1#![allow(clippy::missing_safety_doc)]
2#![allow(clippy::redundant_closure_call)]
3#![allow(clippy::len_zero)]
4#![allow(clippy::excessive_precision)]
5#![allow(clippy::approx_constant)]
6#![allow(clippy::manual_is_multiple_of)]
7#![allow(unexpected_cfgs)]
8#![allow(unused_macros)]
9#[macro_use]
10extern crate derive_new;
11extern crate lazy_static;
12extern crate log;
13extern crate num_traits;
14#[macro_use]
15extern crate pastey;
16#[cfg(test)]
17extern crate proptest;
18
19include!(concat!(env!("OUT_DIR"), "/extern_kernel_macro.rs"));
20
21#[macro_use]
22mod frame;
23pub mod generic;
24pub mod multithread;
25pub use frame::weights::WeightType;
26pub use generic::{ScaleShiftAndRound, Scaler};
27use lazy_static::lazy_static;
28use mmm::{MMMInputFormat, MatMatMul, PanelExtractor};
29use tract_data::internal::TensorView;
30#[cfg(target_arch = "x86_64")]
31pub mod x86_64_fma;
32
33pub mod hwbench;
34
35#[cfg(target_arch = "aarch64")]
36pub mod arm64;
37
38#[cfg(target_arch = "aarch64")]
39pub use arm64::has_fp16;
40use tract_itertools::Itertools;
41
42#[cfg(not(target_arch = "aarch64"))]
43pub fn has_fp16() -> bool {
44 false
45}
46
47#[cfg(any(target_arch = "arm", target_arch = "armv7", target_arch = "arm"))]
48pub mod arm32;
49
50#[cfg(all(target_family = "wasm", target_feature = "simd128"))]
51pub mod wasm;
52
53pub use self::frame::*;
54
55use tract_data::prelude::*;
56
57pub type MMMImpl = Box<
58 dyn Fn(Option<usize>, Option<usize>, Option<usize>) -> Box<dyn mmm::MatMatMul> + Send + Sync,
59>;
60
61type MMVImpl = Box<dyn Fn(Option<usize>, Option<usize>) -> Box<dyn mmm::MatMatMul> + Send + Sync>;
62
63#[allow(clippy::type_complexity)]
64pub struct Ops {
65 mmm_impls: Vec<Box<dyn mmm::MatMatMul>>,
66 panel_extractors: Vec<mmm::PanelExtractor>,
67
68 mmm_f64: MMMImpl,
69 mmv_f64: MMVImpl,
70
71 mmm_f32: MMMImpl,
72 mmv_f32: MMVImpl,
73
74 mmm_f16: MMMImpl,
75 mmv_f16: MMVImpl,
76
77 qmmm_i32: MMMImpl,
78 qmmv_i32: MMVImpl,
79
80 pub leaky_relu_f16: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f16, f16>> + Send + Sync>,
81 pub leaky_relu_f32: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f32, f32>> + Send + Sync>,
82 pub mul_by_scalar_f32:
83 Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f32, f32>> + Send + Sync>,
84 pub mul_by_scalar_f16:
85 Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f16, f16>> + Send + Sync>,
86
87 pub sigmoid_f16: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f16>> + Send + Sync>,
88 pub sigmoid_f32: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f32>> + Send + Sync>,
89 pub tanh_f16: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f16>> + Send + Sync>,
90 pub tanh_f32: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f32>> + Send + Sync>,
91 pub erf_f32: Box<dyn Fn() -> Box<dyn element_wise::ElementWise<f32>> + Send + Sync>,
92 pub lut_u8: Box<dyn Fn(&[u8]) -> Box<dyn lut::Lut> + Send + Sync>,
93
94 pub max_f16: Box<dyn Fn() -> Box<dyn reduce::Reduce<f16>> + Send + Sync>,
95 pub max_f32: Box<dyn Fn() -> Box<dyn reduce::Reduce<f32>> + Send + Sync>,
96
97 pub sum_f16: Box<dyn Fn() -> Box<dyn reduce::Reduce<f16>> + Send + Sync>,
98 pub sum_f32: Box<dyn Fn() -> Box<dyn reduce::Reduce<f32>> + Send + Sync>,
99
100 pub softmax2_fastcompact_f16:
101 Box<dyn Fn() -> Box<dyn reduce::MapReduce<f16, f16>> + Send + Sync>,
102 pub softmax2_fastcompact_f32:
103 Box<dyn Fn() -> Box<dyn reduce::MapReduce<f32, f32>> + Send + Sync>,
104}
105
106impl Ops {
107 pub fn mmm_impls(&self) -> &[Box<dyn mmm::MatMatMul>] {
108 &self.mmm_impls
109 }
110
111 pub fn all_possible_packing(
112 &self,
113 weight_type: impl Into<WeightType>,
114 ) -> impl Iterator<Item = &dyn MMMInputFormat> {
115 let weight_type = weight_type.into();
116 self.mmm_impls
117 .iter()
118 .flat_map(|m| m.packings())
119 .map(|p| &*p.0)
120 .flat_map(move |p| {
121 let mut packs: Vec<&dyn MMMInputFormat> = vec![];
122 if p.precursor() == weight_type {
123 packs.push(p)
124 };
125 for pe in &self.panel_extractors {
126 if pe.from.precursor() == weight_type && pe.to.same_as(p) {
127 packs.push(&*pe.from);
128 }
129 }
130 packs.into_iter()
131 })
132 .sorted_by_key(|p| p.to_string())
133 .dedup()
134 }
135
136 pub fn filter_impls<'o>(
137 &'o self,
138 weight: &'o dyn MMMInputFormat,
139 acc: &[DatumType],
140 act: DatumType,
141 store: DatumType,
142 ) -> impl Iterator<
143 Item = (
144 &'o dyn MatMatMul,
145 usize,
146 &'o dyn MMMInputFormat,
147 Option<&'o PanelExtractor>,
148 &'o dyn MMMInputFormat,
149 ),
150 > {
151 let acc = acc.to_vec();
152 self.mmm_impls
153 .iter()
154 .filter(move |mmm| acc.contains(&mmm.internal_type()) && mmm.stores().contains(&store))
155 .flat_map(|mmm| {
156 mmm.packings()
157 .iter()
158 .enumerate()
159 .map(|(pack_ix, (a, b))| (&**mmm, pack_ix, &**a, &**b))
160 })
161 .filter_map(|(mmm, ix, a, b)| {
162 if a.same_as(weight) {
163 Some((mmm, ix, a, None, b))
164 } else {
165 self.panel_extractors
166 .iter()
167 .find(|pe| pe.from.same_as(weight) && pe.to.same_as(a))
168 .map(|pe| (mmm, ix, a, Some(pe), b))
169 }
170 })
171 .filter(move |(_mmm, _ix, _a, _pe, b)| {
172 b.precursor().as_dt().is_some_and(|dt| dt == act)
173 })
174 }
175
176 pub fn panel_extractors(&self) -> &[mmm::panel_extract::PanelExtractor] {
177 &self.panel_extractors
178 }
179
180 pub fn mmm(
181 &self,
182 accumulator: DatumType,
183 m: Option<usize>,
184 k: Option<usize>,
185 n: Option<usize>,
186 ) -> Option<Box<dyn mmm::MatMatMul>> {
187 use DatumType::*;
188 match accumulator {
189 F64 => Some(if n == Some(1) { (self.mmv_f64)(m, k) } else { (self.mmm_f64)(m, k, n) }),
190 F32 => Some(if n == Some(1) { (self.mmv_f32)(m, k) } else { (self.mmm_f32)(m, k, n) }),
191 F16 => Some(if n == Some(1) { (self.mmv_f16)(m, k) } else { (self.mmm_f16)(m, k, n) }),
192 I32 => {
193 Some(if n == Some(1) { (self.qmmv_i32)(m, k) } else { (self.qmmm_i32)(m, k, n) })
194 }
195 _ => None,
196 }
197 }
198}
199
200pub fn generic() -> Ops {
201 use crate::generic::mmm::*;
202 use element_wise::ElementWiseKer;
203 use reduce::{MapReduceKer, ReduceKer};
204 let mut ops = Ops {
205 mmm_impls: vec![],
206 panel_extractors: vec![],
207 mmm_f64: Box::new(|_, _, _| generic_f64_4x4.mmm()),
208 mmv_f64: Box::new(|_, _| generic_f64_4x1.mmm()),
209 mmm_f32: Box::new(|_, _, _| generic_f32_4x4.mmm()),
210 mmv_f32: Box::new(|_, _| generic_f32_4x1.mmm()),
211 mmm_f16: Box::new(|_, _, _| generic_f16_4x4.mmm()),
212 mmv_f16: Box::new(|_, _| generic_f16_4x1.mmm()),
213 qmmm_i32: Box::new(|_, _, _| generic_i32_4x4.mmm()),
214 qmmv_i32: Box::new(|_, _| generic_i32_4x4.mmm()),
215 leaky_relu_f16: Box::new(|| generic::HLeakyRelu8::ew()),
216 leaky_relu_f32: Box::new(|| generic::SLeakyRelu4::ew()),
217 mul_by_scalar_f16: Box::new(|| generic::HMulByScalar8::ew()),
218 mul_by_scalar_f32: Box::new(|| generic::SMulByScalar4::ew()),
219 sigmoid_f16: Box::new(|| generic::HSigmoid8::ew()),
220 sigmoid_f32: Box::new(|| generic::SSigmoid4::ew()),
221 tanh_f16: Box::new(|| generic::HTanh8::ew()),
222 tanh_f32: Box::new(|| generic::STanh4::ew()),
223 erf_f32: Box::new(|| generic::SErf4::ew()),
224 lut_u8: Box::new(|table: &[u8]| Box::new(lut::LutImpl::<generic::GenericLut8>::new(table))),
225 max_f16: Box::new(|| generic::reduce::max::HMax8::red()),
226 max_f32: Box::new(|| generic::reduce::max::SMax4::red()),
227 sum_f16: Box::new(|| generic::reduce::sum::HSum8::red()),
228 sum_f32: Box::new(|| generic::reduce::sum::SSum4::red()),
229 softmax2_fastcompact_f16: Box::new(|| generic::reduce::softmax_l2::HSoftMaxL2::red()),
233 softmax2_fastcompact_f32: Box::new(|| generic::reduce::softmax_l2::SSoftMaxL2::red()),
234 };
235 crate::generic::mmm::plug(&mut ops);
236 ops
237}
238
239#[allow(unreachable_code, unused_mut, unexpected_cfgs)]
240pub fn best() -> Ops {
241 let mut ops = generic();
242 #[cfg(target_arch = "x86_64")]
243 x86_64_fma::plug(&mut ops);
244 #[cfg(any(target_arch = "arm", target_arch = "armv7"))]
245 arm32::plug(&mut ops);
246 #[cfg(target_arch = "aarch64")]
247 arm64::plug(&mut ops);
248 #[cfg(all(target_family = "wasm", target_feature = "simd128"))]
249 wasm::plug(&mut ops);
250
251 ops
252}
253
254lazy_static::lazy_static! {
255 static ref OPS: Ops = {
256 best()
257 };
258}
259
260#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
261pub enum BinOp {
262 Min,
263 Max,
264 Add,
265 Mul,
266 Sub,
267 SubF,
268}
269
270impl BinOp {
271 pub fn flip(&self) -> BinOp {
272 use BinOp::*;
273 match self {
274 Sub => SubF,
275 SubF => Sub,
276 sym => *sym,
277 }
278 }
279}
280
281fn register_all_unicast(registry: &mut LinalgRegistry) {
282 generic::register_all_unicast(registry);
283 #[cfg(target_arch = "aarch64")]
284 arm64::register_all_unicast(registry);
285}
286
287fn register_all_by_scalar(registry: &mut LinalgRegistry) {
288 generic::register_all_by_scalar(registry);
289 #[cfg(target_arch = "aarch64")]
290 arm64::register_all_by_scalar(registry);
291}
292
293pub type LinalgFn = dyn Fn(&mut TensorView, &TensorView) -> TractResult<()> + Send + Sync;
294type LinalgRegistry = HashMap<(BinOp, DatumType), Box<dyn Fn() -> Box<LinalgFn> + Send + Sync>>;
295lazy_static! {
296 static ref BIN_UNICAST_OPS: Mutex<LinalgRegistry> = {
297 let mut registry = HashMap::default();
298 register_all_unicast(&mut registry);
299 Mutex::new(registry)
300 };
301 static ref BIN_BY_SCALAR_OPS: Mutex<LinalgRegistry> = {
302 let mut registry = HashMap::default();
303 register_all_by_scalar(&mut registry);
304 Mutex::new(registry)
305 };
306}
307
308pub fn bin_by_scalar(dt: DatumType, bin: BinOp) -> Option<Box<LinalgFn>> {
309 let map = BIN_BY_SCALAR_OPS.lock().unwrap();
310 if (dt == DatumType::F16) && !has_fp16() {
311 return None;
312 }
313 map.get(&(bin, dt)).map(|it| (it)())
314}
315
316pub fn bin_unicast(dt: DatumType, bin: BinOp) -> Option<Box<LinalgFn>> {
317 let map = BIN_UNICAST_OPS.lock().unwrap();
318 if (dt == DatumType::F16) && !has_fp16() {
319 return None;
320 }
321 map.get(&(bin, dt)).map(|it| (it)())
322}
323
324pub fn ops() -> &'static Ops {
325 &OPS
326}
327
328use num_traits::*;
329use std::collections::HashMap;
330use std::fmt::Debug;
331use std::ops::*;
332use std::sync::Mutex;
333
334pub trait LADatum:
335 Sized
336 + std::fmt::Display
337 + Debug
338 + Copy
339 + Clone
340 + Zero
341 + One
342 + 'static
343 + Add<Output = Self>
344 + Sub<Output = Self>
345 + Mul
346 + AddAssign
347 + PartialOrd
348 + Bounded
349 + tract_data::prelude::Datum
350{
351 #[cfg(test)]
352 fn strat() -> proptest::prelude::BoxedStrategy<Self>;
353}
354
355#[cfg(test)]
356use proptest::prelude::*;
357
358impl LADatum for f16 {
359 #[cfg(test)]
360 fn strat() -> BoxedStrategy<Self> {
361 f32::strat().prop_map(|f| f.as_()).boxed()
362 }
363}
364
365impl LADatum for f32 {
366 #[cfg(test)]
367 fn strat() -> BoxedStrategy<Self> {
368 (-1000isize..1000).prop_map(|i| i as f32 / 1000.0).boxed()
369 }
370}
371
372impl LADatum for f64 {
373 #[cfg(test)]
374 fn strat() -> BoxedStrategy<Self> {
375 (-1000isize..1000).prop_map(|i| i as f64 / 1000.0).boxed()
376 }
377}
378
379impl LADatum for u8 {
380 #[cfg(test)]
381 fn strat() -> BoxedStrategy<Self> {
382 any::<u8>().boxed()
383 }
384}
385
386impl LADatum for i8 {
387 #[cfg(test)]
388 fn strat() -> BoxedStrategy<Self> {
389 any::<i8>().boxed()
390 }
391}
392
393impl LADatum for i32 {
394 #[cfg(test)]
395 fn strat() -> BoxedStrategy<Self> {
396 any::<i32>().boxed()
397 }
398}
399
400#[cfg(test)]
401#[allow(dead_code)]
402fn setup_test_logger() {
403 let _ = env_logger::Builder::from_env("TRACT_LOG").try_init();
404}