File: kernel_cache.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 (33 lines) | stat: -rw-r--r-- 994 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
#pragma once

#include <torch/csrc/Export.h>
#include <torch/csrc/jit/codegen/fuser/kernel_spec.h>
#include <torch/csrc/jit/ir/ir.h>

#include <cstdint>
#include <functional>
#include <optional>

namespace torch::jit::fuser {

// A thread-safe cache interface.

// Normalizes the graph by canonicalizing and erasing shape information
TORCH_API std::shared_ptr<Graph> normalizeGraphForCache(
    const std::shared_ptr<Graph>& graph);

// Stores the given graph, returning the key used to access it
TORCH_API int64_t store(std::shared_ptr<Graph> graph);

// Given a graph, find a KernelSpec based on it
TORCH_API std::optional<KernelSpec*> lookupGraph(
    const std::shared_ptr<Graph>& graph);

// Returns the graph corresponding to the given key (if it exists)
TORCH_API std::optional<KernelSpec*> retrieve(const int64_t key);

// Returns the size of the fusion key -> KernelSpec cache.
// Only used for testing.
TORCH_API int64_t debugNumCachedKernelSpecs();

} // namespace torch::jit::fuser