Blocked transform#5

Closed

ahurta92 wants to merge 23 commits into

devreal:mainfrom

ahurta92:blocked-transform

Commits on Apr 15, 2026

starting fresh
ahurta92
committed

Commits on Apr 16, 2026

Add multi-level transform benchmarks and validation tests
ahurta92
committed
Update .clangd configuration to enhance HIP file parsing and include necessary flags
ahurta92
committed

Commits on Apr 22, 2026

Add block-distributed 3D transform reference

ahurta92
and
claude
committed
Right-multiply pass 3 to eliminate the post-transform un-shuffle

ahurta92
and
claude
committed
Add L7: block-distributed GPU transform (scalar, single-buffer)

ahurta92
and
claude
committed
L7: add MFMA path and document actual GFX90A f64 16x16x4 layout

ahurta92
and
claude
committed
L7: pad per-wave LDS region to kill 16-way bank conflicts

ahurta92
and
claude
committed
L7: drop barriers that only synced within-wave same-region accesses

ahurta92
and
claude
committed
L7: fuse pass-2 store with corner-turn write

ahurta92
and
claude
committed
L7: swap pass-1 MFMA operands so its output feeds pass 2 in-register

ahurta92
and
claude
committed
L7: cooperative coalesced distribute (2.1-2.7x perf jump)

ahurta92
and
claude
committed
L7: one block per tensor (grid = nfuncs), drop the cube loop

ahurta92
and
claude
committed
L7: use double4 for distribute to emit 128-bit HBM loads

ahurta92
and
claude
committed
L7: wide double4 loads for B cache

ahurta92
and
claude
committed
BLOCKED_TRANSFORM: add measured perf table and lessons learned

ahurta92
and
claude
committed

Commits on Apr 23, 2026