CUDA_ERROR_LAUNCH_FAILED = 719

Hi,

I’m using MMAPI, backend.
When USE_CPU_FOR_INTFLOAT_CONVERSION was set to 0 and executed, CUDA_ERROR_LAUNCH_FAILED = 719 occurred at “mapEGLImage2Float > cuGraphicsEGLRegisterImage”.

According to
https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TYPES.html#group__CUDA__TYPES_1ggc6c391505e117393cc2558fff6bfc2e90fd2e63260c6317ba943af0f7e4b8d21

CUDA_ERROR_LAUNCH_FAILED = 719 is said to occur because of “Common causes include dereferencing an invalid device pointer and accessing out of bounds shared memory”.

Am i missing something?

Hi,

We want to reproduce this issue in our environment first, are you using JetPack4.2.2?

Thanks.

Yes, i install JetPack4.2.2 using Jetson Software with SDK Manager on Jetson Xavier.

And the above result occurred when set “trtinstance_num = 4” in trt_inference.cpp.

Thanks.

Hi,

Thanks for the confirm.

trtinstance_num indicates the inference channel number.
Usually, the input channel is 3 or 1, color or gray scale image respectively.
May I know why you use trtinstance_num=4?

Thanks.

I want to use object detection each 4 input channels.

I set trtinstance_num to 4 in my backend code.
I’ve verified that creating 4 conv1 and detecting objects on each channel.

However, the problem occurs when using mapEGLImage2Float with USE_CPU_FOR_INTFLOAT_CONVERSION = 0.

Thanks Aasta.

Hi,

Guess your input is RGBA, is it correct?
There are some limitation in our MMAPI to set trtinstance_num = 4 since we rarely deals with C=4 case.

Based on the error message shared in #1:

CUDA_ERROR_LAUNCH_FAILED = 719 is said to occur because of "Common causes include dereferencing an invalid device pointer and accessing out of bounds shared memory".

The error may cause by the C=4 allocation is not supported so MMAPI allocate C=3 instead, which leads to out of bound error.
Do you want to file a feature request for C=4 or the WAR in #5 is good for you?

Thanks.