[lemire]

Closes #30.

Adds Sirius64 by Matteo Zapparoli (Apache-2.0): a 64-bit state, 64-bit output PRNG. The state is a Weyl sequence with the golden-ratio increment (same as splitmix64) and the output function is a multiply-xorshift-rotate mix:

uint64_t z = (*s += 0x9e3779b97f4a7c15);
z = 0x9e3779b97f4a7c15 * (z ^ (z >> 17));
z = (z << 32) | (z >> 32);
return 0x9e3779b97f4a7c15 * ((*s) ^ z ^ (z >> 17));

Changes

• source/sirius64.h — generator, following the wyrand.h/w1rand.h pattern (sirius64_seed, sirius64, sirius64_stateless, sirius64_cast32). Verified bit-for-bit against the upstream reference over 1M outputs for several seeds.
• practrand/src/testsirius64.c, testu01/src/testsirius64.c — harnesses (with -H support).
• Registered in speed/src/rng.cpp, practrand/CMakeLists.txt, testu01/CMakeLists.txt, entropy/CMakeLists.txt, practrand/runtests.sh, testu01/testlist.sh.
• README: added to the list of contenders.

Results so far (Apple Silicon, arm64)

Speed (./build/speed/rng, no perf counters):

splitmix64   :   0.11 ns/byte   8.97 GB/s
wyrand       :   0.11 ns/byte   8.78 GB/s
sirius64     :   0.11 ns/byte   8.98 GB/s

TestU01 SmallCrush: all tests passed for lsb, msb (-H), bit-reversed (-z) and byte-reversed msb (-r -H).

PractRand 0.93: clean through 8 GB on both lsb and msb (only "unusual" p-values). Full 512 GB runs and BigCrush (lsb and msb) are in progress; I'll push the logs to practrand/results/ and testu01/results/ when they finish.

[lemire]

PractRand 0.93 at 512 GB: both lsb (testsirius64) and msb (testsirius64 -H) pass with no FAIL or suspicious results, only a handful of "unusual" p-values. Logs pushed to practrand/results/. Added a row to the visual summary; TestU01 column left blank pending BigCrush (running).

[lemire]

TestU01 BigCrush (default seed, default bit order), ~66 min CPU each:

• testsirius64 -b (lsb 32 bits): 159/160 statistics in range; one marginal p-value, CouponCollector, r = 27 at 8.4e-4. This is the same kind of lone borderline value that shows up in the stored splitmix64/pcg64/aesctr logs, not a failure.
• testsirius64 -H -b (msb 32 bits): all tests were passed.

Logs pushed to testu01/results/. I marked the TestU01 column as passing in the visual summary, but note that only 2 of the 24 configurations run by bigcrush.sh (4 seeds × 3 bit orders × lsb/msb) were run here; the remaining ones would take about a day of CPU.