new compressor & finisher

N-R-K · N-R-K · commit 55ecfcff372e · 2024-11-29T02:25:35.000Z
diff --git a/README.md b/README.md
@@ -5,7 +5,7 @@ Things may change at any point without notice.
 
 Some major improvement compared to v1:
 
-- Faster performance on short string (49 cycles/hash vs 35 cycles/hash).
+- Faster performance on short string (49 cycles/hash vs 34 cycles/hash).
   The tail end handling has been reworked entirely with some inspiration from
   wyhash's short input reading.
 - Better seeding. v1 seed only affected 64 bits of the initial state.
@@ -17,8 +17,6 @@ Some major improvement compared to v1:
 Avenue for improvement:
 
 - Faster bulk handling without using 128 bit multiplication.
-- Investigate better/faster compressor for the 256 bit -> 64 bit reduction.
-- Drop moremur in favor of using existing multiplier (digits of e) for the finisher ??
 
 ---
 
diff --git a/chibihash64.h b/chibihash64.h
@@ -22,12 +22,6 @@ static inline uint64_t chibihash64__rotl(uint64_t x, int n)
 {
 	return (x << n) | (x >> (-n & 63));
 }
-static inline uint64_t chibihash64__reduce(uint64_t a, uint64_t b)
-{
-	uint64_t x = a * (chibihash64__rotl(b, 9) | 1);
-	x ^= chibihash64__rotl(b, 41) + chibihash64__rotl(a, 31);
-	return x;
-}
 
 static inline uint64_t
 chibihash64(const void *keyIn, ptrdiff_t len, uint64_t seed)
@@ -68,15 +62,19 @@ chibihash64(const void *keyIn, ptrdiff_t len, uint64_t seed)
 		h[3] ^= p[l/2] | ((uint64_t)p[l-1] << 8);
 	}
 
+	h[0] += chibihash64__rotl(h[2] * K, 31) ^ (h[2] >> 31);
+	h[1] += chibihash64__rotl(h[3] * K, 31) ^ (h[3] >> 31);
+	h[0] *= K; h[0] ^= h[0] >> 31;
+	h[1] += h[0];
+
 	uint64_t x = (uint64_t)len * K;
 	x ^= chibihash64__rotl(x, 29);
-	x ^= chibihash64__reduce(h[0], h[1]);
-	x ^= chibihash64__reduce(h[2], h[3]) + seed;
+	x += seed;
+	x ^= h[1];
 
-	// moremur: https://mostlymangling.blogspot.com/2019/12/stronger-better-morer-moremur-better.html
-	x ^= x >> 27; x *= UINT64_C(0x3C79AC492BA7B653);
-	x ^= x >> 33; x *= UINT64_C(0x1C69B3F74AC4AE35);
-	x ^= x >> 27;
+	x ^= chibihash64__rotl(x, 15) ^ chibihash64__rotl(x, 42);
+	x *= K;
+	x ^= chibihash64__rotl(x, 13) ^ chibihash64__rotl(x, 31);
 
 	return x;
 }
