I was trying to accelerate ka_prob_glocal() as this will be the

bottleneck. After an hour, the only gain is to change division to
multiplication. OK. I will stop.

diff --git a/kaln.c b/kaln.c
index cf03d89349f649d739eb2a748301899aa51a963e..b52d73a4d43b321873b45e30e23c6df6fe4cdd62 100644 (file)
--- a/kaln.c
+++ b/kaln.c
@@ -502,11 +502,11 @@ int ka_prob_glocal(const uint8_t *_ref, int l_ref, const uint8_t *_query, int l_
                for (k = end; k >= beg; --k) {
                        int u, v11, v01, v10;
                        double e;
-                       e = k >= l_ref? 0 : (ref[k+1] > 3 || qyi1 > 3)? 1. : ref[k+1] == qyi1? 1. - qli1 : qli1 / 3.;
                        set_u(u, bw, i, k); set_u(v11, bw, i+1, k+1); set_u(v10, bw, i+1, k); set_u(v01, bw, i, k+1);
-                       bi[u+0] = e * m[0] * bi1[v11+0] + .25 * m[1] * bi1[v10+1] + m[2] * bi[v01+2];
-                       bi[u+1] = e * m[3] * bi1[v11+0] + .25 * m[4] * bi1[v10+1];
-                       bi[u+2] = (e * m[6] * bi1[v11+0] + m[8] * bi[v01+2]) * y;
+                       e = (k >= l_ref? 0 : (ref[k+1] > 3 || qyi1 > 3)? 1. : ref[k+1] == qyi1? 1. - qli1 : qli1 / 3.) * bi1[v11];
+                       bi[u+0] = e * m[0] + .25 * m[1] * bi1[v10+1] + m[2] * bi[v01+2]; // bi1[v11] has been foled into e.
+                       bi[u+1] = e * m[3] + .25 * m[4] * bi1[v10+1];
+                       bi[u+2] = (e * m[6] + m[8] * bi[v01+2]) * y;
 //                     fprintf(stderr, "B (%d,%d;%d): %lg,%lg,%lg\n", i, k, u, bi[u], bi[u+1], bi[u+2]); // DEBUG
                }
                // rescale