diff --git a/bitutil.h b/bitutil.h
index 42d29b3..ec7de21 100644
--- a/bitutil.h
+++ b/bitutil.h
@@ -21,7 +21,7 @@
     - twitter  : https://twitter.com/powturbo
     - email    : powturbo [_AT_] gmail [_DOT_] com
 **/
-//    bitutil.h - "Integer Compression" 
+//     bitutil.h - "Integer Compression" 
 #include <stdint.h>
 
 #define _BITFORZERO(out, n, start, inc) do {\
@@ -34,66 +34,51 @@
  while(i < n) out[i] = start+i*inc,++i;\
 } while(0)
 
+#define BITSIZE(__in, __n, __b, __usize) { typeof(__in[0]) *_ip;\
+  for(__b=0,_ip = __in; _ip != __in+(__n&~(4-1)); )\
+    __b |= *_ip++ | *_ip++ | *_ip++ | *_ip++;\
+  while(_ip != __in+__n) __b |= *_ip++;\
+  __b = TEMPLATE(bsr, __usize)(__b);\
+}
+
   #ifdef __SSE2__
 #include <emmintrin.h>
 
-#define DELTAV(__v, __sv) _mm_sub_epi32(__v, _mm_or_si128(_mm_srli_si128(__sv, 12), _mm_slli_si128(__v, 4)))
+#define DELTA128_32(__v, __sv) _mm_sub_epi32(__v, _mm_or_si128(_mm_srli_si128(__sv, 12), _mm_slli_si128(__v, 4)))
 
-#define VSCAN( __v, __sv) __v = _mm_add_epi32(__v, _mm_slli_si128(__v, 4)); __sv = _mm_add_epi32(_mm_shuffle_epi32(__sv, _MM_SHUFFLE(3, 3, 3, 3)), _mm_add_epi32(_mm_slli_si128(__v, 8), __v) )
-#define VSCANI(__v, __sv, __vi) VSCAN(__v, __sv); __sv = _mm_add_epi32(__sv, __vi)
+#define SCAN128_32( __v, __sv) __v = _mm_add_epi32(__v, _mm_slli_si128(__v, 4)); __sv = _mm_add_epi32(_mm_shuffle_epi32(__sv, _MM_SHUFFLE(3, 3, 3, 3)), _mm_add_epi32(_mm_slli_si128(__v, 8), __v) )
+#define SCANI128_32(__v, __sv, __vi) SCAN128_32(__v, __sv); __sv = _mm_add_epi32(__sv, __vi)
 
 // SIMD Horizontal OR
-#define HOR(__v,__b) __v = _mm_or_si128(__v, _mm_srli_si128(__v, 8)); __v = _mm_or_si128(__v, _mm_srli_si128(__v, 4)); __b = (unsigned)_mm_cvtsi128_si32(__v)
+#define HOR128_32(__v,__b) __v = _mm_or_si128(__v, _mm_srli_si128(__v, 8)); __v = _mm_or_si128(__v, _mm_srli_si128(__v, 4)); __b = (unsigned)_mm_cvtsi128_si32(__v)
 
-#define BSRN(__in, __n, __b) { typeof(__in[0]) *_ip; __m128i v = _mm_setzero_si128();\
+#define BITSIZE32(__in, __n, __b) { typeof(__in[0]) *_ip; __m128i v = _mm_setzero_si128();\
   for(_ip = __in; _ip != __in+(__n&~(4-1)); _ip+=4) v = _mm_or_si128(v, _mm_loadu_si128((__m128i*)_ip));\
-  HOR(v,__b);\
+  HOR128_32(v,__b);\
   while(_ip != __in+__n) __b |= *_ip++;\
   __b = bsr32(__b);\
 }
 
-#define BITZERO(out, n, start) do {\
+#define BITZERO32(out, n, start) do {\
   __m128i sv = _mm_set1_epi32(start), *ov = (__m128i *)(out), *ove = (__m128i *)(out + n);\
   do { _mm_storeu_si128(ov++, sv); } while(ov < ove); \
 } while(0)
 
-#define BITFORZERO(out, n, start, inc) do {\
+#define BITFORZERO32(out, n, start, inc) do {\
   __m128i sv = _mm_set1_epi32(start), *ov=(__m128i *)(out), *ove = (__m128i *)(out + n), cv = _mm_set_epi32(3*inc,2*inc,1*inc,0); \
     sv = _mm_add_epi32(sv, cv);\
     cv = _mm_set1_epi32(4);\
   do { _mm_storeu_si128(ov++, sv); sv = _mm_add_epi32(sv, cv); } while(ov < ove);\
 } while(0)
 
-#if 0
-#define BITUNBLKV32_0(ip, i, __op, __parm) {\
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-  _mm_storeu_si128(__op++, __parm); __parm = _mm_add_epi32(__parm, cv); \
-}
-#define BITUNPACK0(__parm,inc) __parm = _mm_add_epi32(__parm, cv); cv = _mm_set1_epi32(4*inc)
-
-#define BITDIZERO(out, n, start, inc) do {\
-  __m128i sv = _mm_set1_epi32(start), cv = _mm_set_epi32(3+inc,2+inc,1+inc,inc), *ov=(__m128i *)(out), *ove = (__m128i *)(out + n);	BITUNPACK0(sv,inc); do BITUNBLKV32_0( iv, 0, ov, sv) while(ov < ove);\
-} while(0)
-#else
-#define BITDIZERO(out, n, start, inc) do { __m128i sv = _mm_set1_epi32(start), cv = _mm_set_epi32(3+inc,2+inc,1+inc,inc), *ov=(__m128i *)(out), *ove = (__m128i *)(out + n);\
+#define BITDIZERO32(out, n, start, inc) do { __m128i sv = _mm_set1_epi32(start), cv = _mm_set_epi32(3+inc,2+inc,1+inc,inc), *ov=(__m128i *)(out), *ove = (__m128i *)(out + n);\
   sv = _mm_add_epi32(sv, cv); cv = _mm_set1_epi32(4*inc); do { _mm_storeu_si128(ov++, sv), sv = _mm_add_epi32(sv, cv); } while(ov < ove);\
 } while(0)
-#endif
 
   #else																					
-#define BSRN(__in, __n, __b) { typeof(__in[0]) *_ip;\
-  for(__b=0,_ip = __in; _ip != __in+(__n&~(4-1)); )\
-    __b |= *_ip++ | *_ip++ | *_ip++ | *_ip++;\
-  while(_ip != __in+__n) __b |= *_ip++;\
-  __b = bsr32(__b);\
-}
-#define BITFORZERO(out, n, start, inc) _BITFORZERO(out, n, start, inc)
+#define BITSIZE32(__in, __n, __b)        BITSIZE(__in, __n, __b, 32)
+#define BITFORZERO32(out, n, start, inc) _BITFORZERO(out, n, start, inc)
+#define BITZERO32(out, n, start)         _BITFORZERO(out, n, start, 0)
   #endif
 
 #define ZIGZAG( __in, __n, __mode,      __out) { unsigned _v; for(      __out[0]=__in[0],_v = __n-1; _v >   0; --_v) { int _z = ((int)__in[_v] - (int)__in[_v-1]) - __mode; __out[_v] = (_z << 1) ^ (_z >> 31); } }
@@ -110,6 +95,9 @@
 extern "C" {
 #endif
 
+// get maximum bit length of the elements in the integer array
+unsigned bit32( unsigned *in, unsigned n);  
+
 // transform sorted integer array to delta array. inc = increment
 unsigned bitdelta32(unsigned *in, unsigned n, unsigned *out, unsigned start, unsigned inc);
 unsigned bitdelta64(uint64_t *in, unsigned n, uint64_t *out, uint64_t start, unsigned inc);