File: loop_unrolling.h

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (34 lines) | stat: -rw-r--r-- 981 bytes parent folder | download | duplicates (3)
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
#pragma once

#include <torch/csrc/jit/ir/ir.h>

namespace torch::jit {

// return true if graph is modified
TORCH_API bool UnrollLoops(std::shared_ptr<Graph>& graph);

// Only unrolls constant loops. Will unroll them regardless of loop block size
TORCH_API bool UnrollConstantLoops(std::shared_ptr<Graph>& graph);

TORCH_API Node* PeelLoop(Node* n, size_t times);

// return true if graph is modified
TORCH_API bool PeelProfilingLoops(const std::shared_ptr<Graph>& graph);

struct TORCH_API LoopsPeeler {
  LoopsPeeler(std::function<bool(Node* n)> callback, size_t num_iterations = 1)
      : callback_(std::move(callback)), num_iterations_(num_iterations) {}

  bool run(const std::shared_ptr<Graph>& graph);

 private:
  void collectLoop(Node* n);
  void collectLoops(Block* block);
  void peelLoops();

  std::function<bool(Node* n)> callback_ = nullptr;
  Node* in_loop_ = nullptr;
  std::list<Node*> loops_to_peel_;
  size_t num_iterations_ = 1;
};
} // namespace torch::jit