[DO NOT MERGE] Yanwen/release14 quant shaders - #13

Open

xuyanwen2012 wants to merge 28 commits into

release/1.4from

yanwen/release14-quant-shaders

Commits on Aug 14, 2026

Add ET_VK_EXECUTE_NODE_THRESHOLD opt-in GPU-watchdog workaround
xuyanwen2012
committed
a9038c4

Commits on Aug 17, 2026

Fix linear_dq8ca_q4gsw wrong output when M > K

SS-JIA
authored and
xuyanwen2012
committed
745a738

Commits on Aug 18, 2026

[ET-VK] Port dev's shipped 4w/8da4w coopmat tiles + texture-IO WMMA to release/1.4
xuyanwen2012
committed
f60080d
[ET-VK] Ship 8da4w texture-IO tile sweep winner as the new default
xuyanwen2012
committed
26f4557
[ET-VK] Fix q4gsw texture-IO default crash: give it a default variant too
xuyanwen2012
committed
3ceeefc
[ET-VK] Revert 8da4w default to t64x32k32g12s64 -- new tile was numerically wrong
xuyanwen2012
committed
0320ac3
Document flaky coopmat correctness failure, keep 8da4w default unchanged
xuyanwen2012
committed
bcba76d
[ET-VK] Port SDPA cooperative-matrix (WMMA) shaders to release/1.4
xuyanwen2012
committed
8be6567
[ET-VK] Add SDPA coopmat correctness harness (--sdpa-correctness-only)
xuyanwen2012
committed
d7360b7

Commits on Aug 21, 2026

[ET-VK] Port shmem_double_buf4-tr's coopmat-staged A to 8da4w (UNVALIDATED)
xuyanwen2012
committed
7de604d

Commits on Aug 25, 2026

[ET-VK] Cut 8da4w coopmat linear time 33.6% by hoisting the dequant correction
xuyanwen2012
committed
a11830e

Commits on Aug 26, 2026

[ET-VK] Speed up 8da4w coopmat linear: bigger tile, drop B skew, coalesce B store
xuyanwen2012
committed
8cde63e

Commits on Aug 28, 2026

Commits on Sep 2, 2026

[ET-VK] Promote coopMat-staged-A ("-tr" on zpg) as 8da4w coopmat default
xuyanwen2012
committed
009aa29

Commits on Sep 3, 2026

Commits on Sep 8, 2026

[ET-VK] Enable WMMA by default: texture coopmat no longer needs an env var
xuyanwen2012
committed
921b0cf

Commits on Sep 9, 2026

[ET-VK] Six opt-in dq8ca coopmat variants: A-layout and B-staging, all rejected
xuyanwen2012
committed
ce7b22f

Commits on Sep 15, 2026