bulletproofs: speed up the latest changes a bit

7 files changed, 67 insertions, 34 deletions

diff --git a/src/ringct/bulletproofs.cc b/src/ringct/bulletproofs.cc
index 5bcdd968d..165691bfe 100644
--- a/src/ringct/bulletproofs.cc
+++ b/src/ringct/bulletproofs.cc
@@ -92,30 +92,37 @@ static bool is_reduced(const rct::key &scalar)
   return scalar == reduced;
 }
 
-//addKeys3acc_p3
-//aAbB += a*A + b*B where a, b are scalars, A, B are curve points
-//A and B must be input after applying "precomp"
-static void addKeys3acc_p3(ge_p3 *aAbB, const key &a, const ge_dsmp A, const key &b, const ge_dsmp B)
+static void addKeys_acc_p3(ge_p3 *acc_p3, const rct::key &a, const rct::key &point)
 {
-    ge_p3 rv;
+    ge_p3 p3;
+    CHECK_AND_ASSERT_THROW_MES(ge_frombytes_vartime(&p3, point.bytes) == 0, "ge_frombytes_vartime failed");
+    ge_scalarmult_p3(&p3, a.bytes, &p3);
+    ge_cached cached;
+    ge_p3_to_cached(&cached, acc_p3);
     ge_p1p1 p1;
-    ge_p2 p2;
-    ge_double_scalarmult_precomp_vartime2_p3(&rv, a.bytes, A, b.bytes, B);
+    ge_add(&p1, &p3, &cached);
+    ge_p1p1_to_p3(acc_p3, &p1);
+}
+
+static void add_acc_p3(ge_p3 *acc_p3, const rct::key &point)
+{
+    ge_p3 p3;
+    CHECK_AND_ASSERT_THROW_MES(ge_frombytes_vartime(&p3, point.bytes) == 0, "ge_frombytes_vartime failed");
     ge_cached cached;
-    ge_p3_to_cached(&cached, aAbB);
-    ge_add(&p1, &rv, &cached);
-    ge_p1p1_to_p3(aAbB, &p1);
+    ge_p3_to_cached(&cached, &p3);
+    ge_p1p1 p1;
+    ge_add(&p1, acc_p3, &cached);
+    ge_p1p1_to_p3(acc_p3, &p1);
 }
 
-static void addKeys_acc_p3(ge_p3 *acc_p3, const rct::key &a, const rct::key &point)
+static void sub_acc_p3(ge_p3 *acc_p3, const rct::key &point)
 {
     ge_p3 p3;
     CHECK_AND_ASSERT_THROW_MES(ge_frombytes_vartime(&p3, point.bytes) == 0, "ge_frombytes_vartime failed");
-    ge_scalarmult_p3(&p3, a.bytes, &p3);
     ge_cached cached;
-    ge_p3_to_cached(&cached, acc_p3);
+    ge_p3_to_cached(&cached, &p3);
     ge_p1p1 p1;
-    ge_add(&p1, &p3, &cached);
+    ge_sub(&p1, acc_p3, &cached);
     ge_p1p1_to_p3(acc_p3, &p1);
 }
 
@@ -1087,9 +1094,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
       sc_mul(tmp.bytes, zpow[j+2].bytes, EIGHT.bytes);
       multiexp_data.emplace_back(tmp, proof.V[j]);
     }
-    rct::key temp = multiexp(multiexp_data, false);
-
-    rct::addKeys(Y2, Y2, rct::scalarmultKey(temp, weight));
+    rct::addKeys(Y2, Y2, rct::scalarmultKey(multiexp(multiexp_data, false), weight));
     rct::key weight8;
     sc_mul(weight8.bytes, weight.bytes, EIGHT.bytes);
     sc_mul(tmp.bytes, x.bytes, weight8.bytes);
@@ -1103,7 +1108,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
     PERF_TIMER_START_BP(VERIFY_line_62);
     // PAPER LINE 62
     sc_mul(tmp.bytes, x.bytes, EIGHT.bytes);
-    rct::addKeys(Z0, Z0, rct::scalarmultKey(rct::addKeys(rct::scalarmultKey(proof.A, EIGHT), rct::scalarmultKey(proof.S, tmp)), weight));
+    rct::addKeys(Z0, Z0, rct::scalarmultKey(rct::addKeys(rct::scalarmult8(proof.A), rct::scalarmultKey(proof.S, tmp)), weight));
     PERF_TIMER_STOP(VERIFY_line_62);
 
     // Compute the number of rounds for the inner product
@@ -1202,23 +1207,22 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
 
   // now check all proofs at once
   PERF_TIMER_START_BP(VERIFY_step2_check);
-  rct::key check1 = rct::identity();
-  rct::addKeys(check1, check1, rct::scalarmultBase(y0));
-  rct::addKeys(check1, check1, rct::scalarmultH(y1));
-  rct::subKeys(check1, check1, Y2);
-  rct::subKeys(check1, check1, Y3);
-  rct::subKeys(check1, check1, Y4);
-  if (!(check1 == rct::identity()))
+  ge_p3 check1;
+  ge_scalarmult_base(&check1, y0.bytes);
+  addKeys_acc_p3(&check1, y1, rct::H);
+  sub_acc_p3(&check1, Y2);
+  sub_acc_p3(&check1, Y3);
+  sub_acc_p3(&check1, Y4);
+  if (!ge_p3_is_point_at_infinity(&check1))
   {
     MERROR("Verification failure at step 1");
     return false;
   }
-  rct::key check2 = rct::identity();
-  rct::addKeys(check2, check2, Z0);
+  ge_p3 check2;
   sc_sub(tmp.bytes, rct::zero().bytes, z1.bytes);
-  rct::addKeys(check2, check2, rct::scalarmultBase(tmp));
-  rct::addKeys(check2, check2, Z2);
-  rct::addKeys(check2, check2, rct::scalarmultH(z3));
+  ge_double_scalarmult_base_vartime_p3(&check2, z3.bytes, &ge_p3_H, tmp.bytes);
+  add_acc_p3(&check2, Z0);
+  add_acc_p3(&check2, Z2);
 
   std::vector<MultiexpData> multiexp_data;
   multiexp_data.reserve(2 * maxMN);
@@ -1229,10 +1233,10 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
     sc_sub(tmp.bytes, rct::zero().bytes, z5[i].bytes);
     multiexp_data.emplace_back(tmp, Hi_p3[i]);
   }
-  rct::addKeys(check2, check2, multiexp(multiexp_data, true));
+  add_acc_p3(&check2, multiexp(multiexp_data, true));
   PERF_TIMER_STOP(VERIFY_step2_check);
 
-  if (!(check2 == rct::identity()))
+  if (!ge_p3_is_point_at_infinity(&check2))
   {
     MERROR("Verification failure at step 2");
     return false;
diff --git a/src/ringct/rctOps.cpp b/src/ringct/rctOps.cpp
index df027f4b6..6c3c4500e 100644
--- a/src/ringct/rctOps.cpp
+++ b/src/ringct/rctOps.cpp
@@ -211,6 +211,20 @@ namespace rct {
         return aP;
     }
 
+    //Computes 8P
+    key scalarmult8(const key & P) {
+        ge_p3 p3;
+        CHECK_AND_ASSERT_THROW_MES_L1(ge_frombytes_vartime(&p3, P.bytes) == 0, "ge_frombytes_vartime failed at "+boost::lexical_cast<std::string>(__LINE__));
+        ge_p2 p2;
+        ge_p3_to_p2(&p2, &p3);
+        ge_p1p1 p1;
+        ge_mul8(&p1, &p2);
+        ge_p1p1_to_p2(&p2, &p1);
+        rct::key res;
+        ge_tobytes(res.bytes, &p2);
+        return res;
+    }
+
     //Computes aL where L is the curve order
     bool isInMainSubgroup(const key & a) {
         ge_p3 p3;
diff --git a/src/ringct/rctOps.h b/src/ringct/rctOps.h
index 53c8e1d01..50645821c 100644
--- a/src/ringct/rctOps.h
+++ b/src/ringct/rctOps.h
@@ -122,6 +122,8 @@ namespace rct {
     key scalarmultKey(const key &P, const key &a);
     //Computes aH where H= toPoint(cn_fast_hash(G)), G the basepoint
     key scalarmultH(const key & a);
+    // multiplies a point by 8
+    key scalarmult8(const key & P);
     // checks a is in the main subgroup (ie, not a small one)
     bool isInMainSubgroup(const key & a);
 
diff --git a/src/ringct/rctSigs.cpp b/src/ringct/rctSigs.cpp
index 29743c790..fe0cd9c57 100644
--- a/src/ringct/rctSigs.cpp
+++ b/src/ringct/rctSigs.cpp
@@ -768,7 +768,7 @@ namespace rct {
                 #endif
                 for (i = 0; i < outamounts.size(); ++i)
                 {
-                    rv.outPk[i].mask = rct::scalarmultKey(C[i], EIGHT);
+                    rv.outPk[i].mask = rct::scalarmult8(C[i]);
                     outSk[i].mask = masks[i];
                 }
             }
@@ -788,7 +788,7 @@ namespace rct {
             #endif
                 for (i = 0; i < batch_size; ++i)
                 {
-                  rv.outPk[i + amounts_proved].mask = rct::scalarmultKey(C[i], EIGHT);
+                  rv.outPk[i + amounts_proved].mask = rct::scalarmult8(C[i]);
                   outSk[i + amounts_proved].mask = masks[i];
                 }
                 amounts_proved += batch_size;
diff --git a/tests/performance_tests/crypto_ops.h b/tests/performance_tests/crypto_ops.h
index f52faa708..3c68583c5 100644
--- a/tests/performance_tests/crypto_ops.h
+++ b/tests/performance_tests/crypto_ops.h
@@ -45,6 +45,8 @@ enum test_op
   op_addKeys,
   op_scalarmultBase,
   op_scalarmultKey,
+  op_scalarmultH,
+  op_scalarmult8,
   op_ge_double_scalarmult_base_vartime,
   op_ge_double_scalarmult_precomp_vartime,
   op_ge_double_scalarmult_precomp_vartime2,
@@ -97,6 +99,8 @@ public:
       case op_addKeys: rct::addKeys(key, point0, point1); break;
       case op_scalarmultBase: rct::scalarmultBase(scalar0); break;
       case op_scalarmultKey: rct::scalarmultKey(point0, scalar0); break;
+      case op_scalarmultH: rct::scalarmultH(scalar0); break;
+      case op_scalarmult8: rct::scalarmult8(point0); break;
       case op_ge_double_scalarmult_base_vartime: ge_double_scalarmult_base_vartime(&tmp_p2, scalar0.bytes, &p3_0, scalar1.bytes); break;
       case op_ge_double_scalarmult_precomp_vartime: ge_double_scalarmult_precomp_vartime(&tmp_p2, scalar0.bytes, &p3_0, scalar1.bytes, precomp0); break;
       case op_ge_double_scalarmult_precomp_vartime2: ge_double_scalarmult_precomp_vartime2(&tmp_p2, scalar0.bytes, precomp0, scalar1.bytes, precomp1); break;
diff --git a/tests/performance_tests/main.cpp b/tests/performance_tests/main.cpp
index d83486296..7c5135c65 100644
--- a/tests/performance_tests/main.cpp
+++ b/tests/performance_tests/main.cpp
@@ -247,6 +247,8 @@ int main(int argc, char** argv)
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_addKeys);
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_scalarmultBase);
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_scalarmultKey);
+  TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_scalarmultH);
+  TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_scalarmult8);
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_ge_double_scalarmult_base_vartime);
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_ge_double_scalarmult_precomp_vartime);
   TEST_PERFORMANCE1(filter, p, test_crypto_ops, op_ge_double_scalarmult_precomp_vartime2);
diff --git a/tests/unit_tests/ringct.cpp b/tests/unit_tests/ringct.cpp
index 360529b9e..3877ef785 100644
--- a/tests/unit_tests/ringct.cpp
+++ b/tests/unit_tests/ringct.cpp
@@ -1093,6 +1093,13 @@ TEST(ringct, H)
   ASSERT_EQ(memcmp(&p3, &ge_p3_H, sizeof(ge_p3)), 0);
 }
 
+TEST(ringct, mul8)
+{
+  ASSERT_EQ(rct::scalarmult8(rct::identity()), rct::identity());
+  ASSERT_EQ(rct::scalarmult8(rct::H), rct::scalarmultKey(rct::H, rct::EIGHT));
+  ASSERT_EQ(rct::scalarmultKey(rct::scalarmultKey(rct::H, rct::INV_EIGHT), rct::EIGHT), rct::H);
+}
+
 TEST(ringct, aggregated)
 {
   static const size_t N_PROOFS = 16;