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
use num_traits::Zero;
use std::fmt::Debug;
use super::{FusedKerSpec, FusedSpec, MatMatMulKer, MatrixStore, Tile};
use downcast_rs::{impl_downcast, Downcast};
use tract_data::prelude::*;
use std::alloc::Layout;
pub trait ScratchSpace: Downcast + Send {}
impl_downcast!(ScratchSpace);
pub struct ScratchSpaceFusedNonLinear<TI: Copy> {
uspecs: TVec<FusedKerSpec<TI>>,
buffers: TVec<(bool, Layout, *mut u8)>,
}
impl<TI: Copy> Default for ScratchSpaceFusedNonLinear<TI> {
fn default() -> ScratchSpaceFusedNonLinear<TI> {
ScratchSpaceFusedNonLinear { uspecs: tvec![], buffers: tvec![] }
}
}
impl<TI: Copy + 'static> ScratchSpace for ScratchSpaceFusedNonLinear<TI> {}
unsafe impl<TI: Copy + 'static> Send for ScratchSpaceFusedNonLinear<TI> {}
impl<TI: Copy> Drop for ScratchSpaceFusedNonLinear<TI> {
fn drop(&mut self) {
unsafe { self.buffers.drain(..).for_each(|(_, lo, buf)| std::alloc::dealloc(buf, lo)) }
}
}
impl<TI: Copy> ScratchSpaceFusedNonLinear<TI> {
#[inline]
pub fn clear(&mut self) {
self.buffers.iter_mut().for_each(|(used, _, _)| *used = false);
}
fn get_raw_buffer(&mut self, bytes: usize) -> *mut u8 {
if let Some(buf) =
self.buffers.iter_mut().find(|(used, layout, _)| !used && layout.size() == bytes)
{
buf.0 = true;
buf.2
} else {
let layout = Layout::from_size_align(bytes, 4).unwrap();
let buf = unsafe { std::alloc::alloc(layout) };
self.buffers.push((true, layout, buf));
buf
}
}
fn get_temp_slice<'a, T: Datum>(&mut self, len: usize) -> &'a mut [T] {
let buf = self.get_raw_buffer(std::mem::size_of::<T>() * len) as *mut T;
unsafe { std::slice::from_raw_parts_mut(buf, len) }
}
#[inline]
pub unsafe fn for_tile<K: MatMatMulKer<TI>>(
&mut self,
specs: &[FusedSpec],
down: usize,
right: usize,
c_store: &MatrixStore,
) -> *const FusedKerSpec<TI>
where
TI: Datum + Copy + Debug + Zero,
{
if specs.is_empty() {
std::ptr::null()
} else {
self.fused_for_tile::<K>(specs, down, right, c_store)
}
}
unsafe fn fused_for_tile<K: MatMatMulKer<TI>>(
&mut self,
specs: &[FusedSpec],
down: usize,
right: usize,
c_store: &MatrixStore,
) -> *const FusedKerSpec<TI>
where
TI: Datum + Copy + Debug + Zero,
{
self.uspecs.clear();
for spec in specs {
let s = match spec {
FusedSpec::Min(m) => FusedKerSpec::Min(*m.to_scalar_unchecked()),
FusedSpec::Max(m) => FusedKerSpec::Max(*m.to_scalar_unchecked()),
FusedSpec::PerRowAdd(v)
| FusedSpec::PerRowMul(v)
| FusedSpec::PerColMul(v)
| FusedSpec::PerColAdd(v) => {
let (dir, r) =
if matches!(spec, FusedSpec::PerColAdd(_) | FusedSpec::PerColMul(_)) {
(right, K::nr())
} else {
(down, K::mr())
};
let have = v.len().saturating_sub(dir * r).min(r);
let ptr = if have < K::mr() {
let buf = self.get_temp_slice(r);
buf[have..].iter_mut().for_each(|x| *x = TI::zero());
if have > 0 {
buf[..have].copy_from_slice(&v.as_slice_unchecked()[dir * r..][..have]);
}
buf.as_ptr()
} else {
v.as_ptr_unchecked::<TI>().add(dir * r)
};
match spec {
FusedSpec::PerRowAdd(_) => FusedKerSpec::PerRowAdd(ptr),
FusedSpec::PerRowMul(_) => FusedKerSpec::PerRowMul(ptr),
FusedSpec::PerColAdd(_) => FusedKerSpec::PerColAdd(ptr),
FusedSpec::PerColMul(_) => FusedKerSpec::PerColMul(ptr),
_ => unreachable!(),
}
}
FusedSpec::AddRowColProducts(rows, cols) => {
let have = rows.len() - down * K::mr();
let row_ptr = if have < K::mr() {
let buf = self.get_temp_slice(K::mr());
buf[..have]
.copy_from_slice(&rows.as_slice_unchecked()[down * K::mr()..][..have]);
buf.as_ptr()
} else {
rows.as_ptr_unchecked::<TI>().add(down * K::mr())
};
let have = cols.len() - right * K::nr();
let col_ptr = if have < K::nr() {
let buf = self.get_temp_slice(K::nr());
buf[..have]
.copy_from_slice(&cols.as_slice_unchecked()[right * K::nr()..][..have]);
buf.as_ptr()
} else {
cols.as_ptr_unchecked::<TI>().add(right * K::nr())
};
FusedKerSpec::AddRowColProducts(row_ptr, col_ptr)
}
FusedSpec::ScalarMul(t) => FusedKerSpec::ScalarMul(*t.to_scalar_unchecked()),
FusedSpec::ScalarAdd(t) => FusedKerSpec::ScalarAdd(*t.to_scalar_unchecked()),
FusedSpec::AddUnicast(tensor) => {
let (rsc, csc, item_count) = match c_store {
MatrixStore::Strides {
row_item_stride,
col_item_stride,
item_count,
..
} => (*row_item_stride, *col_item_stride, *item_count),
_ => panic!(),
};
let tile_offset = rsc * down as isize * K::mr() as isize
+ csc * right as isize * K::nr() as isize;
let tile_ptr: *const TI = tensor.as_ptr_unchecked::<TI>().offset(tile_offset);
let max_inner_offset =
(K::mr() - 1) as isize * rsc + (K::nr() - 1) as isize * csc;
if max_inner_offset + tile_offset > item_count as isize {
let tmp_d_tile = self.get_temp_slice::<TI>(K::mr() * K::nr());
for r in 0..K::mr() as isize {
for c in 0..K::nr() as isize {
let inner_offset = c * csc + r * rsc;
if inner_offset + tile_offset < item_count as isize {
tmp_d_tile[r as usize + c as usize * K::mr()] =
*tile_ptr.offset(inner_offset);
}
}
}
FusedKerSpec::AddUnicast(Tile {
ptr: tmp_d_tile.as_ptr() as _,
row_byte_stride: std::mem::size_of::<TI>() as isize,
col_byte_stride: (std::mem::size_of::<TI>() * K::mr()) as isize,
item_size: std::mem::size_of::<TI>(),
})
} else {
FusedKerSpec::AddUnicast(Tile {
ptr: tile_ptr as _,
row_byte_stride: (rsc as usize * std::mem::size_of::<TI>()) as isize,
col_byte_stride: (csc as usize * std::mem::size_of::<TI>()) as isize,
item_size: std::mem::size_of::<TI>(),
})
}
}
FusedSpec::QScale(s, rp, m) => FusedKerSpec::QScale(*s, *rp, *m),
};
self.uspecs.push(s);
}
self.uspecs.push(FusedKerSpec::Done);
self.uspecs.as_ptr()
}
#[inline]
pub unsafe fn tmp_tile_c(&mut self, item_size: usize, mr: usize, nr: usize) -> Tile {
let ptr = self.get_raw_buffer(mr * nr * item_size);
Tile {
ptr: ptr as _,
item_size,
row_byte_stride: item_size as isize,
col_byte_stride: (item_size * mr) as isize,
}
}
}