diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9edc043..db25d7a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -8,8 +8,14 @@ if ("${CMAKE_BUILD_TYPE}" STREQUAL "")
 endif()
 set(CMAKE_VERBOSE_MAKEFILE OFF)
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native -O2")
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -funroll-loops -ftree-vectorize -ffast-math")
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -fno-strict-overflow -fomit-frame-pointer")
+if ("${CMAKE_C_COMPILER_ID}" STREQUAL "Intel") 
+  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -unroll-aggressive -ipo ")
+  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -fomit-frame-pointer")
+else()
+  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -funroll-loops -ftree-vectorize -ffast-math")
+  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -fno-strict-overflow -fomit-frame-pointer")
+endif()
+
 
 if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION}.${CMAKE_PATCH_VERSION} VERSION_GREATER 2.8.3)
   if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION}.${CMAKE_PATCH_VERSION} VERSION_GREATER 3.0.0)
diff --git a/src/g3c1e.c b/src/g3c1e.c
index ad0e5c6..584cc86 100644
--- a/src/g3c1e.c
+++ b/src/g3c1e.c
@@ -219,6 +219,9 @@ void CINTg3c1e_nuc(double *g, CINTEnvVars *envs, int count, int nuc_id)
         int i, j, k, n, ptr, off;
         __MD crijk[3];
         __MD r0, r1, r2, rt2, fac1, aijk;
+        __MD scale;
+
+        scale = MM_SET1(2./SQRTPI);
 
         aijk = MM_LOAD(envs->ai) + MM_LOAD(envs->aj) + MM_LOAD(envs->ak);
         MM_STORE(tau, aijk);
@@ -227,14 +230,14 @@ void CINTg3c1e_nuc(double *g, CINTEnvVars *envs, int count, int nuc_id)
         }
 
         if (nuc_id < 0) {
-                fac1 = 2./SQRTPI * MM_LOAD(envs->fac) * MM_LOAD(tau) / aijk;
+                fac1 = scale * MM_LOAD(envs->fac) * MM_LOAD(tau) / aijk;
                 cr = env + PTR_RINV_ORIG;
         } else if (atm(NUC_MOD_OF,nuc_id) == FRAC_CHARGE_NUC) {
-                fac1 = 2./SQRTPI * MM_SET1(-env[atm[PTR_FRAC_CHARGE+nuc_id*ATM_SLOTS]]);
+                fac1 = scale * MM_SET1(-env[atm[PTR_FRAC_CHARGE+nuc_id*ATM_SLOTS]]);
                 fac1 = fac1 * MM_LOAD(envs->fac) * MM_LOAD(tau) / aijk;
                 cr = env + atm(PTR_COORD, nuc_id);
         } else {
-                fac1 = 2./SQRTPI * MM_SET1(-fabs(atm[CHARGE_OF+nuc_id*ATM_SLOTS]));
+                fac1 = scale * MM_SET1(-fabs(atm[CHARGE_OF+nuc_id*ATM_SLOTS]));
                 fac1 = fac1 * MM_LOAD(envs->fac) * MM_LOAD(tau) / aijk;
                 cr = env + atm(PTR_COORD, nuc_id);
         }