Message-schedule and consumer-round fusion
Moving W18…W61 assignments directly before their first consuming rounds reduces live ranges, registers or instructions.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
Moving W18…W61 assignments directly before their first consuming rounds reduces live ranges, registers or instructions.
Why the test is meaningful
Manual scheduling is useful only if the compiler does not already perform equivalent dependency and lifetime management.
W_t generated immediately before round thash order and equations unchangedbenefit requires lower SASS/registers and faster timingHow it was tested
Use a deterministic source generator for 44 moves and evaluate four fresh headers in 32 balanced runs.
What happened
Baseline and candidate emitted identical SASS: 2,507 instructions, 1,176 SHF, 629 LOP3, 408 IADD3 and 44 registers. Candidate lost 4/4 with kernel ratio 0.995602.
Exactness and statistical controls
A checksum line-ending issue invalidated the first attempt; checksums were then taken from written bytes and the whole frozen campaign repeated.
What the result means
ptxas already manages schedule lifetimes; manual source fusion is closed.
Limitations
- The result concerns this compiler and source body.
- Other schedule algorithms are separate mechanisms.
- Identical SASS makes the source rewrite non-causal.
Evidence trail
Generate the same moves deterministically, checksum written bytes, compare disassembly and rerun balanced timing.
Canonical variants
CANONICAL-EXP-143PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.