalar,
{
    type Output = A;

    /// Compute the dot product of one-dimensional arrays.
    ///
    /// The dot product is a sum of the elementwise products (no conjugation
    /// of complex operands, and thus not their inner product).
    ///
    /// **Panics** if the arrays are not of the same length.<br>
    /// *Note:* If enabled, uses blas `dot` for elements of `f32, f64` when memory
    /// layout allows.
    fn dot(&self, rhs: &ArrayBase<S2, Ix1>) -> A {
        self.dot_impl(rhs)
    }
}

impl<A, S, S2> Dot<ArrayBase<S2, Ix2>> for ArrayBase<S, Ix1>
where
    S: Data<Elem = A>,
    S2: Data<Elem = A>,
    A: LinalgScalar,
{
    type Output = Array<A, Ix1>;

    /// Perform the matrix multiplication of the row vector `self` and
    /// rectangular matrix `rhs`.
    ///
    /// The array shapes must agree in the way that
    /// if `self` is *M*, then `rhs` is *M* × *N*.
    ///
    /// Return a result array with shape *N*.
    ///
    /// **Panics** if shapes are incompatible.
    fn dot(&self, rhs: &ArrayBase<S2, Ix2>) -> Array<A, Ix1> {
        rhs.t().dot(self)
    }
}

impl<A, S> ArrayBase<S, Ix2>
where
    S: Data<Elem = A>,
{
    /// Perform matrix multiplication of rectangular arrays `self` and `rhs`.
    ///
    /// `Rhs` may be either a one-dimensional or a two-dimensional array.
    ///
    /// If Rhs is two-dimensional, they array shapes must agree in the way that
    /// if `self` is *M* × *N*, then `rhs` is *N* × *K*.
    ///
    /// Return a result array with shape *M* × *K*.
    ///
    /// **Panics** if shapes are incompatible or the number of elements in the
    /// result would overflow `isize`.
    ///
    /// *Note:* If enabled, uses blas `gemv/gemm` for elements of `f32, f64`
    /// when memory layout allows. The default matrixmultiply backend
    /// is otherwise used for `f32, f64` for all memory layouts.
    ///
    /// ```
    /// use ndarray::arr2;
    ///
    /// let a = arr2(&[[1., 2.],
    ///                [0., 1.]]);
    /// let b = arr2(&[[1., 2.],
    ///                [2., 3.]]);
    ///
    /// assert!(
    ///     a.dot(&b) == arr2(&[[5., 8.],
    ///                         [2., 3.]])
    /// );
    /// ```
    pub fn dot<Rhs>(&self, rhs: &Rhs) -> <Self as Dot<Rhs>>::Output
    where
        Self: Dot<Rhs>,
    {
        Dot::dot(self, rhs)
    }
}

impl<A, S, S2> Dot<ArrayBase<S2, Ix2>> for ArrayBase<S, Ix2>
where
    S: Data<Elem = A>,
    S2: Data<Elem = A>,
    A: LinalgScalar,
{
    type Output = Array2<A>;
    fn dot(&self, b: &ArrayBase<S2, Ix2>) -> Array2<A> {
        let a = self.view();
        let b = b.view();
        let ((m, k), (k2, n)) = (a.dim(), b.dim());
        if k != k2 || m.checked_mul(n).is_none() {
            dot_shape_error(m, k, k2, n);
        }

        let lhs_s0 = a.strides()[0];
        let rhs_s0 = b.strides()[0];
        let column_major = lhs_s0 == 1 && rhs_s0 == 1;
        // A is Copy so this is safe
        let mut v = Vec::with_capacity(m * n);
        let mut c;
        unsafe {
            v.set_len(m * n);
            c = Array::from_shape_vec_unchecked((m, n).set_f(column_major), v);
        }
        mat_mul_impl(A::one(), &a, &b, A::zero(), &mut c.view_mut());
        c
    }
}

/// Assumes that `m` and `n` are ≤ `isize::MAX`.
#[cold]
#[inline(never)]
fn dot_shape_error(m: usize, k: usize, k2: usize, n: usize) -> ! {
    match m.checked_mul(n) {
        Some(len) if len <= ::std::isize::MAX as usize => {}
        _ => panic!("ndarray: shape {} × {} overflows isize", m, n),
    }
    panic!(
        "ndarray: inputs {} × {} and {} × {} are not compatible for matrix multiplication",
        m, k, k2, n
    );
}

#[cold]
#[inline(never)]
fn general_dot_shape_error(m: usize, k: usize, k2: usize, n: usize, c1: usize, c2: usize) -> ! {
    panic!("ndarray: inputs {} × {}, {} × {}, and output {} × {} are not compatible for matrix multiplication",
           m, k, k2, n, c1, c2);
}

/// Perform the matrix multiplication of the rectangular array `self` and
/// column vector `rhs`.
///
/// The array shapes must agree in the way that
/// if `self` is *M* × *N*, then `rhs` is *N*.
///
/// Return a result array with shape *M*.
///
/// **Panics** if shapes are incompatible.
impl<A, S, S2> Dot<ArrayBase<S2, Ix1>> for ArrayBase<S, Ix2>
where
    S: Data<Elem = A>,
    S2: Data<Elem = A>,
    A: LinalgScalar,
{
    type Output = Array<A, Ix1>;
    fn dot(&self, rhs: &ArrayBase<S2, Ix1>) -> Array<A, Ix1> {
        let ((m, a), n) = (self.dim(), rhs.dim());
        if a != n {
            dot_shape_error(m, a, n, 1);
        }

        // Avoid initializing the memory in vec -- set it during iteration
        unsafe {
            let mut c = Array::uninitialized(m);
            general_mat_vec_mul(A::one(), self, rhs, A::zero(), &mut c);
            c
        }
    }
}

impl<A, S, D> ArrayBase<S, D>
where
    S: Data<Elem = A>,
    D: Dimension,
{
    /// Perform the operation `self += alpha * rhs` efficiently, where
    /// `alpha` is a scalar and `rhs` is another array. This operation is
    /// also known as `axpy` in BLAS.
    ///
    /// If their shapes disagree, `rhs` is broadcast to the shape of `self`.
    ///
    /// **Panics** if broadcasting isn’t possible.
    pub fn scaled_add<S2, E>(&mut self, alpha: A, rhs: &ArrayBase<S2, E>)
    where
        S: DataMut,
        S2: Data<Elem = A>,
        A: LinalgScalar,
        E: Dimension,
    {
        self.zip_mut_with(rhs, move |y, &x| *y = *y + (alpha * x));
    }
}

// mat_mul_impl uses ArrayView arguments to send all array kinds into
// the same instantiated implementation.
#[cfg(not(feature = "blas"))]
use self::mat_mul_general as mat_mul_impl;

#[cfg(feature = "blas")]
fn mat_mul_impl<A>(
    alpha: A,
    lhs: &ArrayView2<'_, A>,
    rhs: &ArrayView2<'_, A>,
    beta: A,
    c: &mut ArrayViewMut2<'_, A>,
) where
    A: LinalgScalar,
{
    // size cutoff for using BLAS
    let cut = GEMM_BLAS_CUTOFF;
    let ((mut m, a), (_, mut n)) = (lhs.dim(), rhs.dim());
    if !(m > cut || n > cut || a > cut) || !(same_type::<A, f32>() || same_type::<A, f64>()) {
        return mat_mul_general(alpha, lhs, rhs, beta, c);
    }
    {
        // Use `c` for c-order and `f` for an f-order matrix
        // We can handle c * c, f * f generally and
        // c * f and f * c if the `f` matrix is square.
        let mut lhs_ = lhs.view();
        let mut rhs_ = rhs.view();
        let mut c_ = c.view_mut();
        let lhs_s0 = lhs_.strides()[0];
        let rhs_s0 = rhs_.strides()[0];
        let both_f = lhs_s0 == 1 && rhs_s0 == 1;
        let mut lhs_trans = CblasNoTrans;
        let mut rhs_trans = CblasNoTrans;
        if both_f {
            // A^t B^t = C^t => B A = C
            let lhs_t = lhs_.reversed_axes();
            lhs_ = rhs_.reversed_axes();
            rhs_ = lhs_t;
            c_ = c_.reversed_axes();
            swap(&mut m, &mut n);
        } else if lhs_s0 == 1 && m == a {
            lhs_ = lhs_.reversed_axes();
            lhs_trans = CblasTrans;
        } else if rhs_s0 == 1 && a == n {
            rhs_ = rhs_.reversed_axes();
            rhs_trans = CblasTrans;
        }

        macro_rules! gemm {
            ($ty:ty, $gemm:ident) => {
                if blas_row_major_2d::<$ty, _>(&lhs_)
                    && blas_row_major_2d::<$ty, _>(&rhs_)
                    && blas_row_major_2d::<$ty, _>(&c_)
                {
                    let (m, k) = match lhs_trans {
                        CblasNoTrans => lhs_.dim(),
                        _ => {
                            let (rows, cols) = lhs_.dim();
                            (cols, rows)
                        }
                    };
                    let n = match rhs_trans {
                        CblasNoTrans => rhs_.raw_dim()[1],
                        _ => rhs_.raw_dim()[0],
                    };
                    // adjust strides, these may [1, 1] for column matrices
                    let lhs_stride = cmp::max(lhs_.strides()[0] as blas_index, k as blas_index);
                    let rhs_stride = cmp::max(rhs_.strides()[0] as blas_index, n as blas_index);
                    let c_stride = cmp::max(c_.strides()[0] as blas_index, n as blas_index);

                    // gemm is C ← αA^Op B^Op + βC
                    // Where Op is notrans/trans/conjtrans
                    unsafe {
                        blas_sys::$gemm(
                            CblasRowMajor,
                            lhs_trans,
                            rhs_trans,
                            m as blas_index,               // m, rows of Op(a)
                            n as blas_index,               // n, cols of Op(b)
                            k as blas_index,               // k, cols of Op(a)
                            cast_as(&alpha),               // alpha
                            lhs_.ptr.as_ptr() as *const _, // a
                            lhs_stride,                    // lda
                            rhs_.ptr.as_ptr() as *const _, // b
                            rhs_stride,                    // ldb
                            cast_as(&beta),                // beta
                            c_.ptr.as_ptr() as *mut _,     // c
                            c_stride,                      // ldc
                        );
                    }
                    return;
                }
            };
        }
        gemm!(f32, cblas_sgemm);
        gemm!(f64, cblas_dgemm);
    }
    mat_mul_general(alpha, lhs, rhs, beta, c)
}

/// C ← α A B + β C
fn mat_mul_general<A>(
    alpha: A,
    lhs: &ArrayView2<'_, A>,
    rhs: &ArrayView2<'_, A>,
    beta: A,
    c: &mut ArrayViewMut2<'_, A>,
) where
    A: LinalgScalar,
{
    let ((m, k), (_, n)) = (lhs.dim(), rhs.dim());

    // common parameters for gemm
    let ap = lhs.as_ptr();
    let bp = rhs.as_ptr();
    let cp = c.as_mut_ptr();
    let (rsc, csc) = (c.strides()[0], c.strides()[1]);
    if same_type::<A, f32>() {
        unsafe {
            ::matrixmultiply::sgemm(
                m,
                k,
                n,
                cast_as(&alpha),
                ap as *const _,
                lhs.strides()[0],
                lhs.strides()[1],
                bp as *const _,
                rhs.strides()[0],
                rhs.strides()[1],
                cast_as(&beta),
                cp as *mut _,
                rsc,
                csc,
            );
        }
    } else if same_type::<A, f64>() {
        unsafe {
            ::matrixmultiply::dgemm(
                m,
                k,
                n,
                cast_as(&alpha),
                ap as *const _,
                lhs.strides()[0],
                lhs.strides()[1],
                bp as *const _,
                rhs.strides()[0],
                rhs.strides()[1],
                cast_as(&beta),
                cp as *mut _,
                rsc,
                csc,
            );
        }
    } else {
        // It's a no-op if `c` has zero length.
        if c.is_empty() {
            return;
        }

        // initialize memory if beta is zero
        if beta.is_zero() {
            c.fill(beta);
        }

        let mut i = 0;
        let mut j = 0;
        loop {
            unsafe {
                let elt = c.uget_mut((i, j));
                *elt = *elt * beta
                    + alpha
                        * (0..k).fold(A::zero(), move |s, x| {
                            s + *lhs.uget((i, x)) * *rhs.uget((x, j))
                        });
            }
            j += 1;
            if j == n {
                j = 0;
                i += 1;
                if i == m {
                    break;
                }
            }
        }
    }
}

/// General matrix-matrix multiplication.
///
/// Compute C ← α A B + β C
///
/// The array shapes must agree in the way that
/// if `a` is *M* × *N*, then `b` is *N* × *K* and `c` is *M* × *K*.
///
/// ***Panics*** if array shapes are not compatible<br>
/// *Note:* If enabled, uses blas `gemm` for elements of `f32, f64` when memory
/// layout allows.  The default matrixmultiply backend is otherwise used for
/// `f32, f64` for all memory layouts.
pub fn general_mat_mul<A, S1, S2, S3>(
    alpha: A,
    a: &ArrayBase<S1, Ix2>,
    b: &ArrayBase<S2, Ix2>,
    beta: A,
    c: &mut ArrayBase<S3, Ix2>,
) where
    S1: Data<Elem = A>,
    S2: Data<Elem = A>,
    S3: DataMut<Elem = A>,
    A: LinalgScalar,
{
    let ((m, k), (k2, n)) = (a.dim(), b.dim());
    let (m2, n2) = c.dim();
    if k != k2 || m != m2 || n != n2 {
        general_dot_shape_error(m, k, k2, n, m2, n2);
    } else {
        mat_mul_impl(alpha, &a.view(), &b.view(), beta, &mut c.view_mut());
    }
}

/// General matrix-vector multiplication.
///
/// Compute y ← α A x + β y
///
/// where A is a *M* × *N* matrix and x is an *N*-element column vector and
/// y an *M*-element column vector (one dimensional arrays).
///
/// ***Panics*** if array shapes are not compatible<br>
/// *Note:* If enabled, uses blas `gemv` for elements of `f32, f64` when memory
/// layout allows.
#[allow(clippy::collapsible_if)]
pub fn general_mat_vec_mul<A, S1, S2, S3>(
    alpha: A,
    a: &ArrayBase<S1, Ix2>,
    x: &ArrayBase<S2, Ix1>,
    beta: A,
    y: &mut ArrayBase<S3, Ix1>,
) where
    S1: Data<Elem = A>,
    S2: Data<Elem = A>,
    S3: DataMut<Elem = A>,
    A: LinalgScalar,
{
    let ((m, k), k2) = (a.dim(), x.dim());
    let m2 = y.dim();
    if k != k2 || m != m2 {
        general_dot_shape_error(m, k, k2, 1, m2, 1);
    } else {
        #[cfg(feature = "blas")]
        macro_rules! gemv {
            ($ty:ty, $gemv:ident) => {
                if let Some(layout) = blas_layout::<$ty, _>(&a) {
                    if blas_compat_1d::<$ty, _>(&x) && blas_compat_1d::<$ty, _>(&y) {
                        // Determine stride between rows or columns. Note that the stride is
                        // adjusted to at least `k` or `m` to handle the case of a matrix with a
                        // trivial (length 1) dimension, since the stride for the trivial dimension
                        // may be arbitrary.
                        let a_trans = CblasNoTrans;
                        let a_stride = match layout {
                            CBLAS_LAYOUT::CblasRowMajor => {
                                a.strides()[0].max(k as isize) as blas_index
                            }
                            CBLAS_LAYOUT::CblasColMajor => {
                                a.strides()[1].max(m as isize) as blas_index
                            }
                        };

                        let x_stride = x.strides()[0] as blas_index;
                        let y_stride = y.strides()[0] as blas_index;

                        unsafe {
                            blas_sys::$gemv(
                                layout,
                                a_trans,
                                m as blas_index,            // m, rows of Op(a)
                                k as blas_index,            // n, cols of Op(a)
                                cast_as(&alpha),            // alpha
                                a.ptr.as_ptr() as *const _, // a
                                a_stride,                   // lda
                                x.ptr.as_ptr() as *const _, // x
                                x_stride,
                                cast_as(&beta),           // beta
                                y.ptr.as_ptr() as *mut _, // x
                                y_stride,
                            );
                        }
                        return;
                    }
                }
            };
        }
        #[cfg(feature = "blas")]
        gemv!(f32, cblas_sgemv);
        #[cfg(feature = "blas")]
        gemv!(f64, cblas_dgemv);

        /* general */

        if beta.is_zero() {
            Zip::from(a.outer_iter()).and(y).apply(|row, elt| {
                *elt = row.dot(x) * alpha;
            });
        } else {
            Zip::from(a.outer_iter()).and(y).apply(|row, elt| {
                *elt = *elt * beta + row.dot(x) * alpha;
            });
        }
    }
}

#[inline(always)]
/// Return `true` if `A` and `B` are the same type
fn same_type<A: 'static, B: 'static>() -> bool {
    TypeId::of::<A>() == TypeId::of::<B>()
}

// Read pointer to type `A` as type `B`.
//
// **Panics** if `A` and `B` are not the same type
fn cast_as<A: 'static + Copy, B: 'static + Copy>(a: &A) -> B {
    assert!(same_type::<A, B>());
    unsafe { ::std::ptr::read(a as *const _ as *const B) }
}

#[cfg(feature = "blas")]
fn blas_compat_1d<A, S>(a: &ArrayBase<S, Ix1>) -> bool
where
    S: Data,
    A: 'static,
    S::Elem: 'static,
{
    if !same_type::<A, S::Elem>() {
        return false;
    }
    if a.len() > blas_index::max_value() as usize {
        return false;
    }
    let stride = a.strides()[0];
    if stride > blas_index::max_value() as isize || stride < blas_index::min_value() as isize {
        return false;
    }
    true
}

#[cfg(feature = "blas")]
enum MemoryOrder {
    C,
    F,
}

#[cfg(feature = "blas")]
fn blas_row_major_2d<A, S>(a: &ArrayBase<S, Ix2>) -> bool
where
    S: Data,
    A: 'static,
    S::Elem: 'static,
{
    if !same_type::<A, S::Elem>() {
        return false;
    }
    is_blas_2d(&a.dim, &a.strides, MemoryOrder::C)
}

#[cfg(feature = "blas")]
fn blas_column_major_2d<A, S>(a: &ArrayBase<S, Ix2>) -> bool
where
    S: Data,
    A: 'static,
    S::Elem: 'static,
{
    if !same_type::<A, S::Elem>() {
        return false;
    }
    is_blas_2d(&a.dim, &a.strides, MemoryOrder::F)
}

#[cfg(feature = "blas")]
fn is_blas_2d(dim: &Ix2, stride: &Ix2, order: MemoryOrder) -> bool {
    let (m, n) = dim.into_pattern();
    let s0 = stride[0] as isize;
    let s1 = stride[1] as isize;
    let (inner_stride, outer_dim) = match order {
        MemoryOrder::C => (s1, n),
        MemoryOrder::F => (s0, m),
    };
    if !(inner_stride == 1 || outer_dim == 1) {
        return false;
    }
    if s0 < 1 || s1 < 1 {
        return false;
    }
    if (s0 > blas_index::max_value() as isize || s0 < blas_index::min_value() as isize)
        || (s1 > blas_index::max_value() as isize || s1 < blas_index::min_value() as isize)
    {
        return false;
    }
    if m > blas_index::max_value() as usize || n > blas_index::max_value() as usize {
        return false;
    }
    true
}

#[cfg(feature = "blas")]
fn blas_layout<A, S>(a: &ArrayBase<S, Ix2>) -> Option<CBLAS_LAYOUT>
where
    S: Data,
    A: 'static,
    S::Elem: 'static,
{
    if blas_row_major_2d::<A, _>(a) {
        Some(CBLAS_LAYOUT::CblasRowMajor)
    } else if blas_column_major_2d::<A, _>(a) {
        Some(CBLAS_LAYOUT::CblasColMajor)
    } else {
        None
    }
}

#[cfg(test)]
#[cfg(feature = "blas")]
mod blas_tests {
    use super::*;

    #[test]
    fn blas_row_major_2d_normal_matrix() {
        let m: Array2<f32> = Array2::zeros((3, 5));
        assert!(blas_row_major_2d::<f32, _>(&m));
        assert!(!blas_column_major_2d::<f32, _>(&m));
    }

    #[test]
    fn blas_row_major_2d_row_matrix() {
        let m: Array2<f32> = Array2::zeros((1, 5));
        assert!(blas_row_major_2d::<f32, _>(&m));
        assert!(blas_column_major_2d::<f32, _>(&m));
    }

    #[test]
    fn blas_row_major_2d_column_matrix() {
        let m: Array2<f32> = Array2::zeros((5, 1));
        assert!(blas_row_major_2d::<f32, _>(&m));
        assert!(blas_column_major_2d::<f32, _>(&m));
    }

    #[test]
    fn blas_row_major_2d_transposed_row_matrix() {
        let m: Array2<f32> = Array2::zeros((1, 5));
        let m_t = m.t();
        assert!(blas_row_major_2d::<f32, _>(&m_t));
        assert!(blas_column_major_2d::<f32, _>(&m_t));
    }

    #[test]
    fn blas_row_major_2d_transposed_column_matrix() {
        let m: Array2<f32> = Array2::zeros((5, 1));
        let m_t = m.t();
        assert!(blas_row_major_2d::<f32, _>(&m_t));
        assert!(blas_column_major_2d::<f32, _>(&m_t));
    }

    #[test]
    fn blas_column_major_2d_normal_matrix() {
        let m: Array2<f32> = Array2::zeros((3, 5).f());
        assert!(!blas_row_major_2d::<f32, _>(&m));
        assert!(blas_column_major_2d::<f32, _>(&m));
    }
}