arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-23 00:28:11 +00:00
Code Issues Releases Wiki Activity

Cleanup -m 54xx kernels to latest standard

Browse Source
This commit is contained in:
Jens Steube 2016-04-12 16:22:07 +02:00
parent 29dead2039
commit ca9b34277f
3 changed files with 531 additions and 655 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

422
OpenCL/m05400_a0.cl
View File

@ -273,6 +273,24 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_len = pws[gid].pw_len;
/** /**
* salt * salt
*/ */
@ -280,26 +298,6 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
const u32 nr_len = ikepsk_bufs[salt_pos].nr_len; const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
const u32 msg_len = ikepsk_bufs[salt_pos].msg_len; const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
/**
* base
*/
u32 pw_buf0[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
pw_buf1[3] = pws[gid].i[ 7];
const u32 pw_len = pws[gid].pw_len;
/** /**
* loop * loop
*/ */
@ -317,122 +315,103 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
* pads * pads
*/ */
u32x w0_t[4]; w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0_t[0] = swap32 (w0[0]); w0[2] = swap32 (w0[2]);
w0_t[1] = swap32 (w0[1]); w0[3] = swap32 (w0[3]);
w0_t[2] = swap32 (w0[2]); w1[0] = swap32 (w1[0]);
w0_t[3] = swap32 (w0[3]); w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
u32x w1_t[4]; w1[3] = swap32 (w1[3]);
w1_t[0] = swap32 (w1[0]);
w1_t[1] = swap32 (w1[1]);
w1_t[2] = swap32 (w1[2]);
w1_t[3] = swap32 (w1[3]);
u32x w2_t[4];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32x w3_t[4];
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u32x ipad[5]; u32x ipad[5];
u32x opad[5]; u32x opad[5];
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w0_t[0] = w_s[ 0]; w0[0] = w_s[ 0];
w0_t[1] = w_s[ 1]; w0[1] = w_s[ 1];
w0_t[2] = w_s[ 2]; w0[2] = w_s[ 2];
w0_t[3] = w_s[ 3]; w0[3] = w_s[ 3];
w1_t[0] = w_s[ 4]; w1[0] = w_s[ 4];
w1_t[1] = w_s[ 5]; w1[1] = w_s[ 5];
w1_t[2] = w_s[ 6]; w1[2] = w_s[ 6];
w1_t[3] = w_s[ 7]; w1[3] = w_s[ 7];
w2_t[0] = w_s[ 8]; w2[0] = w_s[ 8];
w2_t[1] = w_s[ 9]; w2[1] = w_s[ 9];
w2_t[2] = w_s[10]; w2[2] = w_s[10];
w2_t[3] = w_s[11]; w2[3] = w_s[11];
w3_t[0] = w_s[12]; w3[0] = w_s[12];
w3_t[1] = w_s[13]; w3[1] = w_s[13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + nr_len) * 8; w3[3] = (64 + nr_len) * 8;
u32x digest[5]; u32x digest[5];
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 0; w3[3] = 0;
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
int left; int left;
int off; int off;
for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16) for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
{ {
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform (w0, w1, w2, w3, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + msg_len) * 8; w3[3] = (64 + msg_len) * 8;
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
COMPARE_M_SIMD (digest[3], digest[4], digest[2], digest[1]); COMPARE_M_SIMD (digest[3], digest[4], digest[2], digest[1]);
} }
@ -480,6 +459,24 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_len = pws[gid].pw_len;
/** /**
* salt * salt
*/ */
@ -487,26 +484,6 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
const u32 nr_len = ikepsk_bufs[salt_pos].nr_len; const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
const u32 msg_len = ikepsk_bufs[salt_pos].msg_len; const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
/**
* base
*/
u32 pw_buf0[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
pw_buf1[3] = pws[gid].i[ 7];
const u32 pw_len = pws[gid].pw_len;
/** /**
* digest * digest
*/ */
@ -536,122 +513,103 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
* pads * pads
*/ */
u32x w0_t[4]; w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0_t[0] = swap32 (w0[0]); w0[2] = swap32 (w0[2]);
w0_t[1] = swap32 (w0[1]); w0[3] = swap32 (w0[3]);
w0_t[2] = swap32 (w0[2]); w1[0] = swap32 (w1[0]);
w0_t[3] = swap32 (w0[3]); w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
u32x w1_t[4]; w1[3] = swap32 (w1[3]);
w1_t[0] = swap32 (w1[0]);
w1_t[1] = swap32 (w1[1]);
w1_t[2] = swap32 (w1[2]);
w1_t[3] = swap32 (w1[3]);
u32x w2_t[4];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32x w3_t[4];
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u32x ipad[5]; u32x ipad[5];
u32x opad[5]; u32x opad[5];
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w0_t[0] = w_s[ 0]; w0[0] = w_s[ 0];
w0_t[1] = w_s[ 1]; w0[1] = w_s[ 1];
w0_t[2] = w_s[ 2]; w0[2] = w_s[ 2];
w0_t[3] = w_s[ 3]; w0[3] = w_s[ 3];
w1_t[0] = w_s[ 4]; w1[0] = w_s[ 4];
w1_t[1] = w_s[ 5]; w1[1] = w_s[ 5];
w1_t[2] = w_s[ 6]; w1[2] = w_s[ 6];
w1_t[3] = w_s[ 7]; w1[3] = w_s[ 7];
w2_t[0] = w_s[ 8]; w2[0] = w_s[ 8];
w2_t[1] = w_s[ 9]; w2[1] = w_s[ 9];
w2_t[2] = w_s[10]; w2[2] = w_s[10];
w2_t[3] = w_s[11]; w2[3] = w_s[11];
w3_t[0] = w_s[12]; w3[0] = w_s[12];
w3_t[1] = w_s[13]; w3[1] = w_s[13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + nr_len) * 8; w3[3] = (64 + nr_len) * 8;
u32x digest[5]; u32x digest[5];
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 0; w3[3] = 0;
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
int left; int left;
int off; int off;
for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16) for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
{ {
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform (w0, w1, w2, w3, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + msg_len) * 8; w3[3] = (64 + msg_len) * 8;
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
COMPARE_S_SIMD (digest[3], digest[4], digest[2], digest[1]); COMPARE_S_SIMD (digest[3], digest[4], digest[2], digest[1]);
} }

716
OpenCL/m05400_a1.cl
View File

@ -5,6 +5,8 @@
#define _SHA1_ #define _SHA1_
#define NEW_SIMD_CODE
#include "include/constants.h" #include "include/constants.h"
#include "include/kernel_vendor.h" #include "include/kernel_vendor.h"
@ -16,34 +18,32 @@
#include "include/kernel_functions.c" #include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c" #include "OpenCL/types_ocl.c"
#include "OpenCL/common.c" #include "OpenCL/common.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c" static void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
#define COMPARE_M "OpenCL/check_multi_comp4.c"
static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
{ {
u32 A = digest[0]; u32x A = digest[0];
u32 B = digest[1]; u32x B = digest[1];
u32 C = digest[2]; u32x C = digest[2];
u32 D = digest[3]; u32x D = digest[3];
u32 E = digest[4]; u32x E = digest[4];
u32 w0_t = w0[0]; u32x w0_t = w0[0];
u32 w1_t = w0[1]; u32x w1_t = w0[1];
u32 w2_t = w0[2]; u32x w2_t = w0[2];
u32 w3_t = w0[3]; u32x w3_t = w0[3];
u32 w4_t = w1[0]; u32x w4_t = w1[0];
u32 w5_t = w1[1]; u32x w5_t = w1[1];
u32 w6_t = w1[2]; u32x w6_t = w1[2];
u32 w7_t = w1[3]; u32x w7_t = w1[3];
u32 w8_t = w2[0]; u32x w8_t = w2[0];
u32 w9_t = w2[1]; u32x w9_t = w2[1];
u32 wa_t = w2[2]; u32x wa_t = w2[2];
u32 wb_t = w2[3]; u32x wb_t = w2[3];
u32 wc_t = w3[0]; u32x wc_t = w3[0];
u32 wd_t = w3[1]; u32x wd_t = w3[1];
u32 we_t = w3[2]; u32x we_t = w3[2];
u32 wf_t = w3[3]; u32x wf_t = w3[3];
#undef K #undef K
#define K SHA1C00 #define K SHA1C00
@ -148,7 +148,7 @@ static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], c
digest[4] += E; digest[4] += E;
} }
static void hmac_sha1_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[5], u32 opad[5]) static void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
w0[0] = w0[0] ^ 0x36363636; w0[0] = w0[0] ^ 0x36363636;
w0[1] = w0[1] ^ 0x36363636; w0[1] = w0[1] ^ 0x36363636;
@ -201,7 +201,7 @@ static void hmac_sha1_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[
sha1_transform (w0, w1, w2, w3, opad); sha1_transform (w0, w1, w2, w3, opad);
} }
static void hmac_sha1_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[5], u32 opad[5], u32 digest[5]) static void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
{ {
digest[0] = ipad[0]; digest[0] = ipad[0];
digest[1] = ipad[1]; digest[1] = ipad[1];
@ -255,7 +255,7 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
for (u32 i = lid; i < 16; i += lsz) for (u32 i = lid; i < 16; i += lsz)
{ {
w_s[i] = swap32 (ikepsk_bufs[salt_pos].nr_buf[i]); w_s[i] = swap32_S (ikepsk_bufs[salt_pos].nr_buf[i]);
} }
barrier (CLK_LOCAL_MEM_FENCE); barrier (CLK_LOCAL_MEM_FENCE);
@ -264,13 +264,31 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
for (u32 i = lid; i < 128; i += lsz) for (u32 i = lid; i < 128; i += lsz)
{ {
s_msg_buf[i] = swap32 (ikepsk_bufs[salt_pos].msg_buf[i]); s_msg_buf[i] = swap32_S (ikepsk_bufs[salt_pos].msg_buf[i]);
} }
barrier (CLK_LOCAL_MEM_FENCE); barrier (CLK_LOCAL_MEM_FENCE);
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_l_len = pws[gid].pw_len;
/** /**
* salt * salt
*/ */
@ -278,111 +296,74 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 nr_len = ikepsk_bufs[salt_pos].nr_len; const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
const u32 msg_len = ikepsk_bufs[salt_pos].msg_len; const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
/**
* base
*/
u32 wordl0[4];
wordl0[0] = pws[gid].i[ 0];
wordl0[1] = pws[gid].i[ 1];
wordl0[2] = pws[gid].i[ 2];
wordl0[3] = pws[gid].i[ 3];
u32 wordl1[4];
wordl1[0] = pws[gid].i[ 4];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0]; w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1]; w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2]; w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3]; w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0]; w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1]; w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2]; w3[2] = wordl3[2] | wordr3[2];
@ -392,129 +373,113 @@ __kernel void m05400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* pads * pads
*/ */
u32 w0_t[4]; w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0[2] = swap32 (w0[2]);
w0[3] = swap32 (w0[3]);
w1[0] = swap32 (w1[0]);
w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
w1[3] = swap32 (w1[3]);
w2[0] = swap32 (w2[0]);
w2[1] = swap32 (w2[1]);
w2[2] = swap32 (w2[2]);
w2[3] = swap32 (w2[3]);
w3[0] = swap32 (w3[0]);
w3[1] = swap32 (w3[1]);
w3[2] = swap32 (w3[2]);
w3[3] = swap32 (w3[3]);
w0_t[0] = swap32 (w0[0]); u32x ipad[5];
w0_t[1] = swap32 (w0[1]); u32x opad[5];
w0_t[2] = swap32 (w0[2]);
w0_t[3] = swap32 (w0[3]);
u32 w1_t[4]; hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w1_t[0] = swap32 (w1[0]); w0[0] = w_s[ 0];
w1_t[1] = swap32 (w1[1]); w0[1] = w_s[ 1];
w1_t[2] = swap32 (w1[2]); w0[2] = w_s[ 2];
w1_t[3] = swap32 (w1[3]); w0[3] = w_s[ 3];
w1[0] = w_s[ 4];
w1[1] = w_s[ 5];
w1[2] = w_s[ 6];
w1[3] = w_s[ 7];
w2[0] = w_s[ 8];
w2[1] = w_s[ 9];
w2[2] = w_s[10];
w2[3] = w_s[11];
w3[0] = w_s[12];
w3[1] = w_s[13];
w3[2] = 0;
w3[3] = (64 + nr_len) * 8;
u32 w2_t[4]; u32x digest[5];
w2_t[0] = 0; hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32 w3_t[4]; w0[0] = digest[0];
w0[1] = digest[1];
w0[2] = digest[2];
w0[3] = digest[3];
w1[0] = digest[4];
w1[1] = 0;
w1[2] = 0;
w1[3] = 0;
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
w3_t[0] = 0; hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u32 ipad[5];
u32 opad[5];
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
w0_t[0] = w_s[ 0];
w0_t[1] = w_s[ 1];
w0_t[2] = w_s[ 2];
w0_t[3] = w_s[ 3];
w1_t[0] = w_s[ 4];
w1_t[1] = w_s[ 5];
w1_t[2] = w_s[ 6];
w1_t[3] = w_s[ 7];
w2_t[0] = w_s[ 8];
w2_t[1] = w_s[ 9];
w2_t[2] = w_s[10];
w2_t[3] = w_s[11];
w3_t[0] = w_s[12];
w3_t[1] = w_s[13];
w3_t[2] = 0;
w3_t[3] = (64 + nr_len) * 8;
u32 digest[5];
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
w0_t[0] = digest[0];
w0_t[1] = digest[1];
w0_t[2] = digest[2];
w0_t[3] = digest[3];
w1_t[0] = digest[4];
w1_t[1] = 0;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
int left; int left;
int off; int off;
for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16) for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
{ {
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform (w0, w1, w2, w3, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + msg_len) * 8; w3[3] = (64 + msg_len) * 8;
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
const u32 r0 = digest[3]; COMPARE_M_SIMD (digest[3], digest[4], digest[2], digest[1]);
const u32 r1 = digest[4];
const u32 r2 = digest[2];
const u32 r3 = digest[1];
#include COMPARE_M
} }
} }
@ -544,7 +509,7 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
for (u32 i = lid; i < 16; i += lsz) for (u32 i = lid; i < 16; i += lsz)
{ {
w_s[i] = swap32 (ikepsk_bufs[salt_pos].nr_buf[i]); w_s[i] = swap32_S (ikepsk_bufs[salt_pos].nr_buf[i]);
} }
barrier (CLK_LOCAL_MEM_FENCE); barrier (CLK_LOCAL_MEM_FENCE);
@ -553,13 +518,31 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
for (u32 i = lid; i < 128; i += lsz) for (u32 i = lid; i < 128; i += lsz)
{ {
s_msg_buf[i] = swap32 (ikepsk_bufs[salt_pos].msg_buf[i]); s_msg_buf[i] = swap32_S (ikepsk_bufs[salt_pos].msg_buf[i]);
} }
barrier (CLK_LOCAL_MEM_FENCE); barrier (CLK_LOCAL_MEM_FENCE);
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_l_len = pws[gid].pw_len;
/** /**
* salt * salt
*/ */
@ -567,45 +550,6 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 nr_len = ikepsk_bufs[salt_pos].nr_len; const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
const u32 msg_len = ikepsk_bufs[salt_pos].msg_len; const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
/**
* base
*/
u32 wordl0[4];
wordl0[0] = pws[gid].i[ 0];
wordl0[1] = pws[gid].i[ 1];
wordl0[2] = pws[gid].i[ 2];
wordl0[3] = pws[gid].i[ 3];
u32 wordl1[4];
wordl1[0] = pws[gid].i[ 4];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* digest * digest
*/ */
@ -622,68 +566,70 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0]; w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1]; w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2]; w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3]; w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0]; w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1]; w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2]; w3[2] = wordl3[2] | wordr3[2];
@ -693,129 +639,113 @@ __kernel void m05400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* pads * pads
*/ */
u32 w0_t[4]; w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0[2] = swap32 (w0[2]);
w0[3] = swap32 (w0[3]);
w1[0] = swap32 (w1[0]);
w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
w1[3] = swap32 (w1[3]);
w2[0] = swap32 (w2[0]);
w2[1] = swap32 (w2[1]);
w2[2] = swap32 (w2[2]);
w2[3] = swap32 (w2[3]);
w3[0] = swap32 (w3[0]);
w3[1] = swap32 (w3[1]);
w3[2] = swap32 (w3[2]);
w3[3] = swap32 (w3[3]);
w0_t[0] = swap32 (w0[0]); u32x ipad[5];
w0_t[1] = swap32 (w0[1]); u32x opad[5];
w0_t[2] = swap32 (w0[2]);
w0_t[3] = swap32 (w0[3]);
u32 w1_t[4]; hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w1_t[0] = swap32 (w1[0]); w0[0] = w_s[ 0];
w1_t[1] = swap32 (w1[1]); w0[1] = w_s[ 1];
w1_t[2] = swap32 (w1[2]); w0[2] = w_s[ 2];
w1_t[3] = swap32 (w1[3]); w0[3] = w_s[ 3];
w1[0] = w_s[ 4];
w1[1] = w_s[ 5];
w1[2] = w_s[ 6];
w1[3] = w_s[ 7];
w2[0] = w_s[ 8];
w2[1] = w_s[ 9];
w2[2] = w_s[10];
w2[3] = w_s[11];
w3[0] = w_s[12];
w3[1] = w_s[13];
w3[2] = 0;
w3[3] = (64 + nr_len) * 8;
u32 w2_t[4]; u32x digest[5];
w2_t[0] = 0; hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32 w3_t[4]; w0[0] = digest[0];
w0[1] = digest[1];
w0[2] = digest[2];
w0[3] = digest[3];
w1[0] = digest[4];
w1[1] = 0;
w1[2] = 0;
w1[3] = 0;
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
w3_t[0] = 0; hmac_sha1_pad (w0, w1, w2, w3, ipad, opad);
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u32 ipad[5];
u32 opad[5];
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
w0_t[0] = w_s[ 0];
w0_t[1] = w_s[ 1];
w0_t[2] = w_s[ 2];
w0_t[3] = w_s[ 3];
w1_t[0] = w_s[ 4];
w1_t[1] = w_s[ 5];
w1_t[2] = w_s[ 6];
w1_t[3] = w_s[ 7];
w2_t[0] = w_s[ 8];
w2_t[1] = w_s[ 9];
w2_t[2] = w_s[10];
w2_t[3] = w_s[11];
w3_t[0] = w_s[12];
w3_t[1] = w_s[13];
w3_t[2] = 0;
w3_t[3] = (64 + nr_len) * 8;
u32 digest[5];
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
w0_t[0] = digest[0];
w0_t[1] = digest[1];
w0_t[2] = digest[2];
w0_t[3] = digest[3];
w1_t[0] = digest[4];
w1_t[1] = 0;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
int left; int left;
int off; int off;
for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16) for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
{ {
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform (w0, w1, w2, w3, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
w0_t[1] = s_msg_buf[off + 1]; w0[1] = s_msg_buf[off + 1];
w0_t[2] = s_msg_buf[off + 2]; w0[2] = s_msg_buf[off + 2];
w0_t[3] = s_msg_buf[off + 3]; w0[3] = s_msg_buf[off + 3];
w1_t[0] = s_msg_buf[off + 4]; w1[0] = s_msg_buf[off + 4];
w1_t[1] = s_msg_buf[off + 5]; w1[1] = s_msg_buf[off + 5];
w1_t[2] = s_msg_buf[off + 6]; w1[2] = s_msg_buf[off + 6];
w1_t[3] = s_msg_buf[off + 7]; w1[3] = s_msg_buf[off + 7];
w2_t[0] = s_msg_buf[off + 8]; w2[0] = s_msg_buf[off + 8];
w2_t[1] = s_msg_buf[off + 9]; w2[1] = s_msg_buf[off + 9];
w2_t[2] = s_msg_buf[off + 10]; w2[2] = s_msg_buf[off + 10];
w2_t[3] = s_msg_buf[off + 11]; w2[3] = s_msg_buf[off + 11];
w3_t[0] = s_msg_buf[off + 12]; w3[0] = s_msg_buf[off + 12];
w3_t[1] = s_msg_buf[off + 13]; w3[1] = s_msg_buf[off + 13];
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (64 + msg_len) * 8; w3[3] = (64 + msg_len) * 8;
hmac_sha1_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
const u32 r0 = digest[3]; COMPARE_S_SIMD (digest[3], digest[4], digest[2], digest[1]);
const u32 r1 = digest[4];
const u32 r2 = digest[2];
const u32 r3 = digest[1];
#include COMPARE_S
} }
} }

48
OpenCL/m05400_a3.cl
View File

@ -267,37 +267,31 @@ static void m05400m (__local u32 *w_s, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4
const u32x w0lr = w0l | w0r; const u32x w0lr = w0l | w0r;
/**
* pads
*/
u32x w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = w0lr; w0_t[0] = w0lr;
w0_t[1] = w0[1]; w0_t[1] = w0[1];
w0_t[2] = w0[2]; w0_t[2] = w0[2];
w0_t[3] = w0[3]; w0_t[3] = w0[3];
u32x w1_t[4];
w1_t[0] = w1[0]; w1_t[0] = w1[0];
w1_t[1] = w1[1]; w1_t[1] = w1[1];
w1_t[2] = w1[2]; w1_t[2] = w1[2];
w1_t[3] = w1[3]; w1_t[3] = w1[3];
u32x w2_t[4];
w2_t[0] = w2[0]; w2_t[0] = w2[0];
w2_t[1] = w2[1]; w2_t[1] = w2[1];
w2_t[2] = w2[2]; w2_t[2] = w2[2];
w2_t[3] = w2[3]; w2_t[3] = w2[3];
u32x w3_t[4];
w3_t[0] = w3[0]; w3_t[0] = w3[0];
w3_t[1] = w3[1]; w3_t[1] = w3[1];
w3_t[2] = 0; w3_t[2] = w3[2];
w3_t[3] = 0; w3_t[3] = w3[3];
/**
* pads
*/
u32x ipad[5]; u32x ipad[5];
u32x opad[5]; u32x opad[5];
@ -432,37 +426,31 @@ static void m05400s (__local u32 *w_s, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4
const u32x w0lr = w0l | w0r; const u32x w0lr = w0l | w0r;
/**
* pads
*/
u32x w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = w0lr; w0_t[0] = w0lr;
w0_t[1] = w0[1]; w0_t[1] = w0[1];
w0_t[2] = w0[2]; w0_t[2] = w0[2];
w0_t[3] = w0[3]; w0_t[3] = w0[3];
u32x w1_t[4];
w1_t[0] = w1[0]; w1_t[0] = w1[0];
w1_t[1] = w1[1]; w1_t[1] = w1[1];
w1_t[2] = w1[2]; w1_t[2] = w1[2];
w1_t[3] = w1[3]; w1_t[3] = w1[3];
u32x w2_t[4];
w2_t[0] = w2[0]; w2_t[0] = w2[0];
w2_t[1] = w2[1]; w2_t[1] = w2[1];
w2_t[2] = w2[2]; w2_t[2] = w2[2];
w2_t[3] = w2[3]; w2_t[3] = w2[3];
u32x w3_t[4];
w3_t[0] = w3[0]; w3_t[0] = w3[0];
w3_t[1] = w3[1]; w3_t[1] = w3[1];
w3_t[2] = 0; w3_t[2] = w3[2];
w3_t[3] = 0; w3_t[3] = w3[3];
/**
* pads
*/
u32x ipad[5]; u32x ipad[5];
u32x opad[5]; u32x opad[5];