perf(bench): report host cycles per emulated instruction - #258

Open

doublegate wants to merge 24 commits into

mainfrom

perf/cycles-per-instruction-metric

Commits on Aug 1, 2026

perf(bench): report host cycles per emulated instruction
doublegate
committed
c54d35f
perf(bench): measure a reference emulator on the same host and ROM
doublegate
committed
1cc7dce
perf: profile the fast path, and refute two of the plan's own premises
doublegate
committed
8bb56fc
perf(cpu): census what a direct commit path could actually serve
doublegate
committed
f356e14
perf(cpu): commit the simple 96.81% without building a Latch — 1.22x
doublegate
committed
07c6c63

Commits on Aug 2, 2026

docs(ledger): R-24 — the RDP never writes the color buffer's coverage plane
doublegate
committed
cae5538
fix(rdp): store coverage where the VI reads it back (R-24)
doublegate
committed
d618885
test(rdp): .rvec v3 carries the oracle's hidden coverage plane (R-24)
doublegate
committed
c56ea3f
fix(rdp): the cycle type decides a flat triangle's path, not the shade block
doublegate
committed
27c0374
test(rdp): R-24's coverage is oracle-pinned, and defect 1b did not exist
doublegate
committed
911d389
perf: census what an event scheduler could actually skip — RSP ceiling 1.096x
doublegate
committed
dbdc164
perf: the event scheduler measures neutral — and its ceiling was wrong
doublegate
committed
4c3482e
perf(vi): memoize the coverage fetch, not just the filtered pixel — 1.119x

doublegate
and
claude
committed
6899245
docs(perf): re-derive B2's ceiling by measurement — 2.6-2.7%, not 5.3%

doublegate
and
claude
committed
a9473ee
perf(cpu): skip the N64 idle loop in fast-exec — 1.59x to 2.05x

doublegate
and
claude
committed
fce7153
docs(perf): re-derive the declined backlog against a frame half the size

doublegate
and
claude
committed
860f32d
docs(adr): decide A3, A4 and B2 — 0018, 0019, and 0020 amending 0016

doublegate
and
claude
committed
a2b76b1
docs(perf): size the multi-iteration idle skip at 1.156x, and re-order the queue

doublegate
and
claude
committed
c7b3c9a
docs(perf): the easy multi-iteration idle skip is neutral, and that locates the cost

doublegate
and
claude
committed
1e34ebc
docs(perf): the idle path's interrupt check cannot be bisected by removal

doublegate
and
claude
committed
c0dd5dc
perf(sched): batch idle-loop pairs — 1.064x, and fix the A-B-A harness

doublegate
and
claude
committed
fd7ddfa
perf(cop0): advance `Random` in constant time for the idle batch — 1.028x/1.057x

doublegate
and
claude
committed
28ab3f5
feat(frontend): show FPS in the status bar instead of master ticks

doublegate
and
claude
committed
ff42150
chore(salvaged): rescue 30 hand-written dev artifacts from a volatile /tmp

doublegate
and
claude
committed
b8b7f2b

Navigation Menu

perf(bench): report host cycles per emulated instruction - #258

perf(bench): report host cycles per emulated instruction#258
doublegate wants to merge 24 commits into
mainfrom
perf/cycles-per-instruction-metric

Commits on Aug 1, 2026

perf(bench): report host cycles per emulated instruction

perf(bench): measure a reference emulator on the same host and ROM

perf: profile the fast path, and refute two of the plan's own premises

perf(cpu): census what a direct commit path could actually serve

perf(cpu): commit the simple 96.81% without building a Latch — 1.22x

Commits on Aug 2, 2026

docs(ledger): R-24 — the RDP never writes the color buffer's coverage plane

fix(rdp): store coverage where the VI reads it back (R-24)

test(rdp): .rvec v3 carries the oracle's hidden coverage plane (R-24)

fix(rdp): the cycle type decides a flat triangle's path, not the shade block

test(rdp): R-24's coverage is oracle-pinned, and defect 1b did not exist

perf: census what an event scheduler could actually skip — RSP ceiling 1.096x

perf: the event scheduler measures neutral — and its ceiling was wrong

perf(vi): memoize the coverage fetch, not just the filtered pixel — 1.119x

docs(perf): re-derive B2's ceiling by measurement — 2.6-2.7%, not 5.3%

perf(cpu): skip the N64 idle loop in fast-exec — 1.59x to 2.05x

docs(perf): re-derive the declined backlog against a frame half the size

docs(adr): decide A3, A4 and B2 — 0018, 0019, and 0020 amending 0016

docs(perf): size the multi-iteration idle skip at 1.156x, and re-order the queue

docs(perf): the easy multi-iteration idle skip is neutral, and that locates the cost

docs(perf): the idle path's interrupt check cannot be bisected by removal

perf(sched): batch idle-loop pairs — 1.064x, and fix the A-B-A harness

perf(cop0): advance `Random` in constant time for the idle batch — 1.028x/1.057x

feat(frontend): show FPS in the status bar instead of master ticks

chore(salvaged): rescue 30 hand-written dev artifacts from a volatile /tmp