Support sub-64 MLEN in the emulator (relax 64-byte HBM burst model)

[booth-algo]

Summary

The compiler already compiles at MLEN<64; the emulator's 64-byte HBM burst model blocked end-to-end runs at MLEN=8/16/32. This makes the emulator + Python HBM writer handle tightly-packed sub-64 rows. No compiler changes required, and every change is a no-op at MLEN≥64.

Designed via a 3-phase analyze→synthesize→adversarially-verify workflow; the verifiers caught that a naive assertion-relaxation was unsound (stride/row-layout mismatch), which led to this layout-consistent approach.

Emulator (transactional_emulator/src/main.rs)

• transfer_mx_from_hbm: relax is_multiple_of(8*64) → is_multiple_of(8) (whole-bytes); rewrite the element read to cover [element_addr, +len) by reading the 64-byte-aligned word(s) and extracting the real bytes for any byte offset / boundary span. The addr.is_multiple_of(64) assert is removed (reads are aligned by construction). At MLEN≥64 it reduces to one aligned read per 64 B.
• transfer_mx_to_hbm: symmetric — relax the store assert and read-modify-write the aligned word(s) so tightly-packed sub-64 neighbours aren't clobbered. At MLEN≥64 each store fills whole 64-byte words (RMW == plain write).

HBM writer (testbench/sim_env_utils.py)

• Pack element rows tightly (hbm_row_bytes = element_row_bytes; drop the 64-byte burst floor) to match the compiler's tight HBM stride + scale offset.
• Drop the per-tensor pad-to-64: it shifted the next tensor past its compiler-assigned base whenever a tensor's size wasn't a 64-multiple (only at sub-64 MLEN, e.g. X(16,16)=288 B → 320), making weight prefetches read a zero region. The emulator's MemoryBacked capacity (64-multiple, ~2× preload) zero-covers the final aligned-block read.

Guards (testbench/aten/configurable.py)

• Remove the MLEN%64 guards in setup_hw / from_args.

Test plan

• End-to-end PASS at MLEN=8 (ffn), MLEN=16 (ffn, linear, softmax, rms_norm, embedding_add, rope), MLEN=32 (ffn, linear)
• No regression: ffn(128), softmax(256), linear(256) still pass
• ruff + cargo fmt clean

Note

Stacked on #60 (configurable MLEN + bugfixes). Rebase onto main after #60 merges.

🤖 Generated with Claude Code