Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © <2013>, Intel Corporation.
  3.  *
  4.  * Permission is hereby granted, free of charge, to any person obtaining a
  5.  * copy of this software and associated documentation files (the
  6.  * "Software"), to deal in the Software without restriction, including
  7.  * without limitation the rights to use, copy, modify, merge, publish,
  8.  * distribute, sub license, and/or sell copies of the Software, and to
  9.  * permit persons to whom the Software is furnished to do so, subject to
  10.  * the following conditions:
  11.  *
  12.  * The above copyright notice and this permission notice (including the
  13.  * next paragraph) shall be included in all copies or substantial portions
  14.  * of the Software.
  15.  *
  16.  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  17.  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  18.  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  19.  * IN NO EVENT SHALL PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR
  20.  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  21.  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  22.  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23.  *
  24.  */
  25. // Modual name: mpeg2_inter_gen8.asm
  26. //
  27. // Make inter predition estimation for MPEG2 Inter-frame on gen8
  28. //
  29.  
  30. //
  31. //  Now, begin source code....
  32. //
  33.  
  34. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  35. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  36.  
  37. /*
  38.  * __START
  39.  */
  40. __INTER_START:
  41. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  42. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  43. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  44. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  45.  
  46.        
  47. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  48. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  49.  
  50. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  51. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  52. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
  53. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  54.  
  55.        
  56. shl     (2) pic_ref.0<1>:uw     r4.24<2,2,1>:uw 4:uw    {align1};
  57. mov     (2) pic_ref.16<1>:uw    r4.20<2,2,1>:uw {align1};
  58. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  59. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  60. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  61. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  62. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  63.  
  64. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  65. (f0.0) jmpi (1) __mb_hwdep_end;
  66. /* read back the data for MB A */
  67. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  68. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  69. */
  70. mba_start:
  71. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  72. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  73. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  74. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  75. (f0.0)  jmpi (1)        mbb_start;
  76. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  77. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  78. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  79. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  80. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  81. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  82. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  83.  
  84. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  85. send (16)
  86.        mb_ind
  87.        mb_wb.0<1>:ud
  88.         NULL
  89.        data_port(
  90.                OBR_CACHE_TYPE,
  91.                OBR_MESSAGE_TYPE,
  92.                OBR_CONTROL_4,
  93.                OBR_BIND_IDX,
  94.                OBR_WRITE_COMMIT_CATEGORY,
  95.                OBR_HEADER_PRESENT
  96.        )
  97.        mlen 1
  98.        rlen 2
  99.        {align1};
  100.  
  101. /* TODO: RefID is required after multi-references are added */
  102. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  103. (f0.0)   mov (2)        mba_result.20<1>:w                      -1:w    {align1};
  104. (f0.0)   jmpi   (1)     mbb_start;
  105.  
  106. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  107. /* Read MV for MB A */
  108. /* bind index 3, read 2 oword (32 bytes), msg type: 0(OWord Block Read) */
  109. send (16)
  110.        mb_ind
  111.        mb_mv0.0<1>:ud
  112.         NULL
  113.        data_port(
  114.                OBR_CACHE_TYPE,
  115.                OBR_MESSAGE_TYPE,
  116.                OBR_CONTROL_2,
  117.                OBR_BIND_IDX,
  118.                OBR_WRITE_COMMIT_CATEGORY,
  119.                OBR_HEADER_PRESENT
  120.        )
  121.        mlen 1
  122.        rlen 1
  123.        {align1};
  124. /* TODO: RefID is required after multi-references are added */
  125. /* MV */
  126. mov        (2)          mba_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  127. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  128.  
  129. mbb_start:
  130. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  131. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  132. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  133. /* If MB B doesn't exist, neither MB C nor D exists */
  134. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  135. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  136. (f0.0)  jmpi (1)        mb_mvp_start;
  137. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  138. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  139. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  140. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  141. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  142. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  143. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  144.  
  145. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  146. send (16)
  147.        mb_ind
  148.        mb_wb.0<1>:ud
  149.         NULL
  150.        data_port(
  151.                OBR_CACHE_TYPE,
  152.                OBR_MESSAGE_TYPE,
  153.                OBR_CONTROL_4,
  154.                OBR_BIND_IDX,
  155.                OBR_WRITE_COMMIT_CATEGORY,
  156.                OBR_HEADER_PRESENT
  157.        )
  158.        mlen 1
  159.        rlen 2
  160.        {align1};
  161.  
  162. /* TODO: RefID is required after multi-references are added */
  163. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  164. (f0.0)   mov (2)        mbb_result.20<1>:w                      -1:w    {align1};
  165. (f0.0)   jmpi   (1)     mbc_start;
  166. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  167. /* Read MV for MB B */
  168. /* bind index 3, read 2 oword (32bytes), msg type: 0(OWord Block Read) */
  169. send (16)
  170.        mb_ind
  171.        mb_mv0.0<1>:ud
  172.         NULL
  173.        data_port(
  174.                OBR_CACHE_TYPE,
  175.                OBR_MESSAGE_TYPE,
  176.                OBR_CONTROL_2,
  177.                OBR_BIND_IDX,
  178.                OBR_WRITE_COMMIT_CATEGORY,
  179.                OBR_HEADER_PRESENT
  180.        )
  181.        mlen 1
  182.        rlen 1
  183.        {align1};
  184. /* TODO: RefID is required after multi-references are added */
  185. mov        (2)          mbb_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  186. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  187.  
  188. mbc_start:
  189. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  190. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  191. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  192. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  193. (f0.0)  jmpi (1)        mbd_start;
  194. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  195. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  196. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  197. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  198. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  199. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  200. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  201. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  202.  
  203. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  204. send (16)
  205.        mb_ind
  206.        mb_wb.0<1>:ud
  207.         NULL
  208.        data_port(
  209.                OBR_CACHE_TYPE,
  210.                OBR_MESSAGE_TYPE,
  211.                OBR_CONTROL_4,
  212.                OBR_BIND_IDX,
  213.                OBR_WRITE_COMMIT_CATEGORY,
  214.                OBR_HEADER_PRESENT
  215.        )
  216.        mlen 1
  217.        rlen 2
  218.        {align1};
  219.  
  220. /* TODO: RefID is required after multi-references are added */
  221. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  222. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  223. (f0.0)   jmpi   (1)     mb_mvp_start;
  224. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  225. /* Read MV for MB C */
  226. /* bind index 3, read 2 oword (32bytes), msg type: 0(OWord Block Read) */
  227. send (16)
  228.        mb_ind
  229.        mb_mv0.0<1>:ud
  230.         NULL
  231.        data_port(
  232.                OBR_CACHE_TYPE,
  233.                OBR_MESSAGE_TYPE,
  234.                OBR_CONTROL_2,
  235.                OBR_BIND_IDX,
  236.                OBR_WRITE_COMMIT_CATEGORY,
  237.                OBR_HEADER_PRESENT
  238.        )
  239.        mlen 1
  240.        rlen 1
  241.        {align1};
  242. /* TODO: RefID is required after multi-references are added */
  243. /* Forward MV */
  244. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  245. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  246.  
  247. jmpi   (1)    mb_mvp_start;
  248. mbd_start:
  249. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  250. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  251. (f0.0)  mov (2) mbc_result.20<1>:w      -1:w    {align1};
  252. (f0.0)  jmpi (1)        mb_mvp_start;
  253. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  254. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  255. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  256. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  257. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  258. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  259. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  260.  
  261. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  262. send (16)
  263.        mb_ind
  264.        mb_wb.0<1>:ud
  265.         NULL
  266.        data_port(
  267.                OBR_CACHE_TYPE,
  268.                OBR_MESSAGE_TYPE,
  269.                OBR_CONTROL_4,
  270.                OBR_BIND_IDX,
  271.                OBR_WRITE_COMMIT_CATEGORY,
  272.                OBR_HEADER_PRESENT
  273.        )
  274.        mlen 1
  275.        rlen 2
  276.        {align1};
  277.  
  278. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  279. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  280. (f0.0)   jmpi   (1)     mb_mvp_start;
  281.  
  282. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  283. /* Read MV for MB D */
  284. /* bind index 3, read 2 oword (32bytes), msg type: 0(OWord Block Read) */
  285. send (16)
  286.        mb_ind
  287.        mb_mv0.0<1>:ub
  288.         NULL
  289.        data_port(
  290.                OBR_CACHE_TYPE,
  291.                OBR_MESSAGE_TYPE,
  292.                OBR_CONTROL_2,
  293.                OBR_BIND_IDX,
  294.                OBR_WRITE_COMMIT_CATEGORY,
  295.                OBR_HEADER_PRESENT
  296.        )
  297.        mlen 1
  298.        rlen 1
  299.        {align1};
  300.  
  301. /* TODO: RefID is required after multi-references are added */
  302.  
  303. /* Forward MV */
  304. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  305. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  306.        
  307. mb_mvp_start:
  308. /*TODO: Add the skip prediction */
  309. /* Check whether both MB B and C are inavailable */
  310. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  311. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  312. (-f0.0) jmpi (1)        mb_median_start;
  313. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  314. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  315. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  316. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  317. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  318. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  319. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  320. jmpi    (1)     __mb_hwdep_end;
  321.        
  322. mb_median_start:
  323. /* check whether only one neighbour MB has the same ref ID with the current MB */
  324. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  325. cmp.z.f0.0      (1)     null:d  mba_result.20<0,1,0>:w  0:w     {align1};
  326. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  327. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  328. cmp.z.f0.0      (1)     null:d  mbb_result.20<0,1,0>:w  0:w     {align1};
  329. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  330. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  331. cmp.z.f0.0      (1)     null:d  mbc_result.20<0,1,0>:w  0:w     {align1};
  332. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  333. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  334. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
  335. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  336. (f0.0)  jmpi (1)  __mb_hwdep_end;
  337.  
  338. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  339. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  340. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  341. SAVE_RET        {align1};
  342.  jmpi   (1)     word_imedian;
  343. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  344. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  345. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  346. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  347. SAVE_RET        {align1};
  348. jmpi    (1)     word_imedian;
  349. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  350.  
  351. __mb_hwdep_end:
  352.  
  353. mov     (2)     mv_cc_ref.0<1>:w        mba_result.4<2,2,1>:w   {align1};
  354.  
  355. /* Calibrate the ref window for MPEG2 */
  356. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  357. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  358.  
  359. mov  (1) INPUT_ARG0.0<1>:ud     vme_m0.0<0,1,0>:ud      {align1};
  360. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  361. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  362.  
  363. SAVE_RET        {align1};
  364. jmpi    (1)     ref_boundary_check;
  365. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  366.  
  367. /* m2, get the MV/Mb cost passed from constant buffer when
  368. spawning thread by MEDIA_OBJECT */      
  369. mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
  370.  
  371. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  372.  
  373. /* m3 FWD/BWD cost center*/
  374. mov (8) vme_msg_3<1>:UD         0x0:UD {align1};               
  375.  
  376. /* m4 skip center*/
  377. mov (8) vme_msg_4<1>:UD         0x0:UD {align1};               
  378.  
  379. /* m5 */
  380. mov  (8) vme_msg_5<1>:UD        0x0:UD {align1};
  381.  
  382.  
  383. /* Use the Luma mode */
  384. mov  (1) tmp_reg0.0<1>:UW       LUMA_INTRA_MODE:UW {align1};
  385. mov  (1) vme_msg_5.5<1>:UB      tmp_reg0.0<0,1,0>:UB {align1};
  386.  
  387. mov  (1) tmp_reg0.0<1>:UW       INTRA16_DC_PRED:UW {align1};
  388. mov  (1) vme_msg_5.4<1>:ub      tmp_reg0.0<0,1,0>:UB {align1};
  389.  
  390. /* m6 */        
  391. mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
  392. mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
  393.  
  394. /* the penalty for Intra mode */
  395. mov  (1) vme_msg_6.28<1>:UD     0x010101:UD {align1};
  396.  
  397.  
  398. /* m7 */
  399.  
  400. mov  (8) vme_msg_7.0<1>:ud       0x0:ud {align1};
  401.  
  402. /*
  403. * SIC VME message
  404. */
  405.  
  406. /* Disable Intra8x8/Intra4x4 Intra-prediction */
  407. /* m1 */
  408. mov  (8) vme_m1.0<1>:ud         0x0:UD  {align1};
  409.  
  410. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  411. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  412. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  413. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  414.  
  415. /* assign MB intra struct from the thread payload*/
  416. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  417.  
  418. /* Enable DC HAAR component when calculating HARR SATD block */
  419. mov  (1) tmp_reg0.0<1>:UW       DC_HARR_ENABLE:UW               {align1};
  420. mov  (1) vme_m1.30<1>:UB        tmp_reg0.0<0,1,0>:UB  {align1};
  421. mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
  422.  
  423. /* m0 */        
  424. mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
  425. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  426.  
  427. /* after verification it will be passed by using payload */
  428. send (8)
  429.        vme_msg_ind
  430.        vme_wb<1>:UD
  431.        null
  432.        cre(
  433.                BIND_IDX_VME,
  434.                VME_SIC_MESSAGE_TYPE
  435.        )
  436.        mlen sic_vme_msg_length
  437.        rlen vme_wb_length
  438.        {align1};
  439. /*
  440. * Oword Block Write message
  441. */
  442. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  443.        
  444. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  445. mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
  446. mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
  447. mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
  448.  
  449. /* Distortion, Intra (17-16), */
  450. mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
  451.  
  452. mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
  453. /* VME clock counts */
  454. mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
  455.  
  456. mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
  457.  
  458. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  459. send (16)
  460.        msg_ind
  461.        obw_wb
  462.        null
  463.        data_port(
  464.                OBW_CACHE_TYPE,
  465.                OBW_MESSAGE_TYPE,
  466.                OBW_CONTROL_2,
  467.                OBW_BIND_IDX,
  468.                OBW_WRITE_COMMIT_CATEGORY,
  469.                OBW_HEADER_PRESENT
  470.        )
  471.        mlen 2
  472.        rlen obw_wb_length
  473.        {align1};
  474.  
  475. /* IME search */
  476. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  477. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  478.  
  479. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  480.  
  481. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  482.  
  483. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  484. /* the Max MV number is passed by constant buffer */
  485. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  486. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  487. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  488.  
  489. /* Setup the Cost center */
  490. /* currently four 8x8 share the same cost center */
  491. mov  (4) vme_m3.0<2>:ud         mv_cc_ref.0<0,1,0>:ud   {align1};
  492. mov  (4) vme_m3.4<2>:ud         mv_cc_ref.0<0,1,0>:ud   {align1};
  493.  
  494. mov (8) vme_msg_3<1>:UD         vme_m3.0<8,8,1>:UD {align1};
  495. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  496.  
  497. /* M4/M5 search path */
  498. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  499. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  500. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  501. mov  (1) vme_msg_4.12<1>:UD     0x100F0F0F:UD {align1};
  502. mov  (1) vme_msg_4.16<1>:UD     0x01010101:UD {align1};
  503. mov  (1) vme_msg_4.20<1>:UD     0x10010101:UD {align1};
  504. mov  (1) vme_msg_4.24<1>:UD     0x0F0F0F0F:UD {align1};
  505. mov  (1) vme_msg_4.28<1>:UD     0x100F0F0F:UD {align1};
  506.  
  507. mov  (1) vme_msg_5.0<1>:UD      0x01010101:UD {align1};
  508. mov  (1) vme_msg_5.4<1>:UD      0x10010101:UD {align1};
  509. mov  (1) vme_msg_5.8<1>:UD      0x0F0F0F0F:UD {align1};
  510. mov  (1) vme_msg_5.12<1>:UD     0x000F0F0F:UD {align1};
  511.  
  512. mov  (4) vme_msg_5.16<1>:UD     0x0:UD {align1};
  513.  
  514. send (8)
  515.        vme_msg_ind
  516.        vme_wb<1>:UD
  517.        null
  518.        vme(
  519.                BIND_IDX_VME,
  520.                0,
  521.                0,
  522.                VME_IME_MESSAGE_TYPE
  523.        )
  524.        mlen ime_vme_msg_length
  525.        rlen vme_wb_length {align1};
  526.  
  527. /* Set Macroblock-shape/mode for FBR */
  528.  
  529. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  530. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  531. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  532.  
  533. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  534. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  535.  
  536. /* Send FBR message into CRE */
  537.  
  538. mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  539. mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  540. mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  541. mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  542.  
  543. mov  (1) vme_m0.12<1>:UD        INTER_SAD_HAAR + SUB_PEL_MODE_HALF + FBR_BME_DISABLE:UD {align1};    /* 16x16 Source, 1/2 pixel, harr, BME disable */
  544. /* Bilinear filter */
  545. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  546. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  547.  
  548. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  549. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  550.  
  551. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  552. mov  (8) vme_msg_3.0<1>:UD              vme_m3.0<8,8,1>:UD      {align1};
  553.  
  554. /* after verification it will be passed by using payload */
  555. send (8)
  556.        vme_msg_ind
  557.        vme_wb<1>:UD
  558.        null
  559.        cre(
  560.                BIND_IDX_VME,
  561.                VME_FBR_MESSAGE_TYPE
  562.        )
  563.        mlen fbr_vme_msg_length
  564.        rlen vme_wb_length
  565.        {align1};
  566.  
  567. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  568. (-f0.0) jmpi (1) vme_run_again;
  569. nop;
  570. vme_mv_output:
  571.  
  572. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
  573. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  574. /* write FME info */
  575. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  576.  
  577. mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  578. /* Inter distortion of FME */
  579. mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
  580.  
  581. mov  (1) msg_reg1.12<1>:UD      vme_m2.20<0,1,0>:UD {align1};
  582.  
  583. /* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
  584. send (16)
  585.        msg_ind
  586.        obw_wb
  587.        null
  588.        data_port(
  589.                OBW_CACHE_TYPE,
  590.                OBW_MESSAGE_TYPE,
  591.                OBW_CONTROL_0,
  592.                OBW_BIND_IDX,
  593.                OBW_WRITE_COMMIT_CATEGORY,
  594.                OBW_HEADER_PRESENT
  595.        )
  596.        mlen 2
  597.        rlen obw_wb_length
  598.        {align1};
  599.  
  600. /* Write FME/BME MV */
  601. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
  602. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  603.  
  604.  
  605. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  606. mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  607. mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  608. mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  609. /* bind index 3, write  2 oword (32 bytes), msg type: 8(OWord Block Write) */
  610. send (16)
  611.        msg_ind
  612.        obw_wb
  613.        null
  614.        data_port(
  615.                OBW_CACHE_TYPE,
  616.                OBW_MESSAGE_TYPE,
  617.                OBW_CONTROL_2,
  618.                OBW_BIND_IDX,
  619.                OBW_WRITE_COMMIT_CATEGORY,
  620.                OBW_HEADER_PRESENT
  621.        )
  622.        mlen 2
  623.        rlen obw_wb_length
  624.        {align1};
  625.  
  626. /* Write FME/BME RefID */
  627. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
  628. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  629.  
  630. mov  (8) msg_reg1.0<1>:UD       vme_wb6.0<8,8,1>:UD {align1};
  631.  
  632. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  633. send (16)
  634.        msg_ind
  635.        obw_wb
  636.        null
  637.        data_port(
  638.                OBW_CACHE_TYPE,
  639.                OBW_MESSAGE_TYPE,
  640.                OBW_CONTROL_2,
  641.                OBW_BIND_IDX,
  642.                OBW_WRITE_COMMIT_CATEGORY,
  643.                OBW_HEADER_PRESENT
  644.        )
  645.        mlen 2
  646.        rlen obw_wb_length
  647.        {align1};
  648.  
  649. /* Issue message fence so that the previous write message is committed */
  650. send (16)
  651.        mb_ind
  652.        mb_wb.0<1>:ud
  653.         NULL
  654.        data_port(
  655.                OBR_CACHE_TYPE,
  656.                OBR_MESSAGE_FENCE,
  657.                OBR_MF_COMMIT,
  658.                OBR_BIND_IDX,
  659.                OBR_WRITE_COMMIT_CATEGORY,
  660.                OBR_HEADER_PRESENT
  661.        )
  662.        mlen 1
  663.        rlen 1
  664.        {align1};
  665.  
  666. __EXIT:
  667. /*
  668. * kill thread
  669. */        
  670. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  671. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  672.  
  673.  
  674.         nop             ;
  675.         nop             ;
  676.        
  677. word_imedian:
  678.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  679.         (f0.0)  jmpi (1) cmp_a_ge_b;
  680.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  681.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  682.         (f0.0) jmpi (1) cmp_end;
  683.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  684.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  685.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  686.         jmpi (1) cmp_end;
  687. cmp_a_ge_b:
  688.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  689.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  690.         (f0.0) jmpi (1) cmp_end;
  691.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  692.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  693.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  694. cmp_end:
  695.         RETURN  {align1};
  696.  
  697. nop;
  698. nop;
  699. ref_boundary_check:
  700.  
  701. /* The left/up coordinate of reference window */
  702. add  (2) TEMP_VAR0.0<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG0.0<2,2,1>:w   {align1};
  703. /* The right/bottom coordinate of reference window */
  704. add  (1) TEMP_VAR0.16<1>:w      TEMP_VAR0.0<0,1,0>:w    48:w                    {align1};
  705. add  (1) TEMP_VAR0.18<1>:w      TEMP_VAR0.2<0,1,0>:w    40:w                    {align1};
  706.  
  707. /* Firstly the MV range is checked */
  708. mul  (2) TEMP_VAR1.16<1>:w      INPUT_ARG1.16<2,2,1>:w  -1:w    {align1};
  709. add  (2) TEMP_VAR1.0<1>:w       INPUT_ARG0.8<2,2,1>:w   TEMP_VAR1.16<2,2,1>:w   {align1};
  710. add  (2) TEMP_VAR1.4<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG1.16<2,2,1>:w  {align1};
  711.  
  712. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.0<0,1,0>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  713. (f0.0)  mov     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  714. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   TEMP_VAR1.4<0,1,0>:w    {align1};
  715. (f0.0)  add     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.4<0,1,0>:w    -48:w   {align1};
  716. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.2<0,1,0>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  717. (f0.0)  mov     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  718. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.18<0,1,0>:w   TEMP_VAR1.6<0,1,0>:w    {align1};
  719. (f0.0)  add     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.6<0,1,0>:w    -40:w   {align1};
  720.  
  721. x_left_cmp:
  722.         cmp.l.f0.0 (1)          null:w          TEMP_VAR0.0<0,1,0>:w    0:w     {align1};
  723.         (-f0.0) jmpi    (1)     x_right_cmp;
  724.         (f0.0)  mov     (1)     TEMP_VAR0.0<1>:w        0:w             {align1};
  725.         jmpi    (1)     y_top_cmp;
  726. x_right_cmp:
  727.         cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   INPUT_ARG1.0<0,1,0>:w   {align1};
  728.         (-f0.0) jmpi    (1)     y_top_cmp;     
  729.         (f0.0)  add     (1)     TEMP_VAR0.0<1>:w        INPUT_ARG1.0<0,1,0>:w   -48:w   {align1};
  730. y_top_cmp:
  731.         cmp.l.f0.0  (1)         null:w  TEMP_VAR0.2<0,1,0>:w    0:w             {align1};
  732.         (-f0.0) jmpi    (1)     y_bottom_cmp;
  733.         (f0.0)  mov     (1)     TEMP_VAR0.2<1>:w        0:w             {align1};
  734.         jmpi    (1)     y_bottom_end;
  735. y_bottom_cmp:
  736.         cmp.g.f0.0  (1)         null:w  TEMP_VAR0.18<0,1,0>:w   INPUT_ARG1.2<0,1,0>:w {align1};
  737.         (f0.0)  add     (1)     TEMP_VAR0.2<1>:w                INPUT_ARG1.2<0,1,0>:w   -40:w   {align1};
  738.  
  739. y_bottom_end:
  740. mul     (2)     TEMP_VAR1.0<1>:w        INPUT_ARG0.8<2,2,1>:w   -1:w    {align1};
  741. add     (2)     RET_ARG<1>:w            TEMP_VAR0.0<2,2,1>:w    TEMP_VAR1.0<2,2,1>:w    {align1};      
  742.         RETURN  {align1};
  743. nop;
  744. nop;
  745.  
  746. vme_run_again:
  747.  
  748. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  749. mov     (2)     tmp_reg0.0<1>:w         mb_ref_win.0<2,2,1>:w           {align1};
  750. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  751. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  752.  
  753. cmp.l.f0.0      (1) null:w      tmp_reg0.0<0,1,0>:w     0:w     {align1};
  754. (f0.0)  mul     (1) tmp_reg0.0<1>:w     tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  755. cmp.l.f0.0      (1) null:w      tmp_reg0.2<0,1,0>:w     0:w     {align1};
  756. (f0.0)  mul     (1) tmp_reg0.2<1>:w     tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  757.  
  758. cmp.ge.f0.0     (1) null:w      tmp_reg0.0<0,1,0>:w     4:w     {align1};
  759. (f0.0)  jmpi (1)        vme_start;
  760. cmp.ge.f0.0     (1) null:w      tmp_reg0.2<0,1,0>:w     4:w     {align1};
  761. (f0.0)  jmpi (1)        vme_start;
  762.  
  763. jmpi (1) vme_done;
  764.  
  765. vme_start:
  766.         mov (8) tmp_vme_wb0.0<1>:ud     vme_wb0.0<8,8,1>:ud     {align1};
  767.         mov (8) tmp_vme_wb1.0<1>:ud     vme_wb1.0<8,8,1>:ud     {align1};
  768.  
  769. /* Calibrate the ref window for MPEG2 */
  770. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  771. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  772. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  773. add  (2) INPUT_ARG0.0<1>:w      vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  774. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  775.  
  776. SAVE_RET        {align1};
  777. jmpi    (1)     ref_boundary_check;
  778. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  779.  
  780. /* IME search */
  781. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  782. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  783.  
  784. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  785.  
  786. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  787.  
  788. mov  (8) vme_m1.0<1>:ud         0x0:UD  {align1};
  789.  
  790. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  791. /* the Max MV number is passed by constant buffer */
  792. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  793. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  794. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  795.  
  796. /* Setup the Cost center */
  797. /* currently four 8x8 share the same cost center */
  798. mov  (4) vme_m3.0<2>:ud         mv_cc_ref.0<0,1,0>:ud   {align1};
  799. mov  (4) vme_m3.4<2>:ud         mv_cc_ref.0<0,1,0>:ud   {align1};
  800.  
  801. mov (8) vme_msg_3<1>:UD         vme_m3.0<8,8,1>:UD {align1};
  802. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  803.  
  804. /* M4/M5 search path */
  805. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  806. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  807. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  808. mov  (1) vme_msg_4.12<1>:UD     0x100F0F0F:UD {align1};
  809. mov  (1) vme_msg_4.16<1>:UD     0x01010101:UD {align1};
  810. mov  (1) vme_msg_4.20<1>:UD     0x10010101:UD {align1};
  811. mov  (1) vme_msg_4.24<1>:UD     0x0F0F0F0F:UD {align1};
  812. mov  (1) vme_msg_4.28<1>:UD     0x100F0F0F:UD {align1};
  813.  
  814. mov  (1) vme_msg_5.0<1>:UD      0x01010101:UD {align1};
  815. mov  (1) vme_msg_5.4<1>:UD      0x10010101:UD {align1};
  816. mov  (1) vme_msg_5.8<1>:UD      0x0F0F0F0F:UD {align1};
  817. mov  (1) vme_msg_5.12<1>:UD     0x000F0F0F:UD {align1};
  818.  
  819. mov  (4) vme_msg_5.16<1>:UD     0x0:UD {align1};
  820.  
  821. send (8)
  822.        vme_msg_ind
  823.        vme_wb<1>:UD
  824.        null
  825.        vme(
  826.                BIND_IDX_VME,
  827.                0,
  828.                0,
  829.                VME_IME_MESSAGE_TYPE
  830.        )
  831.        mlen ime_vme_msg_length
  832.        rlen vme_wb_length {align1};
  833.  
  834. /* Set Macroblock-shape/mode for FBR */
  835.  
  836. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  837. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  838. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  839.  
  840. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  841. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  842.  
  843. /* Send FBR message into CRE */
  844.  
  845. mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  846. mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  847. mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  848. mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  849.  
  850. mov  (1) vme_m0.12<1>:UD        INTER_SAD_HAAR + SUB_PEL_MODE_HALF + FBR_BME_DISABLE:UD {align1};    /* 16x16 Source, 1/2 pixel, harr, BME disable */
  851. /* Bilinear filter */
  852. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  853. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  854.  
  855. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  856. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  857.  
  858. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  859. mov  (8) vme_msg_3.0<1>:UD              vme_m3.0<8,8,1>:UD      {align1};
  860.  
  861. /* after verification it will be passed by using payload */
  862. send (8)
  863.        vme_msg_ind
  864.        vme_wb<1>:UD
  865.        null
  866.        cre(
  867.                BIND_IDX_VME,
  868.                VME_FBR_MESSAGE_TYPE
  869.        )
  870.        mlen fbr_vme_msg_length
  871.        rlen vme_wb_length
  872.        {align1};
  873.  
  874. cmp.l.f0.0 (1)  null:uw vme_wb0.8<0,1,0>:uw     tmp_vme_wb0.8<0,1,0>:uw {align1};
  875. (f0.0)  jmpi (1) vme_done;
  876. mov     (8)     vme_wb0.0<1>:ud tmp_vme_wb0.0<8,8,1>:ud {align1};
  877. mov     (8)     vme_wb1.0<1>:ud tmp_vme_wb1.0<8,8,1>:ud {align1};
  878.  
  879. vme_done:
  880.       jmpi (1) vme_mv_output;
  881. nop;
  882. nop;
  883. nop;
  884.  
  885.