Skip to content

Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing#1029

Open
panditsa wants to merge 21 commits intoiree-org:mainfrom
panditsa:sanket/mxfp4_no_valu
Open

Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing#1029
panditsa wants to merge 21 commits intoiree-org:mainfrom
panditsa:sanket/mxfp4_no_valu

Conversation

@panditsa
Copy link
Contributor

@panditsa panditsa commented Mar 3, 2026 •
edited
Loading

  • Hoist per-lane byte offsets before the K-loop and advance K via scalar soffset, eliminating VALU address ops from the loop body
  • Prove stride constancy symbolically: substitute IV = step * j, simplify floor/Mod collapse, extract coeff(j) — valid for all thread/wave/workgroup values without numeric probing
  • Apply to both handle_read (buffer_load) and handle_gather_to_lds (buffer_load_dword_lds) with shared linearization tail

@panditsa panditsa force-pushed the sanket/mxfp4_no_valu branch from 3cb2ebd to 8bdf94a Compare March 3, 2026 18:42
@panditsa panditsa marked this pull request as ready for review March 3, 2026 21:25
@panditsa panditsa force-pushed the sanket/mxfp4_no_valu branch from 71f963c to 4e00fae Compare March 4, 2026 18:47
panditsa added 17 commits March 4, 2026 17:08
@panditsa
working: no valu
b0fdbad 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
remove dead code
1e5dd0e 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
gather_to_lds index splitting
59aea46 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
unify splitting logic
ccd1dab 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
do it purely symbolically
48d8735 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
just get the coefficient
b74e5b3 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
remove wave_id changes
d423191 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
relax conditions
ffb496c 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
fix lit tests
31f4b33 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
fixing lit tests
65d0f6c 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
fixing more lit tests
d830efc 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
fix last test
80f9c1b 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
refactor gather to lds
893b661 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
refactor gather to lds
4f1817c 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
gaurd against buffer_ops, we don't support yet
4011012 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
only iv_split when one induction variable
305de05 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
some cleanup
091aab7 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa panditsa force-pushed the sanket/mxfp4_no_valu branch from e1709f4 to 8744a2b Compare March 4, 2026 23:09
@panditsa panditsa changed the title Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing WIP: Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing Mar 4, 2026
@panditsa panditsa force-pushed the sanket/mxfp4_no_valu branch from 95fe0a0 to 091aab7 Compare March 5, 2026 04:46
panditsa added 4 commits March 4, 2026 22:46
@panditsa
schedule fix
b5be633 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
enable buffer_ops
b378f2b 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
add comments
0f7766c 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa
bump llvm
43b1ab7 
Signed-off-by: Sanket Pandit <sanket.pandit@amd.com>
@panditsa panditsa changed the title WIP: Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing Zero in-loop VALU for MXFP4 preshuffle GEMM via IV-split addressing Mar 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@panditsa