[multi vector] Add tiled MinMax MaxSim kernels - #1394

Open

juchen-ms (partychen) wants to merge 26 commits into

microsoft:mainfrom

partychen:tiled-minmax-maxsim-main

Commits on Sep 12, 2026

Add tiled MinMax MaxSim kernels

juchen-ms (partychen)
and
Copilot
committed
d4a683b
Optimize MinMax8 x MinMax4 MaxSim

juchen-ms (partychen)
and
Copilot
committed
6fc6d82
Keep MinMax SIMD intrinsics local

juchen-ms (partychen)
and
Copilot
committed
510c667
Optimize MinMax SIMD packing

juchen-ms (partychen)
and
Copilot
committed
3267915
Fix MinMax MaxSim reduction and expand kernel coverage

juchen-ms (partychen)
and
Copilot
committed
7dfdb95
Refine MinMax MaxSim packing and ownership

juchen-ms (partychen)
and
Copilot
committed
4b67d7b

Commits on Sep 15, 2026

[multi vector] Align MinMax kernels with paneled traversal

juchen-ms (partychen)
and
Copilot
committed
aef7b43
[multi vector] Bound MinMax tile scratch and avoid small allocations

juchen-ms (partychen)
and
Copilot
committed
ff9a92d
[multi vector] Unify grouped MinMax kernels

juchen-ms (partychen)
and
Copilot
committed
ed75906
[multi vector] Reuse decoded MinMax B tiles

juchen-ms (partychen)
and
Copilot
committed
fc0ee46
[multi vector] Simplify MinMax panel construction

juchen-ms (partychen)
and
Copilot
committed
817e927
[multi vector] Group B source with its stride

juchen-ms (partychen)
and
Copilot
committed
2e4b053

Commits on Sep 16, 2026

[multi vector] Simplify MinMax tiled execution

juchen-ms (partychen)
and
Copilot
committed
3589dc9
Add standard license headers

juchen-ms (partychen)
and
Copilot
committed
3ca0241

Commits on Sep 17, 2026

Refactor MinMax layout and SIMD decoding

juchen-ms (partychen)
and
Copilot
committed
f588b7e
Remove redundant MinMax module indirection

juchen-ms (partychen)
and
Copilot
committed
817c84e

Commits on Sep 18, 2026

Fix Scalar MinMax nibble decoding regression

juchen-ms (partychen)
and
Copilot
committed
fb682b3
update
juchen-ms (partychen)
committed
5a43ff9

Commits on Sep 20, 2026

Merge main and fix MinMax bounds lint expectation

juchen-ms (partychen)
and
Copilot
committed
c6b03fd

Commits on Sep 24, 2026

[multi vector] Support any k in packed views

juchen-ms (partychen)
and
Copilot
committed
8dd1752
[multi vector] Share one block-transposed index helper

juchen-ms (partychen)
and
Copilot
committed
babca7c
[multi vector] Use 16x6 register blocking for MinMax V3

juchen-ms (partychen)
and
Copilot
committed
d9d4379
[multi vector] Use 16-bit products in the MinMax Scalar kernel

juchen-ms (partychen)
and
Copilot
committed
683216d

Commits on Sep 28, 2026

Merge main into tiled MinMax MaxSim

juchen-ms (partychen)
and
Copilot
committed
dfefe31

Commits on Sep 29, 2026