File: events.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 (29 lines) | stat: -rw-r--r-- 1,038 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
#pragma once

#include <array>
#include <cstdint>
#include <cstring>
#include <vector>

namespace torch::profiler {

/* A vector type to hold a list of performance counters */
using perf_counters_t = std::vector<uint64_t>;

/* Standard list of performance events independent of hardware or backend */
constexpr std::array<const char*, 2> ProfilerPerfEvents = {
    /*
     * Number of Processing Elelement (PE) cycles between two points of interest
     * in time. This should correlate positively with wall-time. Measured in
     * uint64_t. PE can be non cpu. TBD reporting behavior for multiple PEs
     * participating (i.e. threadpool).
     */
    "cycles",

    /* Number of PE instructions between two points of interest in time. This
     * should correlate positively with wall time and the amount of computation
     * (i.e. work). Across repeat executions, the number of instructions should
     * be more or less invariant. Measured in uint64_t. PE can be non cpu.
     */
    "instructions"};
} // namespace torch::profiler