Get CUDNN_STATUS_BAD_PARAM while executing cudnnFusedOpsExecute()

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

I was trying out fused op on V100 and got CUDNN_STATUS_BAD_PARAM while executing cudnnFusedOpsExecute():

    Execute line: 43 cuDNN: CUDNN_STATUS_BAD_PARAM
    Execute line: 43 cuDNN: CUDNN_STATUS_BAD_PARAM

According to NVIDIA’s API reference,this indicates the type of cudnnFusedOps_t in the plan descriptor is unsupported.
However,i checked all my settings and they all met the demand.
I turned on CUDNN_LOGERR_DBG/CUDNN_LOGWARN_DBG/CUDNN_LOGINFO_DBG,but didn’t get any err info.

     CuDNN (v8400) function cudnnFusedOpsExecute() called:
     handle: type=cudnnHandle_t; streamId=(nil) (defaultStream);
     varPack: type=cudnnFusedOpsVariantParamPack_t:
         ops: type=cudnnFusedOps_t; val=CUDNN_FUSED_SCALE_BIAS_ACTIVATION_CONV_BNSTATS (0);
         xData: location=dev; addr=0x7f8e4a000000;
         bnEqScale: location=dev; addr=0x7fffd6751a18;
         bnEqBias: location=dev; addr=0x7fffd6751a20;
         wData: location=dev; addr=0x7f8e77c2e800;
         yData: location=dev; addr=0x7f8e3c000000;
         ySum: location=dev; addr=0x7f8e77c36800;
         ySqSum: location=dev; addr=0x7f8e77c36a00;
         workSpace: location=dev; addr=0x7f8e77c2dc00;
         workSpaceSizeInBytes: type=size_t; val=3072;
    Time: 2022-09-06T03:44:32.867190 (0d+0h+0m+27s since start)
    Process=277; Thread=277; GPU=0; Handle=0x559d77e33b60; StreamId=(nil) (defaultStream).


    CuDNN (v8400) function cudnnGetErrorString() called:
     status: type=int; val=3;
    Time: 2022-09-06T03:44:32.867207 (0d+0h+0m+27s since start)
    Process=277; Thread=277; GPU=NULL; Handle=NULL; StreamId=NULL.

I’m wondering is there any means that i can get useful debugging infomation?

2

Hi,

I think you’re already using the available logging levels.

Please make sure, you’re using the api correctly, by referring to the requirements and reason for the CUDNN_STATUS_BAD_PARAM return value.

Are you facing the same issue on 8.5.0 as well?
If yes could you please share with us the minimal issue repro.

Thank you.

3

Here are my settings, I think i got them right.

image
image710×868 27.6 KB

I’m using cudnn 8.5.0.96+cuda 11.7,and this problem still exists.
Here’s my code
repro.tar (230 KB)

4

Hello,if there’s any progress on this problem,please let me know,thank you~

5

looks like these two pointers are not aligned well enough - note in the setup code these are declarered to be 16B aligned

1 Like
6

Thanks very much for the tip! Problem’s now solved.

7

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.