diff --git a/ext/FastPFor b/ext/FastPFor
index 4e484a8..bae51c0 160000
--- a/ext/FastPFor
+++ b/ext/FastPFor
@@ -1 +1 @@
-Subproject commit 4e484a8d6f1b45af0f4a57515d1c31aea1e80a86
+Subproject commit bae51c0f8f7c50c179e10a84d0a3e137bbc062d7
diff --git a/ext/beplug_.h b/ext/beplug_.h
index be75547..18fe680 100644
--- a/ext/beplug_.h
+++ b/ext/beplug_.h
@@ -11,7 +11,13 @@
 #include "FastPFor/headers/simdfastpfor.h"
 #include "FastPFor/headers/optpfor.h"
 #include "FastPFor/headers/simdoptpfor.h"
-//#include "ext/FastPFor/headers/compositecodec.h"
+#ifndef min
+#define min(x,y) (((x)<(y)) ? (x) : (y))
+#endif
+#define _mm_store_si128 _mm_storeu_si128
+#define _mm_load_si128 _mm_loadu_si128
+#include "FastPFor/headers/simdgroupsimple.h"
+#undef min
   #endif
   
   #if C_SIMPLE8B
diff --git a/ext/beplugc_.c b/ext/beplugc_.c
index ab1c759..8b17655 100644
--- a/ext/beplugc_.c
+++ b/ext/beplugc_.c
@@ -36,6 +36,17 @@
       }
 	  ctou32(out) = nvalue;
 	  return out+4+nvalue*4;
+    }
+	case FP_GROUPSIMPLE: {   
+      size_t nvalue = outsize/4;
+      FastPForLib::SIMDGroupSimple<false,false> ic; ic.encodeArray((const int32_t *)in, n & (~127), (uint32_t *)(out+4), nvalue);
+      if(n & 127) {
+        size_t nvalue2 = outsize/4 - nvalue;
+        FastPForLib::VariableByte vc; vc.encodeArray((const int32_t *)(in + (n & (~127))), n & 127, (uint32_t *)(out + 4 + nvalue*4), nvalue2);
+        nvalue += nvalue2;
+      }
+	  ctou32(out) = nvalue;
+	  return out+4+nvalue*4;
     }
     case FP_VBYTE:    	{ size_t nvalue=outsize/4; FastPForLib::VariableByte ic; ic.encodeArray((const int32_t *)in, (const size_t)n, (uint32_t *)(out+4), nvalue); ctou32(out)=nvalue; return out+4+nvalue*4; }
     case FP_SIMPLE8BRLE:{ size_t nvalue=outsize/4; FastPForLib::Simple8b_RLE<true> ic; ic.encodeArray((const int32_t *)in, (const size_t)n, (uint32_t *)(out+4), nvalue); ctou32(out)=nvalue; return out+4+nvalue*4; }
diff --git a/ext/beplugd_.c b/ext/beplugd_.c
index 29ea802..9795a4c 100644
--- a/ext/beplugd_.c
+++ b/ext/beplugd_.c
@@ -38,6 +38,16 @@
       }
       return (unsigned char *)ip; 
     } 
+	case FP_GROUPSIMPLE:  { 	
+	  size_t nvalue = n;
+      FastPForLib::SIMDGroupSimple<false,false> ic; const uint32_t *ip = ic.decodeArray((const int32_t *)(in+4), ctou32(in), out, nvalue);
+      if(n & 127) { 
+        nvalue = n - nvalue;
+	    FastPForLib::VariableByte vc;
+		return (unsigned char *)vc.decodeArray(ip, (const uint32_t *)in+1+ctou32(in) - ip, out + (n&(~127)), nvalue);	  //return vbdec32((unsigned char *)ip, n & 127, out + mynvalue1);
+      }
+      return (unsigned char *)ip; 
+    }
 
     case FP_VBYTE:  //return vbytedec( in, n, out); 
 	  { size_t nvalue=n; FastPForLib::VariableByte ic;       return (unsigned char *)ic.decodeArray((const int32_t *)(in+4), ctou32(in), (uint32_t *)out, nvalue); }
diff --git a/ext/beplugr_.h b/ext/beplugr_.h
index d6380ed..4573f38 100644
--- a/ext/beplugr_.h
+++ b/ext/beplugr_.h
@@ -19,6 +19,7 @@
   { FP_SIMDOPTPFOR,	"FP_SIMDOptPFor",  	C_FASTPFOR,	    0,    	 0,"","OptPFor SIMD" },
   { FP_VBYTE,		"FP_VByte", 		C_FASTPFOR,	    0,    	 0,"","Variable byte" },
   { FP_SIMPLE8BRLE,	"FP_Simple8bRLE", 	C_FASTPFOR,	    0,    	 0,"","Simple-8b + rle" },
+  { FP_GROUPSIMPLE,	"FP_GROUPSIMPLE", 	C_FASTPFOR,	    0,    	 0,"","Group Simple" },
 
   { SC_SIMDPACK128,	"SC_SIMDPack128", 	C_SIMDCOMP,		BLK_V128,0,"","Bit packing (SSE4.1)"},
   { SC_SIMDPACK256,	"SC_SIMDPack256", 	C_SIMDCOMP,		BLK_V256,0,"","Bit packing (SSE4.1)"},
diff --git a/ext/lz4 b/ext/lz4
index d18084c..1525fd1 160000
--- a/ext/lz4
+++ b/ext/lz4
@@ -1 +1 @@
-Subproject commit d18084c65561cddec48f74c3a91363c9f8e6f709
+Subproject commit 1525fd1f52940e9acb0df094b3e6ee5d0a36e4f3