In CUDA, when using the function nvcuda::wmma::load_matrix_sync, if the data stored in shared memory (SMEM) is of type float but the fragment requires type half, will the function automatically perform implicit type conversion from float to half during the loading process? Or do I need to explicitly convert the data to the required type before using this function?
Because I did not see it in doc :)
There is no conversion of data permitted by load_matrix_sync.
1 Like
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.