tlrmchlsmth/momms

Multilevel Optimized Matrix-matrix Multiplication Sandbox

momms's Issues

Efficient packing routines for hierarchically stored matrices

Currently, just cannot pack into hierarchically stored matrices efficiently at all

Once these routines are implemented, can obsolete/remove row and column panel matrices

cleanup loops for kernel_mn and kernel_nm

When m % mr != 0 or n % nr != and C is stored in a Matrix, then junk data could be written to parts of C.

This doesn't happen when using Ukernel. #

Missing dt_on_output

To Whom It May Concern,

I am trying to compile MOMMS with rust-1.28.0-nightly and newest BLIS, but got the following error messages:

error[E0063]: missing field dt_on_output in initializer of kern::ukernel_wrapper::blis_types::auxinfo_t
--> src/kern/ukernel_wrapper.rs:62:28
|
62 | let mut info = auxinfo_t{
| ^^^^^^^^^ missing dt_on_output

error[E0063]: missing field dt_on_output in initializer of kern::ukernel_wrapper::blis_types::auxinfo_t
--> src/kern/ukernel_wrapper.rs:79:28
|
79 | let mut info = auxinfo_t{
| ^^^^^^^^^ missing dt_on_output

error[E0063]: missing field dt_on_output in initializer of kern::ukernel_wrapper::blis_types::auxinfo_t
--> src/kern/ukernel_wrapper.rs:96:28
|
96 | let mut info = auxinfo_t{
| ^^^^^^^^^ missing dt_on_output

error: aborting due to 3 previous errors

For more information about this error, try rustc --explain E0063.
error: Could not compile momms.

Could you please provide some help? Thanks very much.

Specialized packing

For performance reasons, packing at least needs to be specialized to panel width/height and possibly also to row-major/column major.

Not currently possible to do this through the type system in Rust, so it should be done through the control tree

Handle nonmultiples of MR and NR

This library can not yet handle M dimensions that aren't multiples of MR
and N dimensions that aren't multiples of NR when using the BLIS micro-kernel

Memory Leak

Since matrix buffers are now allocated manually on the heap, they need to be manually unallocated

All matrix types should implement the Drop trait and free their buffers there