release/src/router/libsodium/src/libsodium/crypto_stream/aes256estream/hongjun/aes256.h

   1 /* aes256.h */
   2 /* Hongjun Wu, January 2007*/
   3
   4
   5 #include "ecrypt-sync.h"
   6 #include "aes-table.h"
   7
   8 #include <stdio.h>
   9
  10 #define first_round(ctx,x0,y0) { \
  11         u32 z0,t0,tem0;              \
  12         z0 = (x0) ^ ctx->round_key[0][0];        \
  13         t0 = (u8) z0;                \
  14         tem0 = T0[t0];               \
  15         (y0) = tem0 ^ ctx->first_round_output_x0; \
  16 }
  17
  18 #define second_round(ctx,x0,y0,y1,y2,y3) { \
  19         u32 t0,t7,t10,t13; \
  20         u32 tem0,tem7,tem10,tem13;    \
  21         t0 = (u8)(x0);      \
  22         tem0 = T0[t0];      \
  23         (y0) = tem0 ^ ctx->second_round_output[0];   \
  24         t7 = (u8)((x0)>>24);     \
  25         tem7 = T3[t7];    \
  26         (y1) = tem7 ^ ctx->second_round_output[1];   \
  27         t10 = (u8)((x0)>>16);   \
  28         tem10 = T2[t10]; \
  29         (y2) = tem10 ^ ctx->second_round_output[2];  \
  30         t13 = (u8)((x0)>>8);    \
  31         tem13 = T1[t13];\
  32         (y3) = tem13 ^ ctx->second_round_output[3];  \
  33 }
  34
  35 #define round(ctx,x0,x1,x2,x3,y0,y1,y2,y3,r) { \
  36         u32 t0,t1,t2,t3;    \
  37         u32 t4,t5,t6,t7;    \
  38         u32 t8,t9,t10,t11;  \
  39         u32 t12,t13,t14,t15;\
  40         u32 tem0,tem1,tem2,tem3;    \
  41         u32 tem4,tem5,tem6,tem7;    \
  42         u32 tem8,tem9,tem10,tem11;  \
  43         u32 tem12,tem13,tem14,tem15;\
  44         \
  45         t0 = (u8)(x0);      \
  46         tem0 = T0[t0];      \
  47         t1 = (u8)((x1)>>8); \
  48         tem1 = tem0 ^ T1[t1];    \
  49         t2 = (u8)((x2)>>16);     \
  50         tem2 = tem1 ^ T2[t2];    \
  51         t3 = (u8)((x3)>>24);     \
  52         tem3 = tem2 ^ T3[t3];    \
  53         (y0) = tem3 ^ ctx->round_key[r][0];   \
  54         \
  55         t4 = (u8)(x1);      \
  56         tem4 = T0[t4];      \
  57         t5 = (u8)((x2)>>8); \
  58         tem5 = tem4 ^ T1[t5];    \
  59         t6 = (u8)((x3)>>16);     \
  60         tem6 = tem5 ^ T2[t6];    \
  61         t7 = (u8)((x0)>>24);     \
  62         tem7 = tem6 ^ T3[t7];    \
  63         (y1) = tem7 ^ ctx->round_key[r][1];   \
  64         \
  65         t8 = (u8)(x2);          \
  66         tem8 = T0[t8];          \
  67         t9 = (u8)((x3)>>8);     \
  68         tem9 = tem8 ^ T1[t9];   \
  69         t10 = (u8)((x0)>>16);   \
  70         tem10 = tem9 ^ T2[t10]; \
  71         t11 = (u8)((x1)>>24);   \
  72         tem11 = tem10 ^ T3[t11];\
  73         (y2) = tem11 ^ ctx->round_key[r][2]; \
  74         \
  75         t12 = (u8)(x3);         \
  76         tem12 = T0[t12];        \
  77         t13 = (u8)((x0)>>8);    \
  78         tem13 = tem12 ^ T1[t13];\
  79         t14 = (u8)((x1)>>16);   \
  80         tem14 = tem13 ^ T2[t14];\
  81         t15 = (u8)((x2)>>24);   \
  82         tem15 = tem14 ^ T3[t15];\
  83         (y3) = tem15 ^ ctx->round_key[r][3]; \
  84 }
  85
  86 /* 22.14 cycles/byte*/
  87 #define last_round(ctx,x0,x1,x2,x3,output,r) { \
  88         u32 t0,t1,t2,t3;    \
  89         u32 t4,t5,t6,t7;    \
  90         u32 t8,t9,t10,t11;  \
  91         u32 t12,t13,t14,t15;\
  92         \
  93         t0 = (u8)(x0);        \
  94         output[0] = Sbox[t0]; \
  95       t7 = (u8)((x0)>>24);  \
  96         output[7] = Sbox[t7]; \
  97         t10 = (u8)((x0)>>16);   \
  98         output[10] = Sbox[t10]; \
  99         t13 = (u8)((x0)>>8);    \
 100         output[13] = Sbox[t13]; \
 101         \
 102         t1 = (u8)((x1)>>8);   \
 103         output[1] = Sbox[t1]; \
 104         t4 = (u8)(x1);        \
 105         output[4] = Sbox[t4]; \
 106         t11 = (u8)((x1)>>24);     \
 107         output[11] = Sbox[t11]; \
 108         t14 = (u8)((x1)>>16);     \
 109         output[14] = Sbox[t14]; \
 110         \
 111         t2 = (u8)((x2)>>16);  \
 112         output[2] = Sbox[t2]; \
 113         t5 = (u8)((x2)>>8);   \
 114         output[5] = Sbox[t5]; \
 115         t8 = (u8)(x2);            \
 116         output[8] = Sbox[t8]; \
 117         t15 = (u8)((x2)>>24);     \
 118         output[15] = Sbox[t15]; \
 119       \
 120         t3 = (u8)((x3)>>24);  \
 121         output[3] = Sbox[t3]; \
 122         t6 = (u8)((x3)>>16);  \
 123         output[6] = Sbox[t6]; \
 124         t9 = (u8)((x3)>>8);       \
 125         output[9] = Sbox[t9];     \
 126         t12 = (u8)(x3);           \
 127         output[12] = Sbox[t12];   \
 128 }
 129
 130 /*
 131         ((u32*)output)[0] ^= ctx->round_key[r][0]; \
 132         ((u32*)output)[1] ^= ctx->round_key[r][1]; \
 133         ((u32*)output)[2] ^= ctx->round_key[r][2]; \
 134         ((u32*)output)[3] ^= ctx->round_key[r][3]; \
 135 }
 136 */
 137
 138 #define aes256_enc_block(x,output,ctx) {\
 139         u32 y0;\
 140         u32 z0,z1,z2,z3;\
 141         u32 a0,a1,a2,a3;\
 142         u32 b0,b1,b2,b3;\
 143       u32 c0,c1,c2,c3;\
 144       u32 d0,d1,d2,d3;\
 145       u32 e0,e1,e2,e3;\
 146       u32 f0,f1,f2,f3;\
 147       u32 g0,g1,g2,g3;\
 148       u32 h0,h1,h2,h3;\
 149       u32 i0,i1,i2,i3;\
 150       u32 j0,j1,j2,j3;\
 151       u32 k0,k1,k2,k3;\
 152         first_round(ctx,x[0],y0);\
 153         second_round(ctx,y0,z0,z1,z2,z3);\
 154         round(ctx,z0,z1,z2,z3,a0,a1,a2,a3,3);\
 155         round(ctx,a0,a1,a2,a3,b0,b1,b2,b3,4);\
 156         round(ctx,b0,b1,b2,b3,c0,c1,c2,c3,5);\
 157         round(ctx,c0,c1,c2,c3,d0,d1,d2,d3,6);\
 158         round(ctx,d0,d1,d2,d3,e0,e1,e2,e3,7);\
 159         round(ctx,e0,e1,e2,e3,f0,f1,f2,f3,8);\
 160         round(ctx,f0,f1,f2,f3,g0,g1,g2,g3,9);\
 161       round(ctx,g0,g1,g2,g3,h0,h1,h2,h3,10);\
 162       round(ctx,h0,h1,h2,h3,i0,i1,i2,i3,11);\
 163       round(ctx,i0,i1,i2,i3,j0,j1,j2,j3,12);\
 164       round(ctx,j0,j1,j2,j3,k0,k1,k2,k3,13);\
 165       last_round(ctx,k0,k1,k2,k3,(output),14);\
 166 }
 167
 168 /*compute the intermediate values for the first two rounds*/
 169 void partial_precompute_tworounds(ECRYPT_ctx* ctx)
 170 {
 171   u32 x0,x1,x2,x3,y0,y1,y2,y3;
 172
 173   x0 = ctx->counter[0] ^ ctx->round_key[0][0];
 174   x1 = ctx->counter[1] ^ ctx->round_key[0][1];
 175   x2 = ctx->counter[2] ^ ctx->round_key[0][2];
 176   x3 = ctx->counter[3] ^ ctx->round_key[0][3];
 177   x0 &= 0xffffff00;
 178   round(ctx,x0,x1,x2,x3,y0,y1,y2,y3,1);
 179   ctx->first_round_output_x0 = y0 ^ T0[0];
 180   y0 = 0;
 181   round(ctx,y0,y1,y2,y3,x0,x1,x2,x3,2);
 182   ctx->second_round_output[0] = x0 ^ T0[0];
 183   ctx->second_round_output[1] = x1 ^ T3[0];
 184   ctx->second_round_output[2] = x2 ^ T2[0];
 185   ctx->second_round_output[3] = x3 ^ T1[0];
 186 }
 187