-
Notifications
You must be signed in to change notification settings - Fork 99
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
KokkosBatched contraction length bug #258
Labels
Comments
@huttered40 : If you have already fixed it then you can put in a pull request to Kokkos-kernels. |
huttered40
pushed a commit
to huttered40/kokkos-kernels
that referenced
this issue
Jun 18, 2018
huttered40
pushed a commit
to huttered40/kokkos-kernels
that referenced
this issue
Jun 21, 2018
…batched/KokkosBatched_Gemm_Serial_Impl.hpp.
PR #261 fixed this. Batched GEMM supports rectangular matrices now. |
Reopened. @crtrott likes to close it with master update. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
In the partial specialization of the TeamGemm class template located in src/batched/KokkosBatched_Gemm_Team_Impl.hpp, for Trans::Transpose and a Trans::NoTranspose template parameters, the 3rd argument to the TeamGemmInternal<..>::invoke(..) method is A.extent(1), which is used as the contraction length (parameter k) in TeamGemmInternal<..>::invoke(..). I suspect this code has been used primarily for square matrices arising from finite element discretizations, so this bug has gone unnoticed. When calling this routine with unfolded tensors, however, it appears.
For example, in this specialization, A is transposed, so if A is passed in as a 100 x 4 View, then the contraction length will be 100, not 4. A.extent(1) is 4 here, however.
This same issue might be in the Gemm_Serial class template as well.
The text was updated successfully, but these errors were encountered: