2d conv and 3d conv is not same

GPU: RTX 3090
CUDA 11.1
CUDNN 8.0.5
driver 457.30
windows10 vs2019
It is particularly interesting to compare 2d conv with 3d conv in fwd algo 1.
2d conv setting:
b=1,inc = 240,inw=10,inh=10,filter size(1,1),padding(0,0),stride(1,1),dilation(1,1); outc = 80,outw=10,outh=10, fwd algo:1.
3d conv setting
b=1,inc = 240,ind = 1,inw=10,inh=10,filter size(1,1,1),padding(0,0,0),stride(1,1,1),dilation(1,1,1); outc = 80, outd=1,outw=10,outh=10,fwd algo:1.
the value returned by cudnnConvolutionForward is not same, the workspace size is also different , why?

GPU:1080ti,cuda 10.0,cudnn 7.6.5,windows10 vs2015
the returned value is binary same.

Hi @ttt ,
The heuristic may return different kernel for 2d and 3d convolution. Since the kernel is different the requested workspace is also likely to be different. The returned value might not be binary same as the kernel is different, but their difference should be within a range.

Thanks!