I'm running an image filter on GPU and I need to measure the time each part of the program takes for comparison. First I tried time.h library but it always returned zero. Then I read this post
and used the same code in my program before and after calling the kernel but still it is returning zero. Can anyone tell me what the problem could be?
This is my code:
cudaEvent_t start,stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
float Elapsed=0,Cycle;
while(count)
{
cudaEventRecord(start,0);
ImgFilter<<<dimGrid,dimBlock>>>...
cudaEventRecord(stop,0);
cudaElapsedTime(&Cycle,statr,stop);
Elapsed += Cycle;
}
printf("Time = %f",Elapsed);
I also tried printing 'Cycle' but it's always zero.
You miss to call cudaEventSynchronize function
cudaEvent_t start,stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
float Elapsed=0,Cycle;
while(count)
{
cudaEventRecord(start,0);
ImgFilter<<<dimGrid,dimBlock>>>...
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
cudaElapsedTime(&Cycle,statr,stop);
Elapsed += Cycle;
}
printf("Time = %f",Elapsed);
Note, that device function returns before all CUDA threads finished execution and you need to use cudaThreadSynchronize after kernel calling.
Related
I'm trying to implement an error handler using the clock() function from the "time.h" library. The code runs inside an embeeded system (Colibri IMX7 - M4 Processor). The function is used to monitor a current value within a specific range, if the value of the current isn't correct the function should return an error message.
The function will see if the error is ocurring and in the first run it will save the first appearance of the error in a clock_t as reference, and then in the next runs if the error is still there, it will compare the current time using clock() with the previous reference and see if it will be longer than a specific time.
The problem is that the function clock() is always returning -1. What should I do to avoid that? Also, why can't I declare a clock_t variable as static (e.g. static clock_t start_t = clock()?
Please see below the function:
bool CrossLink_check_error_LED_UV_current_clock(int current_state, int current_at_LED_UV)
{
bool has_LED_UV_current_deviated = false;
static int current_number_of_errors_Current_LED_CANNON = 0;
clock_t startTimeError = clock();
const int maximum_operational_current_when_on = 2000;
const int minimum_turned_on_LED_UV_current = 45;
if( (current_at_LED_UV > maximum_operational_current_when_on)
||(current_state!=STATE_EMITTING && (current_at_LED_UV > minimum_turned_on_LED_UV_current))
||(current_state==STATE_EMITTING && (current_at_LED_UV < minimum_turned_on_LED_UV_current)) ){
current_number_of_errors_Current_LED_CANNON++;
if(current_number_of_errors_Current_LED_CANNON > 1) {
if (clock() - startTimeError > 50000){ // 50ms
has_LED_UV_current_deviated = true;
PRINTF("current_at_LED_UV: %d", current_at_LED_UV);
if(current_state==STATE_EMITTING){
PRINTF(" at state emitting");
}
PRINTF("\n\r");
}
}else{
if(startTimeError == -1){
startTimeError = clock();
}
}
}else{
startTimeError = 0;
current_number_of_errors_Current_LED_CANNON = 0;
}
return has_LED_UV_current_deviated;
}
Edit: I forgot to mention before, but we are using GCC 9.3.1 arm-none-eabi compiler with CMake to build the executable file. We have an embedeed system (Colibri IMX7 made by Toradex) that consists in 2 A7 Processors that runs our Linux (more visual interface) and the program that is used to control our device runs in a M4 Processor without an OS, just pure bare-metal.
For a lot of provided functions in the c standard library, if you have the documentation installed (usually it gets installed with the compiler), you can view documentation using the man command in the shell. With man clock, it tells me that:
NAME
clock - determine processor time
SYNOPSIS
#include <time.h>
clock_t clock(void);
DESCRIPTION
The clock() function returns an approximation of processor time used by the program.
RETURN VALUE
The value returned is the CPU time used so far as a clock_t; to get the number of seconds used, divide by
CLOCKS_PER_SEC. If the processor time used is not available or its value cannot be represented, the function
returns the value (clock_t) -1.
etc.
This tells us that -1 means that the processor time (CLOCK_PROCESS_CPUTIME_ID) is unavailable. The solution is to use CLOCK_MONOTONIC instead. We can select the clock we want to use with clock_gettime.
timespec clock_time;
if (clock_gettime(CLOCK_MONOTONIC, &clock_time)) {
printf("CLOCK_MONOTONIC is unavailable!\n");
exit(1);
}
printf("Seconds: %d Nanoseconds: %ld\n", clock_time.tv_sec, clock_time.tv_nsec);
To answer the second part of your question:
static clock_t start_time = clock();
is not allowed because the return value of the function clock() is not known until runtime, but in C the initializer of a static variable must be a compile-time constant.
You can write:
static clock_t start_time = 0;
if (start_time == 0)
{
start_time = clock();
}
But this may or may not be suitable to use in this case, depending on whether zero is a legitimate return value of the function. If it could be, you would need something like:
static bool start_time_initialized = false;
static clock_t start_time;
if (!start_time_initialized)
{
start_time_initialized = true;
start_time = clock();
}
The above is reliable only if you cannot have two copies of this function running at once (it is not re-entrant).
If you have a POSIX library available you could use a pthread_once_t to do the same as the above bool but in a re-entrant way. See man pthread_once for details.
Note that C++ allows more complicated options in this area, but you have asked about C.
Note also that abbreviating "start time" as start_t is a very bad idea, because the suffix _t means "type" and should only be used for type names.
in the end the problem was that since we are running our code on bare metal, the clock() function wasn't working. We ended up using an internal timer on the M4 Processor that we found, so now everything is fine. Thanks for the answers.
I am trying to measure the time period of a square wave on a Beaglebone running Angstrom OS. I have written a kernel driver to register an ISR in which I'm timing the pulses. Everything is working fine, but the time interval being measured is completely wrong. I'm using do_gettimeofday() function to measure the time. When I do the same in userspace program using poll() function, I'm able to achieve correct values (it shows around 1007 us for a 1000us wave), but when I use the driver to measure the pulse, I get the interval as 1923us. I have no idea why the time interval in the kernel is higher than that in user space. I have attached my code below.
I would be grateful if someone can find the mistake in my program.
kernel ISR:
static irqreturn_t ISR ( int irq, void *dev_id)
{
prev = c;
do_gettimeofday(&c);
printk(KERN_ALERT "%ld", (c.tv_usec - prev.tv_usec));
return IRQ_HANDLED;
}
userspace prog:
while(1){
prev = start;
gettimeofday(&start, NULL);
rc = poll(&fdset, 1, 20000);
if(rc < 0){
printf("Error in rc\n");
return -1;
}
if(rc == 0){
printf("Timed out\n");
return -1;
}
if (fdset.revents & POLLPRI) {
len = read(fdset.fd, buf, 2);
printf("%ld\n", (start.tv_usec - prev.tv_usec));
}
}
For profiling interrupt latency, I find it quite useful to be lazy and to set a GPIO pin then measure the time with an oscilloscope. Probably not the answer you want, but it might help you over a hurdle quickly.
I'm launching a set of kernels multiple (30) times.
Every test of these 30 (they are deterministic, at every test a set of kernels is called 10 times and this number is fixed), at the beginning, I do cudaSetDevice(0) and everything gets malloc'd and memcpy'd.
When the test is done and the execution time was taken, everything is cudaFree'd.
Here's a sample output from my program:
avg: 81.7189
times:
213.0105 202.8020 196.8834 202.4001 197.7123 215.4658 199.5302 198.6519 200.8467
203.7865 20.2014 20.1881 21.0537 20.8805 20.1986 20.6036 20.9458 20.9473 20.292
9 20.9167 21.0686 20.4563 24.5359 21.1530 21.7075 23.3320 20.5921 20.6506 19.933
1 20.8211
The first 10 kernels take about 200 ms, while the others take about 20 ms.
Apparently every kernel calculates the same values, they all print the correct one. But since I malloc every test in the same order, couldn't the GPU memory still have the same values from the previous execution?
Also, kernels aren't returning errors because I'm checking them. Every kernel launch has cudaThreadSynchronize() for debugging purposes and error checking right after them with this macro:
#define CUDA_ERROR_CHECK if( (error = cudaGetLastError()) != cudaSuccess) printf("CUDA error: %s\n", cudaGetErrorString(error));
Why is this happening?
I'm getting the execution times from windows functions:
void StartCounter()
{
LARGE_INTEGER li;
if(!QueryPerformanceFrequency(&li))
cout << "QueryPerformanceFrequency failed!\n";
PCFreq = double(li.QuadPart)/1000.0;
QueryPerformanceCounter(&li);
CounterStart = li.QuadPart;
}
void StopCounter()
{
LARGE_INTEGER li;
QueryPerformanceCounter(&li);
double time = double(li.QuadPart-CounterStart)/PCFreq;
v.push_back(time);
}
Edit:
The mallocs, copys and other stuff aren't being timed. I only time the execution time (kernel launch and sync).
Visual Studio 2010's optimizations are turned on. Everything is set to maximize speed. CUDA's optimizations are on as well.
Measuring kernel execution time using QueryPerformanceTime is wrong, because host call device and than they are working in parallel. You ara propably measuring only call time.
To check kernel execution time use as ahmad mentioned cudaEvents:
cudaEvent_t start, stop;
float time;
cudaEventCreate(&start);
cudaEventCreate(&stop);
...
cudaEventRecord(start, 0);
yourkernel <<< n_blocks, block_size >>> (a_d, N);
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
...
cudaEventElapsedTime(&time, start, stop);
printf ("Time for the kernel: %f ms\n", time);
If you want to use QueryPerformanceTime you have to call
cudaDeviceSynchronize();
after kernel call. It will wait until kernel stops.
How can I get the Windows system time with millisecond resolution?
If the above is not possible, then how can I get the operating system start time? I would like to use this value together with timeGetTime() in order to compute a system time with millisecond resolution.
Try this article from MSDN Magazine. It's actually quite complicated.
Implement a Continuously Updating, High-Resolution Time Provider for Windows
(archive link)
This is an elaboration of the above comments to explain the some of the whys.
First, the GetSystemTime* calls are the only Win32 APIs providing the system's time. This time has a fairly coarse granularity, as most applications do not need the overhead required to maintain a higher resolution. Time is (likely) stored internally as a 64-bit count of milliseconds. Calling timeGetTime gets the low order 32 bits. Calling GetSystemTime, etc requests Windows to return this millisecond time, after converting into days, etc and including the system start time.
There are two time sources in a machine: the CPU's clock and an on-board clock (e.g., real-time clock (RTC), Programmable Interval Timers (PIT), and High Precision Event Timer (HPET)). The first has a resolution of around ~0.5ns (2GHz) and the second is generally programmable down to a period of 1ms (though newer chips (HPET) have higher resolution). Windows uses these periodic ticks to perform certain operations, including updating the system time.
Applications can change this period via timerBeginPeriod; however, this affects the entire system. The OS will check / update regular events at the requested frequency. Under low CPU loads / frequencies, there are idle periods for power savings. At high frequencies, there isn't time to put the processor into low power states. See Timer Resolution for further details. Finally, each tick has some overhead and increasing the frequency consumes more CPU cycles.
For higher resolution time, the system time is not maintained to this accuracy, no more than Big Ben has a second hand. Using QueryPerformanceCounter (QPC) or the CPU's ticks (rdtsc) can provide the resolution between the system time ticks. Such an approach was used in the MSDN magazine article Kevin cited. Though these approaches may have drift (e.g., due to frequency scaling), etc and therefore need to be synced to the system time.
In Windows, the base of all time is a function called GetSystemTimeAsFiletime.
It returns a structure that is capable of holding a time with 100ns resoution.
It is kept in UTC
The FILETIME structure records the number of 100ns intervals since January 1, 1600; meaning its resolution is limited to 100ns.
This forms our first function:
A 64-bit number of 100ns ticks since January 1, 1600 is somewhat unwieldy. Windows provides a handy helper function, FileTimeToSystemTime that can decode this 64-bit integer into useful parts:
record SYSTEMTIME {
wYear: Word;
wMonth: Word;
wDayOfWeek: Word;
wDay: Word;
wHour: Word;
wMinute: Word;
wSecond: Word;
wMilliseconds: Word;
}
Notice that SYSTEMTIME has a built-in resolution limitation of 1ms
Now we have a way to go from FILETIME to SYSTEMTIME:
We could write the function to get the current system time as a SYSTEIMTIME structure:
SYSTEMTIME GetSystemTime()
{
//Get the current system time utc in it's native 100ns FILETIME structure
FILETIME ftNow;
GetSytemTimeAsFileTime(ref ft);
//Decode the 100ns intervals into a 1ms resolution SYSTEMTIME for us
SYSTEMTIME stNow;
FileTimeToSystemTime(ref stNow);
return stNow;
}
Except Windows already wrote such a function for you: GetSystemTime
Local, rather than UTC
Now what if you don't want the current time in UTC. What if you want it in your local time? Windows provides a function to convert a FILETIME that is in UTC into your local time: FileTimeToLocalFileTime
You could write a function that returns you a FILETIME in local time already:
FILETIME GetLocalTimeAsFileTime()
{
FILETIME ftNow;
GetSystemTimeAsFileTime(ref ftNow);
//convert to local
FILETIME ftNowLocal
FileTimeToLocalFileTime(ftNow, ref ftNowLocal);
return ftNowLocal;
}
And lets say you want to decode the local FILETIME into a SYSTEMTIME. That's no problem, you can use FileTimeToSystemTime again:
Fortunately, Windows already provides you a function that returns you the value:
Precise
There is another consideration. Before Windows 8, the clock had a resolution of around 15ms. In Windows 8 they improved the clock to 100ns (matching the resolution of FILETIME).
GetSystemTimeAsFileTime (legacy, 15ms resolution)
GetSystemTimeAsPreciseFileTime (Windows 8, 100ns resolution)
This means we should always prefer the new value:
You asked for the time
You asked for the time; but you have some choices.
The timezone:
UTC (system native)
Local timezone
The format:
FILETIME (system native, 100ns resolution)
SYTEMTIME (decoded, 1ms resolution)
Summary
100ns resolution: FILETIME
UTC: GetSytemTimeAsPreciseFileTime (or GetSystemTimeAsFileTime)
Local: (roll your own)
1ms resolution: SYSTEMTIME
UTC: GetSystemTime
Local: GetLocalTime
GetTickCount will not get it done for you.
Look into QueryPerformanceFrequency / QueryPerformanceCounter. The only gotcha here is CPU scaling though, so do your research.
Starting with Windows 8 Microsoft has introduced the new API command GetSystemTimePreciseAsFileTime
Unfortunately you can't use that if you create software which must also run on older operating systems.
My current solution is as follows, but be aware: The determined time is not exact, it is only near to the real time. The result should always be smaller or equal to the real time, but with a fixed error (unless the computer went to standby). The result has a millisecond resolution. For my purpose it is exact enough.
void GetHighResolutionSystemTime(SYSTEMTIME* pst)
{
static LARGE_INTEGER uFrequency = { 0 };
static LARGE_INTEGER uInitialCount;
static LARGE_INTEGER uInitialTime;
static bool bNoHighResolution = false;
if(!bNoHighResolution && uFrequency.QuadPart == 0)
{
// Initialize performance counter to system time mapping
bNoHighResolution = !QueryPerformanceFrequency(&uFrequency);
if(!bNoHighResolution)
{
FILETIME ftOld, ftInitial;
GetSystemTimeAsFileTime(&ftOld);
do
{
GetSystemTimeAsFileTime(&ftInitial);
QueryPerformanceCounter(&uInitialCount);
} while(ftOld.dwHighDateTime == ftInitial.dwHighDateTime && ftOld.dwLowDateTime == ftInitial.dwLowDateTime);
uInitialTime.LowPart = ftInitial.dwLowDateTime;
uInitialTime.HighPart = ftInitial.dwHighDateTime;
}
}
if(bNoHighResolution)
{
GetSystemTime(pst);
}
else
{
LARGE_INTEGER uNow, uSystemTime;
{
FILETIME ftTemp;
GetSystemTimeAsFileTime(&ftTemp);
uSystemTime.LowPart = ftTemp.dwLowDateTime;
uSystemTime.HighPart = ftTemp.dwHighDateTime;
}
QueryPerformanceCounter(&uNow);
LARGE_INTEGER uCurrentTime;
uCurrentTime.QuadPart = uInitialTime.QuadPart + (uNow.QuadPart - uInitialCount.QuadPart) * 10000000 / uFrequency.QuadPart;
if(uCurrentTime.QuadPart < uSystemTime.QuadPart || abs(uSystemTime.QuadPart - uCurrentTime.QuadPart) > 1000000)
{
// The performance counter has been frozen (e. g. after standby on laptops)
// -> Use current system time and determine the high performance time the next time we need it
uFrequency.QuadPart = 0;
uCurrentTime = uSystemTime;
}
FILETIME ftCurrent;
ftCurrent.dwLowDateTime = uCurrentTime.LowPart;
ftCurrent.dwHighDateTime = uCurrentTime.HighPart;
FileTimeToSystemTime(&ftCurrent, pst);
}
}
GetSystemTimeAsFileTime gives the best precision of any Win32 function for absolute time. QPF/QPC as Joel Clark suggested will give better relative time.
Since we all come here for quick snippets instead of boring explanations, I'll write one:
FILETIME t;
GetSystemTimeAsFileTime(&t); // unusable as is
ULARGE_INTEGER i;
i.LowPart = t.dwLowDateTime;
i.HighPart = t.dwHighDateTime;
int64_t ticks_since_1601 = i.QuadPart; // now usable
int64_t us_since_1601 = (i.QuadPart * 1e-1);
int64_t ms_since_1601 = (i.QuadPart * 1e-4);
int64_t sec_since_1601 = (i.QuadPart * 1e-7);
// unix epoch
int64_t unix_us = (i.QuadPart * 1e-1) - 11644473600LL * 1000000;
int64_t unix_ms = (i.QuadPart * 1e-4) - 11644473600LL * 1000;
double unix_sec = (i.QuadPart * 1e-7) - 11644473600LL;
// i.QuadPart is # of 100ns ticks since 1601-01-01T00:00:00Z
// difference to Unix Epoch is 11644473600 seconds (attention to units!)
No idea how drifting performance-counter-based answers went up, don't do slippage bugs, guys.
QueryPerformanceCounter() is built for fine-grained timer resolution.
It is the highest resolution timer that the system has to offer that you can use in your application code to identify performance bottlenecks
Here is a simple implementation for C# devs:
[DllImport("kernel32.dll")]
extern static short QueryPerformanceCounter(ref long x);
[DllImport("kernel32.dll")]
extern static short QueryPerformanceFrequency(ref long x);
private long m_endTime;
private long m_startTime;
private long m_frequency;
public Form1()
{
InitializeComponent();
}
public void Begin()
{
QueryPerformanceCounter(ref m_startTime);
}
public void End()
{
QueryPerformanceCounter(ref m_endTime);
}
private void button1_Click(object sender, EventArgs e)
{
QueryPerformanceFrequency(ref m_frequency);
Begin();
for (long i = 0; i < 1000; i++) ;
End();
MessageBox.Show((m_endTime - m_startTime).ToString());
}
If you are a C/C++ dev, then take a look here: How to use the QueryPerformanceCounter function to time code in Visual C++
Well, this one is very old, yet there is another useful function in Windows C library _ftime, which returns a structure with local time as time_t, milliseconds, timezone, and daylight saving time flag.
In C11 and above (or C++17 and above) you can use timespec_get() to get time with higher precision portably
#include <stdio.h>
#include <time.h>
int main(void)
{
struct timespec ts;
timespec_get(&ts, TIME_UTC);
char buff[100];
strftime(buff, sizeof buff, "%D %T", gmtime(&ts.tv_sec));
printf("Current time: %s.%09ld UTC\n", buff, ts.tv_nsec);
}
If you're using C++ then since C++11 you can use std::chrono::high_resolution_clock, std::chrono::system_clock (wall clock), or std::chrono::steady_clock (monotonic clock) in the new <chrono> header. No need to use Windows-specific APIs anymore
auto start1 = std::chrono::high_resolution_clock::now();
auto start2 = std::chrono::system_clock::now();
auto start3 = std::chrono::steady_clock::now();
// do some work
auto end1 = std::chrono::high_resolution_clock::now();
auto end2 = std::chrono::system_clock::now();
auto end3 = std::chrono::steady_clock::now();
std::chrono::duration<long long, std::milli> diff1 = end1 - start1;
std::chrono::duration<double, std::milli> diff2 = end2 - start2;
auto diff3 = std::chrono::duration_cast<std::chrono::milliseconds>(end3 - start3);
std::cout << diff.count() << ' ' << diff2.count() << ' ' << diff3.count() << '\n';
Is there a common way to get the current time in or with milliseconds?
There is os.time(), but it only provides full seconds.
I use LuaSocket to get more precision.
require "socket"
print("Milliseconds: " .. socket.gettime()*1000)
This adds a dependency of course, but works fine for personal use (in benchmarking scripts for example).
If you want to benchmark, you can use os.clock as shown by the doc:
local x = os.clock()
local s = 0
for i=1,100000 do s = s + i end
print(string.format("elapsed time: %.2f\n", os.clock() - x))
In standard C lua, no. You will have to settle for seconds, unless you are willing to modify the lua interpreter yourself to have os.time use the resolution you want. That may be unacceptable, however, if you are writing code for other people to run on their own and not something like a web application where you have full control of the environment.
Edit: another option is to write your own small DLL in C that extends lua with a new function that would give you the values you want, and require that dll be distributed with your code to whomever is going to be using it.
Get current time in milliseconds.
os.time()
os.time()
return sec // only
posix.clock_gettime(clk)
https://luaposix.github.io/luaposix/modules/posix.time.html#clock_gettime
require'posix'.clock_gettime(0)
return sec, nsec
linux/time.h // man clock_gettime
/*
* The IDs of the various system clocks (for POSIX.1b interval timers):
*/
#define CLOCK_REALTIME 0
#define CLOCK_MONOTONIC 1
#define CLOCK_PROCESS_CPUTIME_ID 2
#define CLOCK_THREAD_CPUTIME_ID 3
#define CLOCK_MONOTONIC_RAW 4
#define CLOCK_REALTIME_COARSE 5
#define CLOCK_MONOTONIC_COARSE 6
socket.gettime()
http://w3.impa.br/~diego/software/luasocket/socket.html#gettime
require'socket'.gettime()
return sec.xxx
as waqas says
compare & test
get_millisecond.lua
local posix=require'posix'
local socket=require'socket'
for i=1,3 do
print( os.time() )
print( posix.clock_gettime(0) )
print( socket.gettime() )
print''
posix.nanosleep(0, 1) -- sec, nsec
end
output
lua get_millisecond.lua
1490186718
1490186718 268570540
1490186718.2686
1490186718
1490186718 268662191
1490186718.2687
1490186718
1490186718 268782765
1490186718.2688
I made a suitable solution for lua on Windows. I basically did what Kevlar suggested, but with a shared library rather than a DLL. This has been tested using cygwin.
I wrote some lua compatible C code, compiled it to a shared library (.so file via gcc in cygwin), and then loaded it up in lua using package.cpath and require" ". Wrote an adapter script for convenience. Here is all of the source:
first the C code, HighResTimer.c
////////////////////////////////////////////////////////////////
//HighResTimer.c by Cody Duncan
//
//compile with: gcc -o Timer.so -shared HighResTimer.c -llua5.1
//compiled in cygwin after installing lua (cant remember if I
// installed via setup or if I downloaded and compiled lua,
// probably the former)
////////////////////////////////////////////////////////////////
#include <windows.h>
typedef unsigned __int64 u64;
double mNanoSecondsPerCount;
#include "lua.h"
#include "lualib.h"
#include "lauxlib.h"
int prevInit = 0;
int currInit = 0;
u64 prevTime = 0;
u64 currTime = 0;
u64 FrequencyCountPerSec;
LARGE_INTEGER frequencyTemp;
static int readHiResTimerFrequency(lua_State *L)
{
QueryPerformanceFrequency(&frequencyTemp);
FrequencyCountPerSec = frequencyTemp.QuadPart;
lua_pushnumber(L, frequencyTemp.QuadPart);
return 1;
}
LARGE_INTEGER timerTemp;
static int storeTime(lua_State *L)
{
QueryPerformanceCounter(&timerTemp);
if(!prevInit)
{
prevInit = 1;
prevTime = timerTemp.QuadPart;
}
else if (!currInit)
{
currInit = 1;
currTime = timerTemp.QuadPart;
}
else
{
prevTime = currTime;
currTime = timerTemp.QuadPart;
}
lua_pushnumber(L, timerTemp.QuadPart);
return 1;
}
static int getNanoElapsed(lua_State *L)
{
double mNanoSecondsPerCount = 1000000000/(double)FrequencyCountPerSec;
double elapsedNano = (currTime - prevTime)*mNanoSecondsPerCount;
lua_pushnumber(L, elapsedNano);
return 1;
}
int luaopen_HighResolutionTimer (lua_State *L) {
static const luaL_reg mylib [] =
{
{"readHiResTimerFrequency", readHiResTimerFrequency},
{"storeTime", storeTime},
{"getNanoElapsed", getNanoElapsed},
{NULL, NULL} /* sentinel */
};
luaL_register(L,"timer",mylib);
return 1;
}
--
--
Now lets get it loaded up in a lua script, HighResTimer.lua .
Note: I compiled the HighResTimer.c to a shared library, Timer.so
#!/bin/lua
------------------------------------
---HighResTimer.lua by Cody Duncan
---Wraps the High Resolution Timer Functions in
--- Timer.so
------------------------------------
package.cpath = "./Timer.so" --assuming Timer.so is in the same directory
require "HighResolutionTimer" --load up the module
timer.readHiResTimerFrequency(); --stores the tickFrequency
--call this before code that is being measured for execution time
function start()
timer.storeTime();
end
--call this after code that is being measured for execution time
function stop()
timer.storeTime();
end
--once the prior two functions have been called, call this to get the
--time elapsed between them in nanoseconds
function getNanosElapsed()
return timer.getNanoElapsed();
end
--
--
and Finally, utilize the timer, TimerTest.lua .
#!/bin/lua
------------------------------------
---TimerTest.lua by Cody Duncan
---
---HighResTimer.lua and Timer.so must
--- be in the same directory as
--- this script.
------------------------------------
require './HighResTimer'
start();
for i = 0, 3000000 do io.write("") end --do essentially nothing 3million times.
stop();
--divide nanoseconds by 1 million to get milliseconds
executionTime = getNanosElapsed()/1000000;
io.write("execution time: ", executionTime, "ms\n");
Note: Any comments were written after pasting the source code into the post editor, so technically this is untested, but hopefully the comments didn't befuddle anything. I will be sure to come back and provide a fix if it does.
If you're using lua with nginx/openresty you could use ngx.now() which returns a float with millisecond precision
If you're using OpenResty then it provides for in-built millisecond time accuracy through the use of its ngx.now() function. Although if you want fine grained millisecond accuracy then you may need to call ngx.update_time() first. Or if you want to go one step further...
If you are using luajit enabled environment, such as OpenResty, then you can also use ffi to access C based time functions such as gettimeofday() e.g: (Note: The pcall check for the existence of struct timeval is only necessary if you're running it repeatedly e.g. via content_by_lua_file in OpenResty - without it you run into errors such as attempt to redefine 'timeval')
if pcall(ffi.typeof, "struct timeval") then
-- check if already defined.
else
-- undefined! let's define it!
ffi.cdef[[
typedef struct timeval {
long tv_sec;
long tv_usec;
} timeval;
int gettimeofday(struct timeval* t, void* tzp);
]]
end
local gettimeofday_struct = ffi.new("struct timeval")
local function gettimeofday()
ffi.C.gettimeofday(gettimeofday_struct, nil)
return tonumber(gettimeofday_struct.tv_sec) * 1000000 + tonumber(gettimeofday_struct.tv_usec)
end
Then the new lua gettimeofday() function can be called from lua to provide the clock time to microsecond level accuracy.
Indeed, one could take a similar approaching using clock_gettime() to obtain nanosecond accuracy.
Kevlar is correct.
An alternative to a custom DLL is Lua Alien
in openresty there is a function ngx.req.start_time.
From the docs:
Returns a floating-point number representing the timestamp (including milliseconds as the decimal part) when the current request was created.
You can use C function gettimeofday :
http://www.opengroup.org/onlinepubs/000095399/functions/gettimeofday.html
Here C library 'ul_time', function sec_usec resides in 'time' global table and returns seconds, useconds. Copy DLL to Lua folder, open it with require 'ul_time'.
http://depositfiles.com/files/3g2fx7dij
If you're on a system with a GNU-compatible implementation of date that you can execute, here's a one-liner to get the Epoch time in milliseconds:
local function gethammertime()
return tonumber(assert(assert(io.popen'date +%s%3N'):read'a'))
end
Note that the assert calls are necessary to ensure that any failures to read or open date will propagate the errors, respectively. Also note that this relies on garbage collection (or finalizers, in Lua 5.4) to close the process handle: if using a pre-5.4 version of Lua and resource exhaustion is a concern, you may wish to extend this to three lines like Klesun's Windows-based answer and close the handle explicitly.
If your environment is Windows and you have access to system commands, you can get time of centiseconds precision with io.popen(command):
local handle = io.popen("echo %time%")
local result = handle:read("*a")
handle:close()
The result will hold string of hh:mm:ss.cc format: (with trailing line break)
"19:56:53.90\n"
Note, it's in local timezone, so you probably want to extract only the .cc part and combine it with epoch seconds from os.time().