#define _MD5_
-#include "include/constants.h"
-#include "include/kernel_vendor.h"
+#include "inc_vendor.cl"
+#include "inc_hash_constants.h"
+#include "inc_hash_functions.cl"
+#include "inc_types.cl"
+#include "inc_common.cl"
-#define DGST_R0 0
-#define DGST_R1 1
-#define DGST_R2 2
-#define DGST_R3 3
+#define COMPARE_S "inc_comp_single.cl"
+#define COMPARE_M "inc_comp_multi.cl"
-#include "include/kernel_functions.c"
-#include "types_ocl.c"
-#include "common.c"
-
-#ifdef VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
-
-static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
+void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 we_t = w3[2];
u32 wf_t = 0;
- u32 tmp2;
-
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w5_t, MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w8_t, MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, wb_t, MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, we_t, MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t, MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w4_t, MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w7_t, MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, wa_t, MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, wd_t, MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t, MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t, MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w6_t, MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w9_t, MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, wc_t, MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, wf_t, MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t, MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
digest[3] += d;
}
-static void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[4])
+void memcat16 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[4])
{
- switch (block_len)
- {
- case 0:
- block0[0] = append[0];
- block0[1] = append[1];
- block0[2] = append[2];
- block0[3] = append[3];
- break;
-
- case 1:
- block0[0] = block0[0] | append[0] << 8;
- block0[1] = append[0] >> 24 | append[1] << 8;
- block0[2] = append[1] >> 24 | append[2] << 8;
- block0[3] = append[2] >> 24 | append[3] << 8;
- block1[0] = append[3] >> 24;
- break;
-
- case 2:
- block0[0] = block0[0] | append[0] << 16;
- block0[1] = append[0] >> 16 | append[1] << 16;
- block0[2] = append[1] >> 16 | append[2] << 16;
- block0[3] = append[2] >> 16 | append[3] << 16;
- block1[0] = append[3] >> 16;
- break;
-
- case 3:
- block0[0] = block0[0] | append[0] << 24;
- block0[1] = append[0] >> 8 | append[1] << 24;
- block0[2] = append[1] >> 8 | append[2] << 24;
- block0[3] = append[2] >> 8 | append[3] << 24;
- block1[0] = append[3] >> 8;
- break;
-
- case 4:
- block0[1] = append[0];
- block0[2] = append[1];
- block0[3] = append[2];
- block1[0] = append[3];
- break;
-
- case 5:
- block0[1] = block0[1] | append[0] << 8;
- block0[2] = append[0] >> 24 | append[1] << 8;
- block0[3] = append[1] >> 24 | append[2] << 8;
- block1[0] = append[2] >> 24 | append[3] << 8;
- block1[1] = append[3] >> 24;
- break;
-
- case 6:
- block0[1] = block0[1] | append[0] << 16;
- block0[2] = append[0] >> 16 | append[1] << 16;
- block0[3] = append[1] >> 16 | append[2] << 16;
- block1[0] = append[2] >> 16 | append[3] << 16;
- block1[1] = append[3] >> 16;
- break;
-
- case 7:
- block0[1] = block0[1] | append[0] << 24;
- block0[2] = append[0] >> 8 | append[1] << 24;
- block0[3] = append[1] >> 8 | append[2] << 24;
- block1[0] = append[2] >> 8 | append[3] << 24;
- block1[1] = append[3] >> 8;
- break;
-
- case 8:
- block0[2] = append[0];
- block0[3] = append[1];
- block1[0] = append[2];
- block1[1] = append[3];
- break;
-
- case 9:
- block0[2] = block0[2] | append[0] << 8;
- block0[3] = append[0] >> 24 | append[1] << 8;
- block1[0] = append[1] >> 24 | append[2] << 8;
- block1[1] = append[2] >> 24 | append[3] << 8;
- block1[2] = append[3] >> 24;
- break;
-
- case 10:
- block0[2] = block0[2] | append[0] << 16;
- block0[3] = append[0] >> 16 | append[1] << 16;
- block1[0] = append[1] >> 16 | append[2] << 16;
- block1[1] = append[2] >> 16 | append[3] << 16;
- block1[2] = append[3] >> 16;
- break;
-
- case 11:
- block0[2] = block0[2] | append[0] << 24;
- block0[3] = append[0] >> 8 | append[1] << 24;
- block1[0] = append[1] >> 8 | append[2] << 24;
- block1[1] = append[2] >> 8 | append[3] << 24;
- block1[2] = append[3] >> 8;
- break;
-
- case 12:
- block0[3] = append[0];
- block1[0] = append[1];
- block1[1] = append[2];
- block1[2] = append[3];
- break;
-
- case 13:
- block0[3] = block0[3] | append[0] << 8;
- block1[0] = append[0] >> 24 | append[1] << 8;
- block1[1] = append[1] >> 24 | append[2] << 8;
- block1[2] = append[2] >> 24 | append[3] << 8;
- block1[3] = append[3] >> 24;
- break;
-
- case 14:
- block0[3] = block0[3] | append[0] << 16;
- block1[0] = append[0] >> 16 | append[1] << 16;
- block1[1] = append[1] >> 16 | append[2] << 16;
- block1[2] = append[2] >> 16 | append[3] << 16;
- block1[3] = append[3] >> 16;
- break;
-
- case 15:
- block0[3] = block0[3] | append[0] << 24;
- block1[0] = append[0] >> 8 | append[1] << 24;
- block1[1] = append[1] >> 8 | append[2] << 24;
- block1[2] = append[2] >> 8 | append[3] << 24;
- block1[3] = append[3] >> 8;
- break;
-
- case 16:
- block1[0] = append[0];
- block1[1] = append[1];
- block1[2] = append[2];
- block1[3] = append[3];
- break;
-
- case 17:
- block1[0] = block1[0] | append[0] << 8;
- block1[1] = append[0] >> 24 | append[1] << 8;
- block1[2] = append[1] >> 24 | append[2] << 8;
- block1[3] = append[2] >> 24 | append[3] << 8;
- block2[0] = append[3] >> 24;
- break;
-
- case 18:
- block1[0] = block1[0] | append[0] << 16;
- block1[1] = append[0] >> 16 | append[1] << 16;
- block1[2] = append[1] >> 16 | append[2] << 16;
- block1[3] = append[2] >> 16 | append[3] << 16;
- block2[0] = append[3] >> 16;
- break;
-
- case 19:
- block1[0] = block1[0] | append[0] << 24;
- block1[1] = append[0] >> 8 | append[1] << 24;
- block1[2] = append[1] >> 8 | append[2] << 24;
- block1[3] = append[2] >> 8 | append[3] << 24;
- block2[0] = append[3] >> 8;
- break;
-
- case 20:
- block1[1] = append[0];
- block1[2] = append[1];
- block1[3] = append[2];
- block2[0] = append[3];
- break;
-
- case 21:
- block1[1] = block1[1] | append[0] << 8;
- block1[2] = append[0] >> 24 | append[1] << 8;
- block1[3] = append[1] >> 24 | append[2] << 8;
- block2[0] = append[2] >> 24 | append[3] << 8;
- block2[1] = append[3] >> 24;
- break;
-
- case 22:
- block1[1] = block1[1] | append[0] << 16;
- block1[2] = append[0] >> 16 | append[1] << 16;
- block1[3] = append[1] >> 16 | append[2] << 16;
- block2[0] = append[2] >> 16 | append[3] << 16;
- block2[1] = append[3] >> 16;
- break;
-
- case 23:
- block1[1] = block1[1] | append[0] << 24;
- block1[2] = append[0] >> 8 | append[1] << 24;
- block1[3] = append[1] >> 8 | append[2] << 24;
- block2[0] = append[2] >> 8 | append[3] << 24;
- block2[1] = append[3] >> 8;
- break;
-
- case 24:
- block1[2] = append[0];
- block1[3] = append[1];
- block2[0] = append[2];
- block2[1] = append[3];
- break;
-
- case 25:
- block1[2] = block1[2] | append[0] << 8;
- block1[3] = append[0] >> 24 | append[1] << 8;
- block2[0] = append[1] >> 24 | append[2] << 8;
- block2[1] = append[2] >> 24 | append[3] << 8;
- block2[2] = append[3] >> 24;
- break;
-
- case 26:
- block1[2] = block1[2] | append[0] << 16;
- block1[3] = append[0] >> 16 | append[1] << 16;
- block2[0] = append[1] >> 16 | append[2] << 16;
- block2[1] = append[2] >> 16 | append[3] << 16;
- block2[2] = append[3] >> 16;
- break;
-
- case 27:
- block1[2] = block1[2] | append[0] << 24;
- block1[3] = append[0] >> 8 | append[1] << 24;
- block2[0] = append[1] >> 8 | append[2] << 24;
- block2[1] = append[2] >> 8 | append[3] << 24;
- block2[2] = append[3] >> 8;
- break;
-
- case 28:
- block1[3] = append[0];
- block2[0] = append[1];
- block2[1] = append[2];
- block2[2] = append[3];
- break;
-
- case 29:
- block1[3] = block1[3] | append[0] << 8;
- block2[0] = append[0] >> 24 | append[1] << 8;
- block2[1] = append[1] >> 24 | append[2] << 8;
- block2[2] = append[2] >> 24 | append[3] << 8;
- block2[3] = append[3] >> 24;
- break;
-
- case 30:
- block1[3] = block1[3] | append[0] << 16;
- block2[0] = append[0] >> 16 | append[1] << 16;
- block2[1] = append[1] >> 16 | append[2] << 16;
- block2[2] = append[2] >> 16 | append[3] << 16;
- block2[3] = append[3] >> 16;
- break;
-
- case 31:
- block1[3] = block1[3] | append[0] << 24;
- block2[0] = append[0] >> 8 | append[1] << 24;
- block2[1] = append[1] >> 8 | append[2] << 24;
- block2[2] = append[2] >> 8 | append[3] << 24;
- block2[3] = append[3] >> 8;
- break;
-
- case 32:
- block2[0] = append[0];
- block2[1] = append[1];
- block2[2] = append[2];
- block2[3] = append[3];
- break;
-
- case 33:
- block2[0] = block2[0] | append[0] << 8;
- block2[1] = append[0] >> 24 | append[1] << 8;
- block2[2] = append[1] >> 24 | append[2] << 8;
- block2[3] = append[2] >> 24 | append[3] << 8;
- block3[0] = append[3] >> 24;
- break;
-
- case 34:
- block2[0] = block2[0] | append[0] << 16;
- block2[1] = append[0] >> 16 | append[1] << 16;
- block2[2] = append[1] >> 16 | append[2] << 16;
- block2[3] = append[2] >> 16 | append[3] << 16;
- block3[0] = append[3] >> 16;
- break;
-
- case 35:
- block2[0] = block2[0] | append[0] << 24;
- block2[1] = append[0] >> 8 | append[1] << 24;
- block2[2] = append[1] >> 8 | append[2] << 24;
- block2[3] = append[2] >> 8 | append[3] << 24;
- block3[0] = append[3] >> 8;
- break;
-
- case 36:
- block2[1] = append[0];
- block2[2] = append[1];
- block2[3] = append[2];
- block3[0] = append[3];
- break;
-
- case 37:
- block2[1] = block2[1] | append[0] << 8;
- block2[2] = append[0] >> 24 | append[1] << 8;
- block2[3] = append[1] >> 24 | append[2] << 8;
- block3[0] = append[2] >> 24 | append[3] << 8;
- block3[1] = append[3] >> 24;
- break;
-
- case 38:
- block2[1] = block2[1] | append[0] << 16;
- block2[2] = append[0] >> 16 | append[1] << 16;
- block2[3] = append[1] >> 16 | append[2] << 16;
- block3[0] = append[2] >> 16 | append[3] << 16;
- block3[1] = append[3] >> 16;
- break;
-
- case 39:
- block2[1] = block2[1] | append[0] << 24;
- block2[2] = append[0] >> 8 | append[1] << 24;
- block2[3] = append[1] >> 8 | append[2] << 24;
- block3[0] = append[2] >> 8 | append[3] << 24;
- block3[1] = append[3] >> 8;
- break;
-
- case 40:
- block2[2] = append[0];
- block2[3] = append[1];
- block3[0] = append[2];
- block3[1] = append[3];
- break;
+ u32 tmp0;
+ u32 tmp1;
+ u32 tmp2;
+ u32 tmp3;
+ u32 tmp4;
- case 41:
- block2[2] = block2[2] | append[0] << 8;
- block2[3] = append[0] >> 24 | append[1] << 8;
- block3[0] = append[1] >> 24 | append[2] << 8;
- block3[1] = append[2] >> 24 | append[3] << 8;
- block3[2] = append[3] >> 24;
- break;
+ #if defined IS_AMD || defined IS_GENERIC
- case 42:
- block2[2] = block2[2] | append[0] << 16;
- block2[3] = append[0] >> 16 | append[1] << 16;
- block3[0] = append[1] >> 16 | append[2] << 16;
- block3[1] = append[2] >> 16 | append[3] << 16;
- block3[2] = append[3] >> 16;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 43:
- block2[2] = block2[2] | append[0] << 24;
- block2[3] = append[0] >> 8 | append[1] << 24;
- block3[0] = append[1] >> 8 | append[2] << 24;
- block3[1] = append[2] >> 8 | append[3] << 24;
- block3[2] = append[3] >> 8;
- break;
+ tmp0 = amd_bytealign (append[0], 0, offset_minus_4);
+ tmp1 = amd_bytealign (append[1], append[0], offset_minus_4);
+ tmp2 = amd_bytealign (append[2], append[1], offset_minus_4);
+ tmp3 = amd_bytealign (append[3], append[2], offset_minus_4);
+ tmp4 = amd_bytealign ( 0, append[3], offset_minus_4);
- case 44:
- block2[3] = append[0];
- block3[0] = append[1];
- block3[1] = append[2];
- block3[2] = append[3];
- break;
+ const u32 mod = block_len & 3;
- case 45:
- block2[3] = block2[3] | append[0] << 8;
- block3[0] = append[0] >> 24 | append[1] << 8;
- block3[1] = append[1] >> 24 | append[2] << 8;
- block3[2] = append[2] >> 24 | append[3] << 8;
- block3[3] = append[3] >> 24;
- break;
-
- case 46:
- block2[3] = block2[3] | append[0] << 16;
- block3[0] = append[0] >> 16 | append[1] << 16;
- block3[1] = append[1] >> 16 | append[2] << 16;
- block3[2] = append[2] >> 16 | append[3] << 16;
- block3[3] = append[3] >> 16;
- break;
-
- case 47:
- block2[3] = block2[3] | append[0] << 24;
- block3[0] = append[0] >> 8 | append[1] << 24;
- block3[1] = append[1] >> 8 | append[2] << 24;
- block3[2] = append[2] >> 8 | append[3] << 24;
- block3[3] = append[3] >> 8;
- break;
-
- case 48:
- block3[0] = append[0];
- block3[1] = append[1];
- block3[2] = append[2];
- block3[3] = append[3];
- break;
+ if (mod == 0)
+ {
+ tmp0 = tmp1;
+ tmp1 = tmp2;
+ tmp2 = tmp3;
+ tmp3 = tmp4;
+ tmp4 = 0;
+ }
- case 49:
- block3[0] = block3[0] | append[0] << 8;
- block3[1] = append[0] >> 24 | append[1] << 8;
- block3[2] = append[1] >> 24 | append[2] << 8;
- block3[3] = append[2] >> 24 | append[3] << 8;
- break;
+ #endif
- case 50:
- block3[0] = block3[0] | append[0] << 16;
- block3[1] = append[0] >> 16 | append[1] << 16;
- block3[2] = append[1] >> 16 | append[2] << 16;
- block3[3] = append[2] >> 16 | append[3] << 16;
- break;
+ #ifdef IS_NV
- case 51:
- block3[0] = block3[0] | append[0] << 24;
- block3[1] = append[0] >> 8 | append[1] << 24;
- block3[2] = append[1] >> 8 | append[2] << 24;
- block3[3] = append[2] >> 8 | append[3] << 24;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 52:
- block3[1] = append[0];
- block3[2] = append[1];
- block3[3] = append[2];
- break;
+ const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
- case 53:
- block3[1] = block3[1] | append[0] << 8;
- block3[2] = append[0] >> 24 | append[1] << 8;
- block3[3] = append[1] >> 24 | append[2] << 8;
- break;
+ tmp0 = __byte_perm ( 0, append[0], selector);
+ tmp1 = __byte_perm (append[0], append[1], selector);
+ tmp2 = __byte_perm (append[1], append[2], selector);
+ tmp3 = __byte_perm (append[2], append[3], selector);
+ tmp4 = __byte_perm (append[3], 0, selector);
- case 54:
- block3[1] = block3[1] | append[0] << 16;
- block3[2] = append[0] >> 16 | append[1] << 16;
- block3[3] = append[1] >> 16 | append[2] << 16;
- break;
+ #endif
- case 55:
- block3[1] = block3[1] | append[0] << 24;
- block3[2] = append[0] >> 8 | append[1] << 24;
- block3[3] = append[1] >> 8 | append[2] << 24;
- break;
+ const u32 div = block_len / 4;
- case 56:
- block3[2] = append[0];
- block3[3] = append[1];
- break;
+ switch (div)
+ {
+ case 0: block0[0] |= tmp0;
+ block0[1] = tmp1;
+ block0[2] = tmp2;
+ block0[3] = tmp3;
+ block1[0] = tmp4;
+ break;
+ case 1: block0[1] |= tmp0;
+ block0[2] = tmp1;
+ block0[3] = tmp2;
+ block1[0] = tmp3;
+ block1[1] = tmp4;
+ break;
+ case 2: block0[2] |= tmp0;
+ block0[3] = tmp1;
+ block1[0] = tmp2;
+ block1[1] = tmp3;
+ block1[2] = tmp4;
+ break;
+ case 3: block0[3] |= tmp0;
+ block1[0] = tmp1;
+ block1[1] = tmp2;
+ block1[2] = tmp3;
+ block1[3] = tmp4;
+ break;
+ case 4: block1[0] |= tmp0;
+ block1[1] = tmp1;
+ block1[2] = tmp2;
+ block1[3] = tmp3;
+ block2[0] = tmp4;
+ break;
+ case 5: block1[1] |= tmp0;
+ block1[2] = tmp1;
+ block1[3] = tmp2;
+ block2[0] = tmp3;
+ block2[1] = tmp4;
+ break;
+ case 6: block1[2] |= tmp0;
+ block1[3] = tmp1;
+ block2[0] = tmp2;
+ block2[1] = tmp3;
+ block2[2] = tmp4;
+ break;
+ case 7: block1[3] |= tmp0;
+ block2[0] = tmp1;
+ block2[1] = tmp2;
+ block2[2] = tmp3;
+ block2[3] = tmp4;
+ break;
+ case 8: block2[0] |= tmp0;
+ block2[1] = tmp1;
+ block2[2] = tmp2;
+ block2[3] = tmp3;
+ block3[0] = tmp4;
+ break;
+ case 9: block2[1] |= tmp0;
+ block2[2] = tmp1;
+ block2[3] = tmp2;
+ block3[0] = tmp3;
+ block3[1] = tmp4;
+ break;
}
}
-static void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[4])
+void memcat16_x80 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[4])
{
- switch (block_len)
- {
- case 0:
- block0[0] = append[0];
- block0[1] = append[1];
- block0[2] = append[2];
- block0[3] = append[3];
- block1[0] = 0x80;
- break;
-
- case 1:
- block0[0] = block0[0] | append[0] << 8;
- block0[1] = append[0] >> 24 | append[1] << 8;
- block0[2] = append[1] >> 24 | append[2] << 8;
- block0[3] = append[2] >> 24 | append[3] << 8;
- block1[0] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 2:
- block0[0] = block0[0] | append[0] << 16;
- block0[1] = append[0] >> 16 | append[1] << 16;
- block0[2] = append[1] >> 16 | append[2] << 16;
- block0[3] = append[2] >> 16 | append[3] << 16;
- block1[0] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 3:
- block0[0] = block0[0] | append[0] << 24;
- block0[1] = append[0] >> 8 | append[1] << 24;
- block0[2] = append[1] >> 8 | append[2] << 24;
- block0[3] = append[2] >> 8 | append[3] << 24;
- block1[0] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 4:
- block0[1] = append[0];
- block0[2] = append[1];
- block0[3] = append[2];
- block1[0] = append[3];
- block1[1] = 0x80;
- break;
-
- case 5:
- block0[1] = block0[1] | append[0] << 8;
- block0[2] = append[0] >> 24 | append[1] << 8;
- block0[3] = append[1] >> 24 | append[2] << 8;
- block1[0] = append[2] >> 24 | append[3] << 8;
- block1[1] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 6:
- block0[1] = block0[1] | append[0] << 16;
- block0[2] = append[0] >> 16 | append[1] << 16;
- block0[3] = append[1] >> 16 | append[2] << 16;
- block1[0] = append[2] >> 16 | append[3] << 16;
- block1[1] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 7:
- block0[1] = block0[1] | append[0] << 24;
- block0[2] = append[0] >> 8 | append[1] << 24;
- block0[3] = append[1] >> 8 | append[2] << 24;
- block1[0] = append[2] >> 8 | append[3] << 24;
- block1[1] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 8:
- block0[2] = append[0];
- block0[3] = append[1];
- block1[0] = append[2];
- block1[1] = append[3];
- block1[2] = 0x80;
- break;
-
- case 9:
- block0[2] = block0[2] | append[0] << 8;
- block0[3] = append[0] >> 24 | append[1] << 8;
- block1[0] = append[1] >> 24 | append[2] << 8;
- block1[1] = append[2] >> 24 | append[3] << 8;
- block1[2] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 10:
- block0[2] = block0[2] | append[0] << 16;
- block0[3] = append[0] >> 16 | append[1] << 16;
- block1[0] = append[1] >> 16 | append[2] << 16;
- block1[1] = append[2] >> 16 | append[3] << 16;
- block1[2] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 11:
- block0[2] = block0[2] | append[0] << 24;
- block0[3] = append[0] >> 8 | append[1] << 24;
- block1[0] = append[1] >> 8 | append[2] << 24;
- block1[1] = append[2] >> 8 | append[3] << 24;
- block1[2] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 12:
- block0[3] = append[0];
- block1[0] = append[1];
- block1[1] = append[2];
- block1[2] = append[3];
- block1[3] = 0x80;
- break;
-
- case 13:
- block0[3] = block0[3] | append[0] << 8;
- block1[0] = append[0] >> 24 | append[1] << 8;
- block1[1] = append[1] >> 24 | append[2] << 8;
- block1[2] = append[2] >> 24 | append[3] << 8;
- block1[3] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 14:
- block0[3] = block0[3] | append[0] << 16;
- block1[0] = append[0] >> 16 | append[1] << 16;
- block1[1] = append[1] >> 16 | append[2] << 16;
- block1[2] = append[2] >> 16 | append[3] << 16;
- block1[3] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 15:
- block0[3] = block0[3] | append[0] << 24;
- block1[0] = append[0] >> 8 | append[1] << 24;
- block1[1] = append[1] >> 8 | append[2] << 24;
- block1[2] = append[2] >> 8 | append[3] << 24;
- block1[3] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 16:
- block1[0] = append[0];
- block1[1] = append[1];
- block1[2] = append[2];
- block1[3] = append[3];
- block2[0] = 0x80;
- break;
-
- case 17:
- block1[0] = block1[0] | append[0] << 8;
- block1[1] = append[0] >> 24 | append[1] << 8;
- block1[2] = append[1] >> 24 | append[2] << 8;
- block1[3] = append[2] >> 24 | append[3] << 8;
- block2[0] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 18:
- block1[0] = block1[0] | append[0] << 16;
- block1[1] = append[0] >> 16 | append[1] << 16;
- block1[2] = append[1] >> 16 | append[2] << 16;
- block1[3] = append[2] >> 16 | append[3] << 16;
- block2[0] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 19:
- block1[0] = block1[0] | append[0] << 24;
- block1[1] = append[0] >> 8 | append[1] << 24;
- block1[2] = append[1] >> 8 | append[2] << 24;
- block1[3] = append[2] >> 8 | append[3] << 24;
- block2[0] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 20:
- block1[1] = append[0];
- block1[2] = append[1];
- block1[3] = append[2];
- block2[0] = append[3];
- block2[1] = 0x80;
- break;
-
- case 21:
- block1[1] = block1[1] | append[0] << 8;
- block1[2] = append[0] >> 24 | append[1] << 8;
- block1[3] = append[1] >> 24 | append[2] << 8;
- block2[0] = append[2] >> 24 | append[3] << 8;
- block2[1] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 22:
- block1[1] = block1[1] | append[0] << 16;
- block1[2] = append[0] >> 16 | append[1] << 16;
- block1[3] = append[1] >> 16 | append[2] << 16;
- block2[0] = append[2] >> 16 | append[3] << 16;
- block2[1] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 23:
- block1[1] = block1[1] | append[0] << 24;
- block1[2] = append[0] >> 8 | append[1] << 24;
- block1[3] = append[1] >> 8 | append[2] << 24;
- block2[0] = append[2] >> 8 | append[3] << 24;
- block2[1] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 24:
- block1[2] = append[0];
- block1[3] = append[1];
- block2[0] = append[2];
- block2[1] = append[3];
- block2[2] = 0x80;
- break;
-
- case 25:
- block1[2] = block1[2] | append[0] << 8;
- block1[3] = append[0] >> 24 | append[1] << 8;
- block2[0] = append[1] >> 24 | append[2] << 8;
- block2[1] = append[2] >> 24 | append[3] << 8;
- block2[2] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 26:
- block1[2] = block1[2] | append[0] << 16;
- block1[3] = append[0] >> 16 | append[1] << 16;
- block2[0] = append[1] >> 16 | append[2] << 16;
- block2[1] = append[2] >> 16 | append[3] << 16;
- block2[2] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 27:
- block1[2] = block1[2] | append[0] << 24;
- block1[3] = append[0] >> 8 | append[1] << 24;
- block2[0] = append[1] >> 8 | append[2] << 24;
- block2[1] = append[2] >> 8 | append[3] << 24;
- block2[2] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 28:
- block1[3] = append[0];
- block2[0] = append[1];
- block2[1] = append[2];
- block2[2] = append[3];
- block2[3] = 0x80;
- break;
-
- case 29:
- block1[3] = block1[3] | append[0] << 8;
- block2[0] = append[0] >> 24 | append[1] << 8;
- block2[1] = append[1] >> 24 | append[2] << 8;
- block2[2] = append[2] >> 24 | append[3] << 8;
- block2[3] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 30:
- block1[3] = block1[3] | append[0] << 16;
- block2[0] = append[0] >> 16 | append[1] << 16;
- block2[1] = append[1] >> 16 | append[2] << 16;
- block2[2] = append[2] >> 16 | append[3] << 16;
- block2[3] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 31:
- block1[3] = block1[3] | append[0] << 24;
- block2[0] = append[0] >> 8 | append[1] << 24;
- block2[1] = append[1] >> 8 | append[2] << 24;
- block2[2] = append[2] >> 8 | append[3] << 24;
- block2[3] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 32:
- block2[0] = append[0];
- block2[1] = append[1];
- block2[2] = append[2];
- block2[3] = append[3];
- block3[0] = 0x80;
- break;
-
- case 33:
- block2[0] = block2[0] | append[0] << 8;
- block2[1] = append[0] >> 24 | append[1] << 8;
- block2[2] = append[1] >> 24 | append[2] << 8;
- block2[3] = append[2] >> 24 | append[3] << 8;
- block3[0] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 34:
- block2[0] = block2[0] | append[0] << 16;
- block2[1] = append[0] >> 16 | append[1] << 16;
- block2[2] = append[1] >> 16 | append[2] << 16;
- block2[3] = append[2] >> 16 | append[3] << 16;
- block3[0] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 35:
- block2[0] = block2[0] | append[0] << 24;
- block2[1] = append[0] >> 8 | append[1] << 24;
- block2[2] = append[1] >> 8 | append[2] << 24;
- block2[3] = append[2] >> 8 | append[3] << 24;
- block3[0] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 36:
- block2[1] = append[0];
- block2[2] = append[1];
- block2[3] = append[2];
- block3[0] = append[3];
- block3[1] = 0x80;
- break;
-
- case 37:
- block2[1] = block2[1] | append[0] << 8;
- block2[2] = append[0] >> 24 | append[1] << 8;
- block2[3] = append[1] >> 24 | append[2] << 8;
- block3[0] = append[2] >> 24 | append[3] << 8;
- block3[1] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 38:
- block2[1] = block2[1] | append[0] << 16;
- block2[2] = append[0] >> 16 | append[1] << 16;
- block2[3] = append[1] >> 16 | append[2] << 16;
- block3[0] = append[2] >> 16 | append[3] << 16;
- block3[1] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 39:
- block2[1] = block2[1] | append[0] << 24;
- block2[2] = append[0] >> 8 | append[1] << 24;
- block2[3] = append[1] >> 8 | append[2] << 24;
- block3[0] = append[2] >> 8 | append[3] << 24;
- block3[1] = append[3] >> 8 | 0x80u << 24;
- break;
-
- case 40:
- block2[2] = append[0];
- block2[3] = append[1];
- block3[0] = append[2];
- block3[1] = append[3];
- block3[2] = 0x80;
- break;
-
- case 41:
- block2[2] = block2[2] | append[0] << 8;
- block2[3] = append[0] >> 24 | append[1] << 8;
- block3[0] = append[1] >> 24 | append[2] << 8;
- block3[1] = append[2] >> 24 | append[3] << 8;
- block3[2] = append[3] >> 24 | 0x80u << 8;
- break;
+ u32 tmp0;
+ u32 tmp1;
+ u32 tmp2;
+ u32 tmp3;
+ u32 tmp4;
- case 42:
- block2[2] = block2[2] | append[0] << 16;
- block2[3] = append[0] >> 16 | append[1] << 16;
- block3[0] = append[1] >> 16 | append[2] << 16;
- block3[1] = append[2] >> 16 | append[3] << 16;
- block3[2] = append[3] >> 16 | 0x80u << 16;
- break;
+ #if defined IS_AMD || defined IS_GENERIC
- case 43:
- block2[2] = block2[2] | append[0] << 24;
- block2[3] = append[0] >> 8 | append[1] << 24;
- block3[0] = append[1] >> 8 | append[2] << 24;
- block3[1] = append[2] >> 8 | append[3] << 24;
- block3[2] = append[3] >> 8 | 0x80u << 24;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 44:
- block2[3] = append[0];
- block3[0] = append[1];
- block3[1] = append[2];
- block3[2] = append[3];
- block3[3] = 0x80;
- break;
+ tmp0 = amd_bytealign (append[0], 0, offset_minus_4);
+ tmp1 = amd_bytealign (append[1], append[0], offset_minus_4);
+ tmp2 = amd_bytealign (append[2], append[1], offset_minus_4);
+ tmp3 = amd_bytealign (append[3], append[2], offset_minus_4);
+ tmp4 = amd_bytealign ( 0x80, append[3], offset_minus_4);
- case 45:
- block2[3] = block2[3] | append[0] << 8;
- block3[0] = append[0] >> 24 | append[1] << 8;
- block3[1] = append[1] >> 24 | append[2] << 8;
- block3[2] = append[2] >> 24 | append[3] << 8;
- block3[3] = append[3] >> 24 | 0x80u << 8;
- break;
-
- case 46:
- block2[3] = block2[3] | append[0] << 16;
- block3[0] = append[0] >> 16 | append[1] << 16;
- block3[1] = append[1] >> 16 | append[2] << 16;
- block3[2] = append[2] >> 16 | append[3] << 16;
- block3[3] = append[3] >> 16 | 0x80u << 16;
- break;
-
- case 47:
- block2[3] = block2[3] | append[0] << 24;
- block3[0] = append[0] >> 8 | append[1] << 24;
- block3[1] = append[1] >> 8 | append[2] << 24;
- block3[2] = append[2] >> 8 | append[3] << 24;
- block3[3] = append[3] >> 8 | 0x80u << 24;
- break;
+ const u32 mod = block_len & 3;
- case 48:
- block3[0] = append[0];
- block3[1] = append[1];
- block3[2] = append[2];
- block3[3] = append[3];
- break;
+ if (mod == 0)
+ {
+ tmp0 = tmp1;
+ tmp1 = tmp2;
+ tmp2 = tmp3;
+ tmp3 = tmp4;
+ tmp4 = 0x80;
+ }
- case 49:
- block3[0] = block3[0] | append[0] << 8;
- block3[1] = append[0] >> 24 | append[1] << 8;
- block3[2] = append[1] >> 24 | append[2] << 8;
- block3[3] = append[2] >> 24 | append[3] << 8;
- break;
+ #endif
- case 50:
- block3[0] = block3[0] | append[0] << 16;
- block3[1] = append[0] >> 16 | append[1] << 16;
- block3[2] = append[1] >> 16 | append[2] << 16;
- block3[3] = append[2] >> 16 | append[3] << 16;
- break;
+ #ifdef IS_NV
- case 51:
- block3[0] = block3[0] | append[0] << 24;
- block3[1] = append[0] >> 8 | append[1] << 24;
- block3[2] = append[1] >> 8 | append[2] << 24;
- block3[3] = append[2] >> 8 | append[3] << 24;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 52:
- block3[1] = append[0];
- block3[2] = append[1];
- block3[3] = append[2];
- break;
+ const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
- case 53:
- block3[1] = block3[1] | append[0] << 8;
- block3[2] = append[0] >> 24 | append[1] << 8;
- block3[3] = append[1] >> 24 | append[2] << 8;
- break;
+ tmp0 = __byte_perm ( 0, append[0], selector);
+ tmp1 = __byte_perm (append[0], append[1], selector);
+ tmp2 = __byte_perm (append[1], append[2], selector);
+ tmp3 = __byte_perm (append[2], append[3], selector);
+ tmp4 = __byte_perm (append[3], 0x80, selector);
- case 54:
- block3[1] = block3[1] | append[0] << 16;
- block3[2] = append[0] >> 16 | append[1] << 16;
- block3[3] = append[1] >> 16 | append[2] << 16;
- break;
+ #endif
- case 55:
- block3[1] = block3[1] | append[0] << 24;
- block3[2] = append[0] >> 8 | append[1] << 24;
- block3[3] = append[1] >> 8 | append[2] << 24;
- break;
+ const u32 div = block_len / 4;
- case 56:
- block3[2] = append[0];
- block3[3] = append[1];
- break;
+ switch (div)
+ {
+ case 0: block0[0] |= tmp0;
+ block0[1] = tmp1;
+ block0[2] = tmp2;
+ block0[3] = tmp3;
+ block1[0] = tmp4;
+ break;
+ case 1: block0[1] |= tmp0;
+ block0[2] = tmp1;
+ block0[3] = tmp2;
+ block1[0] = tmp3;
+ block1[1] = tmp4;
+ break;
+ case 2: block0[2] |= tmp0;
+ block0[3] = tmp1;
+ block1[0] = tmp2;
+ block1[1] = tmp3;
+ block1[2] = tmp4;
+ break;
+ case 3: block0[3] |= tmp0;
+ block1[0] = tmp1;
+ block1[1] = tmp2;
+ block1[2] = tmp3;
+ block1[3] = tmp4;
+ break;
+ case 4: block1[0] |= tmp0;
+ block1[1] = tmp1;
+ block1[2] = tmp2;
+ block1[3] = tmp3;
+ block2[0] = tmp4;
+ break;
+ case 5: block1[1] |= tmp0;
+ block1[2] = tmp1;
+ block1[3] = tmp2;
+ block2[0] = tmp3;
+ block2[1] = tmp4;
+ break;
+ case 6: block1[2] |= tmp0;
+ block1[3] = tmp1;
+ block2[0] = tmp2;
+ block2[1] = tmp3;
+ block2[2] = tmp4;
+ break;
+ case 7: block1[3] |= tmp0;
+ block2[0] = tmp1;
+ block2[1] = tmp2;
+ block2[2] = tmp3;
+ block2[3] = tmp4;
+ break;
+ case 8: block2[0] |= tmp0;
+ block2[1] = tmp1;
+ block2[2] = tmp2;
+ block2[3] = tmp3;
+ block3[0] = tmp4;
+ break;
+ case 9: block2[1] |= tmp0;
+ block2[2] = tmp1;
+ block2[3] = tmp2;
+ block3[0] = tmp3;
+ block3[1] = tmp4;
+ break;
}
}
-static void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[2])
+void memcat8 (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append[2])
{
- switch (block_len)
- {
- case 0:
- block0[0] = append[0];
- block0[1] = append[1];
- break;
-
- case 1:
- block0[0] = block0[0] | append[0] << 8;
- block0[1] = append[0] >> 24 | append[1] << 8;
- block0[2] = append[1] >> 24;
- break;
-
- case 2:
- block0[0] = block0[0] | append[0] << 16;
- block0[1] = append[0] >> 16 | append[1] << 16;
- block0[2] = append[1] >> 16;
- break;
-
- case 3:
- block0[0] = block0[0] | append[0] << 24;
- block0[1] = append[0] >> 8 | append[1] << 24;
- block0[2] = append[1] >> 8;
- break;
-
- case 4:
- block0[1] = append[0];
- block0[2] = append[1];
- break;
-
- case 5:
- block0[1] = block0[1] | append[0] << 8;
- block0[2] = append[0] >> 24 | append[1] << 8;
- block0[3] = append[1] >> 24;
- break;
-
- case 6:
- block0[1] = block0[1] | append[0] << 16;
- block0[2] = append[0] >> 16 | append[1] << 16;
- block0[3] = append[1] >> 16;
- break;
-
- case 7:
- block0[1] = block0[1] | append[0] << 24;
- block0[2] = append[0] >> 8 | append[1] << 24;
- block0[3] = append[1] >> 8;
- break;
-
- case 8:
- block0[2] = append[0];
- block0[3] = append[1];
- break;
-
- case 9:
- block0[2] = block0[2] | append[0] << 8;
- block0[3] = append[0] >> 24 | append[1] << 8;
- block1[0] = append[1] >> 24;
- break;
-
- case 10:
- block0[2] = block0[2] | append[0] << 16;
- block0[3] = append[0] >> 16 | append[1] << 16;
- block1[0] = append[1] >> 16;
- break;
-
- case 11:
- block0[2] = block0[2] | append[0] << 24;
- block0[3] = append[0] >> 8 | append[1] << 24;
- block1[0] = append[1] >> 8;
- break;
-
- case 12:
- block0[3] = append[0];
- block1[0] = append[1];
- break;
-
- case 13:
- block0[3] = block0[3] | append[0] << 8;
- block1[0] = append[0] >> 24 | append[1] << 8;
- block1[1] = append[1] >> 24;
- break;
-
- case 14:
- block0[3] = block0[3] | append[0] << 16;
- block1[0] = append[0] >> 16 | append[1] << 16;
- block1[1] = append[1] >> 16;
- break;
-
- case 15:
- block0[3] = block0[3] | append[0] << 24;
- block1[0] = append[0] >> 8 | append[1] << 24;
- block1[1] = append[1] >> 8;
- break;
-
- case 16:
- block1[0] = append[0];
- block1[1] = append[1];
- break;
-
- case 17:
- block1[0] = block1[0] | append[0] << 8;
- block1[1] = append[0] >> 24 | append[1] << 8;
- block1[2] = append[1] >> 24;
- break;
-
- case 18:
- block1[0] = block1[0] | append[0] << 16;
- block1[1] = append[0] >> 16 | append[1] << 16;
- block1[2] = append[1] >> 16;
- break;
-
- case 19:
- block1[0] = block1[0] | append[0] << 24;
- block1[1] = append[0] >> 8 | append[1] << 24;
- block1[2] = append[1] >> 8;
- break;
-
- case 20:
- block1[1] = append[0];
- block1[2] = append[1];
- break;
-
- case 21:
- block1[1] = block1[1] | append[0] << 8;
- block1[2] = append[0] >> 24 | append[1] << 8;
- block1[3] = append[1] >> 24;
- break;
-
- case 22:
- block1[1] = block1[1] | append[0] << 16;
- block1[2] = append[0] >> 16 | append[1] << 16;
- block1[3] = append[1] >> 16;
- break;
-
- case 23:
- block1[1] = block1[1] | append[0] << 24;
- block1[2] = append[0] >> 8 | append[1] << 24;
- block1[3] = append[1] >> 8;
- break;
-
- case 24:
- block1[2] = append[0];
- block1[3] = append[1];
- break;
-
- case 25:
- block1[2] = block1[2] | append[0] << 8;
- block1[3] = append[0] >> 24 | append[1] << 8;
- block2[0] = append[1] >> 24;
- break;
-
- case 26:
- block1[2] = block1[2] | append[0] << 16;
- block1[3] = append[0] >> 16 | append[1] << 16;
- block2[0] = append[1] >> 16;
- break;
-
- case 27:
- block1[2] = block1[2] | append[0] << 24;
- block1[3] = append[0] >> 8 | append[1] << 24;
- block2[0] = append[1] >> 8;
- break;
-
- case 28:
- block1[3] = append[0];
- block2[0] = append[1];
- break;
-
- case 29:
- block1[3] = block1[3] | append[0] << 8;
- block2[0] = append[0] >> 24 | append[1] << 8;
- block2[1] = append[1] >> 24;
- break;
-
- case 30:
- block1[3] = block1[3] | append[0] << 16;
- block2[0] = append[0] >> 16 | append[1] << 16;
- block2[1] = append[1] >> 16;
- break;
-
- case 31:
- block1[3] = block1[3] | append[0] << 24;
- block2[0] = append[0] >> 8 | append[1] << 24;
- block2[1] = append[1] >> 8;
- break;
-
- case 32:
- block2[0] = append[0];
- block2[1] = append[1];
- break;
-
- case 33:
- block2[0] = block2[0] | append[0] << 8;
- block2[1] = append[0] >> 24 | append[1] << 8;
- block2[2] = append[1] >> 24;
- break;
-
- case 34:
- block2[0] = block2[0] | append[0] << 16;
- block2[1] = append[0] >> 16 | append[1] << 16;
- block2[2] = append[1] >> 16;
- break;
-
- case 35:
- block2[0] = block2[0] | append[0] << 24;
- block2[1] = append[0] >> 8 | append[1] << 24;
- block2[2] = append[1] >> 8;
- break;
-
- case 36:
- block2[1] = append[0];
- block2[2] = append[1];
- break;
-
- case 37:
- block2[1] = block2[1] | append[0] << 8;
- block2[2] = append[0] >> 24 | append[1] << 8;
- block2[3] = append[1] >> 24;
- break;
-
- case 38:
- block2[1] = block2[1] | append[0] << 16;
- block2[2] = append[0] >> 16 | append[1] << 16;
- block2[3] = append[1] >> 16;
- break;
-
- case 39:
- block2[1] = block2[1] | append[0] << 24;
- block2[2] = append[0] >> 8 | append[1] << 24;
- block2[3] = append[1] >> 8;
- break;
-
- case 40:
- block2[2] = append[0];
- block2[3] = append[1];
- break;
-
- case 41:
- block2[2] = block2[2] | append[0] << 8;
- block2[3] = append[0] >> 24 | append[1] << 8;
- block3[0] = append[1] >> 24;
- break;
-
- case 42:
- block2[2] = block2[2] | append[0] << 16;
- block2[3] = append[0] >> 16 | append[1] << 16;
- block3[0] = append[1] >> 16;
- break;
-
- case 43:
- block2[2] = block2[2] | append[0] << 24;
- block2[3] = append[0] >> 8 | append[1] << 24;
- block3[0] = append[1] >> 8;
- break;
+ u32 tmp0;
+ u32 tmp1;
+ u32 tmp2;
- case 44:
- block2[3] = append[0];
- block3[0] = append[1];
- break;
+ #if defined IS_AMD || defined IS_GENERIC
- case 45:
- block2[3] = block2[3] | append[0] << 8;
- block3[0] = append[0] >> 24 | append[1] << 8;
- block3[1] = append[1] >> 24;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 46:
- block2[3] = block2[3] | append[0] << 16;
- block3[0] = append[0] >> 16 | append[1] << 16;
- block3[1] = append[1] >> 16;
- break;
+ tmp0 = amd_bytealign (append[0], 0, offset_minus_4);
+ tmp1 = amd_bytealign (append[1], append[0], offset_minus_4);
+ tmp2 = amd_bytealign ( 0, append[1], offset_minus_4);
- case 47:
- block2[3] = block2[3] | append[0] << 24;
- block3[0] = append[0] >> 8 | append[1] << 24;
- block3[1] = append[1] >> 8;
- break;
+ const u32 mod = block_len & 3;
- case 48:
- block3[0] = append[0];
- block3[1] = append[1];
- break;
+ if (mod == 0)
+ {
+ tmp0 = tmp1;
+ tmp1 = tmp2;
+ tmp2 = 0;
+ }
- case 49:
- block3[0] = block3[0] | append[0] << 8;
- block3[1] = append[0] >> 24 | append[1] << 8;
- block3[2] = append[1] >> 24;
- break;
+ #endif
- case 50:
- block3[0] = block3[0] | append[0] << 16;
- block3[1] = append[0] >> 16 | append[1] << 16;
- block3[2] = append[1] >> 16;
- break;
+ #ifdef IS_NV
- case 51:
- block3[0] = block3[0] | append[0] << 24;
- block3[1] = append[0] >> 8 | append[1] << 24;
- block3[2] = append[1] >> 8;
- break;
+ const int offset_minus_4 = 4 - (block_len & 3);
- case 52:
- block3[1] = append[0];
- block3[2] = append[1];
- break;
+ const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
- case 53:
- block3[1] = block3[1] | append[0] << 8;
- block3[2] = append[0] >> 24 | append[1] << 8;
- block3[3] = append[1] >> 24;
- break;
+ tmp0 = __byte_perm ( 0, append[0], selector);
+ tmp1 = __byte_perm (append[0], append[1], selector);
+ tmp2 = __byte_perm (append[1], 0, selector);
- case 54:
- block3[1] = block3[1] | append[0] << 16;
- block3[2] = append[0] >> 16 | append[1] << 16;
- block3[3] = append[1] >> 16;
- break;
+ #endif
- case 55:
- block3[1] = block3[1] | append[0] << 24;
- block3[2] = append[0] >> 8 | append[1] << 24;
- block3[3] = append[1] >> 8;
- break;
+ const u32 div = block_len / 4;
- case 56:
- block3[2] = append[0];
- block3[3] = append[1];
- break;
+ switch (div)
+ {
+ case 0: block0[0] |= tmp0;
+ block0[1] = tmp1;
+ block0[2] = tmp2;
+ break;
+ case 1: block0[1] |= tmp0;
+ block0[2] = tmp1;
+ block0[3] = tmp2;
+ break;
+ case 2: block0[2] |= tmp0;
+ block0[3] = tmp1;
+ block1[0] = tmp2;
+ break;
+ case 3: block0[3] |= tmp0;
+ block1[0] = tmp1;
+ block1[1] = tmp2;
+ break;
+ case 4: block1[0] |= tmp0;
+ block1[1] = tmp1;
+ block1[2] = tmp2;
+ break;
+ case 5: block1[1] |= tmp0;
+ block1[2] = tmp1;
+ block1[3] = tmp2;
+ break;
+ case 6: block1[2] |= tmp0;
+ block1[3] = tmp1;
+ block2[0] = tmp2;
+ break;
+ case 7: block1[3] |= tmp0;
+ block2[0] = tmp1;
+ block2[1] = tmp2;
+ break;
+ case 8: block2[0] |= tmp0;
+ block2[1] = tmp1;
+ block2[2] = tmp2;
+ break;
+ case 9: block2[1] |= tmp0;
+ block2[2] = tmp1;
+ block2[3] = tmp2;
+ break;
+ case 10: block2[2] |= tmp0;
+ block2[3] = tmp1;
+ block3[0] = tmp2;
+ break;
+ case 11: block2[3] |= tmp0;
+ block3[0] = tmp1;
+ block3[1] = tmp2;
+ break;
}
}
-static void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
+void append_1st (u32 block0[4], u32 block1[4], u32 block2[4], u32 block3[4], const u32 block_len, const u32 append)
{
switch (block_len)
{
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06300_init (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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)
+__kernel void m06300_init (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
/**
* base
tmps[gid].digest_buf[3] = digest[3];
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06300_loop (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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)
+__kernel void m06300_loop (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
/**
* base
tmps[gid].digest_buf[3] = digest[3];
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06300_comp (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global 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)
+__kernel void m06300_comp (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global md5crypt_tmp_t *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{
/**
* modifier
projects / hashcat.git / blobdiff