Exactly how fast is this baby? (pretty fast, and don't call me baby)

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hey guys, I’d like to know something about kernel launches:

If I launch my kernel as follows:

kernel<<<A, B, 0>>>;
cudaThreadSynchronize();

Will there be always EXACTLY A*B kernel executions, no matter what the device?

I thought it would, but this one algorithm I made is absurdly fast and apparently loops through 57 billion kernel launches per second on a gtx 260, and thats not even optimised code! I’m actually wondering if all of those kernels are actually executing :wacko:

I’ll check the profiler output in the meantime

Thx!

#2

57 billion kernel launches per second

wow, it is 20 times per your cpu clock cycle;) maybe you have some error in speed calculation?

#3

Well here is my code:

////////////////////////////////////////////////////////////////////////////////////////////////////

#define GRID_SIZE 1024*64-1

#define THREAD_SIZE 512

#define LOOP_N_SIZE 1024

//kernel declaration:

global static void x_gen(word temp) { /* code */ … }

//main code snippet:

unsigned int timer = 0;

CUT_SAFE_CALL( cutCreateTimer( &timer));

CUT_SAFE_CALL( cutStartTimer( timer));

for (int i=0; i<LOOP_N_SIZE; i++)

{

	x_gen<<<GRID_SIZE, THREAD_SIZE, 0>>>(0);

	CUT_CHECK_ERROR("Kernel execution failed\n");

	CUDA_SAFE_CALL( cudaThreadSynchronize() );

}

CUT_SAFE_CALL( cutStopTimer( timer));

cout << "Processing time: " << cutGetTimerValue( timer ) << "ms\n";

cout << fixed << showpoint << setprecision(2) <<

	"x's generated per millisecond: " << 

	double(GRID_SIZE)*double(THREAD_SIZE)*double(LOOP_N_SIZE)/

	(double(cutGetTimerValue( timer) ) ) <<

	endl;

////////////////////////////////////////////////////////////////////////////////////////////////////

// output: 35m per MILLISECOND!!?

Processing time: 970.456ms

x’s generated per millisecond: 35405237.44

#4

ok, i didn’t understand you in your first post, as i thought that one kernel launch is one call x_gen<<<GRID_SIZE, THREAD_SIZE, 0>>>(0) for all threads. 35G thread launches per second is possible, gxt260 perform 192 * 1.2G clock cycles if you count all thre cores, so if thread don’t do to much and it writes not much it can process threads that fast;)

Could you post your kernel code?

#5

OK I found out what the problem was, (I think). I commented out my entire kernel code and the exec time was the same - essentially I wasn’t returning any values to the kernel calling code because I just wanted to benchmark, so I’m guessing the smart Nvidia compiler compiled out my entire kernel because no results were actually being sent through!!! Who would have thought…

#6

Don’t worry, almost everyone faces this problem sooner or later :)

Yes, if your kernel writes nothing to global memory then it will be optimized out and your code will just launch emtpy kernels.