[QST] [CuTeDSL] Persistent dense gemm on Hopper

[simveit]

I sucessfully implemented a persistent SGEMM kernel in CuTeDSL for H100 using the Static scheduler.

However now I want to apply similar approach to dense_gemm.py but apply the pipeline abstraction.
I get

Mismatched elements: 8388608 / 16777216 (50.0%)
Greatest absolute difference: inf at index (0, 9, 0) (up to 0.1 allowed)
Greatest relative difference: inf at index (0, 9, 0) (up to 0.001 allowed)

Which looks to me like there is something wrong with the way I write out the accumulators.
Could someone take a look and help me what is wrong?
https://gist.github.com/simveit/37b62b563e17ef47785c7fbb07cf8586

I should add that I launch with --tile_shape_mnk 128,256,64 to have appropriate atom size.