1
0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-22 22:58:30 +00:00

Cleanup -m 56xx kernels to latest standard

This commit is contained in:
Jens Steube 2016-04-12 21:21:35 +02:00
parent d6e5f7e829
commit afb394f152
3 changed files with 217 additions and 267 deletions

View File

@ -329,14 +329,12 @@ __kernel void m05600_m04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
*/
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];
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];
@ -368,6 +366,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
make_unicode (w1, w2_t, w3_t);
w3_t[2] = out_len * 8 * 2;
w3_t[3] = 0;
u32x digest[4];
@ -569,14 +568,12 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
*/
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];
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];
@ -620,6 +617,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t * rules_bu
make_unicode (w1, w2_t, w3_t);
w3_t[2] = out_len * 8 * 2;
w3_t[3] = 0;
u32x digest[4];

View File

@ -5,6 +5,8 @@
#define _NETNTLMV2_
#define NEW_SIMD_CODE
#include "include/constants.h"
#include "include/kernel_vendor.h"
@ -16,33 +18,31 @@
#include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c"
#include "OpenCL/common.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
static void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
static void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32 w0_t = w0[0];
u32 w1_t = w0[1];
u32 w2_t = w0[2];
u32 w3_t = w0[3];
u32 w4_t = w1[0];
u32 w5_t = w1[1];
u32 w6_t = w1[2];
u32 w7_t = w1[3];
u32 w8_t = w2[0];
u32 w9_t = w2[1];
u32 wa_t = w2[2];
u32 wb_t = w2[3];
u32 wc_t = w3[0];
u32 wd_t = w3[1];
u32 we_t = w3[2];
u32 wf_t = w3[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
@ -101,29 +101,29 @@ static void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], co
digest[3] += d;
}
static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32 w0_t = w0[0];
u32 w1_t = w0[1];
u32 w2_t = w0[2];
u32 w3_t = w0[3];
u32 w4_t = w1[0];
u32 w5_t = w1[1];
u32 w6_t = w1[2];
u32 w7_t = w1[3];
u32 w8_t = w2[0];
u32 w9_t = w2[1];
u32 wa_t = w2[2];
u32 wb_t = w2[3];
u32 wc_t = w3[0];
u32 wd_t = w3[1];
u32 we_t = w3[2];
u32 wf_t = w3[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
@ -199,7 +199,7 @@ static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], co
digest[3] += d;
}
static void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{
w0[0] = w0[0] ^ 0x36363636;
w0[1] = w0[1] ^ 0x36363636;
@ -250,7 +250,7 @@ static void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4
md5_transform (w0, w1, w2, w3, opad);
}
static void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
static void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
{
digest[0] = ipad[0];
digest[1] = ipad[1];
@ -326,125 +326,105 @@ __kernel void m05600_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* base
*/
u32 wordl0[4];
u32 pw_buf0[4];
u32 pw_buf1[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;
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;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
append_0x80_2x4 (wordl0, wordl1, pw_l_len);
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/**
* 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];
wordr0[1] = combs_buf[il_pos].i[1];
wordr0[2] = combs_buf[il_pos].i[2];
wordr0[3] = combs_buf[il_pos].i[3];
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
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];
wordr1[1] = combs_buf[il_pos].i[5];
wordr1[2] = combs_buf[il_pos].i[6];
wordr1[3] = combs_buf[il_pos].i[7];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
u32 wordr2[4];
wordr2[0] = 0;
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
u32 wordr3[4];
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
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[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
make_unicode (w0, w0_t, w1_t);
make_unicode (w1, w2_t, w3_t);
w3_t[2] = pw_len * 8 * 2;
w3_t[3] = 0;
u32 digest[4];
u32x digest[4];
digest[0] = MD4M_A;
digest[1] = MD4M_B;
@ -475,8 +455,8 @@ __kernel void m05600_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
u32 ipad[4];
u32 opad[4];
u32x ipad[4];
u32x opad[4];
hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
@ -589,12 +569,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
const u32 r0 = digest[0];
const u32 r1 = digest[3];
const u32 r2 = digest[2];
const u32 r3 = digest[1];
#include COMPARE_M
COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
}
}
@ -648,43 +623,20 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* base
*/
u32 wordl0[4];
u32 pw_buf0[4];
u32 pw_buf1[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;
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;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
append_0x80_2x4 (wordl0, wordl1, pw_l_len);
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/**
* digest
*/
@ -701,84 +653,87 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* 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];
wordr0[1] = combs_buf[il_pos].i[1];
wordr0[2] = combs_buf[il_pos].i[2];
wordr0[3] = combs_buf[il_pos].i[3];
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
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];
wordr1[1] = combs_buf[il_pos].i[5];
wordr1[2] = combs_buf[il_pos].i[6];
wordr1[3] = combs_buf[il_pos].i[7];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
u32 wordr2[4];
wordr2[0] = 0;
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
u32 wordr3[4];
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
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[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
make_unicode (w0, w0_t, w1_t);
make_unicode (w1, w2_t, w3_t);
w3_t[2] = pw_len * 8 * 2;
w3_t[3] = 0;
u32 digest[4];
u32x digest[4];
digest[0] = MD4M_A;
digest[1] = MD4M_B;
@ -809,8 +764,8 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
u32 ipad[4];
u32 opad[4];
u32x ipad[4];
u32x opad[4];
hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
@ -923,12 +878,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
const u32 r0 = digest[0];
const u32 r1 = digest[3];
const u32 r2 = digest[2];
const u32 r3 = digest[1];
#include COMPARE_S
COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
}
}

View File

@ -317,6 +317,10 @@ static void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x w0lr = w0l | w0r;
/**
* pads
*/
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
@ -326,17 +330,14 @@ static void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w0_t[1] = w0[1];
w0_t[2] = w0[2];
w0_t[3] = w0[3];
w1_t[0] = w1[0];
w1_t[1] = w1[1];
w1_t[2] = w1[2];
w1_t[3] = w1[3];
w2_t[0] = w2[0];
w2_t[1] = w2[1];
w2_t[2] = w2[2];
w2_t[3] = w2[3];
w3_t[0] = w3[0];
w3_t[1] = w3[1];
w3_t[2] = w3[2];
@ -500,6 +501,16 @@ static void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
/**
* prepare
*/
const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
+ netntlm_bufs[salt_pos].domain_len;
const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
+ netntlm_bufs[salt_pos].clichall_len;
/**
* digest
*/
@ -512,16 +523,6 @@ static void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/**
* prepare
*/
const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
+ netntlm_bufs[salt_pos].domain_len;
const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
+ netntlm_bufs[salt_pos].clichall_len;
/**
* loop
*/
@ -534,6 +535,10 @@ static void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x w0lr = w0l | w0r;
/**
* pads
*/
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
@ -543,17 +548,14 @@ static void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w0_t[1] = w0[1];
w0_t[2] = w0[2];
w0_t[3] = w0[3];
w1_t[0] = w1[0];
w1_t[1] = w1[1];
w1_t[2] = w1[2];
w1_t[3] = w1[3];
w2_t[0] = w2[0];
w2_t[1] = w2[1];
w2_t[2] = w2[2];
w2_t[3] = w2[3];
w3_t[0] = w3[0];
w3_t[1] = w3[1];
w3_t[2] = w3[2];