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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565
use super::Compiler;
use crate::{
ir::{
Binding, CubeDim, Elem, Item, KernelDefinition, Location, ReadingStrategy, Scope, Variable,
Vectorization, Visibility,
},
Runtime,
};
/// The kernel integrator allows you to create a [kernel definition](KernelDefinition) based on
/// [kernel expansion](KernelExpansion) and [kernel settings](KernelSettings).
#[derive(Clone)]
pub struct KernelIntegrator {
expansion: KernelExpansion,
input_bindings: Vec<Binding>,
output_bindings: Vec<Binding>,
named_bindings: Vec<(String, Binding)>,
}
/// The information necessary to compile a [kernel definition](KernelDefinition).
#[derive(Clone)]
pub struct KernelExpansion {
pub inputs: Vec<InputInfo>,
pub outputs: Vec<OutputInfo>,
pub scope: Scope,
}
/// Simply indicate the output that can be replaced by the input.
#[derive(new, Default, Clone, Debug, Hash, PartialEq, Eq)]
pub struct InplaceMapping {
/// Input position.
pub pos_input: usize,
/// Output position.
pub pos_output: usize,
}
#[derive(Clone, Debug, Hash, PartialEq, Eq)]
enum VectorizationPartial {
Input {
pos: usize,
vectorization: Vectorization,
},
Output {
pos: usize,
vectorization: Vectorization,
},
}
#[derive(Default, Clone, Debug, Hash, PartialEq, Eq)]
pub struct KernelSettings {
pub mappings: Vec<InplaceMapping>,
vectorization_global: Option<Vectorization>,
vectorization_partial: Vec<VectorizationPartial>,
cube_dim: CubeDim,
pub reading_strategy: Vec<(u16, ReadingStrategy)>,
}
impl core::fmt::Display for KernelSettings {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
// The goal of this implementation is to generate the shortest representation
// that won't clash with any other compilation settings. This is crucial since we rely on
// this representation to know when to compile a new version of a kernel.
//
// Each main section starts with a letter that can't be used by other main sections:
//
// * Mapping: m
// * Input: i
// * Output: o
//
// * Reading Strategy: r
// * Output layout: o
// * Plain: p
//
// * Vectorization Global: vg{factor}
// * Vectorization Partial Input: v{factor}i{pos}
// * Vectorization Partial Output: vo
// * Cube Dim X: x
// * Cube Dim Y: y
// * Cube Dim Z: z
f.write_str("m")?;
for mapping in self.mappings.iter() {
f.write_fmt(format_args!(
"i{}o{}",
mapping.pos_input, mapping.pos_output
))?;
}
f.write_str("r")?;
for (input, strategy) in self.reading_strategy.iter() {
match strategy {
ReadingStrategy::OutputLayout => f.write_fmt(format_args!("i{}o", input)),
ReadingStrategy::Plain => f.write_fmt(format_args!("i{}p", input)),
}?;
}
match self.vectorization_global {
Some(vectorization) => f.write_fmt(format_args!("vg{}", vectorization))?,
None => f.write_str("vn")?,
};
for vectorization in self.vectorization_partial.iter() {
match vectorization {
VectorizationPartial::Input { pos, vectorization } => {
f.write_fmt(format_args!("v{vectorization}i{pos}"))?
}
VectorizationPartial::Output { pos, vectorization } => {
f.write_fmt(format_args!("v{vectorization}o{pos}"))?
}
};
}
f.write_fmt(format_args!(
"x{}y{}z{}",
self.cube_dim.x, self.cube_dim.y, self.cube_dim.x
))
}
}
impl KernelSettings {
/// Compile the shader with vectorization enabled for all inputs and outputs.
#[allow(dead_code)]
pub fn vectorize_global(mut self, vectorization: Vectorization) -> Self {
self.vectorization_global = Some(vectorization);
self
}
/// Compile the shader with vectorization enabled for an input.
#[allow(dead_code)]
pub fn vectorize_input(mut self, position: usize, vectorization: Vectorization) -> Self {
// Not setting the vectorization factor when it's the default value reduces the kernel id
// size.
if vectorization == 1 {
return self;
}
self.vectorization_partial
.push(VectorizationPartial::Input {
pos: position,
vectorization,
});
self
}
/// Compile the shader with vectorization enabled for an output.
#[allow(dead_code)]
pub fn vectorize_output(mut self, position: usize, vectorization: Vectorization) -> Self {
// Not setting the vectorization factor when it's the default value reduces the kernel id
// size.
if vectorization == 1 {
return self;
}
self.vectorization_partial
.push(VectorizationPartial::Output {
pos: position,
vectorization,
});
self
}
/// Fetch the vectorization for the provided input position.
pub fn vectorization_input(&self, position: usize) -> Vectorization {
if let Some(vec) = self.vectorization_global {
return vec;
}
for partial in self.vectorization_partial.iter() {
if let VectorizationPartial::Input { pos, vectorization } = partial {
if *pos == position {
return *vectorization;
}
}
}
1
}
/// Fetch the vectorization for the provided output position.
pub fn vectorization_output(&self, position: usize) -> Vectorization {
if let Some(vec) = self.vectorization_global {
return vec;
}
for partial in self.vectorization_partial.iter() {
if let VectorizationPartial::Output { pos, vectorization } = partial {
if *pos == position {
return *vectorization;
}
}
}
1
}
/// Compile the shader with inplace enabled by the given [mapping](InplaceMapping).
///
/// Notes:
///
/// You should favor using `dynamic_settings` when using fusion, since the mapping is going to
/// be created from the runtime information.
pub fn inplace(mut self, mappings: Vec<InplaceMapping>) -> Self {
self.mappings = mappings;
self
}
/// Set cube dimension.
#[allow(dead_code)]
pub fn cube_dim(mut self, cube_dim: CubeDim) -> Self {
self.cube_dim = cube_dim;
self
}
}
#[allow(dead_code)]
fn is_contiguous(strides: &[usize]) -> bool {
let mut current = 0;
for stride in strides.iter().rev() {
if current > *stride {
return false;
}
current = *stride;
}
true
}
/// Information related to an input.
#[derive(Clone, Debug)]
pub enum InputInfo {
Array { item: Item, visibility: Visibility },
Scalar { elem: Elem, size: usize },
}
impl InputInfo {
/// The item type of the input.
#[allow(dead_code)]
pub fn item(&self) -> Item {
match self {
InputInfo::Array {
item,
visibility: _,
} => *item,
InputInfo::Scalar { elem, size: _ } => Item::new(*elem),
}
}
}
impl OutputInfo {
/// The item type of the input.
#[allow(dead_code)]
pub fn item(&self) -> Item {
match self {
OutputInfo::ArrayWrite {
item,
local: _,
position: _,
} => *item,
OutputInfo::InputArrayWrite {
item,
input: _,
local: _,
position: _,
} => *item,
OutputInfo::Array { item } => *item,
}
}
}
/// Information related to an output.
#[derive(Clone, Debug)]
pub enum OutputInfo {
/// Write the local variable to a new array.
///
/// This will create a new binding in the [kernel definition](KernelDefinition).
ArrayWrite {
item: Item,
local: u16,
position: Variable,
},
/// Write the local variable to an existing input binding.
InputArrayWrite {
item: Item,
input: u16,
local: u16,
position: Variable,
},
/// Simply register the output, but don't automatically add a write to it.
///
/// Useful when a procedure writes to the output using operations.
Array { item: Item },
}
impl OutputInfo {
#[allow(dead_code)]
pub fn elem_size<R: Runtime>(&self) -> usize {
let elem = match self {
OutputInfo::ArrayWrite {
item,
local: _,
position: _,
} => bool_elem(item.elem()),
OutputInfo::InputArrayWrite {
item,
input: _,
local: _,
position: _,
} => bool_elem(item.elem()),
OutputInfo::Array { item } => bool_elem(item.elem()),
};
<R::Compiler as Compiler>::elem_size(elem)
}
}
impl KernelIntegrator {
/// Starts a new compilation.
pub fn new(info: KernelExpansion) -> Self {
Self {
expansion: info,
input_bindings: Default::default(),
output_bindings: Default::default(),
named_bindings: Default::default(),
}
}
/// Performs the compilation with the provided [settings](KernelSettings).
pub fn integrate(mut self, mut settings: KernelSettings) -> KernelDefinition {
if let Some(vectorization) = settings.vectorization_global {
self.expansion.scope.vectorize(vectorization);
}
self.register_inputs(&settings);
self.register_outputs(&mut settings);
let inputs = self.input_bindings;
let outputs = self.output_bindings;
let mut named = Vec::with_capacity(2);
named.push((
"info".to_string(),
Binding {
item: Item::new(Elem::UInt),
visibility: Visibility::Read,
location: Location::Storage,
size: None, // We avoid putting the length here since it will force a new kernel
// for each tensor rank.
},
));
for (name, binding) in self.named_bindings.into_iter() {
named.push((name, binding));
}
KernelDefinition {
inputs,
outputs,
named,
cube_dim: settings.cube_dim,
body: self.expansion.scope,
}
}
fn register_inputs(&mut self, settings: &KernelSettings) {
for (id, strategy) in settings.reading_strategy.iter() {
self.expansion.scope.update_read(*id, *strategy);
}
for input in self.expansion.inputs.drain(..) {
match input {
InputInfo::Array { item, visibility } => {
let item = if let Some(vectorization) = settings.vectorization_global {
item.vectorize(vectorization)
} else {
item
};
self.input_bindings.push(Binding {
item: bool_item(item),
visibility,
location: Location::Storage,
size: None,
});
}
InputInfo::Scalar { elem, size } => {
let elem = bool_elem(elem);
self.named_bindings.push((
format!("scalars_{}", elem),
Binding {
item: Item::new(elem),
visibility: Visibility::Read,
location: Location::Storage,
size: Some(size),
},
));
}
}
}
}
fn register_outputs(&mut self, settings: &mut KernelSettings) {
let mut index = 0;
if !settings.mappings.is_empty() {
let mut mappings = Vec::new();
core::mem::swap(&mut settings.mappings, &mut mappings);
for mapping in mappings {
self.register_inplace_mapping(mapping);
}
}
for array in self.expansion.outputs.drain(..) {
match array {
OutputInfo::ArrayWrite {
item,
local,
position,
} => {
let item = if let Some(vectorization) = settings.vectorization_global {
item.vectorize(vectorization)
} else {
item
};
let item_adapted = bool_item(item);
self.output_bindings.push(Binding {
item: item_adapted,
visibility: Visibility::ReadWrite,
location: Location::Storage,
size: None,
});
self.expansion.scope.write_global(
Variable::Local {
id: local,
item,
depth: self.expansion.scope.depth,
},
Variable::GlobalOutputArray {
id: index,
item: item_adapted,
},
position,
);
index += 1;
}
OutputInfo::InputArrayWrite {
item,
input,
local,
position,
} => {
let item = if let Some(vectorization) = settings.vectorization_global {
item.vectorize(vectorization)
} else {
item
};
self.expansion.scope.write_global(
Variable::Local {
id: local,
item,
depth: self.expansion.scope.depth,
},
Variable::GlobalInputArray {
id: input,
item: bool_item(item),
},
position,
);
}
OutputInfo::Array { item } => {
let item = if let Some(vectorization) = settings.vectorization_global {
item.vectorize(vectorization)
} else {
item
};
let elem_adapted = bool_item(item);
self.output_bindings.push(Binding {
item: elem_adapted,
visibility: Visibility::ReadWrite,
location: Location::Storage,
size: None,
});
index += 1;
}
}
}
}
fn register_inplace_mapping(&mut self, mapping: InplaceMapping) {
let output = match self.expansion.outputs.get_mut(mapping.pos_output) {
Some(output) => output,
None => {
if let Some(binding) = self.input_bindings.get_mut(mapping.pos_input) {
// Update input visibility.
binding.visibility = Visibility::ReadWrite;
}
// The mapping is handled differently, normally by cube itself.
return;
}
};
let (item, local, position) = match output {
OutputInfo::ArrayWrite { item, local, position } => (item, local, position),
OutputInfo::InputArrayWrite {
item: _,
input,
local: _,
position: _,
} => {
assert_eq!(
*input, mapping.pos_input as u16,
"Can't use different inputs for the same output."
);
return;
}
OutputInfo::Array { item: _ } => panic!("Can't register an inplace operation for an array that isn't using a defined writing strategy."),
};
let item = match self.input_bindings.get_mut(mapping.pos_input) {
Some(binding) => {
// Update input visibility.
binding.visibility = Visibility::ReadWrite;
// Inputs modified inplace should be read without any specified layout.
self.expansion
.scope
.update_read(mapping.pos_input as u16, ReadingStrategy::Plain);
// Use the same item as the input.
//
// The output can be different (i.e inplace boolean operations on float bindings).
binding.item
}
None => *item,
};
// Update the output.
*output = OutputInfo::InputArrayWrite {
item,
input: mapping.pos_input as u16,
local: *local,
position: *position,
};
}
}
fn bool_item(ty: Item) -> Item {
Item {
elem: bool_elem(ty.elem),
vectorization: ty.vectorization,
}
}
pub fn bool_elem(elem: Elem) -> Elem {
match elem {
// U32 are used for bool tensors
Elem::Bool => Elem::UInt,
_ => elem,
}
}