diff --git a/lib/bitpack_.h b/lib/bitpack_.h
index aa262c2..6480df8 100644
--- a/lib/bitpack_.h
+++ b/lib/bitpack_.h
@@ -2258,17 +2258,22 @@
   BITBLK64_64(ip, 31, op, parm);  IPI(ip); op += 64*4/sizeof(op[0]);\
 }
 
+#define BP(_b_,_usize_) unsigned char *out_=out+PAD8(n*_b_),*op, bout[PAD8(64*_b_)]; T3(uint,_usize_,_t) bin[64],*ip,*in_=in+n, v,x; \
+  do { ip = in+32; op = out+PAD8(32*_b_); if(ip > in_) { memcpy(bin, in, (in_-in)*(_usize_/8)); in = bin; out = bout; } \
+    T2(BITPACK64_,_b_)(in, out, start); in = ip; out = op; PREFETCH(in+384,0);\
+  } while(in<in_); if(in>in_) { out -= PAD8(32*_b_); memcpy(out,bout,PAD8((in_-(in-32))*_b_)); }  return out_
+
 #ifndef DELTA
 #define USIZE 8
-unsigned char *T2(_BITPACK_,8_0)( uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
-unsigned char *T2(_BITPACK_,8_1)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*1); uint8_t v,x;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_2)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*2); uint8_t v,x;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_3)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*3); uint8_t v,x;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_4)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*4); uint8_t v,x;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_5)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*5); uint8_t v,x;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_6)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*6); uint8_t v,x;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_7)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*7); uint8_t v,x;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_8)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*8); uint8_t v,x;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,8_0)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
+unsigned char *T2(_BITPACK_,8_1)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(1,8);}
+unsigned char *T2(_BITPACK_,8_2)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(2,8);}
+unsigned char *T2(_BITPACK_,8_3)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(3,8);}
+unsigned char *T2(_BITPACK_,8_4)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(4,8);}
+unsigned char *T2(_BITPACK_,8_5)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(5,8);}
+unsigned char *T2(_BITPACK_,8_6)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(6,8);}
+unsigned char *T2(_BITPACK_,8_7)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(7,8);}
+unsigned char *T2(_BITPACK_,8_8)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(8,8);}
 BITPACK_F8 T2(_BITPACK_,a8)[] = {
   &T2(_BITPACK_,8_0),
   &T2(_BITPACK_,8_1),
@@ -2281,25 +2286,26 @@ BITPACK_F8 T2(_BITPACK_,a8)[] = {
   &T2(_BITPACK_,8_8)
 };
 unsigned char *T2(_BITPACK_,8)( uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , unsigned b) { return T2(_BITPACK_,a8)[ b](in, n, out); }
+#undef USIZE
 
 #define USIZE 16
-unsigned char *T2(_BITPACK_,16_0)( uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
-unsigned char *T2(_BITPACK_,16_1)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*1); uint16_t v,x;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_2)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*2); uint16_t v,x;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_3)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*3); uint16_t v,x;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_4)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*4); uint16_t v,x;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_5)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*5); uint16_t v,x;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_6)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*6); uint16_t v,x;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_7)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*7); uint16_t v,x;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_8)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*8); uint16_t v,x;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_9)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*9); uint16_t v,x;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_10)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*10); uint16_t v,x;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_11)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*11); uint16_t v,x;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_12)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*12); uint16_t v,x;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_13)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*13); uint16_t v,x;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_14)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*14); uint16_t v,x;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_15)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*15); uint16_t v,x;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_16)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*16); uint16_t v,x;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,16_0 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
+unsigned char *T2(_BITPACK_,16_1 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(1,16);}
+unsigned char *T2(_BITPACK_,16_2 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(2,16);}
+unsigned char *T2(_BITPACK_,16_3 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(3,16);}
+unsigned char *T2(_BITPACK_,16_4 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(4,16);}
+unsigned char *T2(_BITPACK_,16_5 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(5,16);}
+unsigned char *T2(_BITPACK_,16_6 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(6,16);}
+unsigned char *T2(_BITPACK_,16_7 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(7,16);}
+unsigned char *T2(_BITPACK_,16_8 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(8,16);}
+unsigned char *T2(_BITPACK_,16_9 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(9,16);}
+unsigned char *T2(_BITPACK_,16_10)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(10,16);}
+unsigned char *T2(_BITPACK_,16_11)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(11,16);}
+unsigned char *T2(_BITPACK_,16_12)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(12,16);}
+unsigned char *T2(_BITPACK_,16_13)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(13,16);}
+unsigned char *T2(_BITPACK_,16_14)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(14,16);}
+unsigned char *T2(_BITPACK_,16_15)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(15,16);}
+unsigned char *T2(_BITPACK_,16_16)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(16,16);}
 BITPACK_F16 T2(_BITPACK_,a16)[] = {
   &T2(_BITPACK_,16_0),
   &T2(_BITPACK_,16_1),
@@ -2320,41 +2326,42 @@ BITPACK_F16 T2(_BITPACK_,a16)[] = {
   &T2(_BITPACK_,16_16)
 };
 unsigned char *T2(_BITPACK_,16)( uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , unsigned b) { return T2(_BITPACK_,a16)[ b](in, n, out); }
+#undef USIZE
 
 #define USIZE 32
-unsigned char *T2(_BITPACK_,32_0)( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
-unsigned char *T2(_BITPACK_,32_1)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*1); uint32_t v,x;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_2)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*2); uint32_t v,x;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_3)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*3); uint32_t v,x;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_4)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*4); uint32_t v,x;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_5)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*5); uint32_t v,x;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_6)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*6); uint32_t v,x;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_7)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*7); uint32_t v,x;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_8)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*8); uint32_t v,x;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_9)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*9); uint32_t v,x;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_10)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*10); uint32_t v,x;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_11)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*11); uint32_t v,x;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_12)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*12); uint32_t v,x;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_13)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*13); uint32_t v,x;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_14)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*14); uint32_t v,x;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_15)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*15); uint32_t v,x;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_16)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*16); uint32_t v,x;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_17)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*17); uint32_t v,x;do { BITPACK64_17( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_18)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*18); uint32_t v,x;do { BITPACK64_18( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_19)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*19); uint32_t v,x;do { BITPACK64_19( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_20)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*20); uint32_t v,x;do { BITPACK64_20( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_21)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*21); uint32_t v,x;do { BITPACK64_21( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_22)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*22); uint32_t v,x;do { BITPACK64_22( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_23)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*23); uint32_t v,x;do { BITPACK64_23( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_24)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*24); uint32_t v,x;do { BITPACK64_24( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_25)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*25); uint32_t v,x;do { BITPACK64_25( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_26)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*26); uint32_t v,x;do { BITPACK64_26( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_27)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*27); uint32_t v,x;do { BITPACK64_27( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_28)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*28); uint32_t v,x;do { BITPACK64_28( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_29)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*29); uint32_t v,x;do { BITPACK64_29( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_30)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*30); uint32_t v,x;do { BITPACK64_30( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_31)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*31); uint32_t v,x;do { BITPACK64_31( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_32)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*32); uint32_t v,x;do { BITPACK64_32( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,32_0 )( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
+unsigned char *T2(_BITPACK_,32_1 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 1,32);}
+unsigned char *T2(_BITPACK_,32_2 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 2,32);}
+unsigned char *T2(_BITPACK_,32_3 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 3,32);}
+unsigned char *T2(_BITPACK_,32_4 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 4,32);}
+unsigned char *T2(_BITPACK_,32_5 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 5,32);}
+unsigned char *T2(_BITPACK_,32_6 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 6,32);}
+unsigned char *T2(_BITPACK_,32_7 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 7,32);}
+unsigned char *T2(_BITPACK_,32_8 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 8,32);}
+unsigned char *T2(_BITPACK_,32_9 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 9,32);}
+unsigned char *T2(_BITPACK_,32_10)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(10,32);}
+unsigned char *T2(_BITPACK_,32_11)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(11,32);}
+unsigned char *T2(_BITPACK_,32_12)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(12,32);}
+unsigned char *T2(_BITPACK_,32_13)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(13,32);}
+unsigned char *T2(_BITPACK_,32_14)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(14,32);}
+unsigned char *T2(_BITPACK_,32_15)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(15,32);}
+unsigned char *T2(_BITPACK_,32_16)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(16,32);}
+unsigned char *T2(_BITPACK_,32_17)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(17,32);}
+unsigned char *T2(_BITPACK_,32_18)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(18,32);}
+unsigned char *T2(_BITPACK_,32_19)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(19,32);}
+unsigned char *T2(_BITPACK_,32_20)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(20,32);}
+unsigned char *T2(_BITPACK_,32_21)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(21,32);}
+unsigned char *T2(_BITPACK_,32_22)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(22,32);}
+unsigned char *T2(_BITPACK_,32_23)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(23,32);}
+unsigned char *T2(_BITPACK_,32_24)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(24,32);}
+unsigned char *T2(_BITPACK_,32_25)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(25,32);}
+unsigned char *T2(_BITPACK_,32_26)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(26,32);}
+unsigned char *T2(_BITPACK_,32_27)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(27,32);}
+unsigned char *T2(_BITPACK_,32_28)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(28,32);}
+unsigned char *T2(_BITPACK_,32_29)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(29,32);}
+unsigned char *T2(_BITPACK_,32_30)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(30,32);}
+unsigned char *T2(_BITPACK_,32_31)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(31,32);}
+unsigned char *T2(_BITPACK_,32_32)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(32,32);}
 BITPACK_F32 T2(_BITPACK_,a32)[] = {
   &T2(_BITPACK_,32_0),
   &T2(_BITPACK_,32_1),
@@ -2391,73 +2398,74 @@ BITPACK_F32 T2(_BITPACK_,a32)[] = {
   &T2(_BITPACK_,32_32)
 };
 unsigned char *T2(_BITPACK_,32)( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , unsigned b) { return T2(_BITPACK_,a32)[ b](in, n, out); }
+#undef USIZE
 
 #define USIZE 64
-unsigned char *T2(_BITPACK_,64_0)( uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
-unsigned char *T2(_BITPACK_,64_1)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*1); uint64_t v,x;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_2)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*2); uint64_t v,x;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_3)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*3); uint64_t v,x;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_4)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*4); uint64_t v,x;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_5)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*5); uint64_t v,x;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_6)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*6); uint64_t v,x;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_7)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*7); uint64_t v,x;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_8)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*8); uint64_t v,x;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_9)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*9); uint64_t v,x;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_10)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*10); uint64_t v,x;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_11)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*11); uint64_t v,x;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_12)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*12); uint64_t v,x;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_13)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*13); uint64_t v,x;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_14)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*14); uint64_t v,x;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_15)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*15); uint64_t v,x;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_16)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*16); uint64_t v,x;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_17)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*17); uint64_t v,x;do { BITPACK64_17( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_18)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*18); uint64_t v,x;do { BITPACK64_18( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_19)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*19); uint64_t v,x;do { BITPACK64_19( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_20)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*20); uint64_t v,x;do { BITPACK64_20( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_21)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*21); uint64_t v,x;do { BITPACK64_21( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_22)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*22); uint64_t v,x;do { BITPACK64_22( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_23)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*23); uint64_t v,x;do { BITPACK64_23( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_24)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*24); uint64_t v,x;do { BITPACK64_24( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_25)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*25); uint64_t v,x;do { BITPACK64_25( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_26)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*26); uint64_t v,x;do { BITPACK64_26( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_27)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*27); uint64_t v,x;do { BITPACK64_27( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_28)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*28); uint64_t v,x;do { BITPACK64_28( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_29)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*29); uint64_t v,x;do { BITPACK64_29( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_30)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*30); uint64_t v,x;do { BITPACK64_30( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_31)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*31); uint64_t v,x;do { BITPACK64_31( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_32)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*32); uint64_t v,x;do { BITPACK64_32( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_33)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*33); uint64_t v,x;do { BITPACK64_33( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_34)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*34); uint64_t v,x;do { BITPACK64_34( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_35)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*35); uint64_t v,x;do { BITPACK64_35( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_36)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*36); uint64_t v,x;do { BITPACK64_36( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_37)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*37); uint64_t v,x;do { BITPACK64_37( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_38)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*38); uint64_t v,x;do { BITPACK64_38( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_39)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*39); uint64_t v,x;do { BITPACK64_39( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_40)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*40); uint64_t v,x;do { BITPACK64_40( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_41)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*41); uint64_t v,x;do { BITPACK64_41( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_42)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*42); uint64_t v,x;do { BITPACK64_42( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_43)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*43); uint64_t v,x;do { BITPACK64_43( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_44)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*44); uint64_t v,x;do { BITPACK64_44( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_45)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*45); uint64_t v,x;do { BITPACK64_45( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_46)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*46); uint64_t v,x;do { BITPACK64_46( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_47)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*47); uint64_t v,x;do { BITPACK64_47( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_48)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*48); uint64_t v,x;do { BITPACK64_48( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_49)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*49); uint64_t v,x;do { BITPACK64_49( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_50)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*50); uint64_t v,x;do { BITPACK64_50( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_51)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*51); uint64_t v,x;do { BITPACK64_51( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_52)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*52); uint64_t v,x;do { BITPACK64_52( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_53)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*53); uint64_t v,x;do { BITPACK64_53( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_54)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*54); uint64_t v,x;do { BITPACK64_54( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_55)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*55); uint64_t v,x;do { BITPACK64_55( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_56)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*56); uint64_t v,x;do { BITPACK64_56( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_57)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*57); uint64_t v,x;do { BITPACK64_57( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_58)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*58); uint64_t v,x;do { BITPACK64_58( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_59)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*59); uint64_t v,x;do { BITPACK64_59( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_60)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*60); uint64_t v,x;do { BITPACK64_60( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_61)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*61); uint64_t v,x;do { BITPACK64_61( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_62)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*62); uint64_t v,x;do { BITPACK64_62( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_63)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*63); uint64_t v,x;do { BITPACK64_63( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_64)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { unsigned char *out_=out+PAD8(n*64); uint64_t v,x;do { BITPACK64_64( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,64_0 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { return out; }
+unsigned char *T2(_BITPACK_,64_1 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 1,64);}
+unsigned char *T2(_BITPACK_,64_2 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 2,64);}
+unsigned char *T2(_BITPACK_,64_3 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 3,64);}
+unsigned char *T2(_BITPACK_,64_4 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 4,64);}
+unsigned char *T2(_BITPACK_,64_5 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 5,64);}
+unsigned char *T2(_BITPACK_,64_6 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 6,64);}
+unsigned char *T2(_BITPACK_,64_7 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 7,64);}
+unsigned char *T2(_BITPACK_,64_8 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 8,64);}
+unsigned char *T2(_BITPACK_,64_9 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP( 9,64);}
+unsigned char *T2(_BITPACK_,64_10)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(10,64);}
+unsigned char *T2(_BITPACK_,64_11)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(11,64);}
+unsigned char *T2(_BITPACK_,64_12)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(12,64);}
+unsigned char *T2(_BITPACK_,64_13)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(13,64);}
+unsigned char *T2(_BITPACK_,64_14)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(14,64);}
+unsigned char *T2(_BITPACK_,64_15)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(15,64);}
+unsigned char *T2(_BITPACK_,64_16)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(16,64);}
+unsigned char *T2(_BITPACK_,64_17)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(17,64);}
+unsigned char *T2(_BITPACK_,64_18)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(18,64);}
+unsigned char *T2(_BITPACK_,64_19)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(19,64);}
+unsigned char *T2(_BITPACK_,64_20)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(20,64);}
+unsigned char *T2(_BITPACK_,64_21)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(21,64);}
+unsigned char *T2(_BITPACK_,64_22)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(22,64);}
+unsigned char *T2(_BITPACK_,64_23)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(23,64);}
+unsigned char *T2(_BITPACK_,64_24)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(24,64);}
+unsigned char *T2(_BITPACK_,64_25)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(25,64);}
+unsigned char *T2(_BITPACK_,64_26)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(26,64);}
+unsigned char *T2(_BITPACK_,64_27)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(27,64);}
+unsigned char *T2(_BITPACK_,64_28)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(28,64);}
+unsigned char *T2(_BITPACK_,64_29)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(29,64);}
+unsigned char *T2(_BITPACK_,64_30)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(30,64);}
+unsigned char *T2(_BITPACK_,64_31)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(31,64);}
+unsigned char *T2(_BITPACK_,64_32)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(32,64);}
+unsigned char *T2(_BITPACK_,64_33)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(33,64);}
+unsigned char *T2(_BITPACK_,64_34)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(34,64);}
+unsigned char *T2(_BITPACK_,64_35)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(35,64);}
+unsigned char *T2(_BITPACK_,64_36)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(36,64);}
+unsigned char *T2(_BITPACK_,64_37)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(37,64);}
+unsigned char *T2(_BITPACK_,64_38)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(38,64);}
+unsigned char *T2(_BITPACK_,64_39)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(39,64);}
+unsigned char *T2(_BITPACK_,64_40)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(40,64);}
+unsigned char *T2(_BITPACK_,64_41)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(41,64);}
+unsigned char *T2(_BITPACK_,64_42)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(42,64);}
+unsigned char *T2(_BITPACK_,64_43)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(43,64);}
+unsigned char *T2(_BITPACK_,64_44)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(44,64);}
+unsigned char *T2(_BITPACK_,64_45)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(45,64);}
+unsigned char *T2(_BITPACK_,64_46)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(46,64);}
+unsigned char *T2(_BITPACK_,64_47)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(47,64);}
+unsigned char *T2(_BITPACK_,64_48)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(48,64);}
+unsigned char *T2(_BITPACK_,64_49)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(49,64);}
+unsigned char *T2(_BITPACK_,64_50)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(50,64);}
+unsigned char *T2(_BITPACK_,64_51)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(51,64);}
+unsigned char *T2(_BITPACK_,64_52)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(52,64);}
+unsigned char *T2(_BITPACK_,64_53)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(53,64);}
+unsigned char *T2(_BITPACK_,64_54)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(54,64);}
+unsigned char *T2(_BITPACK_,64_55)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(55,64);}
+unsigned char *T2(_BITPACK_,64_56)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(56,64);}
+unsigned char *T2(_BITPACK_,64_57)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(57,64);}
+unsigned char *T2(_BITPACK_,64_58)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(58,64);}
+unsigned char *T2(_BITPACK_,64_59)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(59,64);}
+unsigned char *T2(_BITPACK_,64_60)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(60,64);}
+unsigned char *T2(_BITPACK_,64_61)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(61,64);}
+unsigned char *T2(_BITPACK_,64_62)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(62,64);}
+unsigned char *T2(_BITPACK_,64_63)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(63,64);}
+unsigned char *T2(_BITPACK_,64_64)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out  ) { BP(64,64);}
 BITPACK_F64 T2(_BITPACK_,a64)[] = {
   &T2(_BITPACK_,64_0),
   &T2(_BITPACK_,64_1),
@@ -2526,18 +2534,19 @@ BITPACK_F64 T2(_BITPACK_,a64)[] = {
   &T2(_BITPACK_,64_64)
 };
 unsigned char *T2(_BITPACK_,64)( uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , unsigned b) { return T2(_BITPACK_,a64)[ b](in, n, out); }
+#undef USIZE
 
 #else
 #define USIZE 8
 unsigned char *T2(_BITPACK_,8_0)( uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { return out; }
-unsigned char *T2(_BITPACK_,8_1)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*1); uint8_t v,x=0;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_2)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*2); uint8_t v,x=0;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_3)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*3); uint8_t v,x=0;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_4)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*4); uint8_t v,x=0;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_5)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*5); uint8_t v,x=0;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_6)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*6); uint8_t v,x=0;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_7)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*7); uint8_t v,x=0;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,8_8)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { unsigned char *out_=out+PAD8(n*8); uint8_t v,x=0;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,8_1)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(1,8);}
+unsigned char *T2(_BITPACK_,8_2)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(2,8);}
+unsigned char *T2(_BITPACK_,8_3)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(3,8);}
+unsigned char *T2(_BITPACK_,8_4)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(4,8);}
+unsigned char *T2(_BITPACK_,8_5)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(5,8);}
+unsigned char *T2(_BITPACK_,8_6)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(6,8);}
+unsigned char *T2(_BITPACK_,8_7)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(7,8);}
+unsigned char *T2(_BITPACK_,8_8)(uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start ) { BP(8,8);}
 BITPACK_D8 T2(_BITPACK_,a8)[] = {
   &T2(_BITPACK_,8_0),
   &T2(_BITPACK_,8_1),
@@ -2550,25 +2559,26 @@ BITPACK_D8 T2(_BITPACK_,a8)[] = {
   &T2(_BITPACK_,8_8)
 };
 unsigned char *T2(_BITPACK_,8)( uint8_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint8_t start, unsigned b) { return T2(_BITPACK_,a8)[ b](in, n, out, start); }
+#undef USIZE
 
 #define USIZE 16
-unsigned char *T2(_BITPACK_,16_0)( uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { return out; }
-unsigned char *T2(_BITPACK_,16_1)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*1); uint16_t v,x=0;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_2)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*2); uint16_t v,x=0;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_3)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*3); uint16_t v,x=0;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_4)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*4); uint16_t v,x=0;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_5)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*5); uint16_t v,x=0;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_6)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*6); uint16_t v,x=0;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_7)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*7); uint16_t v,x=0;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_8)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*8); uint16_t v,x=0;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_9)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*9); uint16_t v,x=0;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_10)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*10); uint16_t v,x=0;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_11)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*11); uint16_t v,x=0;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_12)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*12); uint16_t v,x=0;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_13)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*13); uint16_t v,x=0;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_14)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*14); uint16_t v,x=0;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_15)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*15); uint16_t v,x=0;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,16_16)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { unsigned char *out_=out+PAD8(n*16); uint16_t v,x=0;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,16_0 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { return out; }
+unsigned char *T2(_BITPACK_,16_1 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 1,16);}
+unsigned char *T2(_BITPACK_,16_2 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 2,16);}
+unsigned char *T2(_BITPACK_,16_3 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 3,16);}
+unsigned char *T2(_BITPACK_,16_4 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 4,16);}
+unsigned char *T2(_BITPACK_,16_5 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 5,16);}
+unsigned char *T2(_BITPACK_,16_6 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 6,16);}
+unsigned char *T2(_BITPACK_,16_7 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 7,16);}
+unsigned char *T2(_BITPACK_,16_8 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 8,16);}
+unsigned char *T2(_BITPACK_,16_9 )(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP( 9,16);}
+unsigned char *T2(_BITPACK_,16_10)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(10,16);}
+unsigned char *T2(_BITPACK_,16_11)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(11,16);}
+unsigned char *T2(_BITPACK_,16_12)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(12,16);}
+unsigned char *T2(_BITPACK_,16_13)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(13,16);}
+unsigned char *T2(_BITPACK_,16_14)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(14,16);}
+unsigned char *T2(_BITPACK_,16_15)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(15,16);}
+unsigned char *T2(_BITPACK_,16_16)(uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start ) { BP(16,16);}
 BITPACK_D16 T2(_BITPACK_,a16)[] = {
   &T2(_BITPACK_,16_0),
   &T2(_BITPACK_,16_1),
@@ -2589,41 +2599,42 @@ BITPACK_D16 T2(_BITPACK_,a16)[] = {
   &T2(_BITPACK_,16_16)
 };
 unsigned char *T2(_BITPACK_,16)( uint16_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint16_t start, unsigned b) { return T2(_BITPACK_,a16)[ b](in, n, out, start); }
+#undef USIZE
 
 #define USIZE 32
-unsigned char *T2(_BITPACK_,32_0)( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { return out; }
-unsigned char *T2(_BITPACK_,32_1)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*1); uint32_t v,x=0;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_2)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*2); uint32_t v,x=0;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_3)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*3); uint32_t v,x=0;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_4)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*4); uint32_t v,x=0;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_5)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*5); uint32_t v,x=0;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_6)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*6); uint32_t v,x=0;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_7)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*7); uint32_t v,x=0;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_8)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*8); uint32_t v,x=0;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_9)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*9); uint32_t v,x=0;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_10)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*10); uint32_t v,x=0;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_11)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*11); uint32_t v,x=0;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_12)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*12); uint32_t v,x=0;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_13)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*13); uint32_t v,x=0;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_14)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*14); uint32_t v,x=0;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_15)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*15); uint32_t v,x=0;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_16)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*16); uint32_t v,x=0;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_17)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*17); uint32_t v,x=0;do { BITPACK64_17( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_18)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*18); uint32_t v,x=0;do { BITPACK64_18( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_19)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*19); uint32_t v,x=0;do { BITPACK64_19( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_20)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*20); uint32_t v,x=0;do { BITPACK64_20( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_21)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*21); uint32_t v,x=0;do { BITPACK64_21( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_22)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*22); uint32_t v,x=0;do { BITPACK64_22( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_23)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*23); uint32_t v,x=0;do { BITPACK64_23( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_24)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*24); uint32_t v,x=0;do { BITPACK64_24( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_25)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*25); uint32_t v,x=0;do { BITPACK64_25( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_26)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*26); uint32_t v,x=0;do { BITPACK64_26( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_27)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*27); uint32_t v,x=0;do { BITPACK64_27( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_28)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*28); uint32_t v,x=0;do { BITPACK64_28( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_29)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*29); uint32_t v,x=0;do { BITPACK64_29( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_30)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*30); uint32_t v,x=0;do { BITPACK64_30( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_31)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*31); uint32_t v,x=0;do { BITPACK64_31( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,32_32)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { unsigned char *out_=out+PAD8(n*32); uint32_t v,x=0;do { BITPACK64_32( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,32_0 )( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { return out; }
+unsigned char *T2(_BITPACK_,32_1 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 1,32);}
+unsigned char *T2(_BITPACK_,32_2 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 2,32);}
+unsigned char *T2(_BITPACK_,32_3 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 3,32);}
+unsigned char *T2(_BITPACK_,32_4 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 4,32);}
+unsigned char *T2(_BITPACK_,32_5 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 5,32);}
+unsigned char *T2(_BITPACK_,32_6 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 6,32);}
+unsigned char *T2(_BITPACK_,32_7 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 7,32);}
+unsigned char *T2(_BITPACK_,32_8 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 8,32);}
+unsigned char *T2(_BITPACK_,32_9 )(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP( 9,32);}
+unsigned char *T2(_BITPACK_,32_10)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(10,32);}
+unsigned char *T2(_BITPACK_,32_11)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(11,32);}
+unsigned char *T2(_BITPACK_,32_12)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(12,32);}
+unsigned char *T2(_BITPACK_,32_13)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(13,32);}
+unsigned char *T2(_BITPACK_,32_14)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(14,32);}
+unsigned char *T2(_BITPACK_,32_15)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(15,32);}
+unsigned char *T2(_BITPACK_,32_16)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(16,32);}
+unsigned char *T2(_BITPACK_,32_17)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(17,32);}
+unsigned char *T2(_BITPACK_,32_18)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(18,32);}
+unsigned char *T2(_BITPACK_,32_19)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(19,32);}
+unsigned char *T2(_BITPACK_,32_20)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(20,32);}
+unsigned char *T2(_BITPACK_,32_21)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(21,32);}
+unsigned char *T2(_BITPACK_,32_22)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(22,32);}
+unsigned char *T2(_BITPACK_,32_23)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(23,32);}
+unsigned char *T2(_BITPACK_,32_24)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(24,32);}
+unsigned char *T2(_BITPACK_,32_25)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(25,32);}
+unsigned char *T2(_BITPACK_,32_26)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(26,32);}
+unsigned char *T2(_BITPACK_,32_27)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(27,32);}
+unsigned char *T2(_BITPACK_,32_28)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(28,32);}
+unsigned char *T2(_BITPACK_,32_29)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(29,32);}
+unsigned char *T2(_BITPACK_,32_30)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(30,32);}
+unsigned char *T2(_BITPACK_,32_31)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(31,32);}
+unsigned char *T2(_BITPACK_,32_32)(uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start ) { BP(32,32);}
 BITPACK_D32 T2(_BITPACK_,a32)[] = {
   &T2(_BITPACK_,32_0),
   &T2(_BITPACK_,32_1),
@@ -2660,73 +2671,74 @@ BITPACK_D32 T2(_BITPACK_,a32)[] = {
   &T2(_BITPACK_,32_32)
 };
 unsigned char *T2(_BITPACK_,32)( uint32_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint32_t start, unsigned b) { return T2(_BITPACK_,a32)[ b](in, n, out, start); }
+#undef USIZE
 
 #define USIZE 64
-unsigned char *T2(_BITPACK_,64_0)( uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { return out; }
-unsigned char *T2(_BITPACK_,64_1)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*1); uint64_t v,x=0;do { BITPACK64_1( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_2)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*2); uint64_t v,x=0;do { BITPACK64_2( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_3)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*3); uint64_t v,x=0;do { BITPACK64_3( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_4)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*4); uint64_t v,x=0;do { BITPACK64_4( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_5)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*5); uint64_t v,x=0;do { BITPACK64_5( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_6)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*6); uint64_t v,x=0;do { BITPACK64_6( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_7)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*7); uint64_t v,x=0;do { BITPACK64_7( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_8)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*8); uint64_t v,x=0;do { BITPACK64_8( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_9)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*9); uint64_t v,x=0;do { BITPACK64_9( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_10)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*10); uint64_t v,x=0;do { BITPACK64_10( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_11)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*11); uint64_t v,x=0;do { BITPACK64_11( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_12)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*12); uint64_t v,x=0;do { BITPACK64_12( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_13)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*13); uint64_t v,x=0;do { BITPACK64_13( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_14)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*14); uint64_t v,x=0;do { BITPACK64_14( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_15)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*15); uint64_t v,x=0;do { BITPACK64_15( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_16)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*16); uint64_t v,x=0;do { BITPACK64_16( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_17)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*17); uint64_t v,x=0;do { BITPACK64_17( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_18)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*18); uint64_t v,x=0;do { BITPACK64_18( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_19)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*19); uint64_t v,x=0;do { BITPACK64_19( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_20)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*20); uint64_t v,x=0;do { BITPACK64_20( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_21)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*21); uint64_t v,x=0;do { BITPACK64_21( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_22)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*22); uint64_t v,x=0;do { BITPACK64_22( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_23)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*23); uint64_t v,x=0;do { BITPACK64_23( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_24)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*24); uint64_t v,x=0;do { BITPACK64_24( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_25)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*25); uint64_t v,x=0;do { BITPACK64_25( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_26)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*26); uint64_t v,x=0;do { BITPACK64_26( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_27)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*27); uint64_t v,x=0;do { BITPACK64_27( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_28)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*28); uint64_t v,x=0;do { BITPACK64_28( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_29)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*29); uint64_t v,x=0;do { BITPACK64_29( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_30)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*30); uint64_t v,x=0;do { BITPACK64_30( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_31)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*31); uint64_t v,x=0;do { BITPACK64_31( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_32)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*32); uint64_t v,x=0;do { BITPACK64_32( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_33)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*33); uint64_t v,x=0;do { BITPACK64_33( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_34)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*34); uint64_t v,x=0;do { BITPACK64_34( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_35)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*35); uint64_t v,x=0;do { BITPACK64_35( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_36)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*36); uint64_t v,x=0;do { BITPACK64_36( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_37)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*37); uint64_t v,x=0;do { BITPACK64_37( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_38)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*38); uint64_t v,x=0;do { BITPACK64_38( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_39)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*39); uint64_t v,x=0;do { BITPACK64_39( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_40)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*40); uint64_t v,x=0;do { BITPACK64_40( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_41)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*41); uint64_t v,x=0;do { BITPACK64_41( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_42)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*42); uint64_t v,x=0;do { BITPACK64_42( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_43)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*43); uint64_t v,x=0;do { BITPACK64_43( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_44)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*44); uint64_t v,x=0;do { BITPACK64_44( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_45)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*45); uint64_t v,x=0;do { BITPACK64_45( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_46)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*46); uint64_t v,x=0;do { BITPACK64_46( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_47)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*47); uint64_t v,x=0;do { BITPACK64_47( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_48)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*48); uint64_t v,x=0;do { BITPACK64_48( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_49)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*49); uint64_t v,x=0;do { BITPACK64_49( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_50)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*50); uint64_t v,x=0;do { BITPACK64_50( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_51)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*51); uint64_t v,x=0;do { BITPACK64_51( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_52)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*52); uint64_t v,x=0;do { BITPACK64_52( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_53)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*53); uint64_t v,x=0;do { BITPACK64_53( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_54)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*54); uint64_t v,x=0;do { BITPACK64_54( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_55)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*55); uint64_t v,x=0;do { BITPACK64_55( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_56)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*56); uint64_t v,x=0;do { BITPACK64_56( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_57)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*57); uint64_t v,x=0;do { BITPACK64_57( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_58)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*58); uint64_t v,x=0;do { BITPACK64_58( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_59)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*59); uint64_t v,x=0;do { BITPACK64_59( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_60)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*60); uint64_t v,x=0;do { BITPACK64_60( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_61)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*61); uint64_t v,x=0;do { BITPACK64_61( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_62)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*62); uint64_t v,x=0;do { BITPACK64_62( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_63)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*63); uint64_t v,x=0;do { BITPACK64_63( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
-unsigned char *T2(_BITPACK_,64_64)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { unsigned char *out_=out+PAD8(n*64); uint64_t v,x=0;do { BITPACK64_64( in, out, start); PREFETCH(in+512,0); } while(out<out_); return out_; }
+unsigned char *T2(_BITPACK_,64_0 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { return out; }
+unsigned char *T2(_BITPACK_,64_1 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 1,64);}
+unsigned char *T2(_BITPACK_,64_2 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 2,64);}
+unsigned char *T2(_BITPACK_,64_3 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 3,64);}
+unsigned char *T2(_BITPACK_,64_4 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 4,64);}
+unsigned char *T2(_BITPACK_,64_5 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 5,64);}
+unsigned char *T2(_BITPACK_,64_6 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 6,64);}
+unsigned char *T2(_BITPACK_,64_7 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 7,64);}
+unsigned char *T2(_BITPACK_,64_8 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 8,64);}
+unsigned char *T2(_BITPACK_,64_9 )(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP( 9,64);}
+unsigned char *T2(_BITPACK_,64_10)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(10,64);}
+unsigned char *T2(_BITPACK_,64_11)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(11,64);}
+unsigned char *T2(_BITPACK_,64_12)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(12,64);}
+unsigned char *T2(_BITPACK_,64_13)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(13,64);}
+unsigned char *T2(_BITPACK_,64_14)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(14,64);}
+unsigned char *T2(_BITPACK_,64_15)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(15,64);}
+unsigned char *T2(_BITPACK_,64_16)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(16,64);}
+unsigned char *T2(_BITPACK_,64_17)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(17,64);}
+unsigned char *T2(_BITPACK_,64_18)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(18,64);}
+unsigned char *T2(_BITPACK_,64_19)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(19,64);}
+unsigned char *T2(_BITPACK_,64_20)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(20,64);}
+unsigned char *T2(_BITPACK_,64_21)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(21,64);}
+unsigned char *T2(_BITPACK_,64_22)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(22,64);}
+unsigned char *T2(_BITPACK_,64_23)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(23,64);}
+unsigned char *T2(_BITPACK_,64_24)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(24,64);}
+unsigned char *T2(_BITPACK_,64_25)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(25,64);}
+unsigned char *T2(_BITPACK_,64_26)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(26,64);}
+unsigned char *T2(_BITPACK_,64_27)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(27,64);}
+unsigned char *T2(_BITPACK_,64_28)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(28,64);}
+unsigned char *T2(_BITPACK_,64_29)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(29,64);}
+unsigned char *T2(_BITPACK_,64_30)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(30,64);}
+unsigned char *T2(_BITPACK_,64_31)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(31,64);}
+unsigned char *T2(_BITPACK_,64_32)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(32,64);}
+unsigned char *T2(_BITPACK_,64_33)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(33,64);}
+unsigned char *T2(_BITPACK_,64_34)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(34,64);}
+unsigned char *T2(_BITPACK_,64_35)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(35,64);}
+unsigned char *T2(_BITPACK_,64_36)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(36,64);}
+unsigned char *T2(_BITPACK_,64_37)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(37,64);}
+unsigned char *T2(_BITPACK_,64_38)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(38,64);}
+unsigned char *T2(_BITPACK_,64_39)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(39,64);}
+unsigned char *T2(_BITPACK_,64_40)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(40,64);}
+unsigned char *T2(_BITPACK_,64_41)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(41,64);}
+unsigned char *T2(_BITPACK_,64_42)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(42,64);}
+unsigned char *T2(_BITPACK_,64_43)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(43,64);}
+unsigned char *T2(_BITPACK_,64_44)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(44,64);}
+unsigned char *T2(_BITPACK_,64_45)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(45,64);}
+unsigned char *T2(_BITPACK_,64_46)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(46,64);}
+unsigned char *T2(_BITPACK_,64_47)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(47,64);}
+unsigned char *T2(_BITPACK_,64_48)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(48,64);}
+unsigned char *T2(_BITPACK_,64_49)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(49,64);}
+unsigned char *T2(_BITPACK_,64_50)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(50,64);}
+unsigned char *T2(_BITPACK_,64_51)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(51,64);}
+unsigned char *T2(_BITPACK_,64_52)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(52,64);}
+unsigned char *T2(_BITPACK_,64_53)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(53,64);}
+unsigned char *T2(_BITPACK_,64_54)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(54,64);}
+unsigned char *T2(_BITPACK_,64_55)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(55,64);}
+unsigned char *T2(_BITPACK_,64_56)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(56,64);}
+unsigned char *T2(_BITPACK_,64_57)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(57,64);}
+unsigned char *T2(_BITPACK_,64_58)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(58,64);}
+unsigned char *T2(_BITPACK_,64_59)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(59,64);}
+unsigned char *T2(_BITPACK_,64_60)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(60,64);}
+unsigned char *T2(_BITPACK_,64_61)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(61,64);}
+unsigned char *T2(_BITPACK_,64_62)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(62,64);}
+unsigned char *T2(_BITPACK_,64_63)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(63,64);}
+unsigned char *T2(_BITPACK_,64_64)(uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start ) { BP(64,64);}
 BITPACK_D64 T2(_BITPACK_,a64)[] = {
   &T2(_BITPACK_,64_0),
   &T2(_BITPACK_,64_1),
@@ -2795,6 +2807,7 @@ BITPACK_D64 T2(_BITPACK_,a64)[] = {
   &T2(_BITPACK_,64_64)
 };
 unsigned char *T2(_BITPACK_,64)( uint64_t *__restrict in, unsigned n, const unsigned char *__restrict out , uint64_t start, unsigned b) { return T2(_BITPACK_,a64)[ b](in, n, out, start); }
+#undef USIZE
 
 #endif
 #endif //IP9