From 2b28ebca4ad3d5461f91c6580748b04e5a7eb1e6 Mon Sep 17 00:00:00 2001
From: couturie <couturie@extinction>
Date: Wed, 6 Sep 2017 16:14:53 +0200
Subject: [PATCH 1/1] new

---
 OneRoundIoT/OneRound/one_round_par2.cpp | 417 +++++++++++++-----------
 1 file changed, 224 insertions(+), 193 deletions(-)

diff --git a/OneRoundIoT/OneRound/one_round_par2.cpp b/OneRoundIoT/OneRound/one_round_par2.cpp
index e42a6d5..f579622 100644
--- a/OneRoundIoT/OneRound/one_round_par2.cpp
+++ b/OneRoundIoT/OneRound/one_round_par2.cpp
@@ -12,9 +12,9 @@
 #include <omp.h>
 
 /*#include <cryptopp/hex.h>
-#include <cryptopp/sha.h>
-#include <cryptopp/osrng.h>
-#include <cryptopp/secblock.h>
+  #include <cryptopp/sha.h>
+  #include <cryptopp/osrng.h>
+  #include <cryptopp/secblock.h>
 */
 
 
@@ -136,11 +136,11 @@ void prga(uchar *sc, int ldata, uchar *r) {
 void rotate(uchar *RM1, uchar *RM2, int size, int n)
 {
   int i;
-   for (i = 0; i< size-n; i++)
-     RM2[i+n] = RM1[i];
+  for (i = 0; i< size-n; i++)
+    RM2[i+n] = RM1[i];
 
-   for (i = 0; i< n; i++)
-     RM2[i] = RM1[size-n-1+i];
+  for (i = 0; i< n; i++)
+    RM2[i] = RM1[size-n-1+i];
 }
 
 
@@ -148,16 +148,6 @@ template<int h2>
 void encrypt_ctr(uchar* seq_in, uchar *seq_out, int len,uchar* RM1,int *Pbox, int *PboxRM, uchar *Sbox1, uchar *Sbox2, int enc, int num) {
 
 
-  uchar X[h2];
-  uchar fX[h2*num];
-
-
-  
-
-
-  for(int a=0;a<h2;a++) {
-     X[a]=Sbox1[a&0xFF];           //Warning according to the size of h2, we can be outsize of Sbox1[a]
-   }
 
 
   int loc_len=len/num;
@@ -166,125 +156,135 @@ void encrypt_ctr(uchar* seq_in, uchar *seq_out, int len,uchar* RM1,int *Pbox, in
   for(int p=0;p<num;p++) {
 
     int id=omp_get_thread_num();
+    uchar fX[h2];
     uchar RM2[h2];
+    
     for(int a=0;a<h2;a++) {
       RM2[a]=RM1[id*h2+a];           //Warning according to the size of h2, we can be outsize of Sbox1[a]
     }
 
-  
+
+    uchar X[h2];
+
+    for(int a=0;a<h2;a++) {
+      X[a]=Sbox1[a&0xFF];           //Warning according to the size of h2, we can be outsize of Sbox1[a]
+    }
+
+
+    
     int offset=p*loc_len;
 
     
     for(int it=offset;it<offset+loc_len;it++) {
-    int ind1,ind2;
+      int ind1,ind2;
 
 
-    //cout<<id<<" "<<it<<endl;
+      //cout<<id<<" "<<it<<endl;
     
-    if(enc) {
-      ind1=it*h2;
-      ind2=Pbox[it]*h2;
-    }
-    else {
-      ind2=it*h2;
-      ind1=Pbox[it]*h2;
-    }
+      if(enc) {
+	ind1=it*h2;
+	ind2=Pbox[it]*h2;
+      }
+      else {
+	ind2=it*h2;
+	ind1=Pbox[it]*h2;
+      }
   
 
-    /*for(int a=0;a<h2;a+=4){
-      fX[a]=RM1[X[a]];
-      fX[a+1]=RM1[X[a+1]];
-      fX[a+2]=RM1[X[a+2]];
-      fX[a+3]=RM1[X[a+3]];
-      }*/
-
-    for(int a=0;a<h2;a+=4){
-      fX[id*h2+a]=X[a];
-      fX[id*h2+a+1]=X[a+1];
-      fX[id*h2+a+2]=X[a+2];
-      fX[id*h2+a+3]=X[a+3];
-    }
+      /*for(int a=0;a<h2;a+=4){
+	fX[a]=RM1[X[a]];
+	fX[a+1]=RM1[X[a+1]];
+	fX[a+2]=RM1[X[a+2]];
+	fX[a+3]=RM1[X[a+3]];
+	}*/
+
+      for(int a=0;a<h2;a+=4){
+	fX[a]=X[a];
+	fX[a+1]=X[a+1];
+	fX[a+2]=X[a+2];
+	fX[a+3]=X[a+3];
+      }
 
 
 
     
-    /*   if(it<513) {
-      for(int a=0;a<h2;a++)
-	printf("%d ",fX[a]);
-      printf("\n");
-      }*/
+      /*   if(it<513) {
+	   for(int a=0;a<h2;a++)
+	   printf("%d ",fX[a]);
+	   printf("\n");
+	   }*/
     
-    *(int*)&fX[id*h2+0]^=it;
+      *(int*)&fX[0]^=it;
 
-    /* if(it<513) {
-      for(int a=0;a<h2;a++)
-	printf("%d ",fX[a]);
-      printf("\n");
-      }*/
+      /* if(it<513) {
+	 for(int a=0;a<h2;a++)
+	 printf("%d ",fX[a]);
+	 printf("\n");
+	 }*/
 
 
   
     
-    /*for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=fX[id*h2+a]^RM1[id*h2+a];
-      fX[id*h2+a+1]=fX[id*h2+a+1]^RM1[id*h2+a+1];
-      fX[id*h2+a+2]=fX[id*h2+a+2]^RM1[id*h2+a+2];
-      fX[id*h2+a+3]=fX[id*h2+a+3]^RM1[id*h2+a+3];
-      }*/
+      /*for(int a=0;a<h2;a+=4) {
+	fX[id*h2+a]=fX[id*h2+a]^RM1[id*h2+a];
+	fX[id*h2+a+1]=fX[id*h2+a+1]^RM1[id*h2+a+1];
+	fX[id*h2+a+2]=fX[id*h2+a+2]^RM1[id*h2+a+2];
+	fX[id*h2+a+3]=fX[id*h2+a+3]^RM1[id*h2+a+3];
+	}*/
 
 
 
     
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=Sbox2[fX[id*h2+a]];
-      fX[id*h2+a+1]=Sbox2[fX[id*h2+a+1]];
-      fX[id*h2+a+2]=Sbox2[fX[id*h2+a+2]];
-      fX[id*h2+a+3]=Sbox2[fX[id*h2+a+3]];
-    }
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=Sbox2[fX[a]];
+	fX[a+1]=Sbox2[fX[a+1]];
+	fX[a+2]=Sbox2[fX[a+2]];
+	fX[a+3]=Sbox2[fX[a+3]];
+      }
 
 //    rotate(RM1, &RM2[id*h2], h2, Pbox[it]%h2);
     
-    for(int a=0;a<h2;a+=4) {
-      RM2[a]=RM2[PboxRM[a]];
-      RM2[a+1]=RM2[PboxRM[a+1]];
-      RM2[a+2]=RM2[PboxRM[a+2]];
-      RM2[a+3]=RM2[PboxRM[a+3]];
-    }
+      for(int a=0;a<h2;a+=4) {
+	RM2[a]=RM2[PboxRM[a]];
+	RM2[a+1]=RM2[PboxRM[a+1]];
+	RM2[a+2]=RM2[PboxRM[a+2]];
+	RM2[a+3]=RM2[PboxRM[a+3]];
+      }
     
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=fX[id*h2+a]^RM2[id*h2+a];
-      fX[id*h2+a+1]=fX[id*h2+a+1]^RM2[id*h2+a+1];
-      fX[id*h2+a+2]=fX[id*h2+a+2]^RM2[id*h2+a+2];
-      fX[id*h2+a+3]=fX[id*h2+a+3]^RM2[id*h2+a+3];
-    }
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=fX[a]^RM2[a];
+	fX[a+1]=fX[a+1]^RM2[a+1];
+	fX[a+2]=fX[a+2]^RM2[a+2];
+	fX[a+3]=fX[a+3]^RM2[a+3];
+      }
 
 
     
-     for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=fX[id*h2+a]^seq_in[ind2+a];
-      fX[id*h2+a+1]=fX[id*h2+a+1]^seq_in[ind2+a+1];
-      fX[id*h2+a+2]=fX[id*h2+a+2]^seq_in[ind2+a+2];
-      fX[id*h2+a+3]=fX[id*h2+a+3]^seq_in[ind2+a+3];
-    }
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=fX[a]^seq_in[ind2+a];
+	fX[a+1]=fX[a+1]^seq_in[ind2+a+1];
+	fX[a+2]=fX[a+2]^seq_in[ind2+a+2];
+	fX[a+3]=fX[a+3]^seq_in[ind2+a+3];
+      }
 
  
-    for(int a=0;a<h2;a+=4) {
-      seq_out[ind1+a]=fX[id*h2+a];
-      seq_out[ind1+a+1]=fX[id*h2+a+1];
-      seq_out[ind1+a+2]=fX[id*h2+a+2];
-      seq_out[ind1+a+3]=fX[id*h2+a+3];
-    }
+      for(int a=0;a<h2;a+=4) {
+	seq_out[ind1+a]=fX[a];
+	seq_out[ind1+a+1]=fX[a+1];
+	seq_out[ind1+a+2]=fX[a+2];
+	seq_out[ind1+a+3]=fX[a+3];
+      }
     
-    /*for(int a=0;a<h2;a+=4) {
-      RM1[id*h2+a]=RM1[id*h2+PboxRM[a]];
-      RM1[id*h2+a+1]=RM1[id*h2+PboxRM[a+1]];
-      RM1[id*h2+a+2]=RM1[id*h2+PboxRM[a+2]];
-      RM1[id*h2+a+3]=RM1[id*h2+PboxRM[a+3]];
-    }
-    */
+      /*for(int a=0;a<h2;a+=4) {
+	RM1[id*h2+a]=RM1[id*h2+PboxRM[a]];
+	RM1[id*h2+a+1]=RM1[id*h2+PboxRM[a+1]];
+	RM1[id*h2+a+2]=RM1[id*h2+PboxRM[a+2]];
+	RM1[id*h2+a+3]=RM1[id*h2+PboxRM[a+3]];
+	}
+      */
 
     
-  }
+    }
   }
 
 }
@@ -297,146 +297,177 @@ void encrypt(uchar* seq_in, uchar *seq_out, int len,uchar* RM1,int *Pbox, int *P
 
 
 /*  uchar *X=new uchar[h2];
-  uchar *fX=new uchar[h2];
-  unsigned int *lX=(unsigned int*)X;
-  unsigned int *lseq_in=(unsigned int*)seq_in;
+    uchar *fX=new uchar[h2];
+    unsigned int *lX=(unsigned int*)X;
+    unsigned int *lseq_in=(unsigned int*)seq_in;
 */
-  uchar X[h2*num];
-  uchar fX[h2*num];
-  uchar RM2[h2*num];
+
 //  unsigned int *lX=(unsigned int*)X;
 //  unsigned int *lseq_in=(unsigned int*)seq_in;
+
+
+  int loc_len=len/num;
   
-#pragma omp parallel for 
-  for(int it=0;it<len;it++) {
-    int ind1=it*h2;
-    int ind2=Pbox[it]*h2;
+#pragma omp parallel for
+  for(int p=0;p<num;p++) {
+
     int id=omp_get_thread_num();
- 
-    for(int a=0;a<h2;a+=4) {
-      X[id*h2+a]=seq_in[ind2+a];
-      X[id*h2+a+1]=seq_in[ind2+a+1];
-      X[id*h2+a+2]=seq_in[ind2+a+2];
-      X[id*h2+a+3]=seq_in[ind2+a+3];
-    }
 
-    for(int a=0;a<h2;a+=4){
-      fX[id*h2+a]=Sbox1[X[id*h2+a]];
-      fX[id*h2+a+1]=Sbox1[X[id*h2+a+1]];
-      fX[id*h2+a+2]=Sbox1[X[id*h2+a+2]];
-      fX[id*h2+a+3]=Sbox1[X[id*h2+a+3]];
+    uchar fX[h2];
+
+    
+    uchar RM2[h2];
+    for(int a=0;a<h2;a++) {
+      RM2[a]=RM1[id*h2+a];           
     }
 
+  
+    int offset=p*loc_len;
 
-/*    for(int a=0;a<h2;a+=4) {
-      fX[a]=fX[a]^RM1[a];
-      fX[a+1]=fX[a+1]^RM1[a+1];
-      fX[a+2]=fX[a+2]^RM1[a+2];
-      fX[a+3]=fX[a+3]^RM1[a+3];
-      }*/
-    rotate(RM1, &RM2[id*h2], h2, Pbox[it]%h2);
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=fX[id*h2+a]^RM2[id*h2+a];
-      fX[id*h2+a+1]=fX[id*h2+a+1]^RM2[id*h2+a+1];
-      fX[id*h2+a+2]=fX[id*h2+a+2]^RM2[id*h2+a+2];
-      fX[id*h2+a+3]=fX[id*h2+a+3]^RM2[id*h2+a+3];
-    }
+    
+    for(int it=offset;it<offset+loc_len;it++) {
 
-    for(int a=0;a<h2;a+=4) {
-      seq_out[ind1+a]=Sbox2[fX[id*h2+a]];
-      seq_out[ind1+a+1]=Sbox2[fX[id*h2+a+1]];
-      seq_out[ind1+a+2]=Sbox2[fX[id*h2+a+2]];
-      seq_out[ind1+a+3]=Sbox2[fX[id*h2+a+3]];
-    }
+  
 
-    /* for(int a=0;a<h2;a+=4) {
-      RM1[a]=RM1[PboxRM[a]];
-      RM1[a+1]=RM1[PboxRM[a+1]];
-      RM1[a+2]=RM1[PboxRM[a+2]];
-      RM1[a+3]=RM1[PboxRM[a+3]];
+      int ind1=it*h2;
+      int ind2=Pbox[it]*h2;
 
-      }*/
+ 
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=seq_in[ind2+a];
+	fX[a+1]=seq_in[ind2+a+1];
+	fX[a+2]=seq_in[ind2+a+2];
+	fX[a+3]=seq_in[ind2+a+3];
+      }
 
 
-  }
+      for(int a=0;a<h2;a+=4){
+	fX[a]=Sbox1[fX[a]];
+	fX[a+1]=Sbox1[fX[a+1]];
+	fX[a+2]=Sbox1[fX[a+2]];
+	fX[a+3]=Sbox1[fX[a+3]];
+      }
+
+
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=fX[a]^RM2[a];
+	fX[a+1]=fX[a+1]^RM2[a+1];
+	fX[a+2]=fX[a+2]^RM2[a+2];
+	fX[a+3]=fX[a+3]^RM2[a+3];
+      }
 
 
 
 
+      for(int a=0;a<h2;a+=4) {
+	seq_out[ind1+a]=Sbox2[fX[a]];
+	seq_out[ind1+a+1]=Sbox2[fX[a+1]];
+	seq_out[ind1+a+2]=Sbox2[fX[a+2]];
+	seq_out[ind1+a+3]=Sbox2[fX[a+3]];
+      }
+
+
+    for(int a=0;a<h2;a+=4) {
+	RM2[a]=RM2[PboxRM[a]];
+	RM2[a+1]=RM2[PboxRM[a+1]];
+	RM2[a+2]=RM2[PboxRM[a+2]];
+	RM2[a+3]=RM2[PboxRM[a+3]];
+      }
+      
+
+
+
+    }
+    
+  }
 }
 
 
+
+
 template<int h2>
 void decrypt(uchar* seq_in, uchar *seq_out, int len,uchar* RM1,int *Pbox, int *PboxRM, uchar *Inv_Sbox1, uchar *Inv_Sbox2, int debug, int num) {
 
 
   /*uchar *fX=new uchar[h2];
-  uchar *Inv_Sbox1=new uchar[256];
-  uchar *Inv_Sbox2=new uchar[256];
+    uchar *Inv_Sbox1=new uchar[256];
+    uchar *Inv_Sbox2=new uchar[256];
   */
-  uchar fX[h2*num];
-  uchar RM2[h2*num];
 
-#pragma omp parallel for 
-  for(int it=0;it<len;it++) {
 
-    int ind1=it*h2;
-    int ind2=Pbox[it]*h2;
+
+
+
+  int loc_len=len/num;
+  
+#pragma omp parallel for
+  for(int p=0;p<num;p++) {
+
     int id=omp_get_thread_num();
+    uchar RM2[h2];
+    for(int a=0;a<h2;a++) {
+      RM2[a]=RM1[id*h2+a];           //Warning according to the size of h2, we can be outsize of Sbox1[a]
+    }
+    uchar fX[h2];
+  
+    int offset=p*loc_len;
 
+    
+    for(int it=offset;it<offset+loc_len;it++) {
 
 
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=seq_in[ind1+a];
-      fX[id*h2+a+1]=seq_in[ind1+a+1];
-      fX[id*h2+a+2]=seq_in[ind1+a+2];
-      fX[id*h2+a+3]=seq_in[ind1+a+3];
+      int ind1=it*h2;
+      int ind2=Pbox[it]*h2;
+
+
+
+
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=seq_in[ind1+a];
+	fX[a+1]=seq_in[ind1+a+1];
+	fX[a+2]=seq_in[ind1+a+2];
+	fX[a+3]=seq_in[ind1+a+3];
 	    
-    }
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=Inv_Sbox2[fX[id*h2+a]];
-      fX[id*h2+a+1]=Inv_Sbox2[fX[id*h2+a+1]];
-      fX[id*h2+a+2]=Inv_Sbox2[fX[id*h2+a+2]];
-      fX[id*h2+a+3]=Inv_Sbox2[fX[id*h2+a+3]];
-    }
-    /*for(int a=0;a<h2;a+=4) {
-      fX[a]=fX[a]^RM1[a];
-      fX[a+1]=fX[a+1]^RM1[a+1];
-      fX[a+2]=fX[a+2]^RM1[a+2];
-      fX[a+3]=fX[a+3]^RM1[a+3];
+      }
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=Inv_Sbox2[fX[a]];
+	fX[a+1]=Inv_Sbox2[fX[a+1]];
+	fX[a+2]=Inv_Sbox2[fX[a+2]];
+	fX[a+3]=Inv_Sbox2[fX[a+3]];
       }
 
-    for(int a=0;a<h2;a+=4) {
-      RM1[a]=RM1[PboxRM[a]];
-      RM1[a+1]=RM1[PboxRM[a+1]];
-      RM1[a+2]=RM1[PboxRM[a+2]];
-      RM1[a+3]=RM1[PboxRM[a+3]];
-      }*/
 
 
-    rotate(RM1, &RM2[id*h2], h2, Pbox[it]%h2);
-    for(int a=0;a<h2;a+=4) {
-      fX[id*h2+a]=fX[id*h2+a]^RM2[id*h2+a];
-      fX[id*h2+a+1]=fX[id*h2+a+1]^RM2[id*h2+a+1];
-      fX[id*h2+a+2]=fX[id*h2+a+2]^RM2[id*h2+a+2];
-      fX[id*h2+a+3]=fX[id*h2+a+3]^RM2[id*h2+a+3];      
-    }
+      for(int a=0;a<h2;a+=4) {
+	fX[a]=fX[a]^RM2[a];
+	fX[a+1]=fX[a+1]^RM2[a+1];
+	fX[a+2]=fX[a+2]^RM2[a+2];
+	fX[a+3]=fX[a+3]^RM2[a+3];      
+      }
 
     
-    for(int a=0;a<h2;a+=4) {
-      seq_out[ind2+a]=Inv_Sbox1[fX[id*h2+a]];
-      seq_out[ind2+a+1]=Inv_Sbox1[fX[id*h2+a+1]];
-      seq_out[ind2+a+2]=Inv_Sbox1[fX[id*h2+a+2]];
-      seq_out[ind2+a+3]=Inv_Sbox1[fX[id*h2+a+3]];
-    }
+      for(int a=0;a<h2;a+=4) {
+	seq_out[ind2+a]=Inv_Sbox1[fX[a]];
+	seq_out[ind2+a+1]=Inv_Sbox1[fX[a+1]];
+	seq_out[ind2+a+2]=Inv_Sbox1[fX[a+2]];
+	seq_out[ind2+a+3]=Inv_Sbox1[fX[a+3]];
+      }
+      for(int a=0;a<h2;a+=4) {
+	RM2[a]=RM2[PboxRM[a]];
+	RM2[a+1]=RM2[PboxRM[a+1]];
+	RM2[a+2]=RM2[PboxRM[a+2]];
+	RM2[a+3]=RM2[PboxRM[a+3]];
+	}
 
      
-  }
+    }
 
 
+  }
 }
 
 
+
 int main(int argc, char** argv) {
 
 
@@ -455,10 +486,10 @@ int main(int argc, char** argv) {
   }
 
 /*  printf("nb times %d\n",nb_test);
-  printf("ctr %d\n",ctr);
-  printf("h %d\n",h);
-  printf("lena %d\n",lena);
-  printf("size_buf %d\n",size_buf);
+    printf("ctr %d\n",ctr);
+    printf("h %d\n",h);
+    printf("lena %d\n",lena);
+    printf("size_buf %d\n",size_buf);
 */
   int h2=h*h;
   
-- 
2.39.5