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
use num_traits::{AsPrimitive, Zero};
use std::fmt;
use std::ops::{Add, Mul};
use crate::internal::*;
use ndarray::*;
use tract_linalg::mmm::{FusedSpec, MatMatMul};
use tract_linalg::frame::PackB;
#[derive(Debug, Clone)]
pub struct MatMatMulPackB<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
{
pub(crate) pack_b: PackB<T>,
pub(crate) row_stride: isize,
pub(crate) col_stride: isize,
pub(crate) output_shape: TVec<usize>,
}
impl<T> Op for MatMatMulPackB<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
{
fn name(&self) -> Cow<str> {
"MatMatMulPackB".into()
}
op_as_typed_op!();
not_a_pulsed_op!();
}
impl<T> StatelessOp for MatMatMulPackB<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
{
fn eval(&self, mut inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
let b = args_1!(inputs);
let mut packed = unsafe {
Tensor::uninitialized_aligned::<T>(&*self.output_shape, self.pack_b.alignment())
.unwrap()
};
if b.shape()[..b.shape().len() - 2].iter().any(|d| *d > 1) {
let b = b.to_array_view::<T>()?;
let b_prefix = &b.shape()[..b.shape().len() - 2];
for prefix in indices(b_prefix).into_iter() {
let mut b = b.view();
let mut p = packed.to_array_view_mut()?;
for &dim in prefix.slice() {
b.index_axis_inplace(Axis(0), dim);
p.index_axis_inplace(Axis(0), dim);
}
self.pack_b.pack(p.as_mut_ptr(), b.as_ptr(), self.row_stride, self.col_stride)
}
} else {
self.pack_b.pack(packed.as_ptr_mut()?, b.as_ptr()?, self.row_stride, self.col_stride)
}
Ok(tvec!(packed.into_arc_tensor()))
}
}
impl<T> TypedOp for MatMatMulPackB<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
{
fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
Ok(tvec!(TypedFact::dt_shape(inputs[0].datum_type, &*self.output_shape)?))
}
typed_op_as_op!();
}
#[derive(Debug, Clone)]
pub(crate) struct MatMatMulUnaryFinite<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
f32: ::num_traits::AsPrimitive<T>,
{
pub(crate) c_shape: TVec<usize>,
pub(crate) c_prefix_dim_and_stride: Option<(TVec<usize>, TVec<isize>)>,
pub(crate) packed_as: ArrayD<Tensor>,
pub(crate) mmm: Box<dyn MatMatMul<T>>,
pub(crate) non_linear: Vec<FusedSpec<T>>,
}
impl<T> Op for MatMatMulUnaryFinite<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike + fmt::Display,
f32: AsPrimitive<T>,
{
fn name(&self) -> Cow<str> {
"MatMatMul".into()
}
fn info(&self) -> TractResult<Vec<String>> {
let mut infos = vec![format!(
"c_prefix: {:?} m:{} k:{} n:{}",
self.c_prefix_dim_and_stride,
self.mmm.m(),
self.mmm.k(),
self.mmm.n(),
)];
infos.push(format!("{}", self.mmm));
if self.non_linear.len() > 0 {
infos.push(format!("{:?}", self.non_linear))
}
Ok(infos)
}
fn fuse(&self, model: &TypedModel, node: &TypedNode) -> TractResult<Option<TypedModelPatch>> {
use crate::ops;
if let Some(succ) = model.single_succ(node.id)? {
let fused_micro_op = (|| -> TractResult<Option<TVec<FusedSpec<T>>>> {
if let Some(op) = succ.op_as::<ops::binary::UnaryOp>() {
if op.a.shape() == &[self.mmm.m()] {
if op.mini_op.is::<ops::math::Mul>() {
return Ok(Some(tvec!(FusedSpec::PerRowMul(
op.a.as_slice::<T>()?.to_vec(),
))));
} else if op.mini_op.is::<ops::math::Add>() {
return Ok(Some(tvec!(FusedSpec::PerRowAdd(
op.a.as_slice::<T>()?.to_vec(),
))));
}
}
} else if let Some(op) = succ.op_as::<ops::element_wise::ElementWiseOp>() {
if let Some(op) = op.0.downcast_ref::<ops::math::ScalarMax>() {
return Ok(Some(tvec!(FusedSpec::Max(op.max.as_()))));
} else if let Some(op) = op.0.downcast_ref::<ops::math::ScalarMin>() {
return Ok(Some(tvec!(FusedSpec::Min(op.min.as_()))));
} else if let Some(op) = op.0.downcast_ref::<ops::math::ScalarMinMax>() {
return Ok(Some(tvec!(
FusedSpec::Min(op.min.as_()),
FusedSpec::Max(op.max.as_()),
)));
}
}
Ok(None)
})()?;
if let Some(op) = fused_micro_op {
let mut ops = self.non_linear.clone();
ops.extend(op.into_iter());
return Ok(Some(TypedModelPatch::fuse_with_next(
model,
&node,
Self { non_linear: ops, ..self.clone() },
)?));
}
}
Ok(None)
}
op_as_typed_op!();
not_a_pulsed_op!();
}
impl<T> StatelessOp for MatMatMulUnaryFinite<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
f32: ::num_traits::AsPrimitive<T>,
{
fn eval(&self, mut inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
unsafe {
let b = args_1!(inputs);
let mut c = Tensor::uninitialized::<T>(&*self.c_shape)?;
if let Some((prefix_dim, prefix_strides)) = &self.c_prefix_dim_and_stride {
let b = b.to_array_view::<T>()?;
let mut c = c.to_array_view_mut::<T>()?;
for prefix in indices(&**prefix_dim).into_iter() {
let mut a = self.packed_as.view();
let mut b = b.view();
let mut c: *mut T = c.as_mut_ptr();
for (ix, &dim) in prefix.slice().iter().enumerate() {
let d = dim.min(a.shape()[0] - 1);
a.index_axis_inplace(Axis(0), d);
let d = dim.min(b.shape()[0] - 1);
b.index_axis_inplace(Axis(0), d);
c = c.offset(prefix_strides[ix] * dim as isize);
}
let pa: &Tensor = a.iter().next().unwrap();
self.mmm.run(pa.as_ptr()?, b.as_ptr(), c, &self.non_linear);
}
} else {
self.mmm.run(
self.packed_as.as_slice().unwrap()[0].as_ptr()?,
b.as_ptr()?,
c.as_ptr_mut()?,
&self.non_linear,
);
}
Ok(tvec!(c.into_arc_tensor()))
}
}
}
impl<T> TypedOp for MatMatMulUnaryFinite<T>
where
T: Copy + Datum + Add + Mul + Zero + FloatLike,
f32: ::num_traits::AsPrimitive<T>,
{
fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
Ok(tvec!(TypedFact::dt_shape(inputs[0].datum_type, &*self.c_shape)?))
}
typed_op_as_op!();
}