NVDLA pads it output to nearest multiple of 64(or 32).
For general usage, output has to be reformatted, see reformat function for details.
reformat
For example, if output is 1 * 3 * 60 * 60 of __half format, then the actual output dimension of cuDLA would be 1 * 3 * 60 * 64
1 * 3 * 60 * 60
__half
1 * 3 * 60 * 64