cudnnBackend api memory leak in calculate 1x1 convolution workspace size with ENGINEHEUR

[1004553826]

Description

when i using ENGINEHEUR to do 1x1 no padding convolution, i found memory leak in calculate workspace size

if using ENGINE, no memory leak

Environment

GPU Type:RTX 3090, RTX4070, RTX 3070, RTX Titan(all of then reproduced the memory leak)
Nvidia Driver Version: 591.44
CUDA Version:12.9
CUDNN Version:9.17
Operating System + Version:win11 pro 23H2 22631.6199

Relevant Files

cudnnTest.zip (3.1 KB)

Steps To Reproduce

1. create vs 2022 sln for this code file
2. include cudnn headers and link cudnn lib
3. compile this project
4. open visual studio diagnostic tool, switch to “memory usage”
5. Set breakpoint in loop for (int i = 0; i < 100; i++)
6. start this program, until the program execute the loop at least once, pause and take a snapshot
7. execute next loop and pause, take another snapshot
8. vs will auto compare this 2 snapshots and report memory leak
9. change the bool variable use_heur to false, repeat step 6-8, no memory leak reported in vs

you can see 0.54KB memory leak in one loop

maybe just i make mistakes in using cudnn api? Please help me, thanks

[1004553826]

well, i update cuda version from 12.9 to 13.1

the memory leak problem solved

[Aharpster]

Thanks for posting your solution!

Best,

Aharpster