From aba596ef0d173780bee4ecc17557cc8a71c7f207 Mon Sep 17 00:00:00 2001
From: x <powturbo@gmail.com>
Date: Wed, 16 Oct 2019 19:47:30 +0200
Subject: [PATCH] TurboPFor: Bit UnPack include

---
 bitunpack_.h | 545 +++++++++++++++++++++++----------------------------
 1 file changed, 249 insertions(+), 296 deletions(-)

diff --git a/bitunpack_.h b/bitunpack_.h
index b876add..de5738b 100644
--- a/bitunpack_.h
+++ b/bitunpack_.h
@@ -3105,15 +3105,15 @@
 
 #ifndef DELTA
 #define USIZE 8
-unsigned char *TEMPLATE2(_BITUNPACK_,8_0)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint8_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_1)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_2)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_3)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_4)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_5)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_6)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_7)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_8)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_0)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint8_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_1)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_2)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_3)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_4)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_5)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_6)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_7)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_8)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_F8 TEMPLATE2(_BITUNPACK_,a8)[] = {
   &TEMPLATE2(_BITUNPACK_,8_0),
   &TEMPLATE2(_BITUNPACK_,8_1),
@@ -3128,23 +3128,23 @@ BITUNPACK_F8 TEMPLATE2(_BITUNPACK_,a8)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,8)( const unsigned char *__restrict in, unsigned n, uint8_t  *__restrict out , unsigned b) { return TEMPLATE2(_BITUNPACK_,a8)[ b](in, n, out); }
 
 #define USIZE 16
-unsigned char *TEMPLATE2(_BITUNPACK_,16_0)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint16_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_1)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_2)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_3)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_4)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_5)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_6)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_7)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_8)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_9)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_10)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_11)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_12)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_13)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_14)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_15)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_16)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_0)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint16_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_1)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_2)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_3)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_4)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_5)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_6)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_7)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_8)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_9)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_10)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_11)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_12)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_13)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_14)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_15)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_16)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_F16 TEMPLATE2(_BITUNPACK_,a16)[] = {
   &TEMPLATE2(_BITUNPACK_,16_0),
   &TEMPLATE2(_BITUNPACK_,16_1),
@@ -3167,39 +3167,39 @@ BITUNPACK_F16 TEMPLATE2(_BITUNPACK_,a16)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,16)( const unsigned char *__restrict in, unsigned n, uint16_t  *__restrict out , unsigned b) { return TEMPLATE2(_BITUNPACK_,a16)[ b](in, n, out); }
 
 #define USIZE 32
-unsigned char *TEMPLATE2(_BITUNPACK_,32_0)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint32_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_1)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_2)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_3)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_4)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_5)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_6)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_7)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_8)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_9)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_10)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_11)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_12)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_13)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_14)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_15)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_16)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_17)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*17); do { BITUNPACK64_17( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_18)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*18); do { BITUNPACK64_18( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_19)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*19); do { BITUNPACK64_19( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_20)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*20); do { BITUNPACK64_20( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_21)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*21); do { BITUNPACK64_21( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_22)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*22); do { BITUNPACK64_22( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_23)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*23); do { BITUNPACK64_23( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_24)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*24); do { BITUNPACK64_24( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_25)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*25); do { BITUNPACK64_25( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_26)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*26); do { BITUNPACK64_26( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_27)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*27); do { BITUNPACK64_27( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_28)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*28); do { BITUNPACK64_28( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_29)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*29); do { BITUNPACK64_29( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_30)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*30); do { BITUNPACK64_30( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_31)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*31); do { BITUNPACK64_31( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_32)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*32); do { BITUNPACK64_32( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_0)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint32_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_1)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_2)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_3)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_4)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_5)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_6)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_7)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_8)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_9)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_10)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_11)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_12)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_13)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_14)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_15)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_16)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_17)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*17); do { BITUNPACK64_17( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_18)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*18); do { BITUNPACK64_18( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_19)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*19); do { BITUNPACK64_19( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_20)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*20); do { BITUNPACK64_20( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_21)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*21); do { BITUNPACK64_21( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_22)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*22); do { BITUNPACK64_22( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_23)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*23); do { BITUNPACK64_23( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_24)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*24); do { BITUNPACK64_24( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_25)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*25); do { BITUNPACK64_25( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_26)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*26); do { BITUNPACK64_26( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_27)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*27); do { BITUNPACK64_27( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_28)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*28); do { BITUNPACK64_28( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_29)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*29); do { BITUNPACK64_29( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_30)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*30); do { BITUNPACK64_30( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_31)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*31); do { BITUNPACK64_31( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_32)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*32); do { BITUNPACK64_32( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_F32 TEMPLATE2(_BITUNPACK_,a32)[] = {
   &TEMPLATE2(_BITUNPACK_,32_0),
   &TEMPLATE2(_BITUNPACK_,32_1),
@@ -3238,71 +3238,71 @@ BITUNPACK_F32 TEMPLATE2(_BITUNPACK_,a32)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,32)( const unsigned char *__restrict in, unsigned n, uint32_t  *__restrict out , unsigned b) { return TEMPLATE2(_BITUNPACK_,a32)[ b](in, n, out); }
 
 #define USIZE 64
-unsigned char *TEMPLATE2(_BITUNPACK_,64_0)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint64_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_1)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_2)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_3)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_4)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_5)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_6)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_7)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_8)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_9)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_10)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_11)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_12)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_13)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_14)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_15)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_16)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_17)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*17); do { BITUNPACK64_17( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_18)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*18); do { BITUNPACK64_18( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_19)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*19); do { BITUNPACK64_19( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_20)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*20); do { BITUNPACK64_20( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_21)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*21); do { BITUNPACK64_21( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_22)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*22); do { BITUNPACK64_22( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_23)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*23); do { BITUNPACK64_23( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_24)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*24); do { BITUNPACK64_24( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_25)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*25); do { BITUNPACK64_25( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_26)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*26); do { BITUNPACK64_26( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_27)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*27); do { BITUNPACK64_27( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_28)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*28); do { BITUNPACK64_28( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_29)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*29); do { BITUNPACK64_29( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_30)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*30); do { BITUNPACK64_30( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_31)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*31); do { BITUNPACK64_31( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_32)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*32); do { BITUNPACK64_32( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_33)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*33); do { BITUNPACK64_33( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_34)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*34); do { BITUNPACK64_34( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_35)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*35); do { BITUNPACK64_35( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_36)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*36); do { BITUNPACK64_36( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_37)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*37); do { BITUNPACK64_37( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_38)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*38); do { BITUNPACK64_38( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_39)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*39); do { BITUNPACK64_39( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_40)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*40); do { BITUNPACK64_40( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_41)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*41); do { BITUNPACK64_41( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_42)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*42); do { BITUNPACK64_42( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_43)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*43); do { BITUNPACK64_43( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_44)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*44); do { BITUNPACK64_44( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_45)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*45); do { BITUNPACK64_45( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_46)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*46); do { BITUNPACK64_46( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_47)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*47); do { BITUNPACK64_47( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_48)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*48); do { BITUNPACK64_48( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_49)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*49); do { BITUNPACK64_49( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_50)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*50); do { BITUNPACK64_50( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_51)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*51); do { BITUNPACK64_51( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_52)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*52); do { BITUNPACK64_52( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_53)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*53); do { BITUNPACK64_53( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_54)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*54); do { BITUNPACK64_54( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_55)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*55); do { BITUNPACK64_55( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_56)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*56); do { BITUNPACK64_56( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_57)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*57); do { BITUNPACK64_57( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_58)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*58); do { BITUNPACK64_58( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_59)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*59); do { BITUNPACK64_59( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_60)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*60); do { BITUNPACK64_60( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_61)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*61); do { BITUNPACK64_61( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_62)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*62); do { BITUNPACK64_62( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_63)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*63); do { BITUNPACK64_63( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_64)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*64); do { BITUNPACK64_64( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_0)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*0); const uint64_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_1)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*1); do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_2)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*2); do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_3)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*3); do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_4)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*4); do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_5)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*5); do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_6)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*6); do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_7)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*7); do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_8)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*8); do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_9)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*9); do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_10)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*10); do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_11)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*11); do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_12)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*12); do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_13)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*13); do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_14)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*14); do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_15)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*15); do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_16)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*16); do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_17)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*17); do { BITUNPACK64_17( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_18)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*18); do { BITUNPACK64_18( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_19)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*19); do { BITUNPACK64_19( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_20)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*20); do { BITUNPACK64_20( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_21)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*21); do { BITUNPACK64_21( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_22)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*22); do { BITUNPACK64_22( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_23)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*23); do { BITUNPACK64_23( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_24)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*24); do { BITUNPACK64_24( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_25)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*25); do { BITUNPACK64_25( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_26)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*26); do { BITUNPACK64_26( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_27)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*27); do { BITUNPACK64_27( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_28)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*28); do { BITUNPACK64_28( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_29)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*29); do { BITUNPACK64_29( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_30)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*30); do { BITUNPACK64_30( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_31)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*31); do { BITUNPACK64_31( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_32)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*32); do { BITUNPACK64_32( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_33)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*33); do { BITUNPACK64_33( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_34)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*34); do { BITUNPACK64_34( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_35)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*35); do { BITUNPACK64_35( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_36)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*36); do { BITUNPACK64_36( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_37)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*37); do { BITUNPACK64_37( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_38)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*38); do { BITUNPACK64_38( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_39)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*39); do { BITUNPACK64_39( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_40)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*40); do { BITUNPACK64_40( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_41)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*41); do { BITUNPACK64_41( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_42)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*42); do { BITUNPACK64_42( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_43)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*43); do { BITUNPACK64_43( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_44)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*44); do { BITUNPACK64_44( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_45)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*45); do { BITUNPACK64_45( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_46)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*46); do { BITUNPACK64_46( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_47)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*47); do { BITUNPACK64_47( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_48)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*48); do { BITUNPACK64_48( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_49)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*49); do { BITUNPACK64_49( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_50)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*50); do { BITUNPACK64_50( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_51)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*51); do { BITUNPACK64_51( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_52)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*52); do { BITUNPACK64_52( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_53)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*53); do { BITUNPACK64_53( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_54)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*54); do { BITUNPACK64_54( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_55)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*55); do { BITUNPACK64_55( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_56)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*56); do { BITUNPACK64_56( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_57)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*57); do { BITUNPACK64_57( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_58)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*58); do { BITUNPACK64_58( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_59)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*59); do { BITUNPACK64_59( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_60)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*60); do { BITUNPACK64_60( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_61)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*61); do { BITUNPACK64_61( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_62)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*62); do { BITUNPACK64_62( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_63)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*63); do { BITUNPACK64_63( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_64)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out  ) { unsigned char *in_=in+PAD8(n*64); do { BITUNPACK64_64( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_F64 TEMPLATE2(_BITUNPACK_,a64)[] = {
   &TEMPLATE2(_BITUNPACK_,64_0),
   &TEMPLATE2(_BITUNPACK_,64_1),
@@ -3374,15 +3374,15 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #else
 #define USIZE 8
-unsigned char *TEMPLATE2(_BITUNPACK_,8_0)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint8_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_1)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_2)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_3)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_4)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_5)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_6)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_7)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,8_8)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_0)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint8_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_1)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_2)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_3)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_4)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_5)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_6)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_7)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,8_8)(const unsigned char *__restrict in, unsigned n, uint8_t *__restrict out , uint8_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_D8 TEMPLATE2(_BITUNPACK_,a8)[] = {
   &TEMPLATE2(_BITUNPACK_,8_0),
   &TEMPLATE2(_BITUNPACK_,8_1),
@@ -3397,23 +3397,23 @@ BITUNPACK_D8 TEMPLATE2(_BITUNPACK_,a8)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,8)( const unsigned char *__restrict in, unsigned n, uint8_t  *__restrict out , uint8_t start, unsigned b) { return TEMPLATE2(_BITUNPACK_,a8)[ b](in, n, out, start); }
 
 #define USIZE 16
-unsigned char *TEMPLATE2(_BITUNPACK_,16_0)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint16_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_1)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_2)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_3)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_4)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_5)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_6)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_7)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_8)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_9)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_10)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_11)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_12)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_13)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_14)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_15)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,16_16)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_0)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint16_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_1)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_2)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_3)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_4)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_5)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_6)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_7)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_8)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_9)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_10)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_11)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_12)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_13)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_14)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_15)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,16_16)(const unsigned char *__restrict in, unsigned n, uint16_t *__restrict out , uint16_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_D16 TEMPLATE2(_BITUNPACK_,a16)[] = {
   &TEMPLATE2(_BITUNPACK_,16_0),
   &TEMPLATE2(_BITUNPACK_,16_1),
@@ -3436,39 +3436,39 @@ BITUNPACK_D16 TEMPLATE2(_BITUNPACK_,a16)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,16)( const unsigned char *__restrict in, unsigned n, uint16_t  *__restrict out , uint16_t start, unsigned b) { return TEMPLATE2(_BITUNPACK_,a16)[ b](in, n, out, start); }
 
 #define USIZE 32
-unsigned char *TEMPLATE2(_BITUNPACK_,32_0)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint32_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_1)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_2)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_3)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_4)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_5)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_6)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_7)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_8)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_9)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_10)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_11)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_12)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_13)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_14)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_15)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_16)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_17)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*17),x=0; do { BITUNPACK64_17( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_18)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*18),x=0; do { BITUNPACK64_18( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_19)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*19),x=0; do { BITUNPACK64_19( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_20)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*20),x=0; do { BITUNPACK64_20( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_21)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*21),x=0; do { BITUNPACK64_21( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_22)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*22),x=0; do { BITUNPACK64_22( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_23)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*23),x=0; do { BITUNPACK64_23( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_24)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*24),x=0; do { BITUNPACK64_24( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_25)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*25),x=0; do { BITUNPACK64_25( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_26)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*26),x=0; do { BITUNPACK64_26( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_27)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*27),x=0; do { BITUNPACK64_27( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_28)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*28),x=0; do { BITUNPACK64_28( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_29)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*29),x=0; do { BITUNPACK64_29( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_30)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*30),x=0; do { BITUNPACK64_30( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_31)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*31),x=0; do { BITUNPACK64_31( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,32_32)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*32),x=0; do { BITUNPACK64_32( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_0)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint32_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_1)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_2)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_3)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_4)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_5)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_6)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_7)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_8)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_9)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_10)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_11)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_12)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_13)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_14)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_15)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_16)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_17)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*17),x=0; do { BITUNPACK64_17( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_18)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*18),x=0; do { BITUNPACK64_18( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_19)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*19),x=0; do { BITUNPACK64_19( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_20)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*20),x=0; do { BITUNPACK64_20( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_21)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*21),x=0; do { BITUNPACK64_21( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_22)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*22),x=0; do { BITUNPACK64_22( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_23)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*23),x=0; do { BITUNPACK64_23( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_24)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*24),x=0; do { BITUNPACK64_24( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_25)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*25),x=0; do { BITUNPACK64_25( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_26)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*26),x=0; do { BITUNPACK64_26( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_27)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*27),x=0; do { BITUNPACK64_27( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_28)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*28),x=0; do { BITUNPACK64_28( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_29)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*29),x=0; do { BITUNPACK64_29( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_30)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*30),x=0; do { BITUNPACK64_30( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_31)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*31),x=0; do { BITUNPACK64_31( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,32_32)(const unsigned char *__restrict in, unsigned n, uint32_t *__restrict out , uint32_t start ) { unsigned char *in_=in+PAD8(n*32),x=0; do { BITUNPACK64_32( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_D32 TEMPLATE2(_BITUNPACK_,a32)[] = {
   &TEMPLATE2(_BITUNPACK_,32_0),
   &TEMPLATE2(_BITUNPACK_,32_1),
@@ -3507,71 +3507,71 @@ BITUNPACK_D32 TEMPLATE2(_BITUNPACK_,a32)[] = {
 unsigned char *TEMPLATE2(_BITUNPACK_,32)( const unsigned char *__restrict in, unsigned n, uint32_t  *__restrict out , uint32_t start, unsigned b) { return TEMPLATE2(_BITUNPACK_,a32)[ b](in, n, out, start); }
 
 #define USIZE 64
-unsigned char *TEMPLATE2(_BITUNPACK_,64_0)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint64_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in); } while(out<out_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_1)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_2)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_3)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_4)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_5)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_6)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_7)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_8)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_9)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_10)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_11)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_12)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_13)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_14)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_15)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_16)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_17)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*17),x=0; do { BITUNPACK64_17( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_18)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*18),x=0; do { BITUNPACK64_18( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_19)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*19),x=0; do { BITUNPACK64_19( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_20)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*20),x=0; do { BITUNPACK64_20( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_21)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*21),x=0; do { BITUNPACK64_21( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_22)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*22),x=0; do { BITUNPACK64_22( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_23)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*23),x=0; do { BITUNPACK64_23( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_24)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*24),x=0; do { BITUNPACK64_24( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_25)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*25),x=0; do { BITUNPACK64_25( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_26)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*26),x=0; do { BITUNPACK64_26( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_27)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*27),x=0; do { BITUNPACK64_27( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_28)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*28),x=0; do { BITUNPACK64_28( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_29)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*29),x=0; do { BITUNPACK64_29( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_30)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*30),x=0; do { BITUNPACK64_30( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_31)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*31),x=0; do { BITUNPACK64_31( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_32)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*32),x=0; do { BITUNPACK64_32( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_33)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*33),x=0; do { BITUNPACK64_33( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_34)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*34),x=0; do { BITUNPACK64_34( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_35)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*35),x=0; do { BITUNPACK64_35( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_36)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*36),x=0; do { BITUNPACK64_36( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_37)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*37),x=0; do { BITUNPACK64_37( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_38)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*38),x=0; do { BITUNPACK64_38( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_39)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*39),x=0; do { BITUNPACK64_39( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_40)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*40),x=0; do { BITUNPACK64_40( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_41)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*41),x=0; do { BITUNPACK64_41( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_42)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*42),x=0; do { BITUNPACK64_42( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_43)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*43),x=0; do { BITUNPACK64_43( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_44)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*44),x=0; do { BITUNPACK64_44( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_45)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*45),x=0; do { BITUNPACK64_45( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_46)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*46),x=0; do { BITUNPACK64_46( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_47)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*47),x=0; do { BITUNPACK64_47( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_48)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*48),x=0; do { BITUNPACK64_48( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_49)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*49),x=0; do { BITUNPACK64_49( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_50)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*50),x=0; do { BITUNPACK64_50( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_51)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*51),x=0; do { BITUNPACK64_51( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_52)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*52),x=0; do { BITUNPACK64_52( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_53)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*53),x=0; do { BITUNPACK64_53( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_54)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*54),x=0; do { BITUNPACK64_54( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_55)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*55),x=0; do { BITUNPACK64_55( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_56)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*56),x=0; do { BITUNPACK64_56( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_57)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*57),x=0; do { BITUNPACK64_57( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_58)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*58),x=0; do { BITUNPACK64_58( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_59)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*59),x=0; do { BITUNPACK64_59( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_60)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*60),x=0; do { BITUNPACK64_60( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_61)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*61),x=0; do { BITUNPACK64_61( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_62)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*62),x=0; do { BITUNPACK64_62( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_63)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*63),x=0; do { BITUNPACK64_63( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
-unsigned char *TEMPLATE2(_BITUNPACK_,64_64)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*64),x=0; do { BITUNPACK64_64( in, out, start); PREFETCH(in); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_0)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*0),x=0; const uint64_t *out_ = out+n; do { BITUNPACK64_0( in, out, start); PREFETCH(in+512,0); } while(out<out_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_1)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*1),x=0; do { BITUNPACK64_1( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_2)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*2),x=0; do { BITUNPACK64_2( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_3)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*3),x=0; do { BITUNPACK64_3( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_4)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*4),x=0; do { BITUNPACK64_4( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_5)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*5),x=0; do { BITUNPACK64_5( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_6)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*6),x=0; do { BITUNPACK64_6( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_7)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*7),x=0; do { BITUNPACK64_7( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_8)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*8),x=0; do { BITUNPACK64_8( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_9)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*9),x=0; do { BITUNPACK64_9( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_10)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*10),x=0; do { BITUNPACK64_10( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_11)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*11),x=0; do { BITUNPACK64_11( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_12)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*12),x=0; do { BITUNPACK64_12( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_13)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*13),x=0; do { BITUNPACK64_13( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_14)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*14),x=0; do { BITUNPACK64_14( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_15)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*15),x=0; do { BITUNPACK64_15( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_16)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*16),x=0; do { BITUNPACK64_16( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_17)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*17),x=0; do { BITUNPACK64_17( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_18)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*18),x=0; do { BITUNPACK64_18( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_19)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*19),x=0; do { BITUNPACK64_19( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_20)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*20),x=0; do { BITUNPACK64_20( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_21)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*21),x=0; do { BITUNPACK64_21( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_22)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*22),x=0; do { BITUNPACK64_22( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_23)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*23),x=0; do { BITUNPACK64_23( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_24)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*24),x=0; do { BITUNPACK64_24( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_25)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*25),x=0; do { BITUNPACK64_25( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_26)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*26),x=0; do { BITUNPACK64_26( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_27)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*27),x=0; do { BITUNPACK64_27( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_28)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*28),x=0; do { BITUNPACK64_28( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_29)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*29),x=0; do { BITUNPACK64_29( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_30)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*30),x=0; do { BITUNPACK64_30( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_31)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*31),x=0; do { BITUNPACK64_31( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_32)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*32),x=0; do { BITUNPACK64_32( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_33)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*33),x=0; do { BITUNPACK64_33( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_34)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*34),x=0; do { BITUNPACK64_34( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_35)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*35),x=0; do { BITUNPACK64_35( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_36)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*36),x=0; do { BITUNPACK64_36( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_37)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*37),x=0; do { BITUNPACK64_37( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_38)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*38),x=0; do { BITUNPACK64_38( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_39)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*39),x=0; do { BITUNPACK64_39( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_40)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*40),x=0; do { BITUNPACK64_40( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_41)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*41),x=0; do { BITUNPACK64_41( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_42)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*42),x=0; do { BITUNPACK64_42( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_43)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*43),x=0; do { BITUNPACK64_43( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_44)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*44),x=0; do { BITUNPACK64_44( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_45)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*45),x=0; do { BITUNPACK64_45( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_46)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*46),x=0; do { BITUNPACK64_46( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_47)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*47),x=0; do { BITUNPACK64_47( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_48)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*48),x=0; do { BITUNPACK64_48( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_49)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*49),x=0; do { BITUNPACK64_49( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_50)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*50),x=0; do { BITUNPACK64_50( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_51)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*51),x=0; do { BITUNPACK64_51( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_52)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*52),x=0; do { BITUNPACK64_52( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_53)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*53),x=0; do { BITUNPACK64_53( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_54)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*54),x=0; do { BITUNPACK64_54( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_55)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*55),x=0; do { BITUNPACK64_55( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_56)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*56),x=0; do { BITUNPACK64_56( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_57)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*57),x=0; do { BITUNPACK64_57( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_58)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*58),x=0; do { BITUNPACK64_58( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_59)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*59),x=0; do { BITUNPACK64_59( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_60)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*60),x=0; do { BITUNPACK64_60( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_61)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*61),x=0; do { BITUNPACK64_61( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_62)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*62),x=0; do { BITUNPACK64_62( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_63)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*63),x=0; do { BITUNPACK64_63( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
+unsigned char *TEMPLATE2(_BITUNPACK_,64_64)(const unsigned char *__restrict in, unsigned n, uint64_t *__restrict out , uint64_t start ) { unsigned char *in_=in+PAD8(n*64),x=0; do { BITUNPACK64_64( in, out, start); PREFETCH(in+512,0); } while(in<in_); return in_; }
 BITUNPACK_D64 TEMPLATE2(_BITUNPACK_,a64)[] = {
   &TEMPLATE2(_BITUNPACK_,64_0),
   &TEMPLATE2(_BITUNPACK_,64_1),
@@ -3668,7 +3668,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_1(ip,  op, parm) {\
   BITUNBLK128V16_1(ip,  0, op, parm);\
-  BITUNBLK128V16_1(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_2(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3685,8 +3684,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 #define BITUNPACK128V16_2(ip,  op, parm) {\
   BITUNBLK128V16_2(ip,  0, op, parm);\
   BITUNBLK128V16_2(ip,  1, op, parm);\
-  BITUNBLK128V16_2(ip,  2, op, parm);\
-  BITUNBLK128V16_2(ip,  3, op, parm);\
 }
 
 #define BITUNBLK128V16_3(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3710,7 +3707,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_3(ip,  op, parm) {\
   BITUNBLK128V16_3(ip,  0, op, parm);\
-  BITUNBLK128V16_3(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_4(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3725,10 +3721,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
   BITUNBLK128V16_4(ip,  1, op, parm);\
   BITUNBLK128V16_4(ip,  2, op, parm);\
   BITUNBLK128V16_4(ip,  3, op, parm);\
-  BITUNBLK128V16_4(ip,  4, op, parm);\
-  BITUNBLK128V16_4(ip,  5, op, parm);\
-  BITUNBLK128V16_4(ip,  6, op, parm);\
-  BITUNBLK128V16_4(ip,  7, op, parm);\
 }
 
 #define BITUNBLK128V16_5(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3752,7 +3744,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_5(ip,  op, parm) {\
   BITUNBLK128V16_5(ip,  0, op, parm);\
-  BITUNBLK128V16_5(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_6(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3769,8 +3760,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 #define BITUNPACK128V16_6(ip,  op, parm) {\
   BITUNBLK128V16_6(ip,  0, op, parm);\
   BITUNBLK128V16_6(ip,  1, op, parm);\
-  BITUNBLK128V16_6(ip,  2, op, parm);\
-  BITUNBLK128V16_6(ip,  3, op, parm);\
 }
 
 #define BITUNBLK128V16_7(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3794,7 +3783,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_7(ip,  op, parm) {\
   BITUNBLK128V16_7(ip,  0, op, parm);\
-  BITUNBLK128V16_7(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_8(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3811,14 +3799,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
   BITUNBLK128V16_8(ip,  5, op, parm);\
   BITUNBLK128V16_8(ip,  6, op, parm);\
   BITUNBLK128V16_8(ip,  7, op, parm);\
-  BITUNBLK128V16_8(ip,  8, op, parm);\
-  BITUNBLK128V16_8(ip,  9, op, parm);\
-  BITUNBLK128V16_8(ip, 10, op, parm);\
-  BITUNBLK128V16_8(ip, 11, op, parm);\
-  BITUNBLK128V16_8(ip, 12, op, parm);\
-  BITUNBLK128V16_8(ip, 13, op, parm);\
-  BITUNBLK128V16_8(ip, 14, op, parm);\
-  BITUNBLK128V16_8(ip, 15, op, parm);\
 }
 
 #define BITUNBLK128V16_9(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3842,7 +3822,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_9(ip,  op, parm) {\
   BITUNBLK128V16_9(ip,  0, op, parm);\
-  BITUNBLK128V16_9(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_10(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3859,8 +3838,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 #define BITUNPACK128V16_10(ip,  op, parm) {\
   BITUNBLK128V16_10(ip,  0, op, parm);\
   BITUNBLK128V16_10(ip,  1, op, parm);\
-  BITUNBLK128V16_10(ip,  2, op, parm);\
-  BITUNBLK128V16_10(ip,  3, op, parm);\
 }
 
 #define BITUNBLK128V16_11(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3884,7 +3861,7 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_11(ip,  op, parm) {\
   BITUNBLK128V16_11(ip,  0, op, parm);\
-  BITUNBLK128V16_11(ip,  1, op, parm);\
+  /*BITUNBLK128V16_11(ip,  1, op, parm);*/\
 }
 
 #define BITUNBLK128V16_12(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3899,10 +3876,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
   BITUNBLK128V16_12(ip,  1, op, parm);\
   BITUNBLK128V16_12(ip,  2, op, parm);\
   BITUNBLK128V16_12(ip,  3, op, parm);\
-  BITUNBLK128V16_12(ip,  4, op, parm);\
-  BITUNBLK128V16_12(ip,  5, op, parm);\
-  BITUNBLK128V16_12(ip,  6, op, parm);\
-  BITUNBLK128V16_12(ip,  7, op, parm);\
 }
 
 #define BITUNBLK128V16_13(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3926,7 +3899,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_13(ip,  op, parm) {\
   BITUNBLK128V16_13(ip,  0, op, parm);\
-  BITUNBLK128V16_13(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_14(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3943,8 +3915,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 #define BITUNPACK128V16_14(ip,  op, parm) {\
   BITUNBLK128V16_14(ip,  0, op, parm);\
   BITUNBLK128V16_14(ip,  1, op, parm);\
-  BITUNBLK128V16_14(ip,  2, op, parm);\
-  BITUNBLK128V16_14(ip,  3, op, parm);\
 }
 
 #define BITUNBLK128V16_15(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3968,7 +3938,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
 
 #define BITUNPACK128V16_15(ip,  op, parm) {\
   BITUNBLK128V16_15(ip,  0, op, parm);\
-  BITUNBLK128V16_15(ip,  1, op, parm);\
 }
 
 #define BITUNBLK128V16_16(ip, i, op, parm) { __m128i ov,iv = _mm_loadu_si128((__m128i *)ip++);\
@@ -3992,22 +3961,6 @@ unsigned char *TEMPLATE2(_BITUNPACK_,64)( const unsigned char *__restrict in, un
   BITUNBLK128V16_16(ip, 13, op, parm);\
   BITUNBLK128V16_16(ip, 14, op, parm);\
   BITUNBLK128V16_16(ip, 15, op, parm);\
-  BITUNBLK128V16_16(ip, 16, op, parm);\
-  BITUNBLK128V16_16(ip, 17, op, parm);\
-  BITUNBLK128V16_16(ip, 18, op, parm);\
-  BITUNBLK128V16_16(ip, 19, op, parm);\
-  BITUNBLK128V16_16(ip, 20, op, parm);\
-  BITUNBLK128V16_16(ip, 21, op, parm);\
-  BITUNBLK128V16_16(ip, 22, op, parm);\
-  BITUNBLK128V16_16(ip, 23, op, parm);\
-  BITUNBLK128V16_16(ip, 24, op, parm);\
-  BITUNBLK128V16_16(ip, 25, op, parm);\
-  BITUNBLK128V16_16(ip, 26, op, parm);\
-  BITUNBLK128V16_16(ip, 27, op, parm);\
-  BITUNBLK128V16_16(ip, 28, op, parm);\
-  BITUNBLK128V16_16(ip, 29, op, parm);\
-  BITUNBLK128V16_16(ip, 30, op, parm);\
-  BITUNBLK128V16_16(ip, 31, op, parm);\
 }
 
 #define BITUNPACK128V32_0(ip,  op, parm) {\