[BUG] rfactor + cache_write resulting incorrect schedules and buffer shape

[vinx13]

Two issues I found that might be related to rfactor :
https://gist.github.com/vinx13/049c51021b701c133e8aca2664c9a55d

• the first schedule schedule_sparse_dense_cuda_0: calling rfactor after cache_write makes the copying-back block lost
• the second schedule schedule_sparse_dense_cuda1: the shape of BSRmm_rf_local is not inferred correctly, it doesn’t affect performance though. Read/write region looks correct

@spectrometerHBH @MasterJH5574

[junrushao]

Is this bug fixed on mainline @MasterJH5574?

[MasterJH5574]

Is this bug fixed on mainline @MasterJH5574?

@junrushao1994 I'm not sure because I knew nothing about rfactor at that time. Will check it soon.

[MasterJH5574]

@junrushao1994 Bug 1 was fixed before. As for Bug 2, since cross-thread reduction hasn't been upstreamed, we may only check whether it's fixed after cross-thread reduction upstreming.

[junrushao]

Sounds great! I'm going to close this issue and wait for cross-thread reduction stuff upstreamed 😄 Perhaps it is a good unittest to add

[MasterJH5574]

Bug 2 should be a problem of other lowering passes like CompactBufferAllocation (I guess so). Anyway, I will keep an eye on this issue when doing upstreaming.

[junrushao]

Would be nice if we can fix the corresponding passes. Don’t hesitate to ask Siyuan or me if there is any question on these passes