rocsparse_spmm dense matrices have to have same order
jakub-homola opened this issue · comments
I was having trouble using the rocsparse_spmm function, the solve stage (only the solve stage) kept returning an error rocsparse_status_invalid_value. After a while of searching, I found this if in the source code, only allowing dense matrices B and C with the same storage order (row major or column major). Okay, this might make sense, since it's another case to implement (although cuSPARSE has no problem with that). But I could not find any info about this restriction in the documentation.
Please add a note about this to the docs (and please take a look at other functions if the same problem isn't there too).
Also, as a sidenote, please improve the error reporting (across all ROCm libraries). As far as I remember, CUDA libraries print an actually helpful error message before returning the status/error code.
@jakub-homola Thanks for bringing this to our attention.I am currently investigating allowing different rocsparse_order for both B and C matrices. Ill report back once I have a working PR.
I agree on your side note about improving the error reporting. Ill bring this up at our next group meeting.
@jakub-homola I have merged the fix (b13e637) that allows using mixed rocsparse_order for the dense B and C matrices in SpMM to the develop branch. We are targeting this for rocm-6.0 release. I can also report that we are currently working on improving the logging also targetted for the rocm-6.0 release.
Awesome, thanks :)