[QST]int4b_t dataptr question

[liuyt929]

What is your question?
hi~
excuse me,I have some problems when using s4t_s4n_s4t_sm80 gemm, can I create a cutlass::int4_b* data_ptr ,and malloc elements_number/2 byte for it. And create a tensorref for it, then directly use it as the output_ref for the gemm op.
like this
cutlass::int4b_t* output_int4; cudaMalloc((void**)&output_int4, M*N/2); cutlass::TensorRef<cutlass::int4b_t,cutlass::layout::ColumnMajor> out_ref( output_int4,cutlass::layout::ColumnMajor::packed(cutlass::MatrixCoord(M,N)));

Can I get the right answer if I do so?
I have tried a lot, but the result is always wrong, I wonder how I can receive the result correctly.
Thanks!

[thakkarV]

We can't help you debug without proper repro steps