bitcoin/bitcoin
1
1
Fork 0
mirror of https://github.com/bitcoin/bitcoin.git synced 2024-05-17 23:56:39 +00:00
Code Issues Releases Activity

bench: update nanobench from 4.3.4 to 4.3.6

Browse Source 
Most importantly, this update fixes a bug in nanobench that always
disabled performance counters on linux.

It also adds another sanitizer suppression that is caught in clang++ 12.
This commit is contained in:
Martin Ankerl 2021-09-17 11:58:57 +02:00
parent 8f022a59b8
commit eed99cf272
No known key found for this signature in database
GPG Key ID: FBEAAD7FC6FFFE81
1 changed files with 24 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
src/bench/nanobench.h
View File

@ -33,7 +33,7 @@
// see https://semver.org/ // see https://semver.org/
#define ANKERL_NANOBENCH_VERSION_MAJOR 4 // incompatible API changes #define ANKERL_NANOBENCH_VERSION_MAJOR 4 // incompatible API changes
#define ANKERL_NANOBENCH_VERSION_MINOR 3 // backwards-compatible changes #define ANKERL_NANOBENCH_VERSION_MINOR 3 // backwards-compatible changes
#define ANKERL_NANOBENCH_VERSION_PATCH 4 // backwards-compatible bug fixes #define ANKERL_NANOBENCH_VERSION_PATCH 6 // backwards-compatible bug fixes
/////////////////////////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////////////////////////
// public facing api - as minimal as possible // public facing api - as minimal as possible
@ -88,13 +88,15 @@
} while (0) } while (0)
#endif #endif
#if defined(__linux__) && defined(PERF_EVENT_IOC_ID) && defined(PERF_COUNT_HW_REF_CPU_CYCLES) && defined(PERF_FLAG_FD_CLOEXEC) && \ #define ANKERL_NANOBENCH_PRIVATE_PERF_COUNTERS() 0
!defined(ANKERL_NANOBENCH_DISABLE_PERF_COUNTERS) #if defined(__linux__) && !defined(ANKERL_NANOBENCH_DISABLE_PERF_COUNTERS)
// only enable perf counters on kernel 3.14 which seems to have all the necessary defines. The three PERF_... defines are not in # include <linux/version.h>
// kernel 2.6.32 (all others are). # if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 14, 0)
# define ANKERL_NANOBENCH_PRIVATE_PERF_COUNTERS() 1 // PERF_COUNT_HW_REF_CPU_CYCLES only available since kernel 3.3
#else // PERF_FLAG_FD_CLOEXEC since kernel 3.14
# define ANKERL_NANOBENCH_PRIVATE_PERF_COUNTERS() 0 # undef ANKERL_NANOBENCH_PRIVATE_PERF_COUNTERS
# define ANKERL_NANOBENCH_PRIVATE_PERF_COUNTERS() 1
# endif
#endif #endif
#if defined(__clang__) #if defined(__clang__)
@ -2210,20 +2212,20 @@ struct IterationLogic::Impl {
columns.emplace_back(10, 1, "err%", "%", rErrorMedian * 100.0); columns.emplace_back(10, 1, "err%", "%", rErrorMedian * 100.0);
double rInsMedian = -1.0; double rInsMedian = -1.0;
if (mResult.has(Result::Measure::instructions)) { if (mBench.performanceCounters() && mResult.has(Result::Measure::instructions)) {
rInsMedian = mResult.median(Result::Measure::instructions); rInsMedian = mResult.median(Result::Measure::instructions);
columns.emplace_back(18, 2, "ins/" + mBench.unit(), "", rInsMedian / mBench.batch()); columns.emplace_back(18, 2, "ins/" + mBench.unit(), "", rInsMedian / mBench.batch());
} }
double rCycMedian = -1.0; double rCycMedian = -1.0;
if (mResult.has(Result::Measure::cpucycles)) { if (mBench.performanceCounters() && mResult.has(Result::Measure::cpucycles)) {
rCycMedian = mResult.median(Result::Measure::cpucycles); rCycMedian = mResult.median(Result::Measure::cpucycles);
columns.emplace_back(18, 2, "cyc/" + mBench.unit(), "", rCycMedian / mBench.batch()); columns.emplace_back(18, 2, "cyc/" + mBench.unit(), "", rCycMedian / mBench.batch());
} }
if (rInsMedian > 0.0 && rCycMedian > 0.0) { if (rInsMedian > 0.0 && rCycMedian > 0.0) {
columns.emplace_back(9, 3, "IPC", "", rCycMedian <= 0.0 ? 0.0 : rInsMedian / rCycMedian); columns.emplace_back(9, 3, "IPC", "", rCycMedian <= 0.0 ? 0.0 : rInsMedian / rCycMedian);
} }
if (mResult.has(Result::Measure::branchinstructions)) { if (mBench.performanceCounters() && mResult.has(Result::Measure::branchinstructions)) {
double rBraMedian = mResult.median(Result::Measure::branchinstructions); double rBraMedian = mResult.median(Result::Measure::branchinstructions);
columns.emplace_back(17, 2, "bra/" + mBench.unit(), "", rBraMedian / mBench.batch()); columns.emplace_back(17, 2, "bra/" + mBench.unit(), "", rBraMedian / mBench.batch());
if (mResult.has(Result::Measure::branchmisses)) { if (mResult.has(Result::Measure::branchmisses)) {
@ -2402,6 +2404,14 @@ public:
return (a + divisor / 2) / divisor; return (a + divisor / 2) / divisor;
} }
ANKERL_NANOBENCH_NO_SANITIZE("integer", "undefined")
static inline uint32_t mix(uint32_t x) noexcept {
x ^= x << 13;
x ^= x >> 17;
x ^= x << 5;
return x;
}
template <typename Op> template <typename Op>
ANKERL_NANOBENCH_NO_SANITIZE("integer", "undefined") ANKERL_NANOBENCH_NO_SANITIZE("integer", "undefined")
void calibrate(Op&& op) { void calibrate(Op&& op) {
@ -2441,15 +2451,10 @@ public:
uint64_t const numIters = 100000U + (std::random_device{}() & 3); uint64_t const numIters = 100000U + (std::random_device{}() & 3);
uint64_t n = numIters; uint64_t n = numIters;
uint32_t x = 1234567; uint32_t x = 1234567;
auto fn = [&]() {
x ^= x << 13;
x ^= x >> 17;
x ^= x << 5;
};
beginMeasure(); beginMeasure();
while (n-- > 0) { while (n-- > 0) {
fn(); x = mix(x);
} }
endMeasure(); endMeasure();
detail::doNotOptimizeAway(x); detail::doNotOptimizeAway(x);
@ -2459,8 +2464,8 @@ public:
beginMeasure(); beginMeasure();
while (n-- > 0) { while (n-- > 0) {
// we now run *twice* so we can easily calculate the overhead // we now run *twice* so we can easily calculate the overhead
fn(); x = mix(x);
fn(); x = mix(x);
} }
endMeasure(); endMeasure();
detail::doNotOptimizeAway(x); detail::doNotOptimizeAway(x);