How to run SpMM with half precision? I see that Sputnik only provides an interface for half2. If I have a CSR matrix with val in half precision, do I need to preprocess it? Also, why is the type of column idx is short2? If the matrix is very large, won't it be impossible to store the column indices?

cudaError_t CudaSpmm(int m, int k, int n, int nonzeros,
                     const int* __restrict__ row_indices,
                     const half2* __restrict__ values,
                     const int* __restrict__ row_offsets,
                     const short2* __restrict__ column_indices,
                     const half2* __restrict__ dense_matrix,
                     half2* __restrict__ output_matrix, cudaStream_t stream)

Any response would be helpful, thank you.

I want to run the code for a single matrix without depending on googletest, e.g., something like "/.spmm [options] tmp.mtx". Is that possible with the current code? P.S. the code runs and passes all the test cases performed by spmm_test.

Thanks for releasing the code.
In your test API, I saw that you are using the Scipy api to convert the sparse output to the dense. However, I guess this is not an efficient way to convert it back. I would like to know, do you have any suggestion to make it more efficient? More precisely, I would like to know how I can get a dense output for SDDMM without sacrificing the performance?

When initializing the sparse_tile_loader，the threadIdx.x should be threadIdx.x%kBlockWidth. Is what I said correct ?

It seems that No user API is provided in your code to create a "SparseMatrix" object from a sparse matrix with a dense format (such as float* A).
Otherwise, I can not test the performance of your kernel with real sparse weight matrices.
I have to write my own codes to generate the CSR format from float* A, and then I have to generate the row indexes according to your row-swizzle policy.

Is my understanding correct or did I miss something in your code?

is there anyone know how to fix it?

Where is the cusparse testcase ? Thanks