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

use crate::Storage;
use std::arch::x86_64::{__m128i, _mm_movemask_epi8, _mm_set_epi16, _mm_setr_epi8, _mm_slli_epi64};
use std::mem;

#[repr(C)]
union __U128 {
    vector: __m128i,
    bytes: [u8; 16],
}

impl Default for __U128 {
    #[inline]
    fn default() -> Self {
        __U128 { bytes: [0u8; 16] }
    }
}

#[inline]
#[allow(unused)]
pub(crate) fn transpose<T: Storage>(input: &[T], nrows: usize, ncols: usize) -> Vec<T> {
    assert_eq!(nrows % 8, 0);
    assert_eq!(ncols % 8, 0);
    assert!(ncols >= 16, "ncols must be >= 16");
    assert!(nrows >= 16, "nrows must be >= 16");
    assert_eq!(
        ncols * nrows,
        mem::size_of_val(input) * 8,
        "input.len() does not match nrows/ncols"
    );
    let mut output = vec![T::zero(); input.len()];
    let input = bytemuck::cast_slice(input);
    let byte_output = bytemuck::cast_slice_mut(output.as_mut_slice());

    let inp = |x: usize, y: usize| -> usize { x * ncols / 8 + y / 8 };
    let out = |x: usize, y: usize| -> usize { y * nrows / 8 + x / 8 };

    unsafe {
        let mut h = [0_u8; 4];
        let mut rr: usize = 0;
        while rr <= nrows - 16 {
            let mut cc = 0;
            while cc < ncols {
                let mut v = _mm_setr_epi8(
                    *input.get_unchecked(inp(rr, cc)) as i8,
                    *input.get_unchecked(inp(rr + 1, cc)) as i8,
                    *input.get_unchecked(inp(rr + 2, cc)) as i8,
                    *input.get_unchecked(inp(rr + 3, cc)) as i8,
                    *input.get_unchecked(inp(rr + 4, cc)) as i8,
                    *input.get_unchecked(inp(rr + 5, cc)) as i8,
                    *input.get_unchecked(inp(rr + 6, cc)) as i8,
                    *input.get_unchecked(inp(rr + 7, cc)) as i8,
                    *input.get_unchecked(inp(rr + 8, cc)) as i8,
                    *input.get_unchecked(inp(rr + 9, cc)) as i8,
                    *input.get_unchecked(inp(rr + 10, cc)) as i8,
                    *input.get_unchecked(inp(rr + 11, cc)) as i8,
                    *input.get_unchecked(inp(rr + 12, cc)) as i8,
                    *input.get_unchecked(inp(rr + 13, cc)) as i8,
                    *input.get_unchecked(inp(rr + 14, cc)) as i8,
                    *input.get_unchecked(inp(rr + 15, cc)) as i8,
                );
                (0..8).rev().for_each(|i| {
                    h = _mm_movemask_epi8(v).to_le_bytes();
                    // TODO maybe this can be optimized by directly writing the
                    // output of the movemask
                    *byte_output.get_unchecked_mut(out(rr, cc + i)) = h[0];
                    *byte_output.get_unchecked_mut(out(rr, cc + i) + 1) = h[1];

                    v = _mm_slli_epi64::<1>(v);
                });
                cc += 8;
            }
            rr += 16;
        }
        if rr == nrows {
            return output;
        }

        // The remainder is a block of 8x(16n+8) bits (n may be 0).
        //  Do a PAIR of 8x8 blocks in each step:
        let mut tmp = __U128::default();
        let mut cc = 0;
        if ncols % 16 != 0 || nrows % 16 != 0 {
            // The fancy optimizations in the else-branch don't work if the above if-condition
            // holds, so we use the simpler non-simd variant for that case.
            while cc <= ncols - 16 {
                for i in 0..8 {
                    let h = *(input.get_unchecked(inp(rr + i, cc)) as *const _ as *const [u8; 2]);
                    tmp.bytes[i] = h[0];
                    tmp.bytes[i + 8] = h[1];
                }
                for i in (0..8).rev() {
                    h = _mm_movemask_epi8(tmp.vector).to_le_bytes();
                    // TODO maybe this can be optimized by directly writing the
                    // output of the movemask
                    *byte_output.get_unchecked_mut(out(rr, cc + i)) = h[0];
                    *byte_output.get_unchecked_mut(out(rr, cc + i + 8)) = h[1];

                    tmp.vector = _mm_slli_epi64::<1>(tmp.vector);
                }
                cc += 16;
            }
        } else {
            while cc <= ncols - 16 {
                let mut v = _mm_set_epi16(
                    *(input.get_unchecked(inp(rr + 7, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 6, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 5, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 4, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 3, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 2, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr + 1, cc)) as *const _ as *const i16),
                    *(input.get_unchecked(inp(rr, cc)) as *const _ as *const i16),
                );
                for i in (0..8).rev() {
                    h = _mm_movemask_epi8(v).to_le_bytes();
                    *byte_output.get_unchecked_mut(out(rr, cc + i)) = h[0];
                    *byte_output.get_unchecked_mut(out(rr, cc + i + 8)) = h[1];
                    v = _mm_slli_epi64::<1>(v);
                }
                cc += 16;
            }
        }
        if cc == ncols {
            return output;
        }
        for i in 0..8 {
            tmp.bytes[i] = *input.get_unchecked(inp(rr + i, cc));
        }
        for i in (0..8).rev() {
            h = _mm_movemask_epi8(tmp.vector).to_le_bytes();
            *byte_output.get_unchecked_mut(out(rr, cc + i)) = h[0];
            tmp.vector = _mm_slli_epi64::<1>(tmp.vector);
        }
    };
    output
}

#[cfg(test)]
mod tests {
    use proptest::prelude::*;

    fn arbitrary_bitmat<T: Arbitrary + Clone>(
        max_row: usize,
        max_col: usize,
    ) -> BoxedStrategy<(Vec<T>, usize, usize)>
    where
        T::Strategy: Clone,
    {
        (
            (16..max_row).prop_map(|row| row / 8 * 8),
            (16..max_col).prop_map(|col| col / 8 * 8),
        )
            .prop_flat_map(|(rows, cols)| {
                (
                    vec![any::<T>(); rows * cols / (std::mem::size_of::<T>() * 8)],
                    Just(rows),
                    Just(cols),
                )
            })
            .boxed()
    }

    proptest! {
        #[test]
        fn test_byte_transpose((v, rows, cols) in arbitrary_bitmat::<u8>(16 * 30, 16 * 30)) {
            let transposed = super::transpose(v.as_slice(), rows, cols);
            let expected = crate::simple::transpose(v.as_slice(), rows, cols);
            prop_assert_eq!(transposed, expected);
        }

        #[test]
        fn test_u64_transpose((v, rows, cols) in arbitrary_bitmat::<u64>(16 * 30, 16 * 30)) {
            let transposed = super::transpose(v.as_slice(), rows, cols);
            let expected = crate::simple::transpose(v.as_slice(), rows, cols);
            prop_assert_eq!(transposed, expected);
        }
    }
}