From 4ec867f9bb65cec9ad6e622a3a7a6e897a259930 Mon Sep 17 00:00:00 2001
From: Jens Steube <jens.steube@gmail.com>
Date: Thu, 3 Mar 2016 16:33:39 +0100
Subject: [PATCH] Cleanup -m 1430 kernels to latest standard

---
 OpenCL/m01430_a0.cl | 329 ++++++++++++++++++++++----------------------
 OpenCL/m01430_a1.cl |   4 +-
 OpenCL/m01430_a3.cl |  29 ++--
 3 files changed, 175 insertions(+), 187 deletions(-)

diff --git a/OpenCL/m01430_a0.cl b/OpenCL/m01430_a0.cl
index a2247aa..905beed 100644
--- a/OpenCL/m01430_a0.cl
+++ b/OpenCL/m01430_a0.cl
@@ -39,18 +39,16 @@ __kernel void m01430_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   if (gid >= gid_max) return;
 
   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];
+  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;
 
@@ -59,18 +57,26 @@ __kernel void m01430_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   u32 salt_buf0[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-
   u32 salt_buf1[4];
-
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+
+  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
+  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
+  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
+  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -87,89 +93,78 @@ __kernel void m01430_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x out_len2 = out_len * 2;
+
     /**
      * append salt
      */
 
     u32x s0[4];
+    u32x s1[4];
+    u32x s2[4];
+    u32x s3[4];
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, (out_len * 2));
-
-    const u32x out_salt_len = (out_len * 2) + salt_len;
-
-    u32x w0_t2[4];
-    u32x w1_t2[4];
-    u32x w2_t2[4];
-    u32x w3_t2[4];
-
-    make_unicode (w0, w0_t2, w1_t2);
-    make_unicode (w1, w2_t2, w3_t2);
-
-    w0_t2[0] |= s0[0];
-    w0_t2[1] |= s0[1];
-    w0_t2[2] |= s0[2];
-    w0_t2[3] |= s0[3];
-    w1_t2[0] |= s1[0];
-    w1_t2[1] |= s1[1];
-    w1_t2[2] |= s1[2];
-    w1_t2[3] |= s1[3];
-    w2_t2[0] |= s2[0];
-    w2_t2[1] |= s2[1];
-    w2_t2[2] |= s2[2];
-    w2_t2[3] |= s2[3];
-    w3_t2[0] |= s3[0];
-    w3_t2[1] |= s3[1];
-    w3_t2[2] |= s3[2];
-    w3_t2[3] |= s3[3];
-
-    append_0x80_4x4_VV (w0_t2, w1_t2, w2_t2, w3_t2, out_salt_len);
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
+
+    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len2);
+
+    const u32x pw_salt_len = out_len2 + salt_len;
+
+    w0[0] |= s0[0];
+    w0[1] |= s0[1];
+    w0[2] |= s0[2];
+    w0[3] |= s0[3];
+    w1[0] |= s1[0];
+    w1[1] |= s1[1];
+    w1[2] |= s1[2];
+    w1[3] |= s1[3];
+    w2[0] |= s2[0];
+    w2[1] |= s2[1];
+    w2[2] |= s2[2];
+    w2[3] |= s2[3];
+    w3[0] |= s3[0];
+    w3[1] |= s3[1];
+    w3[2] |= s3[2];
+    w3[3] |= s3[3];
 
     /**
      * sha256
      */
 
-    u32x w0_t = swap32 (w0_t2[0]);
-    u32x w1_t = swap32 (w0_t2[1]);
-    u32x w2_t = swap32 (w0_t2[2]);
-    u32x w3_t = swap32 (w0_t2[3]);
-    u32x w4_t = swap32 (w1_t2[0]);
-    u32x w5_t = swap32 (w1_t2[1]);
-    u32x w6_t = swap32 (w1_t2[2]);
-    u32x w7_t = swap32 (w1_t2[3]);
-    u32x w8_t = swap32 (w2_t2[0]);
-    u32x w9_t = swap32 (w2_t2[1]);
-    u32x wa_t = swap32 (w2_t2[2]);
-    u32x wb_t = swap32 (w2_t2[3]);
-    u32x wc_t = swap32 (w3_t2[0]);
-    u32x wd_t = swap32 (w3_t2[1]);
+    u32x w0_t = swap32 (w0[0]);
+    u32x w1_t = swap32 (w0[1]);
+    u32x w2_t = swap32 (w0[2]);
+    u32x w3_t = swap32 (w0[3]);
+    u32x w4_t = swap32 (w1[0]);
+    u32x w5_t = swap32 (w1[1]);
+    u32x w6_t = swap32 (w1[2]);
+    u32x w7_t = swap32 (w1[3]);
+    u32x w8_t = swap32 (w2[0]);
+    u32x w9_t = swap32 (w2[1]);
+    u32x wa_t = swap32 (w2[2]);
+    u32x wb_t = swap32 (w2[3]);
+    u32x wc_t = swap32 (w3[0]);
+    u32x wd_t = swap32 (w3[1]);
     u32x we_t = 0;
-    u32x wf_t = out_salt_len * 8;
+    u32x wf_t = pw_salt_len * 8;
 
     u32x a = SHA256M_A;
     u32x b = SHA256M_B;
@@ -277,18 +272,16 @@ __kernel void m01430_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   if (gid >= gid_max) return;
 
   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];
+  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;
 
@@ -297,18 +290,26 @@ __kernel void m01430_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   u32 salt_buf0[4];
-
-  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
-  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
-  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
-  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-
   u32 salt_buf1[4];
-
-  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
-  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
-  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
-  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+
+  salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
+  salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
+  salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
+  salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
+  salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
@@ -337,89 +338,78 @@ __kernel void m01430_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    make_unicode (w1, w2, w3);
+    make_unicode (w0, w0, w1);
+
+    const u32x out_len2 = out_len * 2;
+
     /**
      * append salt
      */
 
     u32x s0[4];
+    u32x s1[4];
+    u32x s2[4];
+    u32x s3[4];
 
     s0[0] = salt_buf0[0];
     s0[1] = salt_buf0[1];
     s0[2] = salt_buf0[2];
     s0[3] = salt_buf0[3];
-
-    u32x s1[4];
-
     s1[0] = salt_buf1[0];
     s1[1] = salt_buf1[1];
     s1[2] = salt_buf1[2];
     s1[3] = salt_buf1[3];
-
-    u32x s2[4];
-
-    s2[0] = 0;
-    s2[1] = 0;
-    s2[2] = 0;
-    s2[3] = 0;
-
-    u32x s3[4];
-
-    s3[0] = 0;
-    s3[1] = 0;
-    s3[2] = 0;
-    s3[3] = 0;
-
-    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, (out_len * 2));
-
-    const u32x out_salt_len = (out_len * 2) + salt_len;
-
-    u32x w0_t2[4];
-    u32x w1_t2[4];
-    u32x w2_t2[4];
-    u32x w3_t2[4];
-
-    make_unicode (w0, w0_t2, w1_t2);
-    make_unicode (w1, w2_t2, w3_t2);
-
-    w0_t2[0] |= s0[0];
-    w0_t2[1] |= s0[1];
-    w0_t2[2] |= s0[2];
-    w0_t2[3] |= s0[3];
-    w1_t2[0] |= s1[0];
-    w1_t2[1] |= s1[1];
-    w1_t2[2] |= s1[2];
-    w1_t2[3] |= s1[3];
-    w2_t2[0] |= s2[0];
-    w2_t2[1] |= s2[1];
-    w2_t2[2] |= s2[2];
-    w2_t2[3] |= s2[3];
-    w3_t2[0] |= s3[0];
-    w3_t2[1] |= s3[1];
-    w3_t2[2] |= s3[2];
-    w3_t2[3] |= s3[3];
-
-    append_0x80_4x4_VV (w0_t2, w1_t2, w2_t2, w3_t2, out_salt_len);
+    s2[0] = salt_buf2[0];
+    s2[1] = salt_buf2[1];
+    s2[2] = salt_buf2[2];
+    s2[3] = salt_buf2[3];
+    s3[0] = salt_buf3[0];
+    s3[1] = salt_buf3[1];
+    s3[2] = salt_buf3[2];
+    s3[3] = salt_buf3[3];
+
+    switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len2);
+
+    const u32x pw_salt_len = out_len2 + salt_len;
+
+    w0[0] |= s0[0];
+    w0[1] |= s0[1];
+    w0[2] |= s0[2];
+    w0[3] |= s0[3];
+    w1[0] |= s1[0];
+    w1[1] |= s1[1];
+    w1[2] |= s1[2];
+    w1[3] |= s1[3];
+    w2[0] |= s2[0];
+    w2[1] |= s2[1];
+    w2[2] |= s2[2];
+    w2[3] |= s2[3];
+    w3[0] |= s3[0];
+    w3[1] |= s3[1];
+    w3[2] |= s3[2];
+    w3[3] |= s3[3];
 
     /**
      * sha256
      */
 
-    u32x w0_t = swap32 (w0_t2[0]);
-    u32x w1_t = swap32 (w0_t2[1]);
-    u32x w2_t = swap32 (w0_t2[2]);
-    u32x w3_t = swap32 (w0_t2[3]);
-    u32x w4_t = swap32 (w1_t2[0]);
-    u32x w5_t = swap32 (w1_t2[1]);
-    u32x w6_t = swap32 (w1_t2[2]);
-    u32x w7_t = swap32 (w1_t2[3]);
-    u32x w8_t = swap32 (w2_t2[0]);
-    u32x w9_t = swap32 (w2_t2[1]);
-    u32x wa_t = swap32 (w2_t2[2]);
-    u32x wb_t = swap32 (w2_t2[3]);
-    u32x wc_t = swap32 (w3_t2[0]);
-    u32x wd_t = swap32 (w3_t2[1]);
+    u32x w0_t = swap32 (w0[0]);
+    u32x w1_t = swap32 (w0[1]);
+    u32x w2_t = swap32 (w0[2]);
+    u32x w3_t = swap32 (w0[3]);
+    u32x w4_t = swap32 (w1[0]);
+    u32x w5_t = swap32 (w1[1]);
+    u32x w6_t = swap32 (w1[2]);
+    u32x w7_t = swap32 (w1[3]);
+    u32x w8_t = swap32 (w2[0]);
+    u32x w9_t = swap32 (w2[1]);
+    u32x wa_t = swap32 (w2[2]);
+    u32x wb_t = swap32 (w2[3]);
+    u32x wc_t = swap32 (w3[0]);
+    u32x wd_t = swap32 (w3[1]);
     u32x we_t = 0;
-    u32x wf_t = out_salt_len * 8;
+    u32x wf_t = pw_salt_len * 8;
 
     u32x a = SHA256M_A;
     u32x b = SHA256M_B;
@@ -494,6 +484,9 @@ __kernel void m01430_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     wa_t = SHA256_EXPAND (w8_t, w3_t, wb_t, wa_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, g, h, a, b, c, d, e, f, wa_t, SHA256C3a);
     wb_t = SHA256_EXPAND (w9_t, w4_t, wc_t, wb_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, f, g, h, a, b, c, d, e, wb_t, SHA256C3b);
     wc_t = SHA256_EXPAND (wa_t, w5_t, wd_t, wc_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, e, f, g, h, a, b, c, d, wc_t, SHA256C3c);
+
+    if (MATCHES_NONE_VS (d, search[0])) continue;
+
     wd_t = SHA256_EXPAND (wb_t, w6_t, we_t, wd_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, d, e, f, g, h, a, b, c, wd_t, SHA256C3d);
     we_t = SHA256_EXPAND (wc_t, w7_t, wf_t, we_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, c, d, e, f, g, h, a, b, we_t, SHA256C3e);
     wf_t = SHA256_EXPAND (wd_t, w8_t, w0_t, wf_t); SHA256_STEP (SHA256_F0o, SHA256_F1o, b, c, d, e, f, g, h, a, wf_t, SHA256C3f);
diff --git a/OpenCL/m01430_a1.cl b/OpenCL/m01430_a1.cl
index 11728b3..897878f 100644
--- a/OpenCL/m01430_a1.cl
+++ b/OpenCL/m01430_a1.cl
@@ -204,7 +204,7 @@ __kernel void m01430_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[3] |= s3[3];
 
     /**
-     * sha1
+     * sha256
      */
 
     u32x w0_t = swap32 (w0[0]);
@@ -509,7 +509,7 @@ __kernel void m01430_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w3[3] |= s3[3];
 
     /**
-     * sha1
+     * sha256
      */
 
     u32x w0_t = swap32 (w0[0]);
diff --git a/OpenCL/m01430_a3.cl b/OpenCL/m01430_a3.cl
index f60e3a8..2a743a0 100644
--- a/OpenCL/m01430_a3.cl
+++ b/OpenCL/m01430_a3.cl
@@ -29,37 +29,32 @@ static void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
   const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
+
   /**
    * salt
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
 
   salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
   salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
   salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
   salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
-
-  u32 salt_buf1[4];
-
   salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
   salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
   salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
   salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
-
-  u32 salt_buf2[4];
-
   salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
-  salt_buf2[1] = 0;
-  salt_buf2[2] = 0;
-  salt_buf2[3] = 0;
-
-  u32 salt_buf3[4];
-
-  salt_buf3[0] = 0;
-  salt_buf3[1] = 0;
-  salt_buf3[2] = 0;
-  salt_buf3[3] = 0;
+  salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
+  salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
+  salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
+  salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
+  salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
+  salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
 
   switch_buffer_by_offset_le_S (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
@@ -86,7 +81,7 @@ static void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
 
   w[15] = pw_salt_len * 8;
 
- /**
+  /**
    * loop
    */
 
-- 
2.43.0