add optimizedSplitForBLS12_381

include/mcl/bn.hpp

@@ -663,18 +663,22 @@ struct GLV1 : mcl::GLV1T<G1, Fr> {
 		(void)b;
 		rw = -(rw + 1) / 2;
 		rBitSize = Fr::getOp().bitSize;
-//		rBitSize = (rBitSize + UnitBitSize - 1) & ~(UnitBitSize - 1);// a little better size
 		if (isBLS12) {
 			/*
 				BLS12
 				L = z^4
 				(-z^2+1) + L = 0
 				1 + z^2 L = 0
 			*/
-			B[0][0] = -z * z + 1;
-			B[0][1] = 1;
-			B[1][0] = 1;
-			B[1][1] = z * z;
+			// only B[0][0] and v0 are used
+			const mpz_class& r = Fr::getOp().mp;
+			B[0][0] = z * z - 1; // L
+			v0 = (B[0][0] << rBitSize) / r;
+			if (curveType == BLS12_381.curveType && MCL_SIZEOF_UNIT == 8) {
+				optimizedSplit = optimizedSplitForBLS12_381;
+			} else {
+				optimizedSplit = splitForBLS12;
+			}
 		} else {
 			/*
 				BN
@@ -686,18 +690,21 @@ struct GLV1 : mcl::GLV1T<G1, Fr> {
 			B[0][1] = -2 * z - 1;
 			B[1][0] = -2 * z - 1;
 			B[1][1] = -6 * z * z - 4 * z - 1;
-		}
-		// [v0 v1] = [r 0] * B^(-1)
-		const mpz_class& r = Fr::getOp().mp;
-		v0 = ((-B[1][1]) << rBitSize) / r;
-		v1 = ((B[1][0]) << rBitSize) / r;
-		if (curveType == BLS12_381.curveType) {
-			optimizedSplit = optimizedSplitForBLS12_381;
-			v0 = -v0;
-			B[0][0] = -B[0][0];
+			// [v0 v1] = [r 0] * B^(-1)
+			const mpz_class& r = Fr::getOp().mp;
+			v0 = ((-B[1][1]) << rBitSize) / r;
+			v1 = ((B[1][0]) << rBitSize) / r;
 		}
 	}
 	// x = (a + b L) mod r
+	static inline void splitForBLS12(mpz_class u[2], const mpz_class& x)
+	{
+		mpz_class& a = u[0];
+		mpz_class& b = u[1];
+		mpz_class t;
+		b = (x * v0) >> rBitSize;
+		a = x - b * B[0][0];
+	}
 	static inline void optimizedSplitForBLS12_381(mpz_class u[2], const mpz_class& x)
 	{
 		/*
@@ -709,7 +716,6 @@ struct GLV1 : mcl::GLV1T<G1, Fr> {
 		*/
 		mpz_class& a = u[0];
 		mpz_class& b = u[1];
-#if MCL_SIZEOF_UNIT == 8
 		static const uint64_t Lv[] = { 0x00000000ffffffff, 0xac45a4010001a402 };
 		static const uint64_t vv[] = { 0xb1fb72917b67f718, 0xbe35f678f00fd56e };
 		static const size_t n = 128 / mcl::UnitBitSize;
@@ -722,11 +728,6 @@ struct GLV1 : mcl::GLV1T<G1, Fr> {
 		mcl::bint::subT<n>(t, x.getUnit(), t);
 		a.setArray(&dummy, t, n);
 		(void)dummy;
-#else
-		mpz_class t;
-		b = (x * v0) >> 255;
-		a = x - b * B[0][0];
-#endif
 	}
 };
 

misc/internal.md

@@ -7,13 +7,15 @@
 ### Definition of parameters
 
 ```python
+M = 1<<256
+H = 1<<128
 z = -0xd201000000010000
 L = z*z - 1
 r = L*L + L + 1
 s = r.bit_length()
 S = 1<<s
-v = (L+1)*S // r
-r0 = (L+1)*S % r
+v = S // L
+r0 = S % L
 ```
 
 variables|z|L|r|S|v
@@ -28,34 +30,25 @@ def split(x):
     a = x - b * L
     return (a, b)
 ```
-
-- x in [0, r-1].
 - a + b L = x for (a, b) = split(x).
 
 ### Theorem
-a, b < 2^128 for all x in [0, r-1].
+0 <= a, b < H for all x in [0, M-r].
 
 ### Proof
 
 ```
-Let r0 := (L+1)S % r, then (L+1)S=v r + r0 and r0 in [0, r-1].
+Let r0 := L S % r, then S=v L + r0 and r0 in [0, L-1].
 Let r1 := x v % S, then x v = b S + r1 and r1 in [0, S-1].
 ```
 
 ```
-b = floor((xv)/S) <= (xv)/S = (xvr)/(Sr) = x((L+1)S - r0) / (Sr) = (x/r)(L+1) - (r0/Sr)
-  <= (L+1) < 2^128.
+b <=  xv / S < (M-r) (S/L)/S = (M-r)/L < H.
 ```
 
 ```
-arS = (x - bL)rS = xrS - bSrL = xrS - (xv - r1)rL = xrS - x(vr)L + r1 rL
-    = xrS - xL((L+1)S - r0) + r1 rL = xrS - xL(L+1)S + x r0 L + r1 rL
-    = xS(r - L^2-L) + (x r0 + r1 r)L
-    = x(S + r0 L) + r1 r L.
-```
-Then,
-```
-a = (x/r)(1 + (r0/S)L) + (r1/r)L
-  < 1 + (r0/S)L + L.
+aS = (x - bL)S = xS - bSL = xS - (xv - r1)L = x(S - vL) + r1 L = r0 x + r1 L
+     <= r0 (M-r) + (S-1)L < S H.
 ```
-r0/S < 0.10017, then a < 2^128.
+Then, a < H.
+So for x in [0, M-1], set x = x - r if x >= H and apply split() to x.