2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
28 #include "gpu_aes256_nv.c"
29 #include "gpu_twofish256_nv.c"
30 #include "gpu_serpent256_nv.c"
32 __device__ static void ripemd160_transform (const u32x w[16], u32x dgst[5])
40 RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C00, RIPEMD160S00);
41 RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 1], RIPEMD160C00, RIPEMD160S01);
42 RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 2], RIPEMD160C00, RIPEMD160S02);
43 RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 3], RIPEMD160C00, RIPEMD160S03);
44 RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C00, RIPEMD160S04);
45 RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 5], RIPEMD160C00, RIPEMD160S05);
46 RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 6], RIPEMD160C00, RIPEMD160S06);
47 RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 7], RIPEMD160C00, RIPEMD160S07);
48 RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C00, RIPEMD160S08);
49 RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 9], RIPEMD160C00, RIPEMD160S09);
50 RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[10], RIPEMD160C00, RIPEMD160S0A);
51 RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[11], RIPEMD160C00, RIPEMD160S0B);
52 RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[12], RIPEMD160C00, RIPEMD160S0C);
53 RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[13], RIPEMD160C00, RIPEMD160S0D);
54 RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[14], RIPEMD160C00, RIPEMD160S0E);
55 RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[15], RIPEMD160C00, RIPEMD160S0F);
57 RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 7], RIPEMD160C10, RIPEMD160S10);
58 RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 4], RIPEMD160C10, RIPEMD160S11);
59 RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[13], RIPEMD160C10, RIPEMD160S12);
60 RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[ 1], RIPEMD160C10, RIPEMD160S13);
61 RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[10], RIPEMD160C10, RIPEMD160S14);
62 RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 6], RIPEMD160C10, RIPEMD160S15);
63 RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[15], RIPEMD160C10, RIPEMD160S16);
64 RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 3], RIPEMD160C10, RIPEMD160S17);
65 RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[12], RIPEMD160C10, RIPEMD160S18);
66 RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[ 0], RIPEMD160C10, RIPEMD160S19);
67 RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 9], RIPEMD160C10, RIPEMD160S1A);
68 RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 5], RIPEMD160C10, RIPEMD160S1B);
69 RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C10, RIPEMD160S1C);
70 RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[14], RIPEMD160C10, RIPEMD160S1D);
71 RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[11], RIPEMD160C10, RIPEMD160S1E);
72 RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 8], RIPEMD160C10, RIPEMD160S1F);
74 RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 3], RIPEMD160C20, RIPEMD160S20);
75 RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[10], RIPEMD160C20, RIPEMD160S21);
76 RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[14], RIPEMD160C20, RIPEMD160S22);
77 RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 4], RIPEMD160C20, RIPEMD160S23);
78 RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 9], RIPEMD160C20, RIPEMD160S24);
79 RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[15], RIPEMD160C20, RIPEMD160S25);
80 RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C20, RIPEMD160S26);
81 RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[ 1], RIPEMD160C20, RIPEMD160S27);
82 RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C20, RIPEMD160S28);
83 RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 7], RIPEMD160C20, RIPEMD160S29);
84 RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 0], RIPEMD160C20, RIPEMD160S2A);
85 RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 6], RIPEMD160C20, RIPEMD160S2B);
86 RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[13], RIPEMD160C20, RIPEMD160S2C);
87 RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[11], RIPEMD160C20, RIPEMD160S2D);
88 RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C20, RIPEMD160S2E);
89 RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[12], RIPEMD160C20, RIPEMD160S2F);
91 RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 1], RIPEMD160C30, RIPEMD160S30);
92 RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[ 9], RIPEMD160C30, RIPEMD160S31);
93 RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[11], RIPEMD160C30, RIPEMD160S32);
94 RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[10], RIPEMD160C30, RIPEMD160S33);
95 RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 0], RIPEMD160C30, RIPEMD160S34);
96 RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 8], RIPEMD160C30, RIPEMD160S35);
97 RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[12], RIPEMD160C30, RIPEMD160S36);
98 RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[ 4], RIPEMD160C30, RIPEMD160S37);
99 RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[13], RIPEMD160C30, RIPEMD160S38);
100 RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 3], RIPEMD160C30, RIPEMD160S39);
101 RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 7], RIPEMD160C30, RIPEMD160S3A);
102 RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[15], RIPEMD160C30, RIPEMD160S3B);
103 RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[14], RIPEMD160C30, RIPEMD160S3C);
104 RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[ 5], RIPEMD160C30, RIPEMD160S3D);
105 RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 6], RIPEMD160C30, RIPEMD160S3E);
106 RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C30, RIPEMD160S3F);
108 RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C40, RIPEMD160S40);
109 RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C40, RIPEMD160S41);
110 RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C40, RIPEMD160S42);
111 RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 9], RIPEMD160C40, RIPEMD160S43);
112 RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 7], RIPEMD160C40, RIPEMD160S44);
113 RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[12], RIPEMD160C40, RIPEMD160S45);
114 RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C40, RIPEMD160S46);
115 RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[10], RIPEMD160C40, RIPEMD160S47);
116 RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[14], RIPEMD160C40, RIPEMD160S48);
117 RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 1], RIPEMD160C40, RIPEMD160S49);
118 RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 3], RIPEMD160C40, RIPEMD160S4A);
119 RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 8], RIPEMD160C40, RIPEMD160S4B);
120 RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[11], RIPEMD160C40, RIPEMD160S4C);
121 RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 6], RIPEMD160C40, RIPEMD160S4D);
122 RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[15], RIPEMD160C40, RIPEMD160S4E);
123 RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[13], RIPEMD160C40, RIPEMD160S4F);
131 //RIPEMD160_STEP_WORKAROUND_BUG (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 5], RIPEMD160C50, RIPEMD160S50);
132 RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 5], RIPEMD160C50, RIPEMD160S50);
133 RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[14], RIPEMD160C50, RIPEMD160S51);
134 RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 7], RIPEMD160C50, RIPEMD160S52);
135 RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[ 0], RIPEMD160C50, RIPEMD160S53);
136 RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C50, RIPEMD160S54);
137 RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 2], RIPEMD160C50, RIPEMD160S55);
138 RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[11], RIPEMD160C50, RIPEMD160S56);
139 RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C50, RIPEMD160S57);
140 RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[13], RIPEMD160C50, RIPEMD160S58);
141 RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 6], RIPEMD160C50, RIPEMD160S59);
142 RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[15], RIPEMD160C50, RIPEMD160S5A);
143 RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C50, RIPEMD160S5B);
144 RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 1], RIPEMD160C50, RIPEMD160S5C);
145 RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[10], RIPEMD160C50, RIPEMD160S5D);
146 RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 3], RIPEMD160C50, RIPEMD160S5E);
147 RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[12], RIPEMD160C50, RIPEMD160S5F);
149 RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 6], RIPEMD160C60, RIPEMD160S60);
150 RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[11], RIPEMD160C60, RIPEMD160S61);
151 RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 3], RIPEMD160C60, RIPEMD160S62);
152 RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 7], RIPEMD160C60, RIPEMD160S63);
153 RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C60, RIPEMD160S64);
154 RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[13], RIPEMD160C60, RIPEMD160S65);
155 RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[ 5], RIPEMD160C60, RIPEMD160S66);
156 RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[10], RIPEMD160C60, RIPEMD160S67);
157 RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[14], RIPEMD160C60, RIPEMD160S68);
158 RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[15], RIPEMD160C60, RIPEMD160S69);
159 RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 8], RIPEMD160C60, RIPEMD160S6A);
160 RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[12], RIPEMD160C60, RIPEMD160S6B);
161 RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 4], RIPEMD160C60, RIPEMD160S6C);
162 RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 9], RIPEMD160C60, RIPEMD160S6D);
163 RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 1], RIPEMD160C60, RIPEMD160S6E);
164 RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 2], RIPEMD160C60, RIPEMD160S6F);
166 RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[15], RIPEMD160C70, RIPEMD160S70);
167 RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 5], RIPEMD160C70, RIPEMD160S71);
168 RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 1], RIPEMD160C70, RIPEMD160S72);
169 RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 3], RIPEMD160C70, RIPEMD160S73);
170 RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C70, RIPEMD160S74);
171 RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[14], RIPEMD160C70, RIPEMD160S75);
172 RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 6], RIPEMD160C70, RIPEMD160S76);
173 RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C70, RIPEMD160S77);
174 RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[11], RIPEMD160C70, RIPEMD160S78);
175 RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C70, RIPEMD160S79);
176 RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[12], RIPEMD160C70, RIPEMD160S7A);
177 RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C70, RIPEMD160S7B);
178 RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[10], RIPEMD160C70, RIPEMD160S7C);
179 RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 0], RIPEMD160C70, RIPEMD160S7D);
180 RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 4], RIPEMD160C70, RIPEMD160S7E);
181 RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[13], RIPEMD160C70, RIPEMD160S7F);
183 RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 8], RIPEMD160C80, RIPEMD160S80);
184 RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[ 6], RIPEMD160C80, RIPEMD160S81);
185 RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 4], RIPEMD160C80, RIPEMD160S82);
186 RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 1], RIPEMD160C80, RIPEMD160S83);
187 RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[ 3], RIPEMD160C80, RIPEMD160S84);
188 RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[11], RIPEMD160C80, RIPEMD160S85);
189 RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[15], RIPEMD160C80, RIPEMD160S86);
190 RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C80, RIPEMD160S87);
191 RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 5], RIPEMD160C80, RIPEMD160S88);
192 RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[12], RIPEMD160C80, RIPEMD160S89);
193 RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 2], RIPEMD160C80, RIPEMD160S8A);
194 RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[13], RIPEMD160C80, RIPEMD160S8B);
195 RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 9], RIPEMD160C80, RIPEMD160S8C);
196 RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 7], RIPEMD160C80, RIPEMD160S8D);
197 RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[10], RIPEMD160C80, RIPEMD160S8E);
198 RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[14], RIPEMD160C80, RIPEMD160S8F);
200 RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[12], RIPEMD160C90, RIPEMD160S90);
201 RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[15], RIPEMD160C90, RIPEMD160S91);
202 RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[10], RIPEMD160C90, RIPEMD160S92);
203 RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C90, RIPEMD160S93);
204 RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 1], RIPEMD160C90, RIPEMD160S94);
205 RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[ 5], RIPEMD160C90, RIPEMD160S95);
206 RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[ 8], RIPEMD160C90, RIPEMD160S96);
207 RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C90, RIPEMD160S97);
208 RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 6], RIPEMD160C90, RIPEMD160S98);
209 RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C90, RIPEMD160S99);
210 RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[13], RIPEMD160C90, RIPEMD160S9A);
211 RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[14], RIPEMD160C90, RIPEMD160S9B);
212 RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 0], RIPEMD160C90, RIPEMD160S9C);
213 RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 3], RIPEMD160C90, RIPEMD160S9D);
214 RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 9], RIPEMD160C90, RIPEMD160S9E);
215 RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[11], RIPEMD160C90, RIPEMD160S9F);
217 const u32x a = dgst[1] + c1 + d2;
218 const u32x b = dgst[2] + d1 + e2;
219 const u32x c = dgst[3] + e1 + a2;
220 const u32x d = dgst[4] + a1 + b2;
221 const u32x e = dgst[0] + b1 + c2;
230 __device__ static void hmac_run2 (const u32x w1[16], const u32x w2[16], const u32x ipad[5], const u32x opad[5], u32x dgst[5])
238 ripemd160_transform (w1, dgst);
239 ripemd160_transform (w2, dgst);
257 w[14] = (64 + 20) * 8;
266 ripemd160_transform (w, dgst);
269 __device__ static void hmac_run (u32x w[16], const u32x ipad[5], const u32x opad[5], u32x dgst[5])
277 ripemd160_transform (w, dgst);
293 w[14] = (64 + 20) * 8;
302 ripemd160_transform (w, dgst);
305 __device__ static void hmac_init (u32x w[16], u32x ipad[5], u32x opad[5])
324 ipad[0] = RIPEMD160M_A;
325 ipad[1] = RIPEMD160M_B;
326 ipad[2] = RIPEMD160M_C;
327 ipad[3] = RIPEMD160M_D;
328 ipad[4] = RIPEMD160M_E;
330 ripemd160_transform (w, ipad);
349 opad[0] = RIPEMD160M_A;
350 opad[1] = RIPEMD160M_B;
351 opad[2] = RIPEMD160M_C;
352 opad[3] = RIPEMD160M_D;
353 opad[4] = RIPEMD160M_E;
355 ripemd160_transform (w, opad);
358 __device__ static u32 u8add (const u32 a, const u32 b)
360 const u32 a1 = (a >> 0) & 0xff;
361 const u32 a2 = (a >> 8) & 0xff;
362 const u32 a3 = (a >> 16) & 0xff;
363 const u32 a4 = (a >> 24) & 0xff;
365 const u32 b1 = (b >> 0) & 0xff;
366 const u32 b2 = (b >> 8) & 0xff;
367 const u32 b3 = (b >> 16) & 0xff;
368 const u32 b4 = (b >> 24) & 0xff;
370 const u32 r1 = (a1 + b1) & 0xff;
371 const u32 r2 = (a2 + b2) & 0xff;
372 const u32 r3 = (a3 + b3) & 0xff;
373 const u32 r4 = (a4 + b4) & 0xff;
375 const u32 r = r1 << 0
383 extern "C" __global__ void __launch_bounds__ (256, 1) m06211_init (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, tc_tmp_t *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const tc_t *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
389 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
391 if (gid >= gid_max) return;
395 w0[0] = pws[gid].i[ 0];
396 w0[1] = pws[gid].i[ 1];
397 w0[2] = pws[gid].i[ 2];
398 w0[3] = pws[gid].i[ 3];
402 w1[0] = pws[gid].i[ 4];
403 w1[1] = pws[gid].i[ 5];
404 w1[2] = pws[gid].i[ 6];
405 w1[3] = pws[gid].i[ 7];
409 w2[0] = pws[gid].i[ 8];
410 w2[1] = pws[gid].i[ 9];
411 w2[2] = pws[gid].i[10];
412 w2[3] = pws[gid].i[11];
416 w3[0] = pws[gid].i[12];
417 w3[1] = pws[gid].i[13];
418 w3[2] = pws[gid].i[14];
419 w3[3] = pws[gid].i[15];
425 w0[0] = u8add (w0[0], esalt_bufs[salt_pos].keyfile_buf[ 0]);
426 w0[1] = u8add (w0[1], esalt_bufs[salt_pos].keyfile_buf[ 1]);
427 w0[2] = u8add (w0[2], esalt_bufs[salt_pos].keyfile_buf[ 2]);
428 w0[3] = u8add (w0[3], esalt_bufs[salt_pos].keyfile_buf[ 3]);
429 w1[0] = u8add (w1[0], esalt_bufs[salt_pos].keyfile_buf[ 4]);
430 w1[1] = u8add (w1[1], esalt_bufs[salt_pos].keyfile_buf[ 5]);
431 w1[2] = u8add (w1[2], esalt_bufs[salt_pos].keyfile_buf[ 6]);
432 w1[3] = u8add (w1[3], esalt_bufs[salt_pos].keyfile_buf[ 7]);
433 w2[0] = u8add (w2[0], esalt_bufs[salt_pos].keyfile_buf[ 8]);
434 w2[1] = u8add (w2[1], esalt_bufs[salt_pos].keyfile_buf[ 9]);
435 w2[2] = u8add (w2[2], esalt_bufs[salt_pos].keyfile_buf[10]);
436 w2[3] = u8add (w2[3], esalt_bufs[salt_pos].keyfile_buf[11]);
437 w3[0] = u8add (w3[0], esalt_bufs[salt_pos].keyfile_buf[12]);
438 w3[1] = u8add (w3[1], esalt_bufs[salt_pos].keyfile_buf[13]);
439 w3[2] = u8add (w3[2], esalt_bufs[salt_pos].keyfile_buf[14]);
440 w3[3] = u8add (w3[3], esalt_bufs[salt_pos].keyfile_buf[15]);
448 salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
449 salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
450 salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[ 2];
451 salt_buf1[ 3] = esalt_bufs[salt_pos].salt_buf[ 3];
452 salt_buf1[ 4] = esalt_bufs[salt_pos].salt_buf[ 4];
453 salt_buf1[ 5] = esalt_bufs[salt_pos].salt_buf[ 5];
454 salt_buf1[ 6] = esalt_bufs[salt_pos].salt_buf[ 6];
455 salt_buf1[ 7] = esalt_bufs[salt_pos].salt_buf[ 7];
456 salt_buf1[ 8] = esalt_bufs[salt_pos].salt_buf[ 8];
457 salt_buf1[ 9] = esalt_bufs[salt_pos].salt_buf[ 9];
458 salt_buf1[10] = esalt_bufs[salt_pos].salt_buf[10];
459 salt_buf1[11] = esalt_bufs[salt_pos].salt_buf[11];
460 salt_buf1[12] = esalt_bufs[salt_pos].salt_buf[12];
461 salt_buf1[13] = esalt_bufs[salt_pos].salt_buf[13];
462 salt_buf1[14] = esalt_bufs[salt_pos].salt_buf[14];
463 salt_buf1[15] = esalt_bufs[salt_pos].salt_buf[15];
468 salt_buf2[ 1] = 0x80;
481 salt_buf2[14] = (64 + 64 + 4) * 8;
484 const u32 truecrypt_mdlen = salt_bufs[0].truecrypt_mdlen;
508 hmac_init (w, ipad, opad);
510 tmps[gid].ipad[0] = ipad[0];
511 tmps[gid].ipad[1] = ipad[1];
512 tmps[gid].ipad[2] = ipad[2];
513 tmps[gid].ipad[3] = ipad[3];
514 tmps[gid].ipad[4] = ipad[4];
516 tmps[gid].opad[0] = opad[0];
517 tmps[gid].opad[1] = opad[1];
518 tmps[gid].opad[2] = opad[2];
519 tmps[gid].opad[3] = opad[3];
520 tmps[gid].opad[4] = opad[4];
522 for (u32 i = 0, j = 1; i < (truecrypt_mdlen / 8 / 4); i += 5, j += 1)
524 salt_buf2[0] = swap_workaround (j);
528 hmac_run2 (salt_buf1, salt_buf2, ipad, opad, dgst);
530 tmps[gid].dgst[i + 0] = dgst[0];
531 tmps[gid].dgst[i + 1] = dgst[1];
532 tmps[gid].dgst[i + 2] = dgst[2];
533 tmps[gid].dgst[i + 3] = dgst[3];
534 tmps[gid].dgst[i + 4] = dgst[4];
536 tmps[gid].out[i + 0] = dgst[0];
537 tmps[gid].out[i + 1] = dgst[1];
538 tmps[gid].out[i + 2] = dgst[2];
539 tmps[gid].out[i + 3] = dgst[3];
540 tmps[gid].out[i + 4] = dgst[4];
544 extern "C" __global__ void __launch_bounds__ (256, 1) m06211_loop (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, tc_tmp_t *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const tc_t *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
546 const u32 truecrypt_mdlen = salt_bufs[0].truecrypt_mdlen;
548 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
550 if (gid >= gid_max) return;
555 ipad[0] = tmps[gid].ipad[0];
556 ipad[1] = tmps[gid].ipad[1];
557 ipad[2] = tmps[gid].ipad[2];
558 ipad[3] = tmps[gid].ipad[3];
559 ipad[4] = tmps[gid].ipad[4];
561 opad[0] = tmps[gid].opad[0];
562 opad[1] = tmps[gid].opad[1];
563 opad[2] = tmps[gid].opad[2];
564 opad[3] = tmps[gid].opad[3];
565 opad[4] = tmps[gid].opad[4];
567 for (u32 i = 0; i < (truecrypt_mdlen / 8 / 4); i += 5)
572 dgst[0] = tmps[gid].dgst[i + 0];
573 dgst[1] = tmps[gid].dgst[i + 1];
574 dgst[2] = tmps[gid].dgst[i + 2];
575 dgst[3] = tmps[gid].dgst[i + 3];
576 dgst[4] = tmps[gid].dgst[i + 4];
578 out[0] = tmps[gid].out[i + 0];
579 out[1] = tmps[gid].out[i + 1];
580 out[2] = tmps[gid].out[i + 2];
581 out[3] = tmps[gid].out[i + 3];
582 out[4] = tmps[gid].out[i + 4];
584 for (u32 j = 0; j < loop_cnt; j++)
602 w[14] = (64 + 20) * 8;
605 hmac_run (w, ipad, opad, dgst);
614 tmps[gid].dgst[i + 0] = dgst[0];
615 tmps[gid].dgst[i + 1] = dgst[1];
616 tmps[gid].dgst[i + 2] = dgst[2];
617 tmps[gid].dgst[i + 3] = dgst[3];
618 tmps[gid].dgst[i + 4] = dgst[4];
620 tmps[gid].out[i + 0] = out[0];
621 tmps[gid].out[i + 1] = out[1];
622 tmps[gid].out[i + 2] = out[2];
623 tmps[gid].out[i + 3] = out[3];
624 tmps[gid].out[i + 4] = out[4];
628 extern "C" __global__ void __launch_bounds__ (256, 1) m06211_comp (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, tc_tmp_t *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const tc_t *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
634 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
635 const u32 lid = threadIdx.x;
637 if (gid >= gid_max) return;
641 ukey1[0] = tmps[gid].out[ 0];
642 ukey1[1] = tmps[gid].out[ 1];
643 ukey1[2] = tmps[gid].out[ 2];
644 ukey1[3] = tmps[gid].out[ 3];
645 ukey1[4] = tmps[gid].out[ 4];
646 ukey1[5] = tmps[gid].out[ 5];
647 ukey1[6] = tmps[gid].out[ 6];
648 ukey1[7] = tmps[gid].out[ 7];
652 ukey2[0] = tmps[gid].out[ 8];
653 ukey2[1] = tmps[gid].out[ 9];
654 ukey2[2] = tmps[gid].out[10];
655 ukey2[3] = tmps[gid].out[11];
656 ukey2[4] = tmps[gid].out[12];
657 ukey2[5] = tmps[gid].out[13];
658 ukey2[6] = tmps[gid].out[14];
659 ukey2[7] = tmps[gid].out[15];
663 data[0] = esalt_bufs[0].data_buf[0];
664 data[1] = esalt_bufs[0].data_buf[1];
665 data[2] = esalt_bufs[0].data_buf[2];
666 data[3] = esalt_bufs[0].data_buf[3];
676 aes256_decrypt_xts (ukey1, ukey2, tmp, tmp);
678 if (((tmp[0] == 0x45555254) && (tmp[3] == 0)) || ((tmp[0] == 0x45555254) && ((tmp[1] >> 16) <= 5)))
680 mark_hash_s0 (plains_buf, hashes_shown, 0, gid, 0);
682 d_return_buf[lid] = 1;
692 serpent256_decrypt_xts (ukey1, ukey2, tmp, tmp);
694 if (((tmp[0] == 0x45555254) && (tmp[3] == 0)) || ((tmp[0] == 0x45555254) && ((tmp[1] >> 16) <= 5)))
696 mark_hash_s0 (plains_buf, hashes_shown, 0, gid, 0);
698 d_return_buf[lid] = 1;
708 twofish256_decrypt_xts (ukey1, ukey2, tmp, tmp);
710 if (((tmp[0] == 0x45555254) && (tmp[3] == 0)) || ((tmp[0] == 0x45555254) && ((tmp[1] >> 16) <= 5)))
712 mark_hash_s0 (plains_buf, hashes_shown, 0, gid, 0);
714 d_return_buf[lid] = 1;