File: graph_fuser.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 (35 lines) | stat: -rw-r--r-- 1,226 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
35
#pragma once

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

namespace torch::jit {

TORCH_API bool canFuseOnCPULegacy();
TORCH_API void overrideCanFuseOnCPULegacy(bool value);

// NB: Be sure to run DCE before fusion, because dead instructions
// can prevent fusion opportunities from being exploited.
// On Windows will noop, NYI
TORCH_API void FuseGraph(
    std::shared_ptr<Graph>& graph,
    bool strict_fuser_check = false);

// \brief Custom fusion pass using a node-level callback to
// determine the inclusion of nodes in a subgraph.
//
// This helper omits aliased inputs and fusion across control flow
// boundaries.
//
// \arg graph The graph to be modified in-place
// \arg is_fusable A callback run on each fusable node in the graph.
// \arg kind The label given to the resultant fused subgraph
// \arg arg_limit The maximum number of args the resultant fused subgraph
//                should have.  Note: This will likely develop into a general
//                post condition on the fused subgraph.
TORCH_API void CustomFuseGraph(
    std::shared_ptr<Graph>& graph,
    const std::function<bool(Node*)>& is_fusable,
    Symbol kind,
    size_t arg_limit = std::numeric_limits<size_t>::max());

} // namespace torch::jit