diff --git a/CMakeLists.txt b/CMakeLists.txt
index b66b286f..e85bfa28 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -555,6 +555,14 @@ target_link_libraries(
   pqclean_dilithium5_clean
 )
 
+install(TARGETS pqclean pqclean_s
+  PERMISSIONS OWNER_READ OWNER_WRITE GROUP_READ GROUP_WRITE WORLD_READ WORLD_WRITE
+  LIBRARY DESTINATION lib
+  ARCHIVE DESTINATION lib)
+install(FILES
+  ${QRS_PUBLIC_INC}
+  DESTINATION include/pqclean)
+
 # TODO: this requires changes to testvectors.c
 # add_executable(
 #   test
diff --git a/src/sign/dilithium/dilithium2/clean/rounding.c b/src/sign/dilithium/dilithium2/clean/rounding.c
index 21376539..b2e5a1d5 100644
--- a/src/sign/dilithium/dilithium2/clean/rounding.c
+++ b/src/sign/dilithium/dilithium2/clean/rounding.c
@@ -38,66 +38,126 @@ int32_t PQCLEAN_DILITHIUM2_CLEAN_power2round(int32_t *a0, int32_t a)  {
 *
 * Returns a1.
 **************************************************/
-int32_t PQCLEAN_DILITHIUM2_CLEAN_decompose(int32_t *a0, int32_t a) {
-    int32_t a1 = 0;
-    uint64_t r;
-
-    int32_t r0, r1;
+int32_t PQCLEAN_DILITHIUM2_CLEAN_decompose_ORG(int32_t *a0, int32_t a) {
+    /* TODO:
+        a % Q is skipped, as it seems a<Q always. In case this
+        needs to be done, then we can use a fact that Q is a
+        Generalized Marsenne Prime, so modular redc is fast
+        (see work by Jerome Solina and Crandall '92 algo).
+    */
 
     assert(a>0); assert(a<Q);
 
-    // mod ALPHA
-    static const uint32_t u = 360800;
-    r = ((uint64_t)a)*u;
-    r >>= 36;
-    r *= 2 * GAMMA2;
-    r = a - r;
-
-    if (r>(2*GAMMA2)) {
-        r -= 2*GAMMA2;
+    // Use Barrett reduction to calculate r0 = r % A. The
+    // code calculates:
+    //     floor(a/A) = floor( (a * R) / 2^M)
+    //       where,
+    //          M is so that 2^M>= A^2
+    //          r = floor(2^M / A)
+    static const uint32_t M = 36;
+    // Precomputed reciprocal r = floor((2^36) / 190464
+    static const uint32_t R = 360800;
+    // As per spec ALPHA, A = 2*GAMMA2 = (Q-1)/88 * 2 (Dilithium2)
+    static const uint32_t A = 2*GAMMA2;
+
+    // a0
+    int32_t r;
+    int32_t v,w,z;
+    // Barrett reduction:
+    //   a0' = a mod A = a - A*floor((a*r) / 2^M)
+    r = (int32_t)((((uint64_t)a)*R) >> M);
+    r = a - r*A;
+
+    v = ((A-r)>>31) & 1;
+    w = ((GAMMA2 - r)>>31) & 1;
+    z = (((A + GAMMA2) -r) >> 31) & 1;
+//    printf("%d %d %d\n", v,w,z);
+    *a0 = r - (((!z)&(v|w))*A) - (z)*2*A;
+
+    /*
+    // REDC
+    if (r>(int32_t)A) {
+        r -= A;
     }
 
-    r1 = ((int32_t)r)*2*GAMMA2;
-
     // centrize
-    if (r > GAMMA2) {
-        *a0 = (int32_t)r - 2*GAMMA2;
+    if (r > (int32_t)GAMMA2) {
+        *a0 = (int32_t)r - A;
     } else {
         *a0 = r;
     }
+    */
 
+    // a1
+    uint64_t a2 = a - *a0;
 
-    // OLD
-    a1  = (a + 127) >> 7;
-    a1  = (a1 * 11275 + (1 << 23)) >> 24;
-    a1 ^= ((43 - a1) >> 31) & a1;
-
+    v = ((int32_t)a2 - Q + 1);
+    //a2 = (!v);
+    *a0 -= !v;
+    a2 = (!!v)*a2;
     // CASE: r-r0 = q-1 => r1=0, r0 = r0-1
-    uint64_t a2 = (uint64_t)a - *a0;
+    #if 0
     if (a2 == (Q-1)) {
         a2 = 0;
-        *a0--;
+        *a0 = *a0-1;
     }
-
-    // divide (r-r0)/alpha
-    // int32_t a2 = ((uint64_t)a-*a0)/(2*GAMMA2);
-    if ( (a2 >= (2*GAMMA2))) {
-        a2 = (a2*u) >> 36;
-        // a2 is divisible by ALPHA=(2*GAMMA2) and hence
+    #endif
+
+    // divide (r-r0)/A
+    // int32_t a2 = ((uint64_t)a-*a0)/(A);
+    v = ((int32_t)a2-A) >> 31;
+    a2 = (!v)*(((a2*R) >> M) + 1) + v*a2;
+    /*
+    if ( (a2 >= (A))) {
+        a2 = (a2*R) >> M;
+        // a2 is divisible by ALPHA=(A) and hence
         // it will always be off by one.
         a2++;
     }
+    */
+    return a2;
+}
 
-
-
-    //if (!a1) a2 = a1;
-
-    //*a0  = a - a1 * 2 * GAMMA2;
-    //*a0 -= (((Q - 1) / 2 - *a0) >> 31) & Q;
-    if (a1 != (int32_t)a2)
-        printf("OZAPTF: (A1=%d, A2=%d, A=%d R=%d)\n",
-            a1, (int32_t)a2, a, (a-(*a0)));
-//    printf("OZAPTF: %d %d %d\n", a, *a0, (a-*a0));
+int32_t PQCLEAN_DILITHIUM2_CLEAN_decompose(int32_t *a0, int32_t a) {
+    /* TODO:
+        a % Q is skipped, as it seems a<Q always. In case this
+        needs to be done, then we can use a fact that Q is a
+        Generalized Marsenne Prime, so modular redc is fast
+        (see work by Jerome Solina and Crandall '92 algo).
+    */
+
+    // Use Barrett reduction to calculate r0 = r % A. The
+    // code calculates:
+    //     floor(a/A) = floor( (a * R) / 2^M)
+    //       where,
+    //          M is so that 2^M>= A^2
+    //          r = floor(2^M / A)
+    static const uint32_t M = 36;
+    // Precomputed reciprocal r = floor((2^36) / 190464
+    static const uint32_t R = 360800;
+    // As per spec ALPHA, A = 2*GAMMA2 = (Q-1)/88 * 2 (Dilithium2)
+    static const uint32_t A = 2*GAMMA2;
+
+    // a0
+    int32_t r;
+    int32_t v,w,z;
+    // Barrett reduction:
+    //   a0' = a mod A = a - A*floor((a*r) / 2^M)
+    r = (int32_t)((((uint64_t)a)*R) >> M);
+    r = a - r*A;
+
+    v = ((A-r)>>31) & 1;
+    w = ((GAMMA2 - r)>>31) & 1;
+    z = (((A + GAMMA2) -r) >> 31) & 1;
+    *a0 = r - (((!z)&(v|w))*A) - (z)*2*A;
+
+    // a1
+    uint64_t a2 = a - *a0;
+    v = ((int32_t)a2 - Q + 1);
+    *a0 -= !v;
+    a2 = (!!v)*a2;
+    v = ((int32_t)a2-A) >> 31;
+    a2 = (!v)*(((a2*R) >> M) + 1) + v*a2;
     return a2;
 }