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

pub mod lir_unary;
pub mod mir;
pub mod mir_quant;
pub mod mir_quant_unary;
pub mod mir_unary;
pub mod pack;

use crate::internal::*;
use tract_itertools::Itertools;
use tract_ndarray::prelude::*;

pub use self::mir::MatMul;
pub use self::mir_quant::{MatMulQParams, QMatMul};
pub use self::mir_unary::MatMulUnary;
use self::pack::MatMatMulPack;

pub fn compute_shape<D: DimLike>(
    ashape: &[D],
    bshape: &[D],
    a_trans: bool,
    b_trans: bool,
    c_trans: bool,
) -> TractResult<(D, D, D, TVec<D>)> {
    let mut c_shape = crate::broadcast::multi_broadcast(&[
        &ashape[..(ashape.len() - 2)],
        &bshape[..(bshape.len() - 2)],
    ])
    .ok_or_else(|| format_err!("Could not broadcast"))?;
    let (mut m, mut ka) = (ashape[ashape.len() - 2].clone(), ashape[ashape.len() - 1].clone());
    let (mut kb, mut n) = (bshape[bshape.len() - 2].clone(), bshape[bshape.len() - 1].clone());
    if a_trans {
        std::mem::swap(&mut m, &mut ka);
    }
    if b_trans {
        std::mem::swap(&mut kb, &mut n);
    }
    if ka != kb {
        bail!(
            "Inconsistent matmul: a: {} b: {}, a_trans: {} b_trans: {} c_trans: {}",
            ashape.iter().join(","),
            bshape.iter().join(","),
            a_trans,
            b_trans,
            c_trans
        );
    }
    if c_trans {
        c_shape.push(n.clone());
        c_shape.push(m.clone());
    } else {
        c_shape.push(m.clone());
        c_shape.push(n.clone());
    }
    Ok((m, ka, n, c_shape))
}

pub fn output_type(input: DatumType) -> DatumType {
    if input.is_float() {
        input
    } else {
        i32::datum_type()
    }
}

pub(super) fn eval(
    a: &Tensor,
    b: &Tensor,
    a_trans: bool,
    b_trans: bool,
    c_trans: bool,
) -> TractResult<Tensor> {
    unsafe {
        let rank = a.rank();
        let (m, k, n, c_shape) = compute_shape(a.shape(), b.shape(), a_trans, b_trans, c_trans)?;
        let dt = output_type(a.datum_type());
        let mm = tract_linalg::ops()
            .mmm(a.datum_type(), b.datum_type(), dt, Some(m), Some(k), Some(n))
            .with_context(|| {
                format!(
                    "No matrix multiplier for {:?}x{:?} to {:?}",
                    a.datum_type(),
                    b.datum_type(),
                    dt
                )
            })?;
        let c_storage = mm.c_from_data_and_strides(
            dt.size_of(),
            if c_trans { 1 } else { c_shape[rank - 1] as isize },
            if !c_trans { 1 } else { c_shape[rank - 1] as isize },
        );

        let mut c = Tensor::uninitialized_dt(dt, &c_shape)?;

        let a_pack = mm.a_pack(k);
        let b_pack = mm.b_pack(k);

        let mut packed_a =
            Tensor::uninitialized_aligned_dt(a.datum_type(), &[a_pack.len(m)], a_pack.alignment())?;
        let mut packed_b =
            Tensor::uninitialized_aligned_dt(b.datum_type(), &[b_pack.len(n)], b_pack.alignment())?;

        for prefix in tract_ndarray::indices(&c_shape[..rank - 2]).into_iter() {
            let mut a_prefix = tvec!();
            let mut b_prefix = tvec!();
            for (axis, &dim) in prefix.slice().iter().enumerate() {
                a_prefix.push(dim.min(a.shape()[axis] - 1));
                b_prefix.push(dim.min(b.shape()[axis] - 1));
            }
            a_pack.pack(
                packed_a.view_mut(),
                &a.view_at_prefix(&a_prefix)?,
                !a_trans as usize,
                a_trans as usize,
            );
            b_pack.pack(
                packed_b.view_mut(),
                &b.view_at_prefix(&b_prefix)?,
                b_trans as usize,
                !b_trans as usize,
            );
            mm.run(
                m,
                k,
                n,
                &mm.a_packed(a.datum_type().size_of(), k).wrap(&packed_a.view()),
                &mm.b_packed(b.datum_type().size_of(), k).wrap(&packed_b.view()),
                &mut c_storage.wrap(&c.view_at_prefix_mut(prefix.slice())?),
                &[],
            )?;
        }
        Ok(c)
    }
}

pub(super) fn cost<A: DimLike + Clone, B: DimLike + Clone>(
    a: &[A],
    b: &[B],
    dt: DatumType,
    a_trans: bool,
    b_trans: bool,
) -> TractResult<TVec<(Cost, TDim)>> {
    let (m, k, n, c_shape) = compute_shape(
        &a.iter().map(|d| d.clone().to_dim()).collect::<TVec<_>>(),
        &b.iter().map(|d| d.clone().to_dim()).collect::<TVec<_>>(),
        a_trans,
        b_trans,
        false,
    )?;
    let mul = c_shape.iter().rev().skip(2).cloned().product();
    Ok(tvec!((Cost::FMA(dt), [mul, m.to_dim(), k.to_dim(), n.to_dim()].iter().product())))
}