CPU Usage

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi All, When a function is executed in GPU and we use cudathreadsynchronise() function after gpu function, then does CPU is in sleep condition? CPU usage is 100 % when any function is executing in CPU. Why???

2

You are saying to host: “wait till I’ve finished executing this kernel!”
cudaThreadSynchronize blocks CPU until the device has completed all previous calls and returns an error if one of the preceding tasks fails.
Cheers,

luca

3

Thank you for your answer.

Then is there any process to call cpu after finishing gpu work without blocking cpu???

4

I think I cannot understand you… After launching kernel, control is immediately return back to CPU (asynchronous call). You block CPU if using cudaThreadSynchronize, that let you to find any error.

Cheers,

luca

5

Yes. Syncrhonize to an event with the flag cudaEventBlockingSync or set the global device flag cudaDeviceBlockingSync to enable all cudaThreadSynchronize and implicit syncs. Note that this significantly increases latency to detect when the GPU completes its work. Using these flags slows my app down by 10%.

6

Thank you MisterAnderson42. Your solution works properly. Using these flags slows my app down by 4-10%.

I’ve tried cudaDeviceScheduleYield. But cpu usage was 100%. Why that did’nt work?

What does cudaDeviceBlockingSync do? Why it takes a minimum cpu percentage?

7

These flags do exactly what they say in their names, and as specified in the documentation. *Yield puts a yield in the spin-wait loop, allowing other threads to potentially run while waiting. *BlockingSync blocks the thread from execution until an event from the GPU (i.e. interrupt) wakes it up.