Repository navigation

[HIP] [FlyDSL] [JIT] Feat/topk per row gfx950 - #5686

Open

MHYangAMD wants to merge 157 commits into

ROCm:mainfrom

AFDEAPAC:feat/topk-per-row-gfx950-avo

Open

[HIP] [FlyDSL] [JIT] Feat/topk per row gfx950#5686
MHYangAMD wants to merge 157 commits into
ROCm:mainfrom
AFDEAPAC:feat/topk-per-row-gfx950-avo

Commits on Sep 21, 2026
Commits on Sep 22, 2026
- Optimize terminal-window stream top-k
  ClementLinCF
  committed
  7d9c2d1
- Optimize 32K plain radix tail
  ClementLinCF
  committed
  80973ce
- style: black the lds_plan call that 7d9c2d128 left unformatted
  MHYangAMD
  committed
  f74510c
- topk: stop rebuilding the sampled dispatch row bounds on every call
  MHYangAMD
  committed
  438e832
- topk_shape: require cap headroom when the S-rule picks a sample count
  MHYangAMD
  committed
  00826e4
- topk_select: route sampled by width, not only by rows*width
  MHYangAMD
  committed
  cd486fb
- Optimize dense near-full stream top-k
  ClementLinCF
  committed
  c49762e
- Generalize dense small-reject stream top-k
  ClementLinCF
  committed
  3a3d00a
- Guard zero-reject stream specialization
  ClementLinCF
  committed
  d6c27b5
Commits on Sep 23, 2026
Commits on Sep 24, 2026
Commits on Sep 25, 2026
- Route measured k2048 short rows to plain
  ClementLinCF
  committed
  3f7a067
- Use 32-bit LDS-tail sample quotient
  ClementLinCF
  committed
  615cbdd
- Vectorize LDS-tail histogram clears
  ClementLinCF
  committed
  2f23145
- Use DPP for cross-wave bucket scan
  ClementLinCF
  committed
  9563347
Commits on Sep 29, 2026
- Optimize gfx950 plain top-k specializations
  ClementLinCF
  committed
  b8b8918
- Prefer general paths over static-only 4K dispatches
  ClementLinCF
  committed
  cbcf442
- Remove remaining static-only 4K extents
  ClementLinCF
  committed
  61249e7
- Make plain top-k row lengths runtime-only
  ClementLinCF
  committed
  85f85df
- topk sampled: freeze round-3 v6 baseline
  MHYangAMD
  committed
  efa060b
- Optimize sampled phase A/C residency on gfx950
  MHYangAMD
  committed
  54e35d9
Commits on Sep 30, 2026
Commits on Oct 1, 2026
Commits on Oct 2, 2026
Commits on Oct 3, 2026
- topk plain: stay one-block through 80K columns for k=2048 on gfx950
  ClementLinCF
  committed
  018b7dc
- topk_select: yield k=2048 plain to small-reject only for 1-2 rejects
  ClementLinCF
  committed
  d41159e
Commits on Oct 4, 2026
- topk_select: route k=2048 rows >= 3000 at 131072..131840 columns to plain
  ClementLinCF
  committed
  534f5e7
- topk_select: write whole-row k=2048 selections in one launch on gfx950
  ClementLinCF
  committed
  c305f49
Commits on Oct 5, 2026
- topk plain: predict the >80K LDS-tail threshold from a 4096-value slice
  ClementLinCF
  committed
  08352b1
- op_tests/test_topk_select: add --include-large for shapes of 4 GiB or more
  ClementLinCF
  committed
  9144f77
Commits on Oct 6, 2026
- topk sampled: serve gfx950 only, and check the launches
  MHYangAMD
  committed
  3f67d32
- op_tests/test_topk_select: skip routes this device does not serve
  MHYangAMD
  committed
  812874e
- topk_select: black
  MHYangAMD
  committed
  a3a2ce3
- topk plain: build on gfx942 and gfx1250 again
  MHYangAMD
  committed
  e36e79a
Commits on Oct 7, 2026
- topk plain: route the DPP adds through opus::upd_dpp
  ClementLinCF
  committed
  1884637
- topk sampled: more room in the candidate cap where rows overflowed 4096
  MHYangAMD
  committed
  256c903
- topk sampled: drop the per-source -foffload-uniform-block override
  MHYangAMD
  committed
  e3d3927
- topk sampled: remove dead code and stale comments
  MHYangAMD
  committed
  788a762
- topk prefill: key the sampled floor by arch behind one predicate
  MHYangAMD
  committed
  7854d0c
- topk sampled: route only from k = 96
  MHYangAMD
  committed
  dacf425
- topk plain: drop the unreachable exact-width specializations
  ClementLinCF
  committed
  3431bc4
- topk plain: drop scan_and_choose_bucket
  ClementLinCF
  committed
  34b9744
Commits on Oct 8, 2026
Commits on Oct 12, 2026
Commits on Oct 13, 2026
Commits on Oct 15, 2026
Commits on Oct 19, 2026
Commits on Oct 23, 2026

Commits

Commits on Sep 21, 2026

Commits on Sep 22, 2026

Commits on Sep 23, 2026

Commits on Sep 24, 2026

Commits on Sep 25, 2026

Commits on Sep 29, 2026

Commits on Sep 30, 2026

Commits on Oct 1, 2026

Commits on Oct 2, 2026

Commits on Oct 3, 2026

Commits on Oct 4, 2026

Commits on Oct 5, 2026

Commits on Oct 6, 2026

Commits on Oct 7, 2026

Commits on Oct 8, 2026

Commits on Oct 12, 2026

Commits on Oct 13, 2026

Commits on Oct 15, 2026