Add 16x16->32 word Comba multiply and square

2 files changed, 716 insertions, 176 deletions

diff --git a/src/mp_comba.cpp b/src/mp_comba.cpp
index 589a4037c..635a7547a 100644
--- a/src/mp_comba.cpp
+++ b/src/mp_comba.cpp
@@ -11,6 +11,39 @@ namespace Botan {
 extern "C" {
 
 /*************************************************
+* Comba 4x4 Squaring                             *
+*************************************************/
+void bigint_comba_sqr4(word z[8], const word x[4])
+   {
+   word w2 = 0, w1 = 0, w0 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
+   z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
+   z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
+   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
+   z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
+   z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
+   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
+   z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
+   z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
+   z[6] = w0;
+   z[7] = w1;
+   }
+
+/*************************************************
 * Comba 4x4 Multiplication                       *
 *************************************************/
 void bigint_comba_mul4(word z[8], const word x[4], const word y[4])
@@ -50,6 +83,58 @@ void bigint_comba_mul4(word z[8], const word x[4], const word y[4])
    }
 
 /*************************************************
+* Comba 6x6 Squaring                             *
+*************************************************/
+void bigint_comba_sqr6(word z[12], const word x[6])
+   {
+   word w2 = 0, w1 = 0, w0 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
+   z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
+   z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
+   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
+   z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
+   z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[4]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
+   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
+   z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[4]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
+   z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[4]);
+   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
+   z[6] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[4]);
+   z[7] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[5]);
+   word3_muladd(&w2, &w1, &w0, x[4], x[4]);
+   z[8] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[4], x[5]);
+   z[9] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[5], x[5]);
+   z[10] = w0;
+   z[11] = w1;
+   }
+
+/*************************************************
 * Comba 6x6 Multiplication                       *
 *************************************************/
 void bigint_comba_mul6(word z[12], const word x[6], const word y[6])
@@ -117,6 +202,81 @@ void bigint_comba_mul6(word z[12], const word x[6], const word y[6])
    }
 
 /*************************************************
+* Comba 8x8 Squaring                             *
+*************************************************/
+void bigint_comba_sqr8(word z[16], const word x[8])
+   {
+   word w2 = 0, w1 = 0, w0 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
+   z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
+   z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
+   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
+   z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
+   z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[4]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
+   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
+   z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[4]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
+   z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[6]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[4]);
+   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
+   z[6] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[0], x[7]);
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[6]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[5]);
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[4]);
+   z[7] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[1], x[7]);
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[6]);
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[5]);
+   word3_muladd(&w2, &w1, &w0, x[4], x[4]);
+   z[8] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[2], x[7]);
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[6]);
+   word3_muladd_2(&w2, &w1, &w0, x[4], x[5]);
+   z[9] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[3], x[7]);
+   word3_muladd_2(&w2, &w1, &w0, x[4], x[6]);
+   word3_muladd(&w2, &w1, &w0, x[5], x[5]);
+   z[10] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[4], x[7]);
+   word3_muladd_2(&w2, &w1, &w0, x[5], x[6]);
+   z[11] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[5], x[7]);
+   word3_muladd(&w2, &w1, &w0, x[6], x[6]);
+   z[12] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[6], x[7]);
+   z[13] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[7], x[7]);
+   z[14] = w0;
+   z[15] = w1;
+   }
+
+/*************************************************
 * Comba 8x8 Multiplication                       *
 *************************************************/
 void bigint_comba_mul8(word z[16], const word x[8], const word y[8])
@@ -220,163 +380,539 @@ void bigint_comba_mul8(word z[16], const word x[8], const word y[8])
    }
 
 /*************************************************
-* Comba 4x4 Squaring                             *
+* Comba 16x16 Squaring                           *
 *************************************************/
-void bigint_comba_sqr4(word z[8], const word x[4])
+void bigint_comba_sqr16(word z[32], const word x[16])
    {
-   word w2 = 0, w1 = 0, w0 = 0;
-
-   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
-   z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
+   return bigint_comba_mul16(z, x, x);
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
-   z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
-   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
-   z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
-   z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
-   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
-   z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
-   z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
-   z[6] = w0;
-   z[7] = w1;
-   }
-
-/*************************************************
-* Comba 6x6 Squaring                             *
-*************************************************/
-void bigint_comba_sqr6(word z[12], const word x[6])
-   {
    word w2 = 0, w1 = 0, w0 = 0;
 
-   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
-   z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
-   z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
-   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
-   z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
-   z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[4]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
-   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
-   z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[4]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
-   z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
-
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[4]);
-   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
-   z[6] = w0; w0 = w1; w1 = w2; w2 = 0;
+   word3_muladd(&w2, &w1, &w0, x[ 0], x[ 0]);
+   z[ 0] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 1]);
+   z[ 1] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], x[ 1]);
+   z[ 2] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 3]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 2]);
+   z[ 3] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 4]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], x[ 2]);
+   z[ 4] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 5]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 4]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 3]);
+   z[ 5] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 6]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 5]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], x[ 3]);
+   z[ 6] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 7]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 6]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 5]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 4]);
+   z[ 7] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 7]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 6]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], x[ 4]);
+   z[ 8] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 7]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 6]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[ 5]);
+   z[ 9] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 7]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], x[ 5]);
+   z[10] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[4]);
-   z[7] = w0; w0 = w1; w1 = w2; w2 = 0;
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[ 7]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[ 6]);
+   z[11] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[5]);
-   word3_muladd(&w2, &w1, &w0, x[4], x[4]);
-   z[8] = w0; w0 = w1; w1 = w2; w2 = 0;
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], x[ 6]);
+   z[12] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[4], x[5]);
-   z[9] = w0; w0 = w1; w1 = w2; w2 = 0;
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[ 8]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[ 7]);
+   z[13] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd(&w2, &w1, &w0, x[5], x[5]);
-   z[10] = w0;
-   z[11] = w1;
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], x[ 7]);
+   z[14] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 0], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[ 9]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[ 8]);
+   z[15] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 1], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], x[ 8]);
+   z[16] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 2], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[10]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[ 9]);
+   z[17] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 3], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], x[ 9]);
+   z[18] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 4], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[11]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[10]);
+   z[19] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 5], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[11]);
+   word3_muladd(&w2, &w1, &w0, x[10], x[10]);
+   z[20] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 6], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[12]);
+   word3_muladd_2(&w2, &w1, &w0, x[10], x[11]);
+   z[21] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 7], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[10], x[12]);
+   word3_muladd(&w2, &w1, &w0, x[11], x[11]);
+   z[22] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 8], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[10], x[13]);
+   word3_muladd_2(&w2, &w1, &w0, x[11], x[12]);
+   z[23] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[ 9], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[10], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[11], x[13]);
+   word3_muladd(&w2, &w1, &w0, x[12], x[12]);
+   z[24] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[10], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[11], x[14]);
+   word3_muladd_2(&w2, &w1, &w0, x[12], x[13]);
+   z[25] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[11], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[12], x[14]);
+   word3_muladd(&w2, &w1, &w0, x[13], x[13]);
+   z[26] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[12], x[15]);
+   word3_muladd_2(&w2, &w1, &w0, x[13], x[14]);
+   z[27] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[13], x[15]);
+   word3_muladd(&w2, &w1, &w0, x[14], x[14]);
+   z[28] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd_2(&w2, &w1, &w0, x[14], x[15]);
+   z[29] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[15], x[15]);
+   z[30] = w0;
+   z[31] = w1;
    }
 
 /*************************************************
-* Comba 8x8 Squaring                             *
+* Comba 16x16 Multiplication                     *
 *************************************************/
-void bigint_comba_sqr8(word z[16], const word x[8])
+void bigint_comba_mul16(word z[32], const word x[16], const word y[16])
    {
    word w2 = 0, w1 = 0, w0 = 0;
 
-   word3_muladd(&w2, &w1, &w0, x[0], x[0]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 0]);
    z[0] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[1]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 0]);
    z[1] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[2]);
-   word3_muladd(&w2, &w1, &w0, x[1], x[1]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 0]);
    z[2] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[3]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[2]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 0]);
    z[3] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[4]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[3]);
-   word3_muladd(&w2, &w1, &w0, x[2], x[2]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 0]);
    z[4] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[4]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[3]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 0]);
    z[5] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[6]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[4]);
-   word3_muladd(&w2, &w1, &w0, x[3], x[3]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 0]);
    z[6] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[0], x[7]);
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[6]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[5]);
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[4]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 0]);
    z[7] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[1], x[7]);
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[6]);
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[5]);
-   word3_muladd(&w2, &w1, &w0, x[4], x[4]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 0]);
    z[8] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[2], x[7]);
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[6]);
-   word3_muladd_2(&w2, &w1, &w0, x[4], x[5]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 0]);
    z[9] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[3], x[7]);
-   word3_muladd_2(&w2, &w1, &w0, x[4], x[6]);
-   word3_muladd(&w2, &w1, &w0, x[5], x[5]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 0]);
    z[10] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[4], x[7]);
-   word3_muladd_2(&w2, &w1, &w0, x[5], x[6]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 0]);
    z[11] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[5], x[7]);
-   word3_muladd(&w2, &w1, &w0, x[6], x[6]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 0]);
    z[12] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd_2(&w2, &w1, &w0, x[6], x[7]);
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 0]);
    z[13] = w0; w0 = w1; w1 = w2; w2 = 0;
 
-   word3_muladd(&w2, &w1, &w0, x[7], x[7]);
-   z[14] = w0;
-   z[15] = w1;
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 0]);
+   z[14] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 0], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 1]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 0]);
+   z[15] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 1], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 2]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 1]);
+   z[16] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 2], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 3]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 2]);
+   z[17] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 3], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 4]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 3]);
+   z[18] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 4], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 5]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 4]);
+   z[19] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 5], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 6]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 5]);
+   z[20] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 6], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 7]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 6]);
+   z[21] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 7], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 8]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 7]);
+   z[22] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 8], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[ 9]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 8]);
+   z[23] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[ 9], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[10], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[10]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[ 9]);
+   z[24] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[10], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[11], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[11]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[10]);
+   z[25] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[11], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[12], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[12]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[11]);
+   z[26] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[12], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[13], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[13]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[12]);
+   z[27] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[13], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[14], y[14]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[13]);
+   z[28] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[14], y[15]);
+   word3_muladd(&w2, &w1, &w0, x[15], y[14]);
+   z[29] = w0; w0 = w1; w1 = w2; w2 = 0;
+
+   word3_muladd(&w2, &w1, &w0, x[15], y[15]);
+   z[30] = w0;
+   z[31] = w1;
    }
 
 }
diff --git a/src/mp_karat.cpp b/src/mp_karat.cpp
index 770846b4e..15b0551fd 100644
--- a/src/mp_karat.cpp
+++ b/src/mp_karat.cpp
@@ -17,13 +17,13 @@ namespace {
 void karatsuba_mul(word z[], const word x[], const word y[], u32bit N,
                    word workspace[])
    {
-   const u32bit KARATSUBA_MUL_LOWER_SIZE = BOTAN_KARAT_MUL_THRESHOLD;
-
    if(N == 6)
       bigint_comba_mul6(z, x, y);
    else if(N == 8)
       bigint_comba_mul8(z, x, y);
-   else if(N < KARATSUBA_MUL_LOWER_SIZE || N % 2)
+   else if(N == 16)
+      bigint_comba_mul16(z, x, y);
+   else if(N < BOTAN_KARAT_MUL_THRESHOLD || N % 2)
       bigint_simple_mul(z, x, N, y, N);
    else
       {
@@ -96,13 +96,13 @@ void karatsuba_mul(word z[], const word x[], const word y[], u32bit N,
 *************************************************/
 void karatsuba_sqr(word z[], const word x[], u32bit N, word workspace[])
    {
-   const u32bit KARATSUBA_SQR_LOWER_SIZE = BOTAN_KARAT_SQR_THRESHOLD;
-
    if(N == 6)
       bigint_comba_sqr6(z, x);
    else if(N == 8)
       bigint_comba_sqr8(z, x);
-   else if(N < KARATSUBA_SQR_LOWER_SIZE || N % 2)
+   else if(N == 16)
+      bigint_comba_sqr16(z, x);
+   else if(N < BOTAN_KARAT_SQR_THRESHOLD || N % 2)
       bigint_simple_sqr(z, x, N);
    else
       {
@@ -234,97 +234,101 @@ u32bit karatsuba_size(u32bit z_size, u32bit x_size, u32bit x_sw)
    return 0;
    }
 
+}
+
 /*************************************************
-* Handle small operand multiplies                *
+* Multiplication Algorithm Dispatcher            *
 *************************************************/
-void handle_small_mul(word z[], u32bit z_size,
-                      const word x[], u32bit x_size, u32bit x_sw,
-                      const word y[], u32bit y_size, u32bit y_sw)
+void bigint_mul(word z[], u32bit z_size, word workspace[],
+                const word x[], u32bit x_size, u32bit x_sw,
+                const word y[], u32bit y_size, u32bit y_sw)
    {
-   if(x_sw == 1)        bigint_linmul3(z, y, y_sw, x[0]);
-   else if(y_sw == 1)   bigint_linmul3(z, x, x_sw, y[0]);
-
+   if(x_sw == 1)
+      {
+      bigint_linmul3(z, y, y_sw, x[0]);
+      }
+   else if(y_sw == 1)
+      {
+      bigint_linmul3(z, x, x_sw, y[0]);
+      }
    else if(x_sw <= 4 && x_size >= 4 &&
            y_sw <= 4 && y_size >= 4 && z_size >= 8)
+      {
       bigint_comba_mul4(z, x, y);
-
+      }
    else if(x_sw <= 6 && x_size >= 6 &&
            y_sw <= 6 && y_size >= 6 && z_size >= 12)
+      {
       bigint_comba_mul6(z, x, y);
-
+      }
    else if(x_sw <= 8 && x_size >= 8 &&
            y_sw <= 8 && y_size >= 8 && z_size >= 16)
+      {
       bigint_comba_mul8(z, x, y);
-
-   else
+      }
+   else if(x_sw <= 16 && x_size >= 16 &&
+           y_sw <= 16 && y_size >= 16 && z_size >= 32)
+      {
+      bigint_comba_mul16(z, x, y);
+      }
+   else if(x_sw < BOTAN_KARAT_MUL_THRESHOLD || y_sw < BOTAN_KARAT_MUL_THRESHOLD)
       bigint_simple_mul(z, x, x_sw, y, y_sw);
+   else
+      {
+      const u32bit N = karatsuba_size(z_size, x_size, x_sw, y_size, y_sw);
+
+      if(N)
+         {
+         clear_mem(workspace, 2*N);
+         karatsuba_mul(z, x, y, N, workspace);
+         }
+      else
+         bigint_simple_mul(z, x, x_sw, y, y_sw);
+      }
    }
 
 /*************************************************
-* Handle small operand squarings                 *
+* Squaring Algorithm Dispatcher                  *
 *************************************************/
-void handle_small_sqr(word z[], u32bit z_size,
-                      const word x[], u32bit x_size, u32bit x_sw)
+void bigint_sqr(word z[], u32bit z_size, word workspace[],
+                const word x[], u32bit x_size, u32bit x_sw)
    {
    if(x_sw == 1)
+      {
       bigint_linmul3(z, x, x_sw, x[0]);
+      }
    else if(x_sw <= 4 && x_size >= 4 && z_size >= 8)
+      {
       bigint_comba_sqr4(z, x);
+      }
    else if(x_sw <= 6 && x_size >= 6 && z_size >= 12)
+      {
       bigint_comba_sqr6(z, x);
+      }
    else if(x_sw <= 8 && x_size >= 8 && z_size >= 16)
+      {
       bigint_comba_sqr8(z, x);
-   else
-      bigint_simple_sqr(z, x, x_sw);
-   }
-
-}
-
-/*************************************************
-* Multiplication Algorithm Dispatcher            *
-*************************************************/
-void bigint_mul(word z[], u32bit z_size, word workspace[],
-                const word x[], u32bit x_size, u32bit x_sw,
-                const word y[], u32bit y_size, u32bit y_sw)
-   {
-   if(x_size <= 8 || y_size <= 8)
+      }
+   else if(x_sw <= 16 && x_size >= 16 && z_size >= 32)
       {
-      handle_small_mul(z, z_size, x, x_size, x_sw, y, y_size, y_sw);
-      return;
+      bigint_comba_sqr16(z, x);
       }
-
-   const u32bit N = karatsuba_size(z_size, x_size, x_sw, y_size, y_sw);
-
-   if(N)
+   else if(x_size < BOTAN_KARAT_SQR_THRESHOLD)
       {
-      clear_mem(workspace, 2*N);
-      karatsuba_mul(z, x, y, N, workspace);
+      bigint_simple_sqr(z, x, x_sw);
       }
    else
-      bigint_simple_mul(z, x, x_sw, y, y_sw);
-   }
-
-/*************************************************
-* Squaring Algorithm Dispatcher                  *
-*************************************************/
-void bigint_sqr(word z[], u32bit z_size, word workspace[],
-                const word x[], u32bit x_size, u32bit x_sw)
-   {
-   if(x_size <= 8 || x_sw <= 8)
       {
-      handle_small_sqr(z, z_size, x, x_size, x_sw);
-      return;
-      }
-
-   const u32bit N = karatsuba_size(z_size, x_size, x_sw);
+      const u32bit N = karatsuba_size(z_size, x_size, x_sw);
 
-   if(N)
-      {
-      clear_mem(workspace, 2*N);
-      karatsuba_sqr(z, x, N, workspace);
+      if(N)
+         {
+         clear_mem(workspace, 2*N);
+         karatsuba_sqr(z, x, N, workspace);
+         }
+      else
+         bigint_simple_sqr(z, x, x_sw);
       }
-   else
-      bigint_simple_sqr(z, x, x_sw);
    }
 
 }