File: vulkan.cpp

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 (41 lines) | stat: -rw-r--r-- 1,377 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
36
37
38
39
40
41
#include <torch/csrc/profiler/orchestration/vulkan.h>

#include <utility>

namespace torch::profiler::impl::vulkan {
namespace {

GetShaderNameAndDurationNsFn get_shader_name_and_duration_ns_fn;

} // namespace

void registerGetShaderNameAndDurationNs(
    GetShaderNameAndDurationNsFn get_shader_name_and_duration_ns) {
  get_shader_name_and_duration_ns_fn =
      std::move(get_shader_name_and_duration_ns);
}

void deregisterGetShaderNameAndDurationNs() {
  get_shader_name_and_duration_ns_fn = nullptr;
}

std::tuple<std::string, uint64_t> getShaderNameAndDurationNs(
    const vulkan_id_t& vulkan_id) {
  /*
    We don't need to worry about a race condition with
    deregisterGetShaderNameAndDurationNs here currently because
    deregisterGetShaderNameAndDurationNs is only called within the destructor
    of QueryPool, which would only be called after we're done calling
    getShaderNameAndDurationNs
  */
  TORCH_CHECK(
      get_shader_name_and_duration_ns_fn != nullptr,
      "Attempting to get shader duration in ",
      "torch::profiler::impl::vulkan::getShaderNameAndDurationNs, but "
      "get_shader_duration_fn is unregistered. Use "
      "torch::profiler::impl::vulkan::registerGetShaderNameAndDurationNs to register "
      "it first");
  return get_shader_name_and_duration_ns_fn(vulkan_id.value_of());
}

} // namespace torch::profiler::impl::vulkan