Support generic reduction and scan cases.

[ienkovich]

No description provided.

[minjang]

Looks good to me! I just have a naming convention suggestion (ScanReduce -> ReduceScan) and some minor code clarity. I will accept soon.

This is a general question for reduce/scan with vectorization. So, basically is the scan algorithm like std::inclusive_scan? And use vector::shuffle to move the data inside SIMD registers and then finally accumulate.

[minjang]

Looks good. It's like std::fmin and std::min.

[minjang]

Not urgent, maybe we can support F16/BF16 using its raw binary representations for quite_NaN?

[ienkovich]

Yes, this needs to be added. I couldn't yet find examples of how such constants are created, used, and lowered. Did you see any examples?

[minjang]

I just saw one F16 testing cases in the test_core.py regarding reduction. But not urgent at all.

[ienkovich]

This is a general question for reduce/scan with vectorization. So, basically is the scan algorithm like std::inclusive_scan? And use vector::shuffle to move the data inside SIMD registers and then finally accumulate.

There are three general cases here. For scans on trailing dimension, we use shuffle to accumulate neighbors, then increase shuffle stride and use masks to keep already computed elements intact. For reductions on trailing dimensions we use shuffle to swap to vector parts and accumulate, then swap halves of halves, etc. For cases when a scan/reduction goes on a non-trailing dimension, we iterate through all sub-vectors and accumulate. Works like a fully unrolled loop.