diff --git a/OpenCL/m07350_a1-optimized.cl b/OpenCL/m07350_a1-optimized.cl
new file mode 100644
index 000000000..058b6411d
--- /dev/null
+++ b/OpenCL/m07350_a1-optimized.cl
@@ -0,0 +1,644 @@
+/**
+ * Author......: See docs/credits.txt
+ * License.....: MIT
+ */
+
+#define NEW_SIMD_CODE
+
+#ifdef KERNEL_STATIC
+#include M2S(INCLUDE_PATH/inc_vendor.h)
+#include M2S(INCLUDE_PATH/inc_types.h)
+#include M2S(INCLUDE_PATH/inc_platform.cl)
+#include M2S(INCLUDE_PATH/inc_common.cl)
+#include M2S(INCLUDE_PATH/inc_simd.cl)
+#include M2S(INCLUDE_PATH/inc_hash_md5.cl)
+#endif
+
+DECLSPEC void hmac_md5_pad (PRIVATE_AS u32x *w0, PRIVATE_AS u32x *w1, PRIVATE_AS u32x *w2, PRIVATE_AS u32x *w3, PRIVATE_AS u32x *ipad, PRIVATE_AS u32x *opad)
+{
+  w0[0] = w0[0] ^ 0x36363636;
+  w0[1] = w0[1] ^ 0x36363636;
+  w0[2] = w0[2] ^ 0x36363636;
+  w0[3] = w0[3] ^ 0x36363636;
+  w1[0] = w1[0] ^ 0x36363636;
+  w1[1] = w1[1] ^ 0x36363636;
+  w1[2] = w1[2] ^ 0x36363636;
+  w1[3] = w1[3] ^ 0x36363636;
+  w2[0] = w2[0] ^ 0x36363636;
+  w2[1] = w2[1] ^ 0x36363636;
+  w2[2] = w2[2] ^ 0x36363636;
+  w2[3] = w2[3] ^ 0x36363636;
+  w3[0] = w3[0] ^ 0x36363636;
+  w3[1] = w3[1] ^ 0x36363636;
+  w3[2] = w3[2] ^ 0x36363636;
+  w3[3] = w3[3] ^ 0x36363636;
+
+  ipad[0] = MD5M_A;
+  ipad[1] = MD5M_B;
+  ipad[2] = MD5M_C;
+  ipad[3] = MD5M_D;
+
+  md5_transform_vector (w0, w1, w2, w3, ipad);
+
+  w0[0] = w0[0] ^ 0x6a6a6a6a;
+  w0[1] = w0[1] ^ 0x6a6a6a6a;
+  w0[2] = w0[2] ^ 0x6a6a6a6a;
+  w0[3] = w0[3] ^ 0x6a6a6a6a;
+  w1[0] = w1[0] ^ 0x6a6a6a6a;
+  w1[1] = w1[1] ^ 0x6a6a6a6a;
+  w1[2] = w1[2] ^ 0x6a6a6a6a;
+  w1[3] = w1[3] ^ 0x6a6a6a6a;
+  w2[0] = w2[0] ^ 0x6a6a6a6a;
+  w2[1] = w2[1] ^ 0x6a6a6a6a;
+  w2[2] = w2[2] ^ 0x6a6a6a6a;
+  w2[3] = w2[3] ^ 0x6a6a6a6a;
+  w3[0] = w3[0] ^ 0x6a6a6a6a;
+  w3[1] = w3[1] ^ 0x6a6a6a6a;
+  w3[2] = w3[2] ^ 0x6a6a6a6a;
+  w3[3] = w3[3] ^ 0x6a6a6a6a;
+
+  opad[0] = MD5M_A;
+  opad[1] = MD5M_B;
+  opad[2] = MD5M_C;
+  opad[3] = MD5M_D;
+
+  md5_transform_vector (w0, w1, w2, w3, opad);
+}
+
+/*
+DECLSPEC void hmac_md5_run (PRIVATE_AS u32x *w0, PRIVATE_AS u32x *w1, PRIVATE_AS u32x *w2, PRIVATE_AS u32x *w3, PRIVATE_AS u32x *ipad, PRIVATE_AS u32x *opad, PRIVATE_AS u32x *digest)
+{
+  digest[0] = ipad[0];
+  digest[1] = ipad[1];
+  digest[2] = ipad[2];
+  digest[3] = ipad[3];
+
+  md5_transform_vector (w0, w1, w2, w3, digest);
+
+  w0[0] = digest[0];
+  w0[1] = digest[1];
+  w0[2] = digest[2];
+  w0[3] = digest[3];
+  w1[0] = 0x80;
+  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] = (64 + 16) * 8;
+  w3[3] = 0;
+
+  digest[0] = opad[0];
+  digest[1] = opad[1];
+  digest[2] = opad[2];
+  digest[3] = opad[3];
+
+  md5_transform_vector (w0, w1, w2, w3, digest);
+}
+*/
+
+KERNEL_FQ void m07350_m04 (KERN_ATTR_BASIC ())
+{
+  /**
+   * modifier
+   */
+
+  const u64 lid = get_local_id (0);
+
+  /**
+   * base
+   */
+
+  const u64 gid = get_global_id (0);
+
+  if (gid >= GID_CNT) return;
+
+  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 & 63;
+
+  /**
+   * salt
+   */
+
+  u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+  u32 salt_buf4[4];
+  u32 salt_buf5[4];
+  u32 salt_buf6[4];
+  u32 salt_buf7[4];
+
+  salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0];
+  salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1];
+  salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2];
+  salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3];
+  salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10];
+  salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11];
+  salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12];
+  salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13];
+  salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14];
+  salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15];
+  salt_buf4[0] = salt_bufs[SALT_POS_HOST].salt_buf[16];
+  salt_buf4[1] = salt_bufs[SALT_POS_HOST].salt_buf[17];
+  salt_buf4[2] = salt_bufs[SALT_POS_HOST].salt_buf[18];
+  salt_buf4[3] = salt_bufs[SALT_POS_HOST].salt_buf[19];
+  salt_buf5[0] = salt_bufs[SALT_POS_HOST].salt_buf[20];
+  salt_buf5[1] = salt_bufs[SALT_POS_HOST].salt_buf[21];
+  salt_buf5[2] = salt_bufs[SALT_POS_HOST].salt_buf[22];
+  salt_buf5[3] = salt_bufs[SALT_POS_HOST].salt_buf[23];
+  salt_buf6[0] = salt_bufs[SALT_POS_HOST].salt_buf[24];
+  salt_buf6[1] = salt_bufs[SALT_POS_HOST].salt_buf[25];
+  salt_buf6[2] = salt_bufs[SALT_POS_HOST].salt_buf[26];
+  salt_buf6[3] = salt_bufs[SALT_POS_HOST].salt_buf[27];
+  salt_buf7[0] = salt_bufs[SALT_POS_HOST].salt_buf[28];
+  salt_buf7[1] = salt_bufs[SALT_POS_HOST].salt_buf[29];
+  salt_buf7[2] = salt_bufs[SALT_POS_HOST].salt_buf[30];
+  salt_buf7[3] = salt_bufs[SALT_POS_HOST].salt_buf[31];
+
+  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
+
+  /**
+   * loop
+   */
+
+  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
+  {
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
+
+    const u32x pw_len = (pw_l_len + pw_r_len) & 63;
+
+    /**
+     * concat password candidate
+     */
+
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    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];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 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_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
+    }
+
+    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];
+    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];
+
+    /**
+     * pads
+     */
+
+    u32x ipad[4];
+    u32x opad[4];
+
+    hmac_md5_pad (w0, w1, w2, w3, ipad, opad);
+
+    //    hmac_md5_run (w0, w1, w2, w3, ipad, opad, digest);
+
+    u32x digest[4];
+
+    digest[0] = ipad[0];
+    digest[1] = ipad[1];
+    digest[2] = ipad[2];
+    digest[3] = ipad[3];
+
+    if (salt_len < 56)
+    {
+      w0[0] = salt_buf0[0];
+      w0[1] = salt_buf0[1];
+      w0[2] = salt_buf0[2];
+      w0[3] = salt_buf0[3];
+      w1[0] = salt_buf1[0];
+      w1[1] = salt_buf1[1];
+      w1[2] = salt_buf1[2];
+      w1[3] = salt_buf1[3];
+      w2[0] = salt_buf2[0];
+      w2[1] = salt_buf2[1];
+      w2[2] = salt_buf2[2];
+      w2[3] = salt_buf2[3];
+      w3[0] = salt_buf3[0];
+      w3[1] = salt_buf3[1];
+      w3[2] = (64 + salt_len) * 8;
+      w3[3] = 0;
+    }
+    else
+    {
+      w0[0] = salt_buf0[0];
+      w0[1] = salt_buf0[1];
+      w0[2] = salt_buf0[2];
+      w0[3] = salt_buf0[3];
+      w1[0] = salt_buf1[0];
+      w1[1] = salt_buf1[1];
+      w1[2] = salt_buf1[2];
+      w1[3] = salt_buf1[3];
+      w2[0] = salt_buf2[0];
+      w2[1] = salt_buf2[1];
+      w2[2] = salt_buf2[2];
+      w2[3] = salt_buf2[3];
+      w3[0] = salt_buf3[0];
+      w3[1] = salt_buf3[1];
+      w3[2] = salt_buf3[2];
+      w3[3] = salt_buf3[3];
+
+      md5_transform_vector (w0, w1, w2, w3, digest);
+
+      w0[0] = salt_buf4[0];
+      w0[1] = salt_buf4[1];
+      w0[2] = salt_buf4[2];
+      w0[3] = salt_buf4[3];
+      w1[0] = salt_buf5[0];
+      w1[1] = salt_buf5[1];
+      w1[2] = salt_buf5[2];
+      w1[3] = salt_buf5[3];
+      w2[0] = salt_buf6[0];
+      w2[1] = salt_buf6[1];
+      w2[2] = salt_buf6[2];
+      w2[3] = salt_buf6[3];
+      w3[0] = salt_buf7[0];
+      w3[1] = salt_buf7[1];
+      w3[2] = (64 + salt_len) * 8;
+      w3[3] = 0;
+    }
+
+    md5_transform_vector (w0, w1, w2, w3, digest);
+
+    w0[0] = digest[0];
+    w0[1] = digest[1];
+    w0[2] = digest[2];
+    w0[3] = digest[3];
+    w1[0] = 0x80;
+    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] = (64 + 16) * 8;
+    w3[3] = 0;
+
+    digest[0] = opad[0];
+    digest[1] = opad[1];
+    digest[2] = opad[2];
+    digest[3] = opad[3];
+
+    md5_transform_vector (w0, w1, w2, w3, digest);
+
+    COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
+  }
+}
+
+KERNEL_FQ void m07350_m08 (KERN_ATTR_BASIC ())
+{
+}
+
+KERNEL_FQ void m07350_m16 (KERN_ATTR_BASIC ())
+{
+}
+
+KERNEL_FQ void m07350_s04 (KERN_ATTR_BASIC ())
+{
+  /**
+   * modifier
+   */
+
+  const u64 lid = get_local_id (0);
+
+  /**
+   * base
+   */
+
+  const u64 gid = get_global_id (0);
+
+  if (gid >= GID_CNT) return;
+
+  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 & 63;
+
+  /**
+   * salt
+   */
+
+  u32 salt_buf0[4];
+  u32 salt_buf1[4];
+  u32 salt_buf2[4];
+  u32 salt_buf3[4];
+  u32 salt_buf4[4];
+  u32 salt_buf5[4];
+  u32 salt_buf6[4];
+  u32 salt_buf7[4];
+
+  salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0];
+  salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1];
+  salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2];
+  salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3];
+  salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4];
+  salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5];
+  salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6];
+  salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7];
+  salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8];
+  salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9];
+  salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10];
+  salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11];
+  salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12];
+  salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13];
+  salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14];
+  salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15];
+  salt_buf4[0] = salt_bufs[SALT_POS_HOST].salt_buf[16];
+  salt_buf4[1] = salt_bufs[SALT_POS_HOST].salt_buf[17];
+  salt_buf4[2] = salt_bufs[SALT_POS_HOST].salt_buf[18];
+  salt_buf4[3] = salt_bufs[SALT_POS_HOST].salt_buf[19];
+  salt_buf5[0] = salt_bufs[SALT_POS_HOST].salt_buf[20];
+  salt_buf5[1] = salt_bufs[SALT_POS_HOST].salt_buf[21];
+  salt_buf5[2] = salt_bufs[SALT_POS_HOST].salt_buf[22];
+  salt_buf5[3] = salt_bufs[SALT_POS_HOST].salt_buf[23];
+  salt_buf6[0] = salt_bufs[SALT_POS_HOST].salt_buf[24];
+  salt_buf6[1] = salt_bufs[SALT_POS_HOST].salt_buf[25];
+  salt_buf6[2] = salt_bufs[SALT_POS_HOST].salt_buf[26];
+  salt_buf6[3] = salt_bufs[SALT_POS_HOST].salt_buf[27];
+  salt_buf7[0] = salt_bufs[SALT_POS_HOST].salt_buf[28];
+  salt_buf7[1] = salt_bufs[SALT_POS_HOST].salt_buf[29];
+  salt_buf7[2] = salt_bufs[SALT_POS_HOST].salt_buf[30];
+  salt_buf7[3] = salt_bufs[SALT_POS_HOST].salt_buf[31];
+
+  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
+
+  /**
+   * digest
+   */
+
+  const u32 search[4] =
+  {
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
+  };
+
+  /**
+   * loop
+   */
+
+  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
+  {
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
+
+    const u32x pw_len = (pw_l_len + pw_r_len) & 63;
+
+    /**
+     * concat password candidate
+     */
+
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    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];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 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_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
+    }
+
+    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];
+    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];
+
+    /**
+     * pads
+     */
+
+    u32x ipad[4];
+    u32x opad[4];
+
+    hmac_md5_pad (w0, w1, w2, w3, ipad, opad);
+
+    //    hmac_md5_run (w0, w1, w2, w3, ipad, opad, digest);
+
+    u32x digest[4];
+
+    digest[0] = ipad[0];
+    digest[1] = ipad[1];
+    digest[2] = ipad[2];
+    digest[3] = ipad[3];
+
+    if (salt_len < 56)
+    {
+      w0[0] = salt_buf0[0];
+      w0[1] = salt_buf0[1];
+      w0[2] = salt_buf0[2];
+      w0[3] = salt_buf0[3];
+      w1[0] = salt_buf1[0];
+      w1[1] = salt_buf1[1];
+      w1[2] = salt_buf1[2];
+      w1[3] = salt_buf1[3];
+      w2[0] = salt_buf2[0];
+      w2[1] = salt_buf2[1];
+      w2[2] = salt_buf2[2];
+      w2[3] = salt_buf2[3];
+      w3[0] = salt_buf3[0];
+      w3[1] = salt_buf3[1];
+      w3[2] = (64 + salt_len) * 8;
+      w3[3] = 0;
+    }
+    else
+    {
+      w0[0] = salt_buf0[0];
+      w0[1] = salt_buf0[1];
+      w0[2] = salt_buf0[2];
+      w0[3] = salt_buf0[3];
+      w1[0] = salt_buf1[0];
+      w1[1] = salt_buf1[1];
+      w1[2] = salt_buf1[2];
+      w1[3] = salt_buf1[3];
+      w2[0] = salt_buf2[0];
+      w2[1] = salt_buf2[1];
+      w2[2] = salt_buf2[2];
+      w2[3] = salt_buf2[3];
+      w3[0] = salt_buf3[0];
+      w3[1] = salt_buf3[1];
+      w3[2] = salt_buf3[2];
+      w3[3] = salt_buf3[3];
+
+      md5_transform_vector (w0, w1, w2, w3, digest);
+
+      w0[0] = salt_buf4[0];
+      w0[1] = salt_buf4[1];
+      w0[2] = salt_buf4[2];
+      w0[3] = salt_buf4[3];
+      w1[0] = salt_buf5[0];
+      w1[1] = salt_buf5[1];
+      w1[2] = salt_buf5[2];
+      w1[3] = salt_buf5[3];
+      w2[0] = salt_buf6[0];
+      w2[1] = salt_buf6[1];
+      w2[2] = salt_buf6[2];
+      w2[3] = salt_buf6[3];
+      w3[0] = salt_buf7[0];
+      w3[1] = salt_buf7[1];
+      w3[2] = (64 + salt_len) * 8;
+      w3[3] = 0;
+    }
+
+    md5_transform_vector (w0, w1, w2, w3, digest);
+
+    w0[0] = digest[0];
+    w0[1] = digest[1];
+    w0[2] = digest[2];
+    w0[3] = digest[3];
+    w1[0] = 0x80;
+    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] = (64 + 16) * 8;
+    w3[3] = 0;
+
+    digest[0] = opad[0];
+    digest[1] = opad[1];
+    digest[2] = opad[2];
+    digest[3] = opad[3];
+
+    md5_transform_vector (w0, w1, w2, w3, digest);
+
+    COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
+  }
+}
+
+KERNEL_FQ void m07350_s08 (KERN_ATTR_BASIC ())
+{
+}
+
+KERNEL_FQ void m07350_s16 (KERN_ATTR_BASIC ())
+{
+}
diff --git a/OpenCL/m07350_a1-pure.cl b/OpenCL/m07350_a1-pure.cl
new file mode 100644
index 000000000..09b774141
--- /dev/null
+++ b/OpenCL/m07350_a1-pure.cl
@@ -0,0 +1,183 @@
+/**
+ * Author......: See docs/credits.txt
+ * License.....: MIT
+ */
+
+//#define NEW_SIMD_CODE
+
+#ifdef KERNEL_STATIC
+#include M2S(INCLUDE_PATH/inc_vendor.h)
+#include M2S(INCLUDE_PATH/inc_types.h)
+#include M2S(INCLUDE_PATH/inc_platform.cl)
+#include M2S(INCLUDE_PATH/inc_common.cl)
+#include M2S(INCLUDE_PATH/inc_scalar.cl)
+#include M2S(INCLUDE_PATH/inc_hash_md5.cl)
+#endif
+
+KERNEL_FQ void m07350_mxx (KERN_ATTR_BASIC ())
+{
+  /**
+   * modifier
+   */
+
+  const u64 lid = get_local_id (0);
+  const u64 gid = get_global_id (0);
+
+  if (gid >= GID_CNT) return;
+
+  /**
+   * base
+   */
+
+  const u32 pw_len = pws[gid].pw_len;
+
+  u32 w[64] = { 0 };
+
+  for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
+  {
+    w[idx] = pws[gid].i[idx];
+  }
+
+  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
+
+  u32 s[64] = { 0 };
+
+  for (u32 i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
+  {
+    s[idx] = salt_bufs[SALT_POS_HOST].salt_buf[idx];
+  }
+
+  /**
+   * loop
+   */
+
+  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos++)
+  {
+    const u32 comb_len = combs_buf[il_pos].pw_len;
+
+    u32 c[64];
+
+    #ifdef _unroll
+    #pragma unroll
+    #endif
+    for (int idx = 0; idx < 64; idx++)
+    {
+      c[idx] = combs_buf[il_pos].i[idx];
+    }
+
+    switch_buffer_by_offset_1x64_le_S (c, pw_len);
+
+    #ifdef _unroll
+    #pragma unroll
+    #endif
+    for (int i = 0; i < 64; i++)
+    {
+      c[i] |= w[i];
+    }
+
+    md5_hmac_ctx_t ctx;
+
+    md5_hmac_init (&ctx, c, pw_len + comb_len);
+
+    md5_hmac_update (&ctx, s, salt_len);
+
+    md5_hmac_final (&ctx);
+
+    const u32 r0 = ctx.opad.h[DGST_R0];
+    const u32 r1 = ctx.opad.h[DGST_R1];
+    const u32 r2 = ctx.opad.h[DGST_R2];
+    const u32 r3 = ctx.opad.h[DGST_R3];
+
+    COMPARE_M_SCALAR (r0, r1, r2, r3);
+  }
+}
+
+KERNEL_FQ void m07350_sxx (KERN_ATTR_BASIC ())
+{
+  /**
+   * modifier
+   */
+
+  const u64 lid = get_local_id (0);
+  const u64 gid = get_global_id (0);
+
+  if (gid >= GID_CNT) return;
+
+  /**
+   * digest
+   */
+
+  const u32 search[4] =
+  {
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2],
+    digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
+  };
+
+  /**
+   * base
+   */
+
+  const u32 pw_len = pws[gid].pw_len;
+
+  u32 w[64] = { 0 };
+
+  for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
+  {
+    w[idx] = pws[gid].i[idx];
+  }
+
+  const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
+
+  u32 s[64] = { 0 };
+
+  for (u32 i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
+  {
+    s[idx] = salt_bufs[SALT_POS_HOST].salt_buf[idx];
+  }
+
+  /**
+   * loop
+   */
+
+  for (u32 il_pos = 0; il_pos < IL_CNT; il_pos++)
+  {
+    const u32 comb_len = combs_buf[il_pos].pw_len;
+
+    u32 c[64];
+
+    #ifdef _unroll
+    #pragma unroll
+    #endif
+    for (int idx = 0; idx < 64; idx++)
+    {
+      c[idx] = combs_buf[il_pos].i[idx];
+    }
+
+    switch_buffer_by_offset_1x64_le_S (c, pw_len);
+
+    #ifdef _unroll
+    #pragma unroll
+    #endif
+    for (int i = 0; i < 64; i++)
+    {
+      c[i] |= w[i];
+    }
+
+    md5_hmac_ctx_t ctx;
+
+    md5_hmac_init (&ctx, c, pw_len + comb_len);
+
+    md5_hmac_update (&ctx, s, salt_len);
+
+    md5_hmac_final (&ctx);
+
+    const u32 r0 = ctx.opad.h[DGST_R0];
+    const u32 r1 = ctx.opad.h[DGST_R1];
+    const u32 r2 = ctx.opad.h[DGST_R2];
+    const u32 r3 = ctx.opad.h[DGST_R3];
+
+    COMPARE_S_SCALAR (r0, r1, r2, r3);
+  }
+}