diff --git a/vp4d.c b/vp4d.c
index 9dc9d02..69a61e2 100644
--- a/vp4d.c
+++ b/vp4d.c
@@ -29,10 +29,9 @@
 #include "bitunpack.h"
 #include "bitutil.h"
 #include "vp4d.h"
-#include "vint.h"
-#include "vsimple.h"
+#include "vint.h"				//#include "vsimple.h"
+
 #define PAD8(__x) ( (((__x)+8-1)/8) )
-#define VSIZEX 256
 
   #if 0 //defined(__AVX_2__) 
 #include "avx2.h"
@@ -60,18 +59,17 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 };
   #endif
  
-
-#define USIZE 64
-
+#define P4DELTA(a)
 #define _P4DEC      _p4dec
 #define  P4DEC       p4dec
 #define  BITUNPACK   bitunpack  // unpack only
 #define  BITUNPACKD  bitunpack  // integrated unpack
-#define _BITUNPACKD  bitunpack // integrated 
-#define P4START
+#define _BITUNPACKD  bitunpack  // integrated pfor
+
+#define USIZE 64
 #include __FILE__
 
-#define  P4DECX
+#define  P4DECX  // direct access
 
 #define USIZE 16
 #include __FILE__
@@ -79,10 +77,8 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #define USIZE 32
 #include __FILE__
 #undef  P4DECX
-#undef  P4DECX
 
-#define P4START 	start,
-#define P4START_T 	
+#define P4DELTA(a) ,a
 #define _P4DEC      _p4ddec			//delta0
 #define  P4DEC       p4ddec
 #define  BITUNPACKD  bitdunpack
@@ -101,10 +97,10 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #undef  P4DEC
 #undef  BITUNPACK
 #undef  BITUNDD
+#undef  P4DELTA
 
 // SIMD -------------
-#define P4START
-#undef P4START_T
+#define P4DELTA(a) 
 #define  VSIZE 128
 #define _P4DEC        _p4dec128v
 #define  P4DEC         p4dec128v
@@ -113,8 +109,7 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #define  _BITUNPACKD  _bitunpack128v
 #include __FILE__
 
-#define P4START 	start,
-#define P4START_T 	
+#define P4DELTA(a) ,a
 #define _P4DEC        _p4ddec128v
 #define  P4DEC         p4ddec128v
 #define  BITUNPACKD    bitdunpack128v
@@ -129,10 +124,9 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #define  BITUNDD       bitund1
 #include __FILE__
 #undef  BITUNDD
+#undef  P4DELTA
 
     #ifdef __AVX2__
-#define P4START
-#undef P4START_T
 #define  VSIZE 256
 #define _P4DEC        _p4dec256v
 #define  P4DEC         p4dec256v
@@ -141,8 +135,7 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #define  _BITUNPACKD  _bitunpack256v
 #include __FILE__
 
-#define P4START 	start,
-#define P4START_T 	
+#define P4DELTA 
 #define _P4DEC        _p4ddec256v
 #define  P4DEC         p4ddec256v
 #define  BITUNPACKD    bitdunpack256v
@@ -157,24 +150,15 @@ static ALIGNED(char, shuffles[16][16], 16) = {
 #define  BITUNDD       bitund1
 #include __FILE__
 #undef  BITUNDD
-    #endif    
-
+    #endif  
+  
+#undef USIZE
 #else
 #define uint_t TEMPLATE3(uint, USIZE, _t)
 
-//#pragma GCC push_options
-//#pragma GCC optimize ("unroll-loops")
-  #ifdef P4START_T
-unsigned char *TEMPLATE2(_P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out, uint_t start, unsigned b, unsigned bx ) {
-  #else
-unsigned char *TEMPLATE2(_P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out, unsigned b, unsigned bx ) {
-  #endif
-  uint_t ex[VSIZEX+8];
-  if(!(b & 1)) return TEMPLATE2(BITUNPACKD, USIZE)(in, 
-      #ifndef VSIZE
-    n,  
-	  #endif
-  out, P4START b>>1);
+unsigned char *TEMPLATE2(_P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out P4DELTA(uint_t start), unsigned b, unsigned bx ) {
+  uint_t ex[P4D_MAX+8];
+  if(!(b & 1)) return TEMPLATE2(BITUNPACKD, USIZE)(in, n, out P4DELTA(start), b>>1);
 
   b >>= 1;
     #ifdef VSIZE 
@@ -184,17 +168,13 @@ unsigned char *TEMPLATE2(_P4DEC, USIZE)(unsigned char *__restrict in, unsigned n
       #else
   in = TEMPLATE2(bitunpack, USIZE)(in+32, popcnt64(ctou64(in)) + popcnt64(ctou64(in+8)) + popcnt64(ctou64(in+16)) + popcnt64(ctou64(in+24)), ex, bx); 
       #endif
-  return TEMPLATE2(_BITUNPACKD, USIZE)(in, out, P4START b, ex, pb);	  
+  return TEMPLATE2(_BITUNPACKD, USIZE)(in, n, out P4DELTA(start), b, ex, pb);	  
     #else
-  unsigned long long bb[VSIZEX/64]; unsigned num=0,i,p4dn = (n+63)/64;
+  unsigned long long bb[P4D_MAX/64]; unsigned num=0,i,p4dn = (n+63)/64;
   for(i = 0; i < n/64; i++) { bb[i] = ctou64(in+i*8); num += popcnt64(bb[i]); }
   if(n & 0x3f) { bb[i] = ctou64(in+i*8) & ((1ull<<(n&0x3f))-1); num += popcnt64(bb[i]); }
   in = TEMPLATE2(bitunpack, USIZE)(in+PAD8(n), num, ex, bx);
-  in = TEMPLATE2(bitunpack, USIZE)(in,
-      #ifndef VSIZE
-    n, 
-	  #endif
-  out, b);
+  in = TEMPLATE2(bitunpack, USIZE)(in, n, out, b);
       #if 0 //defined(AVX_2__)
   uint_t *op,*pex = ex;
   for(i = 0; i < p4dn; i++) {
@@ -222,27 +202,18 @@ unsigned char *TEMPLATE2(_P4DEC, USIZE)(unsigned char *__restrict in, unsigned n
     #endif
 }
 
-  #ifdef P4START_T
-unsigned char *TEMPLATE2(P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out, uint_t start) {
-  #else
-unsigned char *TEMPLATE2(P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out) {
-  #endif
-	  
+unsigned char *TEMPLATE2(P4DEC, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out P4DELTA(uint_t start) ) {  
   unsigned b = *in++,bx,i;
   
   if(likely(!(b & 0x80))) {
     if(b & 1)
 	  bx = *in++;
-    return TEMPLATE2(_P4DEC, USIZE)(in, n, out, P4START b, bx );
+    return TEMPLATE2(_P4DEC, USIZE)(in, n, out P4DELTA(start), b, bx );
   } else {  
-    uint_t ex[VSIZEX+8]; 
+    uint_t ex[P4D_MAX+8]; 
 	b  = (b & 0x7f)>>1;
     bx = *in++;
-    in = TEMPLATE2(BITUNPACK, USIZE)(in, 
-        #ifndef VSIZE
-      n, 
-	    #endif
-    out, b); 
+    in = TEMPLATE2(BITUNPACK, USIZE)(in, n, out, b); 
     in = TEMPLATE2(vbdec,     USIZE)(in, bx, ex);
     for(i = 0; i != (bx & ~3); i += 4) { 
 	  out[in[i  ]] |= ex[i  ] << b;
@@ -260,7 +231,7 @@ unsigned char *TEMPLATE2(P4DEC, USIZE)(unsigned char *__restrict in, unsigned n,
   }
 }
 
-    #ifdef P4DECX
+    #ifdef P4DECX  
 unsigned char *TEMPLATE2(p4decx, USIZE)(unsigned char *in, unsigned n, uint_t *__restrict out) {
   unsigned b,i; 
   struct p4 p4;