cudnnGetConvolutionForwardWorkspaceSize return zero

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

//intput set
const int in_n = 1;
const int in_c = 3;
const int in_h = 56;
const int in_w = 56;
cudnnTensorDescriptor_t in_desc;
bool input_c = false;

CUDNN_CALL(cudnnCreateTensorDescriptor(&in_desc));
CUDNN_CALL(cudnnSetTensor4dDescriptor(
in_desc, CUDNN_TENSOR_NHWC, CUDNN_DATA_FLOAT,in_n, in_c, in_h, in_w));

//kernel set
const int filt_k = 4;
const int filt_c = 3;
const int filt_h = 3;
const int filt_w = 3;
cudnnFilterDescriptor_t filt_desc;
CUDNN_CALL(cudnnCreateFilterDescriptor(&filt_desc));
CUDNN_CALL(cudnnSetFilter4dDescriptor(
filt_desc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW,
filt_k, filt_c, filt_h, filt_w));

//get workspacesize
size_t ws_size;
CUDNN_CALL(cudnnGetConvolutionForwardWorkspaceSize(
cudnn, in_desc, filt_desc, conv_desc, out_desc, algo, &ws_size));

when is input h and w set 56 ws_size is 16832，and input h and w set 96 ws_size is 0 ，and input h and w set 150 ws_size is 0，input h and w set 151 ws_size is 5.01MB，so i don’t know why this happen.

#2

Hi,

Could you please confirm were you able to compute results successfully even it returns zero, there could be the possibility of less than MB returns zero. And also could you please share the output log for the above h and w inputs.

Thank you.

#3

when is input h and w set 56 ws_size is 16832kb，and input h and w set 96 ws_size is 0 ，and input h and w set 150 ws_size is 0，input h and w set 151 ws_size is 5.01MB(this is turn to MB)
i confirm result is successfully, i use cudnn 7.6.0, There is a range h size and w size generate to ws_size zero

#4

Could you please try on the latest cuDNN version 8.3.2, and let us know if you still face this issue.

Thank you.