ECM faster create xdbl_tail_uv in marin

Author: cherubrock-seb

include/marin/engine.h

@@ -86,7 +86,27 @@ class engine
 		set_multiplicand(src1, src1);
 		mul(dst1, src1);
 	}
-	
+		
+	virtual void mul_pair_prepared(const Reg dst0, const Reg mul_src0,
+								const Reg dst1, const Reg mul_src1,
+								const uint32 a0 = 1, const uint32 a1 = 1) const
+	{
+		mul(dst0, mul_src0, a0);
+		mul(dst1, mul_src1, a1);
+	}
+
+	virtual void xdbl_tail_uv(const Reg x_out, const Reg z_out,
+							const Reg u_work, const Reg v_reg,
+							const Reg a24_mul,
+							const Reg tmp_e_mul, const Reg tmp_v_mul) const
+	{
+		sub_reg(u_work, v_reg);
+		set_multiplicand(tmp_v_mul, v_reg);
+		mul(x_out, tmp_v_mul);
+		set_multiplicand(tmp_e_mul, u_work);
+		mul_add(u_work, a24_mul, v_reg);
+		mul_copy(u_work, tmp_e_mul, z_out);
+	}
 	virtual void addsub_copy(const Reg sum, const Reg diff, const Reg sum_copy, const Reg diff_copy,
 							const Reg a, const Reg b) const
 	{

include/marin/engine_gpu.h

@@ -1272,7 +1272,37 @@ class engine_gpu : public engine
 	{
 		_gpu->addsub_copy((size_t)sum,(size_t)diff,(size_t)sum_copy,(size_t)diff_copy,(size_t)a,(size_t)b);
 	}
+	void mul_pair_prepared(const Reg rdst0, const Reg rsrc0,
+						const Reg rdst1, const Reg rsrc1,
+						const uint32 a0 = 1, const uint32 a1 = 1) const override
+	{
+		if ((a0 != 1) || (a1 != 1))
+		{
+			mul(rdst0, rsrc0, a0);
+			mul(rdst1, rsrc1, a1);
+			return;
+		}
+
+		const size_t dst0 = size_t(rdst0), src0 = size_t(rsrc0);
+		const size_t dst1 = size_t(rdst1), src1 = size_t(rsrc1);
+
+		mul_new_core(dst0, src0);
+		mul_new_core(dst1, src1);
+		_gpu->carry_weight_mul2_unit(dst0, dst1);
+	}
 
+	void xdbl_tail_uv(const Reg x_out, const Reg z_out,
+					const Reg u_work, const Reg v_reg,
+					const Reg a24_mul,
+					const Reg tmp_e_mul, const Reg tmp_v_mul) const override
+	{
+		sub_reg(u_work, v_reg);
+		set_multiplicand(tmp_e_mul, u_work);
+		set_multiplicand(tmp_v_mul, v_reg);
+		mul_pair_prepared(x_out, tmp_v_mul, u_work, a24_mul);
+		add(u_work, v_reg);
+		mul_copy(u_work, tmp_e_mul, z_out);
+	}
 	size_t get_register_data_size() const override { return _reg_count * _n * sizeof(uint64); }
 
 	bool get_data(std::vector<char> & data, const Reg src) const override

src/modes/RunEcm.cpp

@@ -792,7 +792,7 @@ uint32_t s2_idx = 0, s2_cnt = 0; double s2_et = 0.0;
             eng->mul((engine::Reg)Z2,(engine::Reg)13);               // Z2=k13*...
 
             // xDBL part (U,V,E with fused square+copy; no extra copy for Z1)
-            eng->square_mul_copy((engine::Reg)25,(engine::Reg)X1);   // 25=U=(X1+Z1)^2, and X1=U
+            /*eng->square_mul_copy((engine::Reg)25,(engine::Reg)X1);   // 25=U=(X1+Z1)^2, and X1=U
             eng->square_mul((engine::Reg)24);                        // 24=V=(X1−Z1)^2
             eng->sub_reg((engine::Reg)25,(engine::Reg)24);           // 25=E=U−V
             eng->set_multiplicand((engine::Reg)15,(engine::Reg)24);  // *V
@@ -801,7 +801,13 @@ uint32_t s2_idx = 0, s2_cnt = 0; double s2_et = 0.0;
             //eng->mul((engine::Reg)25,(engine::Reg)12);               // 25=A24*E
             //eng->add((engine::Reg)25,(engine::Reg)24);               // 25=A24*E + V
             eng->mul_add((engine::Reg)25,(engine::Reg)12,(engine::Reg)24);
-            eng->mul_copy((engine::Reg)25,(engine::Reg)15,(engine::Reg)Z1); // Z1=(A24*E+V)*E
+            eng->mul_copy((engine::Reg)25,(engine::Reg)15,(engine::Reg)Z1); // Z1=(A24*E+V)*E*/
+            eng->square_mul_copy((engine::Reg)25,(engine::Reg)X1);
+            eng->square_mul((engine::Reg)24);
+            eng->xdbl_tail_uv((engine::Reg)X1, (engine::Reg)Z1,
+                              (engine::Reg)25, (engine::Reg)24,
+                              (engine::Reg)12,
+                              (engine::Reg)15, (engine::Reg)11);
         };
 
         auto xDBLADD_strict2 = [&](size_t X1,size_t Z1, size_t X2,size_t Z2){