Add _fe_half and use in _gej_add_ge

- Trades 1 _half for 3 _mul_int and 2 _normalize_weak - Updated formula and comments in _gej_add_ge - Added internal benchmark for _fe_half
fanquake · Jan 31, 2022 · 925f78d · 925f78d
1 parent d8a2463
commit 925f78d
Show file tree

Hide file tree

Showing 5 changed files with 172 additions and 20 deletions.
diff --git a/src/bench_internal.c b/src/bench_internal.c
@@ -140,6 +140,15 @@ void bench_scalar_inverse_var(void* arg, int iters) {
     CHECK(j <= iters);
 }
 
+void bench_field_half(void* arg, int iters) {
+    int i;
+    bench_inv *data = (bench_inv*)arg;
+
+    for (i = 0; i < iters; i++) {
+        secp256k1_fe_half(&data->fe[0]);
+    }
+}
+
 void bench_field_normalize(void* arg, int iters) {
     int i;
     bench_inv *data = (bench_inv*)arg;
@@ -354,6 +363,7 @@ int main(int argc, char **argv) {
     if (d || have_flag(argc, argv, "scalar") || have_flag(argc, argv, "inverse")) run_benchmark("scalar_inverse", bench_scalar_inverse, bench_setup, NULL, &data, 10, iters);
     if (d || have_flag(argc, argv, "scalar") || have_flag(argc, argv, "inverse")) run_benchmark("scalar_inverse_var", bench_scalar_inverse_var, bench_setup, NULL, &data, 10, iters);
 
+    if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "half")) run_benchmark("field_half", bench_field_half, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "normalize")) run_benchmark("field_normalize", bench_field_normalize, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "normalize")) run_benchmark("field_normalize_weak", bench_field_normalize_weak, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "sqr")) run_benchmark("field_sqr", bench_field_sqr, bench_setup, NULL, &data, 10, iters*10);

diff --git a/src/field.h b/src/field.h
@@ -130,4 +130,9 @@ static void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_f
 /** If flag is true, set *r equal to *a; otherwise leave it. Constant-time.  Both *r and *a must be initialized.*/
 static void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_fe *a, int flag);
 
+/** Halves the value of a field element modulo the field prime. Constant-time.
+ *  For an input magnitude 'm', the output magnitude is set to 'floor(m/2) + 1'.
+ *  The output is not guaranteed to be normalized, regardless of the input. */
+static void secp256k1_fe_half(secp256k1_fe *r);
+
 #endif /* SECP256K1_FIELD_H */
diff --git a/src/field_10x26_impl.h b/src/field_10x26_impl.h
@@ -1133,6 +1133,82 @@ static SECP256K1_INLINE void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_
 #endif
 }
 
+static SECP256K1_INLINE void secp256k1_fe_half(secp256k1_fe *r) {
+    uint32_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4],
+             t5 = r->n[5], t6 = r->n[6], t7 = r->n[7], t8 = r->n[8], t9 = r->n[9];
+    uint32_t one = (uint32_t)1;
+    uint32_t mask = -(t0 & one) >> 6;
+
+#ifdef VERIFY
+    secp256k1_fe_verify(r);
+    VERIFY_CHECK(r->magnitude < 32);
+#endif
+
+    /* Bounds analysis (over the rationals).
+     *
+     * Let m = r->magnitude
+     *     C = 0x3FFFFFFUL * 2
+     *     D = 0x03FFFFFUL * 2
+     *
+     * Initial bounds: t0..t8 <= C * m
+     *                     t9 <= D * m
+     */
+
+    t0 += 0x3FFFC2FUL & mask;
+    t1 += 0x3FFFFBFUL & mask;
+    t2 += mask;
+    t3 += mask;
+    t4 += mask;
+    t5 += mask;
+    t6 += mask;
+    t7 += mask;
+    t8 += mask;
+    t9 += mask >> 4;
+
+    VERIFY_CHECK((t0 & one) == 0);
+
+    /* t0..t8: added <= C/2
+     *     t9: added <= D/2
+     *
+     * Current bounds: t0..t8 <= C * (m + 1/2)
+     *                     t9 <= D * (m + 1/2)
+     */
+
+    r->n[0] = (t0 >> 1) + ((t1 & one) << 25);
+    r->n[1] = (t1 >> 1) + ((t2 & one) << 25);
+    r->n[2] = (t2 >> 1) + ((t3 & one) << 25);
+    r->n[3] = (t3 >> 1) + ((t4 & one) << 25);
+    r->n[4] = (t4 >> 1) + ((t5 & one) << 25);
+    r->n[5] = (t5 >> 1) + ((t6 & one) << 25);
+    r->n[6] = (t6 >> 1) + ((t7 & one) << 25);
+    r->n[7] = (t7 >> 1) + ((t8 & one) << 25);
+    r->n[8] = (t8 >> 1) + ((t9 & one) << 25);
+    r->n[9] = (t9 >> 1);
+
+    /* t0..t8: shifted right and added <= C/4 + 1/2
+     *     t9: shifted right
+     *
+     * Current bounds: t0..t8 <= C * (m/2 + 1/2)
+     *                     t9 <= D * (m/2 + 1/4)
+     */
+
+#ifdef VERIFY
+    /* Therefore the output magnitude (M) has to be set such that:
+     *     t0..t8: C * M >= C * (m/2 + 1/2)
+     *         t9: D * M >= D * (m/2 + 1/4)
+     *
+     * It suffices for all limbs that, for any input magnitude m:
+     *     M >= m/2 + 1/2
+     *
+     * and since we want the smallest such integer value for M:
+     *     M == floor(m/2) + 1
+     */
+    r->magnitude = (r->magnitude >> 1) + 1;
+    r->normalized = 0;
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static SECP256K1_INLINE void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_fe_storage *a, int flag) {
     uint32_t mask0, mask1;
     VG_CHECK_VERIFY(r->n, sizeof(r->n));

diff --git a/src/field_5x52_impl.h b/src/field_5x52_impl.h
@@ -477,6 +477,71 @@ static SECP256K1_INLINE void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_
 #endif
 }
 
+static SECP256K1_INLINE void secp256k1_fe_half(secp256k1_fe *r) {
+    uint64_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4];
+    uint64_t one = (uint64_t)1;
+    uint64_t mask = -(t0 & one) >> 12;
+
+#ifdef VERIFY
+    secp256k1_fe_verify(r);
+    VERIFY_CHECK(r->magnitude < 32);
+#endif
+
+    /* Bounds analysis (over the rationals).
+     *
+     * Let m = r->magnitude
+     *     C = 0xFFFFFFFFFFFFFULL * 2
+     *     D = 0x0FFFFFFFFFFFFULL * 2
+     *
+     * Initial bounds: t0..t3 <= C * m
+     *                     t4 <= D * m
+     */
+
+    t0 += 0xFFFFEFFFFFC2FULL & mask;
+    t1 += mask;
+    t2 += mask;
+    t3 += mask;
+    t4 += mask >> 4;
+
+    VERIFY_CHECK((t0 & one) == 0);
+
+    /* t0..t3: added <= C/2
+     *     t4: added <= D/2
+     *
+     * Current bounds: t0..t3 <= C * (m + 1/2)
+     *                     t4 <= D * (m + 1/2)
+     */
+
+    r->n[0] = (t0 >> 1) + ((t1 & one) << 51);
+    r->n[1] = (t1 >> 1) + ((t2 & one) << 51);
+    r->n[2] = (t2 >> 1) + ((t3 & one) << 51);
+    r->n[3] = (t3 >> 1) + ((t4 & one) << 51);
+    r->n[4] = (t4 >> 1);
+
+    /* t0..t3: shifted right and added <= C/4 + 1/2
+     *     t4: shifted right
+     *
+     * Current bounds: t0..t3 <= C * (m/2 + 1/2)
+     *                     t4 <= D * (m/2 + 1/4)
+     */
+
+#ifdef VERIFY
+    /* Therefore the output magnitude (M) has to be set such that:
+     *     t0..t3: C * M >= C * (m/2 + 1/2)
+     *         t4: D * M >= D * (m/2 + 1/4)
+     *
+     * It suffices for all limbs that, for any input magnitude m:
+     *     M >= m/2 + 1/2
+     *
+     * and since we want the smallest such integer value for M:
+     *     M == floor(m/2) + 1
+     */
+    r->magnitude = (r->magnitude >> 1) + 1;
+    r->normalized = 0;
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static SECP256K1_INLINE void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_fe_storage *a, int flag) {
     uint64_t mask0, mask1;
     VG_CHECK_VERIFY(r->n, sizeof(r->n));

diff --git a/src/group_impl.h b/src/group_impl.h
@@ -492,7 +492,7 @@ static void secp256k1_gej_add_zinv_var(secp256k1_gej *r, const secp256k1_gej *a,
 
 
 static void secp256k1_gej_add_ge(secp256k1_gej *r, const secp256k1_gej *a, const secp256k1_ge *b) {
-    /* Operations: 7 mul, 5 sqr, 4 normalize, 21 mul_int/add/negate/cmov */
+    /* Operations: 7 mul, 5 sqr, 24 add/cmov/half/mul_int/negate/normalize_weak/normalizes_to_zero */
     secp256k1_fe zz, u1, u2, s1, s2, t, tt, m, n, q, rr;
     secp256k1_fe m_alt, rr_alt;
     int infinity, degenerate;
@@ -513,11 +513,11 @@ static void secp256k1_gej_add_ge(secp256k1_gej *r, const secp256k1_gej *a, const
      *    Z = Z1*Z2
      *    T = U1+U2
      *    M = S1+S2
-     *    Q = T*M^2
+     *    Q = -T*M^2
      *    R = T^2-U1*U2
-     *    X3 = 4*(R^2-Q)
-     *    Y3 = 4*(R*(3*Q-2*R^2)-M^4)
-     *    Z3 = 2*M*Z
+     *    X3 = R^2+Q
+     *    Y3 = -(R*(2*X3+Q)+M^4)/2
+     *    Z3 = M*Z
      *  (Note that the paper uses xi = Xi / Zi and yi = Yi / Zi instead.)
      *
      *  This formula has the benefit of being the same for both addition
@@ -581,29 +581,25 @@ static void secp256k1_gej_add_ge(secp256k1_gej *r, const secp256k1_gej *a, const
      * and denominator of lambda; R and M represent the explicit
      * expressions x1^2 + x2^2 + x1x2 and y1 + y2. */
     secp256k1_fe_sqr(&n, &m_alt);                       /* n = Malt^2 (1) */
-    secp256k1_fe_mul(&q, &n, &t);                       /* q = Q = T*Malt^2 (1) */
+    secp256k1_fe_negate(&q, &t, 2);                     /* q = -T (3) */
+    secp256k1_fe_mul(&q, &q, &n);                       /* q = Q = -T*Malt^2 (1) */
     /* These two lines use the observation that either M == Malt or M == 0,
      * so M^3 * Malt is either Malt^4 (which is computed by squaring), or
      * zero (which is "computed" by cmov). So the cost is one squaring
      * versus two multiplications. */
     secp256k1_fe_sqr(&n, &n);
     secp256k1_fe_cmov(&n, &m, degenerate);              /* n = M^3 * Malt (2) */
     secp256k1_fe_sqr(&t, &rr_alt);                      /* t = Ralt^2 (1) */
-    secp256k1_fe_mul(&r->z, &a->z, &m_alt);             /* r->z = Malt*Z (1) */
+    secp256k1_fe_mul(&r->z, &a->z, &m_alt);             /* r->z = Z3 = Malt*Z (1) */
     infinity = secp256k1_fe_normalizes_to_zero(&r->z) & ~a->infinity;
-    secp256k1_fe_mul_int(&r->z, 2);                     /* r->z = Z3 = 2*Malt*Z (2) */
-    secp256k1_fe_negate(&q, &q, 1);                     /* q = -Q (2) */
-    secp256k1_fe_add(&t, &q);                           /* t = Ralt^2-Q (3) */
-    secp256k1_fe_normalize_weak(&t);
-    r->x = t;                                           /* r->x = Ralt^2-Q (1) */
-    secp256k1_fe_mul_int(&t, 2);                        /* t = 2*x3 (2) */
-    secp256k1_fe_add(&t, &q);                           /* t = 2*x3 - Q: (4) */
-    secp256k1_fe_mul(&t, &t, &rr_alt);                  /* t = Ralt*(2*x3 - Q) (1) */
-    secp256k1_fe_add(&t, &n);                           /* t = Ralt*(2*x3 - Q) + M^3*Malt (3) */
-    secp256k1_fe_negate(&r->y, &t, 3);                  /* r->y = Ralt*(Q - 2x3) - M^3*Malt (4) */
-    secp256k1_fe_normalize_weak(&r->y);
-    secp256k1_fe_mul_int(&r->x, 4);                     /* r->x = X3 = 4*(Ralt^2-Q) */
-    secp256k1_fe_mul_int(&r->y, 4);                     /* r->y = Y3 = 4*Ralt*(Q - 2x3) - 4*M^3*Malt (4) */
+    secp256k1_fe_add(&t, &q);                           /* t = Ralt^2 + Q (2) */
+    r->x = t;                                           /* r->x = X3 = Ralt^2 + Q (2) */
+    secp256k1_fe_mul_int(&t, 2);                        /* t = 2*X3 (4) */
+    secp256k1_fe_add(&t, &q);                           /* t = 2*X3 + Q (5) */
+    secp256k1_fe_mul(&t, &t, &rr_alt);                  /* t = Ralt*(2*X3 + Q) (1) */
+    secp256k1_fe_add(&t, &n);                           /* t = Ralt*(2*X3 + Q) + M^3*Malt (3) */
+    secp256k1_fe_negate(&r->y, &t, 3);                  /* r->y = -(Ralt*(2*X3 + Q) + M^3*Malt) (4) */
+    secp256k1_fe_half(&r->y);                           /* r->y = Y3 = -(Ralt*(2*X3 + Q) + M^3*Malt)/2 (3) */
 
     /** In case a->infinity == 1, replace r with (b->x, b->y, 1). */
     secp256k1_fe_cmov(&r->x, &b->x, a->infinity);