Why calling times.cpuTime() is slow in Nim? - nim-lang

Question

In c++, std::chrono::steady_clock::now().time_since_epoch().count() takes about 20 nanos on my machine, which is close to rdtsc instruction. But in Nim, times.cpuTime (it calls clock_gettime in Linux) takes about 350ns. Any body knows why?
proc toNanos(sec: float): int64 =
let SecInNano = 1000000000f64
return (sec * SecInNano).int64
proc testTimes() =
let N = 1000000
var ts : seq[float]
ts.setLen(N+1)
for i in 0..N :
ts[i]= cpuTime()
var dur = toNanos((ts[N-1] - ts[0]) / N.float)
for i in 0..<N :
ts[i] = ts[i+1] - ts[i]
discard ts.pop
ts.sort()
echo fmt"---------- latency of invoking cpuTime (ns): nSamples: ", N, " ----------------"
echo fmt" Avg Min 50% 90% Max"
echo fmt" {dur} {ts[0].toNanos} {ts[N div 2].toNanos} {ts[int(N.float*0.9)].toNanos} {ts[N-1].toNanos}"

Answer 1

I think the reason is that std::chrono::steady_clock::now().time_since_epoch().count() returns an int. Whereas times.cpuTime() converts the result to a float.
Using monotimes.getMonoTime() which returns an int, I get times of about 40ns.
import std/[algorithm, strformat, monotimes]
proc testTimes() =
const N = 1000000
var ts = newSeq[int64](N+1)
for i in 0..N:
ts[i] = getMonoTime().ticks
var dur = int(float(ts[N-1] - ts[0]) / float(N))
for i in 0..<N:
ts[i] = ts[i+1] - ts[i]
discard ts.pop
ts.sort()
echo fmt"---------- latency of invoking cpuTime (ns): nSamples: ", N, " ----------------"
echo fmt" Avg Min 50% 90% Max"
echo fmt" {dur} {ts[0]} {ts[N div 2]} {ts[int(N.float*0.9)]} {ts[N-1]}"
testTimes()
BTW: The next time please post a complete runnable example, it simplifies testing.