Does cudaFree after asynchronous call work?

Question

I want to ask whether calling to cudaFree after some asynchronous calls is valid? For example

int* dev_a;

// prepare dev_a...

// launch a kernel to process dev_a (asynchronously)

cudaFree(dev_a);

In this case, since kernel launch is asynchronous, when the cudaFree part is reached, the kernel may haven't finish running yet. Then will the cudaFree(dev_a) immediately after it destroy the data?

Pretty sure that `cudaFree` will synchronize before it attempts to deallocate the pointer. — Jared Hoberock, Jan 17 '14 at 04:10

talonmies · Accepted Answer · 2014-01-17T09:09:47.203

3

As per Jared's comment, I am about 99% certain that the CUDA driver free/malloc pair are implemented as blocking calls which will synchronize the context on which they operate before they execute the call.

edited Jan 17 '14 at 09:09

answered Jan 17 '14 at 07:11

talonmies

70,661
34
192
269

Thank you! How about the "free" function inside the kernel? If I have a kernel launch immediately proceeding it, does this work? – shaoyl85 Jan 18 '14 at 00:24

score 2 · Answer 2 · answered Dec 02 '21 at 13:29

2

CUDA now provide functions for asynchronous memory management based on streams: cudaMallocAsync, cudaMemcpyAsync, cudaMemcpyAsync.

A short introduction is available here

answered Dec 02 '21 at 13:29

pixelou

748
6
17

Does cudaFree after asynchronous call work?

2 Answers2

Linked