[MLIR-AIR] Apply outer perm for unpack for air in 4-level tiling #1145
Conversation
erwei-xilinx
requested review from
MaheshRavishankar,
nirvedhmeshram,
yzhang93,
Abhishek-Varma and
jtuyls
as code owners
|
I've also seen the flaky failure on linux ci runner compiling third_party/mlir-air today. Could you please double check to see what caused the problem? |
|
A rerun helps, but it did happen very frequently today on several PRs... |
I can try to go to mlir-air repo and see if I can remove some redundant code, to reduce the codebase size. |
Thanks. You think it's OOM? Did you add a lot of codes compared to your previous bump? |
| auto maybePackPeelTiling = | ||
| ParameterSetting::create(linalgOp, /*isObjectFifo=*/true, targetDevice, | ||
| numRows, numCols, /*kPackScaleL1=*/2); | ||
| ParameterSetting::create(linalgOp, isObjectFifo, targetDevice, numRows, |
There was a problem hiding this comment.
Are you sure this is enough to get correct numerics? You have to modify the outer perm similar to this
setRootConfigForPackPeel4LevelTilingPipeline.
There was a problem hiding this comment.
If you like you can add a test in ci to make sure it's running.
There was a problem hiding this comment.
Oh cool, that would be great! Will add a matmul test to make sure.
Yeah I suspect it is OOM, as I notice that the segfault happens when more than one CI gets triggered at the same time. I didn't add a lot of codes recently, but the code base has been pretty big. AIRDependencyScheduleOpt.cpp for example is >5000 lines... |
FWIW I saw this last week once, I'd never seen it before (ever ever). |
pack-peel-4-level-tilingscheme.