mirror of
https://github.com/saymrwulf/pytorch.git
synced 2026-05-15 21:00:47 +00:00
0b1f3bd158
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/73855 Calling the clock is one of the most expensive parts of profiling. We can reduce the profiling overhead by using `rdtsc` instead. The tradeoff is that we have to measure and convert. (shift and scale) Test Plan: I added a cpp unit test with *very* aggressive anti-flake measures. I also ran the overhead benchmark (9 replicates) with `--stressTestKineto` (0.94 -> 0.89 us) and `--stressTestKineto --kinetoProfileMemory` (1.27 -> 1.17 us) Reviewed By: chaekit Differential Revision: D34231071 fbshipit-source-id: e3b3dd7580d93bcc783e87c7f2fc726cb74f4df8 (cherry picked from commit e8be9f8160793c6ee35d5af02bca3e01703e377d)
76 lines
2.5 KiB
C++
76 lines
2.5 KiB
C++
#include <algorithm>
|
|
#include <cmath>
|
|
#include <utility>
|
|
#include <vector>
|
|
|
|
#include <gtest/gtest.h>
|
|
|
|
#include <c10/util/irange.h>
|
|
#include <torch/csrc/profiler/containers.h>
|
|
#include <torch/csrc/profiler/util.h>
|
|
|
|
TEST(ProfilerTest, AppendOnlyList) {
|
|
const int n = 4096;
|
|
torch::profiler::impl::AppendOnlyList<int, 1024> list;
|
|
for (const auto i : c10::irange(n)) {
|
|
list.emplace_back(i);
|
|
ASSERT_EQ(list.size(), i + 1);
|
|
}
|
|
|
|
int expected = 0;
|
|
for (const auto i : list) {
|
|
ASSERT_EQ(i, expected++);
|
|
}
|
|
ASSERT_EQ(expected, n);
|
|
|
|
list.clear();
|
|
ASSERT_EQ(list.size(), 0);
|
|
}
|
|
|
|
TEST(ProfilerTest, AppendOnlyList_ref) {
|
|
const int n = 512;
|
|
torch::profiler::impl::AppendOnlyList<std::pair<int, int>, 64> list;
|
|
std::vector<std::pair<int, int>*> refs;
|
|
for (const auto _ : c10::irange(n)) {
|
|
refs.push_back(list.emplace_back());
|
|
}
|
|
|
|
for (const auto i : c10::irange(n)) {
|
|
*refs.at(i) = {i, 0};
|
|
}
|
|
|
|
int expected = 0;
|
|
for (const auto& i : list) {
|
|
ASSERT_EQ(i.first, expected++);
|
|
}
|
|
}
|
|
|
|
// Test that we can convert TSC measurements back to wall clock time.
|
|
TEST(ProfilerTest, clock_converter) {
|
|
const int n = 10001;
|
|
torch::profiler::impl::ApproximateClockToUnixTimeConverter converter;
|
|
std::vector<torch::profiler::impl::ApproximateClockToUnixTimeConverter::UnixAndApproximateTimePair> pairs;
|
|
for (const auto i : c10::irange(n)) {
|
|
pairs.push_back(torch::profiler::impl::ApproximateClockToUnixTimeConverter::measurePair());
|
|
}
|
|
auto count_to_ns = converter.makeConverter();
|
|
std::vector<int64_t> deltas;
|
|
for (const auto& i : pairs) {
|
|
deltas.push_back(i.t_ - count_to_ns(i.approx_t_));
|
|
}
|
|
std::sort(deltas.begin(), deltas.end());
|
|
|
|
// In general it's not a good idea to put clocks in unit tests as it leads
|
|
// to flakiness. We mitigate this by:
|
|
// 1) Testing the clock itself. While the time to complete a task may
|
|
// vary, two clocks measuring the same time should be much more
|
|
// consistent.
|
|
// 2) Only testing the interquartile range. Context switches between
|
|
// calls to the two timers do occur and can result in hundreds of
|
|
// nanoseconds of noise, but such switches are only a few percent
|
|
// of cases.
|
|
// 3) We're willing to accept a somewhat large bias which can emerge from
|
|
// differences in the cost of calling each clock.
|
|
EXPECT_LT(std::abs(deltas[n / 2]), 200);
|
|
EXPECT_LT(deltas[n * 3 / 4] - deltas[n / 4], 50);
|
|
}
|