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

use num_traits::Zero;
use std::ops::{Add, AddAssign, Mul};

use std::sync::Arc;

use crate::ops::prelude::*;
use ndarray::prelude::*;

use crate::ops::nn::conv::KernelFormat;
use crate::ops::nn::{DataFormat, Patch};

use tract_linalg::MatMul;

/*
 * group=1, N=1         N>1             g>1
 *
 * A: kernel
 *  * O rows            * O rows        * O rows
 *  * I*h*w cols        * I*w*h         * I/g*w*h
 * B: data
 *                      * N blocks
 *  * I*w*h rows        * I*w*h         * I*w*h
 *  * H*W cols          * H*W           * H*W
 * Gemm
 *  * 1 iter            * N iter        * g iter
 *  * m=O               * m=O           * m=O/g
 *  * k=I*h*w           * k=I*h*w       * k=I/g*h*w
 *  * n=H*W             * n=H*W         * n=H*W
 *
 *                                +------------+
 *                                | B input    |
 *                                +------------+
 *              +--------------+  +----------------+
 *              | A kernel g=0 |  | C output  g=0  |
 *              +--------------+  +----------------+
 *              | A kernel g=1 |  | C output  g=1  |
 *              +--------------+  +----------------+
 */

#[derive(CustomDebug, Clone, new)]
pub struct ConvGemm<T>
where
    T: Datum + Add + Mul + Zero + Copy,
{
    pub patch: Patch,
    pub full_output_shape: TVec<usize>,
    pub m: usize,
    pub k: usize,
    pub n: usize,
    pub kernel_fmt: KernelFormat,
    #[debug(skip)]
    pub packed_kernels: Vec<Tensor>,
    pub bias: Option<ArrayD<T>>,
    pub group: usize,
    pub mm: Arc<MatMul<T>>,
}

impl<T> ConvGemm<T>
where
    T: Datum + Add + Mul + Zero + Copy + AddAssign + ndarray::LinalgScalar,
{
    pub(super) fn conv_gemm<'i>(
        &'i self,
        packed_input: &'i ArrayView3<'i, T>,
    ) -> TractResult<ArrayD<T>> {
        let mut output = unsafe { ArrayD::<T>::uninitialized(&*self.full_output_shape) };
        let input_shape = &self.patch.input_shape;

        let co_per_group = self.full_output_shape[input_shape.c_axis()] / self.group;

        for i in 0..input_shape.n_dim() {
            unsafe {
                let output_i = output
                    .as_mut_ptr()
                    .offset(output.strides()[input_shape.n_axis()] * i as isize);
                for g in 0..self.group {
                    let a = &self.packed_kernels[g];
                    let output_i_g = output_i.offset(
                        output.strides()[input_shape.c_axis()] * co_per_group as isize * g as isize,
                    );

                    let (rsc, csc) = match self.patch.input_shape.fmt {
                        DataFormat::NHWC => (1, self.m as isize),
                        DataFormat::NCHW => (self.n as isize, 1),
                    };
                    self.mm.mat_mul_prepacked(
                        a.as_ptr()?,
                        packed_input
                            .as_ptr()
                            .offset(((self.group * i + g) * self.mm.packed_b_len()) as isize),
                        output_i_g,
                        rsc,
                        csc,
                    );
                }
            }
        }

        if let Some(ref bias) = self.bias {
            output += &bias;
        }

        Ok(output)
    }
}

impl<D> Op for ConvGemm<D>
where
    D: Datum + Clone + ::ndarray::LinalgScalar + ::std::ops::AddAssign<D> + PartialEq,
{
    fn name(&self) -> Cow<str> {
        "ConvGemm".into()
    }

    fn info(&self) -> TractResult<Option<String>> {
        Ok(Some(format!("{:?}", self.mm)))
    }
}

impl<D> StatelessOp for ConvGemm<D>
where
    D: Datum + Clone + ::ndarray::LinalgScalar + ::std::ops::AddAssign<D> + PartialEq,
{
    fn eval(&self, mut inputs: TVec<SharedTensor>) -> TractResult<TVec<SharedTensor>> {
        let input = args_1!(inputs);
        let output = self.conv_gemm(&input.to_array_view::<D>()?.into_dimensionality()?)?;
        Ok(tvec!(output.into()))
    }
}

impl<D> InferenceRulesOp for ConvGemm<D>
where
    D: Datum + Clone + ::ndarray::LinalgScalar + ::std::ops::AddAssign<D>,
{
    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        s: &mut Solver<'r>,
        inputs: &'p SharedTensorsProxy,
        outputs: &'p SharedTensorsProxy,
    ) -> InferenceResult {
        s.equals(&inputs.len, 1)?;
        s.equals(&outputs.len, 1)?;
        s.equals(&inputs[0].datum_type, D::datum_type())?;
        s.equals(&outputs[0].datum_type, D::datum_type())?;
        s.equals(&outputs[0].shape, ShapeFact::from(&*self.full_output_shape))?;
        Ok(())
    }
}