Skip to content

[FPSAN] Fix an issue where tmem writes in WS may get dropped#10000

Merged
pawelszczerbuk merged 3 commits into
triton-lang:mainfrom
pawelszczerbuk:pawel/fpsan_ws_zeros
Apr 13, 2026
Merged

[FPSAN] Fix an issue where tmem writes in WS may get dropped#10000
pawelszczerbuk merged 3 commits into
triton-lang:mainfrom
pawelszczerbuk:pawel/fpsan_ws_zeros

Conversation

@pawelszczerbuk
Copy link
Copy Markdown
Contributor

Fpsan emulates tmem with global scratch. Each tmem buffer is being mapped to a scratch allocation. There was an issue in how the mapping was created for WS partitions, with the same tmem buffer being mapped to distinct scratch allocations. We track the "canonical" tmem allocation now, to make sure that tmem aliases in different regions, even though they are represented by different SSA values, are getting mapped to the original allocation correctly.

@lezcano
Copy link
Copy Markdown
Contributor

lezcano commented Apr 11, 2026

10000!!

@pawelszczerbuk pawelszczerbuk merged commit 249e0ac into triton-lang:main Apr 13, 2026
9 checks passed
plognjen pushed a commit to plognjen/triton that referenced this pull request Apr 14, 2026
…lang#10000)

Fpsan emulates tmem with global scratch. Each tmem buffer is being
mapped to a scratch allocation. There was an issue in how the mapping
was created for WS partitions, with the same tmem buffer being mapped to
distinct scratch allocations. We track the "canonical" tmem allocation
now, to make sure that tmem aliases in different regions, even though
they are represented by different SSA values, are getting mapped to the
original allocation correctly.

---------

Co-authored-by: root <root@codex-gb200-0.brix.pawelszczerbuk.svc.cluster.local>
raymondtay pushed a commit to raymondtay/triton that referenced this pull request Apr 18, 2026
…lang#10000)

Fpsan emulates tmem with global scratch. Each tmem buffer is being
mapped to a scratch allocation. There was an issue in how the mapping
was created for WS partitions, with the same tmem buffer being mapped to
distinct scratch allocations. We track the "canonical" tmem allocation
now, to make sure that tmem aliases in different regions, even though
they are represented by different SSA values, are getting mapped to the
original allocation correctly.

---------

Co-authored-by: root <root@codex-gb200-0.brix.pawelszczerbuk.svc.cluster.local>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants