diff --git a/README.md b/README.md
index 6202d5e..401d519 100644
--- a/README.md
+++ b/README.md
@@ -26,13 +26,13 @@ Python script `x25519.py`
 
     Test vectors from https://tools.ietf.org/html/rfc8031#appendix-A
     Test 1: X25519: q = d*u
-    Computatation time: 45 ms
+    Computatation time: 40 ms
     q [hex/dec] = 66c7fb0d9f7090f777fa8493081ce8a4f174dbbbf9a36f16ba571206d4ddd548      46489245826987382655505058740283756869827209462947799117248009944518788765000
     Test 1 passed.
 
     Test 2: X25519 + y-coordinate recovery + transform to Edwards-curve
     (x, y) = Edward(q, r), (q, r) = d*(u, v)
-    Computatation time: 50 ms
+    Computatation time: 45 ms
     x [hex/dec] = 1ce7e6e3a747a25352df2d3155f06427ba389769e37755731dead2b54c5cef03      13074494971479542188989287385397236998770807488645203601973104535274459557635
     y [hex/dec] = 4dd1c7c2001c147333ceedf77ebd48b1100e2a95f88cf1f40d1b74ec7279e657      35198739055214410372845858661063095427357109357427482712729161712065293444695
     Test 2 passed.
@@ -41,23 +41,23 @@ Python script `x25519.py`
 Python script `ed25519.py`
 
     Test 1: Length of message: 0 bytes
-    Computatation time: 101 ms
+    Computatation time: 91 ms
     Test 1 passed.
 
     Test 2: Length of message: 1 byte
-    Computatation time: 101 ms
+    Computatation time: 91 ms
     Test 2 passed.
 
     Test 3: Length of message: 2 bytes
-    Computatation time: 101 ms
+    Computatation time: 91 ms
     Test 3 passed.
 
     Test 4: Length of message: 1023 bytes
-    Computatation time: 110 ms
+    Computatation time: 100 ms
     Test 4 passed.
 
     Test 5: Length of message: 64 bytes
-    Computatation time: 102 ms
+    Computatation time: 92 ms
     Test 5 passed.    
 
 ## Warning
diff --git a/mpy-modules/curve25519/arithmetic.c b/mpy-modules/curve25519/arithmetic.c
index 39f41e9..b3d5c78 100644
--- a/mpy-modules/curve25519/arithmetic.c
+++ b/mpy-modules/curve25519/arithmetic.c
@@ -132,7 +132,40 @@ void mul_zx0y0(uint32_t *z, uint32_t x, uint32_t y) {
     : : "r" (x), "r" (y), "r" (z) : "r5", "r6"
     );
 }
-    
+
+
+void pu(uint32_t *t, uint32_t u) {
+    // computes (2^255 -19) * u
+    __asm__ volatile (
+    "MOV    r3, 19\n"
+    "UMULL  r5, r6, r3, %1\n"
+    "MOV    r2, 0\n"
+    "LSRS   %1, %1, 1\n"
+    "RRXS   r2, r2\n"
+    "MOV    r3, 0\n"
+    "SUBS   r4, r3, r5\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r6\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r3, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, r2, r3\n"
+    "STMIA  %0!, {r4}\n"
+    "SBCS   r4, %1, r3\n"
+    "STMIA  %0, {r4}\n"
+    //"SUBS   %0, %0, 32\n"
+    : : "r" (t), "r" (u) : "r2", "r3", "r4", "r5", "r6"
+    );
+}
+
 
 void mont_mul_zxy_mod_p(uint32_t *z, uint32_t *x, uint32_t *y, uint32_t *p) {
     // see Alg. 14.36 HoAC
@@ -146,7 +179,8 @@ void mont_mul_zxy_mod_p(uint32_t *z, uint32_t *x, uint32_t *y, uint32_t *p) {
         u = (a[0] + x[i] * y[0]) * 678152731;
         mul_zxy(tmp, y, x[i]);
         a[8] += tmp[8] + add_zxy(a, a, tmp);
-        mul_zxy(tmp, p, u);
+        // mul_zxy(tmp, p, u);
+        pu(tmp, u);
         a[8] += tmp[8] + add_zxy(a, a, tmp);  // A <- (A + xi y + u m) / b
         for (int j=0; j<8; j++) a[j] = a[j+1];
         a[8] = 0;
@@ -168,7 +202,8 @@ void mont_mul_zxy0_mod_p(uint32_t *z, uint32_t *x, uint32_t y, uint32_t *p) {
         u = (a[0] + x[i] * y) * 678152731;
         mul_zx0y0(tmp, x[i], y);
         a[8] += tmp[8] + add_zxy(a, a, tmp);
-        mul_zxy(tmp, p, u);
+        // mul_zxy(tmp, p, u);
+        pu(tmp, u);
         a[8] += tmp[8] + add_zxy(a, a, tmp);  // A <- (A + xi y + u m) / b
         for (int j=0; j<8; j++) a[j] = a[j+1];
         a[8] = 0;
diff --git a/mpy-modules/curve25519/main.c b/mpy-modules/curve25519/main.c
index fb06978..5a5d87d 100644
--- a/mpy-modules/curve25519/main.c
+++ b/mpy-modules/curve25519/main.c
@@ -1,10 +1,8 @@
  
 // Include the header file to get access to the MicroPython API
 #include "py/dynruntime.h"
-#include "arithmetic.h"
 #include "ec.h"
 
-
 // This is the function which will be called from Python
 STATIC mp_obj_t x25519(mp_obj_t k, mp_obj_t u) {
     mp_buffer_info_t bufinfo_k, bufinfo_u;