optimize Fp2::add/sub

4 files changed, 44 insertions, 22 deletions

diff --git a/include/mcl/fp_tower.hpp b/include/mcl/fp_tower.hpp
index 5424b37..116bd4b 100644
--- a/include/mcl/fp_tower.hpp
+++ b/include/mcl/fp_tower.hpp
@@ -237,8 +237,8 @@ public:
 		a = a_;
 		b = b_;
 	}
-	static void add(Fp2T& z, const Fp2T& x, const Fp2T& y) { Fp::op_.fp2_add(z.a.v_, x.a.v_, y.a.v_); }
-	static void sub(Fp2T& z, const Fp2T& x, const Fp2T& y) { Fp::op_.fp2_sub(z.a.v_, x.a.v_, y.a.v_); }
+	static void (*add)(Fp2T& z, const Fp2T& x, const Fp2T& y);
+	static void (*sub)(Fp2T& z, const Fp2T& x, const Fp2T& y);
 	static void addPre(Fp2T& z, const Fp2T& x, const Fp2T& y) { Fp::addPre(z.a, x.a, y.a); Fp::addPre(z.b, x.b, y.b); }
 	static void mul(Fp2T& z, const Fp2T& x, const Fp2T& y) { Fp::op_.fp2_mul(z.a.v_, x.a.v_, y.a.v_); }
 	static void inv(Fp2T& y, const Fp2T& x) { Fp::op_.fp2_inv(y.a.v_, x.a.v_); }
@@ -375,8 +375,10 @@ public:
 //		assert(Fp::maxSize <= 256);
 		xi_a_ = xi_a;
 		mcl::fp::Op& op = Fp::op_;
-		op.fp2_add = fp2_addW;
-		op.fp2_sub = fp2_subW;
+		add = (void (*)(Fp2T& z, const Fp2T& x, const Fp2T& y))op.fp2_addA_;
+		if (add == 0) add = fp2_addC;
+		sub = (void (*)(Fp2T& z, const Fp2T& x, const Fp2T& y))op.fp2_subA_;
+		if (sub == 0) sub = fp2_subC;
 		if (op.fp2Dbl_mulPre == 0) {
 			if (op.isFullBit) {
 				op.fp2Dbl_mulPre = FpDblT<Fp>::fp2Dbl_mulPreW;
@@ -475,21 +477,15 @@ private:
 		default Fp2T operator
 		Fp2T = Fp[i]/(i^2 + 1)
 	*/
-	static void fp2_addW(Unit *z, const Unit *x, const Unit *y)
+	static void fp2_addC(Fp2T& z, const Fp2T& x, const Fp2T& y)
 	{
-		const Fp *px = reinterpret_cast<const Fp*>(x);
-		const Fp *py = reinterpret_cast<const Fp*>(y);
-		Fp *pz = reinterpret_cast<Fp*>(z);
-		Fp::add(pz[0], px[0], py[0]);
-		Fp::add(pz[1], px[1], py[1]);
+		Fp::add(z.a, x.a, y.a);
+		Fp::add(z.b, x.b, y.b);
 	}
-	static void fp2_subW(Unit *z, const Unit *x, const Unit *y)
+	static void fp2_subC(Fp2T& z, const Fp2T& x, const Fp2T& y)
 	{
-		const Fp *px = reinterpret_cast<const Fp*>(x);
-		const Fp *py = reinterpret_cast<const Fp*>(y);
-		Fp *pz = reinterpret_cast<Fp*>(z);
-		Fp::sub(pz[0], px[0], py[0]);
-		Fp::sub(pz[1], px[1], py[1]);
+		Fp::sub(z.a, x.a, y.a);
+		Fp::sub(z.b, x.b, y.b);
 	}
 	static void fp2_negW(Unit *y, const Unit *x)
 	{
@@ -621,6 +617,8 @@ private:
 	}
 };
 
+template<class Fp_> void (*Fp2T<Fp_>::add)(Fp2T& z, const Fp2T& x, const Fp2T& y);
+template<class Fp_> void (*Fp2T<Fp_>::sub)(Fp2T& z, const Fp2T& x, const Fp2T& y);
 template<class Fp_> void (*Fp2T<Fp_>::sqr)(Fp2T& y, const Fp2T& x);
 
 template<class Fp>
diff --git a/include/mcl/op.hpp b/include/mcl/op.hpp
index 460c685..69d2987 100644
--- a/include/mcl/op.hpp
+++ b/include/mcl/op.hpp
@@ -183,6 +183,8 @@ struct Op {
 	void3u fp_subA_;
 	void3u fp_mulA_;
 	void2u fp_sqrA_;
+	void3u fp2_addA_;
+	void3u fp2_subA_;
 	void2u fp2_sqrA_;
 	size_t maxN;
 	size_t N;
@@ -219,7 +221,6 @@ struct Op {
 		x = a + bu
 	*/
 	int xi_a; // xi = xi_a + u
-	void3u fp2_add;
 	void3u fp2_sub;
 	void3u fp2_mul;
 	void4u fp2_mulNF;
@@ -264,6 +265,8 @@ struct Op {
 		fp_subA_ = 0;
 		fp_mulA_ = 0;
 		fp_sqrA_ = 0;
+		fp2_addA_ = 0;
+		fp2_subA_ = 0;
 		fp2_sqrA_ = 0;
 		maxN = 0;
 		N = 0;
@@ -297,7 +300,6 @@ struct Op {
 		fp2Dbl_mulPre = 0;
 
 		xi_a = 0;
-		fp2_add = 0;
 		fp2_sub = 0;
 		fp2_mul = 0;
 		fp2_mulNF = 0;
diff --git a/src/fp_generator.hpp b/src/fp_generator.hpp
index c570548..3c79700 100644
--- a/src/fp_generator.hpp
+++ b/src/fp_generator.hpp
@@ -370,14 +370,20 @@ struct Code : Xbyak::CodeGenerator {
 		}
 		if (op.N == 4 && !isFullBit_) {
 			align(16);
+			op.fp2_addA_ = getCurr<void3u>();
+			gen_fp2_add4();
+			align(16);
+			op.fp2_subA_ = getCurr<void3u>();
+			gen_fp2_sub4();
+			align(16);
 			op.fp2Dbl_mulPre = getCurr<void3u>();
 			gen_fp2Dbl_mulPre();
 			align(16);
 			op.fp2_mul = getCurr<void3u>();
-			gen_fp2_mul();
+			gen_fp2_mul4();
 			align(16);
 			op.fp2_sqrA_ = getCurr<void2u>();
-			gen_fp2_sqr();
+			gen_fp2_sqr4();
 		}
 	}
 	void gen_addSubPre(bool isAdd, int n)
@@ -2844,7 +2850,21 @@ private:
 		gen_raw_fp_sub(gp0 + 8 * 4, gp1 + 8 * 4, gp2 + 8 * 4, Pack(gt0, gt1, gt2, gt3, gt4, gt5, gt6, gt7), true);
 	}
 
-	void gen_fp2_mul()
+	void gen_fp2_add4()
+	{
+		assert(!isFullBit_);
+		StackFrame sf(this, 3, 8);
+		gen_raw_fp_add(sf.p[0], sf.p[1], sf.p[2], sf.t, false);
+		gen_raw_fp_add(sf.p[0] + FpByte_, sf.p[1] + FpByte_, sf.p[2] + FpByte_, sf.t, false);
+	}
+	void gen_fp2_sub4()
+	{
+		assert(!isFullBit_);
+		StackFrame sf(this, 3, 8);
+		gen_raw_fp_sub(sf.p[0], sf.p[1], sf.p[2], sf.t, false);
+		gen_raw_fp_sub(sf.p[0] + FpByte_, sf.p[1] + FpByte_, sf.p[2] + FpByte_, sf.t, false);
+	}
+	void gen_fp2_mul4()
 	{
 		assert(!isFullBit_);
 		const RegExp z = rsp + 0 * 8;
@@ -2892,7 +2912,7 @@ private:
 		lea(gp1, ptr[d1]);
 		call(fpDbl_modL_);
 	}
-	void gen_fp2_sqr()
+	void gen_fp2_sqr4()
 	{
 		assert(!isFullBit_);
 		const RegExp y = rsp + 0 * 8;
diff --git a/test/bench.hpp b/test/bench.hpp
index faec4cd..3807bf9 100644
--- a/test/bench.hpp
+++ b/test/bench.hpp
@@ -45,6 +45,7 @@ void testBench(const G1& P, const G2& Q)
 	CYBOZU_BENCH_C("hashAndMapToG2", C, hashAndMapToG2, QQ, "abc", 3);
 #endif
 	CYBOZU_BENCH_C("Fp::add       ", C3, Fp::add, x, x, y);
+	CYBOZU_BENCH_C("Fp::sub       ", C3, Fp::sub, x, x, y);
 	CYBOZU_BENCH_C("Fp::mul       ", C3, Fp::mul, x, x, y);
 	CYBOZU_BENCH_C("Fp::sqr       ", C3, Fp::sqr, x, x);
 	CYBOZU_BENCH_C("Fp::inv       ", C3, Fp::inv, x, x);
@@ -54,6 +55,7 @@ void testBench(const G1& P, const G2& Q)
 	yy.a = y;
 	yy.b = -5;
 	CYBOZU_BENCH_C("Fp2::add      ", C3, Fp2::add, xx, xx, yy);
+	CYBOZU_BENCH_C("Fp2::sub      ", C3, Fp2::sub, xx, xx, yy);
 	CYBOZU_BENCH_C("Fp2::mul      ", C3, Fp2::mul, xx, xx, yy);
 	CYBOZU_BENCH_C("Fp2::sqr      ", C3, Fp2::sqr, xx, xx);
 	CYBOZU_BENCH_C("Fp2::inv      ", C3, Fp2::inv, xx, xx);