Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © <2010>, 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.  *        Author : Zhao Yakui <yakui.zhao@intel.com>
  24.  */
  25. // Modual name: mpeg2_inter_haswell.asm
  26. //
  27. // Make MPEG2 inter predition estimation for Inter-frame on Haswell
  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. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  47. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  48.  
  49. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  50. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  51. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
  52. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  53.        
  54. shl     (2) pic_ref.0<1>:uw     r4.24<2,2,1>:uw 4:uw    {align1};
  55. mov     (2) pic_ref.16<1>:uw    r4.20<2,2,1>:uw {align1};
  56. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  57. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  58. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  59. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  60. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  61.  
  62. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  63. (f0.0) jmpi (1) __mb_hwdep_end;
  64. /* read back the data for MB A */
  65. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  66. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  67. */
  68. mba_start:
  69. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  70. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  71. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  72. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  73. (f0.0)  jmpi (1)        mbb_start;
  74. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  75. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  76. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  77. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  78. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  79. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  80. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  81.  
  82. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  83. send (16)
  84.        mb_ind
  85.        mb_wb.0<1>:ud
  86.         NULL
  87.        data_port(
  88.                OBR_CACHE_TYPE,
  89.                OBR_MESSAGE_TYPE,
  90.                OBR_CONTROL_4,
  91.                OBR_BIND_IDX,
  92.                OBR_WRITE_COMMIT_CATEGORY,
  93.                OBR_HEADER_PRESENT
  94.        )
  95.        mlen 1
  96.        rlen 2
  97.        {align1};
  98.  
  99. /* TODO: RefID is required after multi-references are added */
  100. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  101. (f0.0)   mov (2)        mba_result.20<1>:w                      -1:w    {align1};
  102. (f0.0)   jmpi   (1)     mbb_start;
  103.  
  104. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  105. /* Read MV for MB A */
  106. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  107. send (16)
  108.        mb_ind
  109.        mb_mv0.0<1>:ud
  110.         NULL
  111.        data_port(
  112.                OBR_CACHE_TYPE,
  113.                OBR_MESSAGE_TYPE,
  114.                OBR_CONTROL_2,
  115.                OBR_BIND_IDX,
  116.                OBR_WRITE_COMMIT_CATEGORY,
  117.                OBR_HEADER_PRESENT
  118.        )
  119.        mlen 1
  120.        rlen 1
  121.        {align1};
  122. /* TODO: RefID is required after multi-references are added */
  123. /* MV */
  124. mov        (2)          mba_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  125. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  126.  
  127. mbb_start:
  128. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  129. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  130. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  131. /* If MB B doesn't exist, neither MB C nor D exists */
  132. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  133. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  134. (f0.0)  jmpi (1)        mb_mvp_start;
  135. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  136. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  137. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  138. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  139. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  140. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  141. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  142.  
  143. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  144. send (16)
  145.        mb_ind
  146.        mb_wb.0<1>:ud
  147.         NULL
  148.        data_port(
  149.                OBR_CACHE_TYPE,
  150.                OBR_MESSAGE_TYPE,
  151.                OBR_CONTROL_4,
  152.                OBR_BIND_IDX,
  153.                OBR_WRITE_COMMIT_CATEGORY,
  154.                OBR_HEADER_PRESENT
  155.        )
  156.        mlen 1
  157.        rlen 2
  158.        {align1};
  159.  
  160. /* TODO: RefID is required after multi-references are added */
  161. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  162. (f0.0)   mov (2)        mbb_result.20<1>:w                      -1:w    {align1};
  163. (f0.0)   jmpi   (1)     mbc_start;
  164. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  165. /* Read MV for MB B */
  166. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  167. send (16)
  168.        mb_ind
  169.        mb_mv0.0<1>:ud
  170.         NULL
  171.        data_port(
  172.                OBR_CACHE_TYPE,
  173.                OBR_MESSAGE_TYPE,
  174.                OBR_CONTROL_2,
  175.                OBR_BIND_IDX,
  176.                OBR_WRITE_COMMIT_CATEGORY,
  177.                OBR_HEADER_PRESENT
  178.        )
  179.        mlen 1
  180.        rlen 1
  181.        {align1};
  182. /* TODO: RefID is required after multi-references are added */
  183. mov        (2)          mbb_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  184. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  185.  
  186. mbc_start:
  187. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  188. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  189. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  190. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  191. (f0.0)  jmpi (1)        mbd_start;
  192. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  193. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  194. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  195. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  196. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  197. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  198. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  199. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  200.  
  201. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  202. send (16)
  203.        mb_ind
  204.        mb_wb.0<1>:ud
  205.         NULL
  206.        data_port(
  207.                OBR_CACHE_TYPE,
  208.                OBR_MESSAGE_TYPE,
  209.                OBR_CONTROL_4,
  210.                OBR_BIND_IDX,
  211.                OBR_WRITE_COMMIT_CATEGORY,
  212.                OBR_HEADER_PRESENT
  213.        )
  214.        mlen 1
  215.        rlen 2
  216.        {align1};
  217.  
  218. /* TODO: RefID is required after multi-references are added */
  219. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  220. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  221. (f0.0)   jmpi   (1)     mb_mvp_start;
  222. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  223. /* Read MV for MB C */
  224. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  225. send (16)
  226.        mb_ind
  227.        mb_mv0.0<1>:ud
  228.         NULL
  229.        data_port(
  230.                OBR_CACHE_TYPE,
  231.                OBR_MESSAGE_TYPE,
  232.                OBR_CONTROL_2,
  233.                OBR_BIND_IDX,
  234.                OBR_WRITE_COMMIT_CATEGORY,
  235.                OBR_HEADER_PRESENT
  236.        )
  237.        mlen 1
  238.        rlen 1
  239.        {align1};
  240. /* TODO: RefID is required after multi-references are added */
  241. /* Forward MV */
  242. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  243. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  244.  
  245. jmpi   (1)    mb_mvp_start;
  246. mbd_start:
  247. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  248. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  249. (f0.0)  jmpi (1)        mb_mvp_start;
  250. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  251. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  252. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  253. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  254. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  255. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  256. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  257.  
  258. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  259. send (16)
  260.        mb_ind
  261.        mb_wb.0<1>:ud
  262.         NULL
  263.        data_port(
  264.                OBR_CACHE_TYPE,
  265.                OBR_MESSAGE_TYPE,
  266.                OBR_CONTROL_4,
  267.                OBR_BIND_IDX,
  268.                OBR_WRITE_COMMIT_CATEGORY,
  269.                OBR_HEADER_PRESENT
  270.        )
  271.        mlen 1
  272.        rlen 2
  273.        {align1};
  274.  
  275. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  276. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  277. (f0.0)   jmpi   (1)     mb_mvp_start;
  278.  
  279. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  280. /* Read MV for MB D */
  281. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  282. send (16)
  283.        mb_ind
  284.        mb_mv0.0<1>:ub
  285.         NULL
  286.        data_port(
  287.                OBR_CACHE_TYPE,
  288.                OBR_MESSAGE_TYPE,
  289.                OBR_CONTROL_2,
  290.                OBR_BIND_IDX,
  291.                OBR_WRITE_COMMIT_CATEGORY,
  292.                OBR_HEADER_PRESENT
  293.        )
  294.        mlen 1
  295.        rlen 1
  296.        {align1};
  297.  
  298. /* TODO: RefID is required after multi-references are added */
  299.  
  300. /* Forward MV */
  301. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  302. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  303.        
  304. mb_mvp_start:
  305. /*TODO: Add the skip prediction */
  306. /* Check whether both MB B and C are inavailable */
  307. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  308. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  309. (-f0.0) jmpi (1)        mb_median_start;
  310. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  311. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  312. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  313. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  314. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  315. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  316. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  317. jmpi    (1)     __mb_hwdep_end;
  318.        
  319. mb_median_start:
  320. /* check whether only one neighbour MB has the same ref ID with the current MB */
  321. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  322. cmp.z.f0.0      (1)     null:d  mba_result.20<1>:w      0:w     {align1};
  323. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  324. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  325. cmp.z.f0.0      (1)     null:d  mbb_result.20<1>:w      0:w     {align1};
  326. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  327. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  328. cmp.z.f0.0      (1)     null:d  mbc_result.20<1>:w      0:w     {align1};
  329. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  330. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  331. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  332. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  333. (f0.0)  jmpi (1)  __mb_hwdep_end;
  334.  
  335. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  336. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  337. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  338. SAVE_RET        {align1};
  339.  jmpi   (1)     word_imedian;
  340. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  341. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  342. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  343. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  344. SAVE_RET        {align1};
  345. jmpi    (1)     word_imedian;
  346. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  347.  
  348. __mb_hwdep_end:
  349.        
  350. mov     (2)     mv_cc_ref.0<1>:w        mba_result.4<2,2,1>:w   {align1};
  351.  
  352. /* Calibrate the ref window for MPEG2 */
  353. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  354. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  355.  
  356. mov  (1) INPUT_ARG0.0<1>:ud     vme_m0.0<0,1,0>:ud      {align1};
  357. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  358. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  359.  
  360. SAVE_RET        {align1};
  361. jmpi    (1)     ref_boundary_check;
  362. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  363.  
  364. /* m2, get the MV/Mb cost passed from constant buffer when
  365. spawning thread by MEDIA_OBJECT */      
  366. mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
  367.  
  368. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  369.  
  370. /* m3 */
  371. mov (8) vme_msg_3<1>:UD         0x0:UD {align1};               
  372.  
  373. /* the neighbour pixel is zero for MPEG2 Intra-prediction */
  374.  
  375. /* m4 */
  376. mov  (8) vme_msg_4<1>:UD         0:UD {align1};
  377. mov  (1) tmp_reg0.0<1>:UW       LUMA_INTRA_MODE:UW {align1};
  378. /* Use the Luma mode */
  379. mov  (1) vme_msg_4.5<1>:UB      tmp_reg0.0<0,1,0>:UB {align1};
  380. mov  (1) tmp_reg0.0<1>:UW       INTRA16_DC_PRED:UW {align1};
  381. mov  (1) vme_msg_4.4<1>:ub      tmp_reg0.0<0,1,0>:UB {align1};
  382.  
  383. /* m5 */        
  384. mov  (8) vme_msg_5<1>:UD         0x0:UD {align1};
  385. mov  (1) vme_msg_5.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
  386.  
  387. /* the penalty for Intra mode */
  388. mov  (1) vme_msg_5.28<1>:UD     0x010101:UD {align1};
  389.  
  390.  
  391. /* m6 */
  392. mov  (8) vme_msg_6.0<1>:UD       0:Ud {align1};
  393.  
  394. /*
  395. * SIC VME message
  396. */
  397. /* m0 */        
  398. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  399.  
  400. /* Disable Intra8x8/Intra4x4 Intra-prediction */
  401. /* m1 */
  402. mov  (8) vme_m1.0<1>:ud         0x0:UD  {align1};
  403.  
  404. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  405. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  406. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  407. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  408.  
  409. /* assign MB intra struct from the thread payload*/
  410. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  411.  
  412. /* Enable DC HAAR component when calculating HARR SATD block */
  413. mov  (1) tmp_reg0.0<1>:UW       DC_HARR_ENABLE:UW               {align1};
  414. mov  (1) vme_m1.30<1>:UB        tmp_reg0.0<0,1,0>:UB  {align1};
  415.  
  416. mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
  417. /* m0 */        
  418. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  419. mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
  420.  
  421. /* after verification it will be passed by using payload */
  422. send (8)
  423.        vme_msg_ind
  424.        vme_wb<1>:UD
  425.        null
  426.        cre(
  427.                BIND_IDX_VME,
  428.                VME_SIC_MESSAGE_TYPE
  429.        )
  430.        mlen sic_vme_msg_length
  431.        rlen vme_wb_length
  432.        {align1};
  433.  
  434. /*
  435. * Oword Block Write message
  436. */
  437. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  438.        
  439. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  440. mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
  441. mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
  442. mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
  443.  
  444. /* Distortion, Intra (17-16), */
  445. mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
  446.  
  447. mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
  448. /* VME clock counts */
  449. mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
  450.  
  451. mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
  452.  
  453. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  454. send (16)
  455.        msg_ind
  456.        obw_wb
  457.        null
  458.        data_port(
  459.                OBW_CACHE_TYPE,
  460.                OBW_MESSAGE_TYPE,
  461.                OBW_CONTROL_2,
  462.                OBW_BIND_IDX,
  463.                OBW_WRITE_COMMIT_CATEGORY,
  464.                OBW_HEADER_PRESENT
  465.        )
  466.        mlen 2
  467.        rlen obw_wb_length
  468.        {align1};
  469.  
  470. /* IME search */
  471. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  472. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  473.  
  474. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  475.  
  476. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  477.  
  478. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  479. /* the Max MV number is passed by constant buffer */
  480. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  481. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  482. /* Set the MV cost center */
  483. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  484. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  485.  
  486. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  487.  
  488. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  489. /* M3/M4 search path */
  490.  
  491. mov  (1) vme_msg_3.0<1>:UD      0x01010101:UD {align1};
  492. mov  (1) vme_msg_3.4<1>:UD      0x10010101:UD {align1};
  493. mov  (1) vme_msg_3.8<1>:UD      0x0F0F0F0F:UD {align1};
  494. mov  (1) vme_msg_3.12<1>:UD     0x100F0F0F:UD {align1};
  495. mov  (1) vme_msg_3.16<1>:UD     0x01010101:UD {align1};
  496. mov  (1) vme_msg_3.20<1>:UD     0x10010101:UD {align1};
  497. mov  (1) vme_msg_3.24<1>:UD     0x0F0F0F0F:UD {align1};
  498. mov  (1) vme_msg_3.28<1>:UD     0x100F0F0F:UD {align1};
  499.  
  500. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  501. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  502. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  503. mov  (1) vme_msg_4.12<1>:UD     0x000F0F0F:UD {align1};
  504.  
  505. mov  (4) vme_msg_4.16<1>:UD     0x0:UD {align1};
  506.  
  507. send (8)
  508.        vme_msg_ind
  509.        vme_wb<1>:UD
  510.        null
  511.        vme(
  512.                BIND_IDX_VME,
  513.                0,
  514.                0,
  515.                VME_IME_MESSAGE_TYPE
  516.        )
  517.        mlen ime_vme_msg_length
  518.        rlen vme_wb_length {align1};
  519.  
  520. /* Set Macroblock-shape/mode for FBR */
  521.  
  522. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  523. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  524. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  525.  
  526. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  527. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  528.  
  529. /* Send FBR message into CRE */
  530.  
  531. mov  (8) vme_msg_3.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  532. mov  (8) vme_msg_4.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  533. mov  (8) vme_msg_5.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  534. mov  (8) vme_msg_6.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  535.  
  536. 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 */
  537.  
  538. /* Bilinear filter */
  539. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  540. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  541.  
  542. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  543. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  544.  
  545. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  546.  
  547. /* after verification it will be passed by using payload */
  548. send (8)
  549.        vme_msg_ind
  550.        vme_wb<1>:UD
  551.        null
  552.        cre(
  553.                BIND_IDX_VME,
  554.                VME_FBR_MESSAGE_TYPE
  555.        )
  556.        mlen fbr_vme_msg_length
  557.        rlen vme_wb_length
  558.        {align1};
  559.  
  560. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  561. (-f0.0) jmpi (1) vme_run_again;
  562. nop;
  563. vme_mv_output:
  564.  
  565. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
  566. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  567. /* write FME info */
  568. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  569.  
  570. mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  571. /* Inter distortion of FME */
  572. mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
  573.  
  574. mov  (1) msg_reg1.12<1>:UD      vme_m2.20<0,1,0>:UD {align1};
  575.  
  576. /* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
  577. send (16)
  578.        msg_ind
  579.        obw_wb
  580.        null
  581.        data_port(
  582.                OBW_CACHE_TYPE,
  583.                OBW_MESSAGE_TYPE,
  584.                OBW_CONTROL_0,
  585.                OBW_BIND_IDX,
  586.                OBW_WRITE_COMMIT_CATEGORY,
  587.                OBW_HEADER_PRESENT
  588.        )
  589.        mlen 2
  590.        rlen obw_wb_length
  591.        {align1};
  592.  
  593.  
  594. /* Write FME/BME MV */
  595. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
  596. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  597.  
  598.  
  599. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  600. mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  601. mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  602. mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  603. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  604. send (16)
  605.        msg_ind
  606.        obw_wb
  607.        null
  608.        data_port(
  609.                OBW_CACHE_TYPE,
  610.                OBW_MESSAGE_TYPE,
  611.                OBW_CONTROL_2,
  612.                OBW_BIND_IDX,
  613.                OBW_WRITE_COMMIT_CATEGORY,
  614.                OBW_HEADER_PRESENT
  615.        )
  616.        mlen 2
  617.        rlen obw_wb_length
  618.        {align1};
  619.  
  620. /* Write FME/BME RefID */
  621. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
  622. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  623.  
  624. mov  (8) msg_reg1.0<1>:UD       vme_wb6.0<8,8,1>:UD {align1};
  625.  
  626. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  627. send (16)
  628.        msg_ind
  629.        obw_wb
  630.        null
  631.        data_port(
  632.                OBW_CACHE_TYPE,
  633.                OBW_MESSAGE_TYPE,
  634.                OBW_CONTROL_2,
  635.                OBW_BIND_IDX,
  636.                OBW_WRITE_COMMIT_CATEGORY,
  637.                OBW_HEADER_PRESENT
  638.        )
  639.        mlen 2
  640.        rlen obw_wb_length
  641.        {align1};
  642.  
  643.  
  644. /* Issue message fence so that the previous write message is committed */
  645. send (16)
  646.        mb_ind
  647.        mb_wb.0<1>:ud
  648.         NULL
  649.        data_port(
  650.                OBR_CACHE_TYPE,
  651.                OBR_MESSAGE_FENCE,
  652.                OBR_MF_COMMIT,
  653.                OBR_BIND_IDX,
  654.                OBR_WRITE_COMMIT_CATEGORY,
  655.                OBR_HEADER_PRESENT
  656.        )
  657.        mlen 1
  658.        rlen 1
  659.        {align1};
  660.  
  661. __EXIT:
  662. /*
  663. * kill thread
  664. */        
  665. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  666. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  667.  
  668.         nop             ;
  669.         nop             ;
  670.        
  671. word_imedian:
  672.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  673.         (f0.0)  jmpi (1) cmp_a_ge_b;
  674.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  675.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  676.         (f0.0) jmpi (1) cmp_end;
  677.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  678.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  679.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  680.         jmpi (1) cmp_end;
  681. cmp_a_ge_b:
  682.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  683.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  684.         (f0.0) jmpi (1) cmp_end;
  685.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  686.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  687.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  688. cmp_end:
  689.         RETURN  {align1};
  690.  
  691. nop;
  692. nop;
  693.  
  694. ref_boundary_check:
  695.  
  696. /* The left/up coordinate of reference window */
  697. add  (2) TEMP_VAR0.0<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG0.0<2,2,1>:w   {align1};
  698. /* The right/bottom coordinate of reference window */
  699. add  (1) TEMP_VAR0.16<1>:w      TEMP_VAR0.0<0,1,0>:w    48:w                    {align1};
  700. add  (1) TEMP_VAR0.18<1>:w      TEMP_VAR0.2<0,1,0>:w    40:w                    {align1};
  701.  
  702. /* Firstly the MV range is checked */
  703. mul  (2) TEMP_VAR1.16<1>:w      INPUT_ARG1.16<2,2,1>:w  -1:w    {align1};
  704. add  (2) TEMP_VAR1.0<1>:w       INPUT_ARG0.8<2,2,1>:w   TEMP_VAR1.16<2,2,1>:w   {align1};
  705. add  (2) TEMP_VAR1.4<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG1.16<2,2,1>:w  {align1};
  706.  
  707. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.0<0,1,0>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  708. (f0.0)  mov     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  709. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   TEMP_VAR1.4<0,1,0>:w    {align1};
  710. (f0.0)  add     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.4<0,1,0>:w    -48:w   {align1};
  711. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.2<0,1,0>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  712. (f0.0)  mov     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  713. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.18<0,1,0>:w   TEMP_VAR1.6<0,1,0>:w    {align1};
  714. (f0.0)  add     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.6<0,1,0>:w    -40:w   {align1};
  715.  
  716. x_left_cmp:
  717.         cmp.l.f0.0 (1)          null:w          TEMP_VAR0.0<0,1,0>:w    0:w     {align1};
  718.         (-f0.0) jmpi    (1)     x_right_cmp;
  719.         (f0.0)  mov     (1)     TEMP_VAR0.0<1>:w        0:w             {align1};
  720.         jmpi    (1)     y_top_cmp;
  721. x_right_cmp:
  722.         cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   INPUT_ARG1.0<0,1,0>:w   {align1};
  723.         (-f0.0) jmpi    (1)     y_top_cmp;     
  724.         (f0.0)  add     (1)     TEMP_VAR0.0<1>:w        INPUT_ARG1.0<0,1,0>:w   -48:w   {align1};
  725. y_top_cmp:
  726.         cmp.l.f0.0  (1)         null:w  TEMP_VAR0.2<0,1,0>:w    0:w             {align1};
  727.         (-f0.0) jmpi    (1)     y_bottom_cmp;
  728.         (f0.0)  mov     (1)     TEMP_VAR0.2<1>:w        0:w             {align1};
  729.         jmpi    (1)     y_bottom_end;
  730. y_bottom_cmp:
  731.         cmp.g.f0.0  (1)         null:w  TEMP_VAR0.18<0,1,0>:w   INPUT_ARG1.2<0,1,0>:w {align1};
  732.         (f0.0)  add     (1)     TEMP_VAR0.2<1>:w                INPUT_ARG1.2<0,1,0>:w   -40:w   {align1};
  733.  
  734. y_bottom_end:
  735. mul     (2)     TEMP_VAR1.0<1>:w        INPUT_ARG0.8<2,2,1>:w   -1:w    {align1};
  736. add     (2)     RET_ARG<1>:w            TEMP_VAR0.0<2,2,1>:w    TEMP_VAR1.0<2,2,1>:w    {align1};      
  737.         RETURN  {align1};
  738. nop;
  739. nop;
  740.  
  741. vme_run_again:
  742.  
  743. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  744. mov     (2)     tmp_reg0.0<1>:w         mb_ref_win.0<2,2,1>:w           {align1};
  745. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  746. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  747.  
  748. cmp.l.f0.0      (1) null:w      tmp_reg0.0<0,1,0>:w     0:w     {align1};
  749. (f0.0)  mul     (1) tmp_reg0.0<1>:w     tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  750. cmp.l.f0.0      (1) null:w      tmp_reg0.2<0,1,0>:w     0:w     {align1};
  751. (f0.0)  mul     (1) tmp_reg0.2<1>:w     tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  752.  
  753. cmp.ge.f0.0     (1) null:w      tmp_reg0.0<0,1,0>:w     4:w     {align1};
  754. (f0.0)  jmpi (1)        vme_start;
  755. cmp.ge.f0.0     (1) null:w      tmp_reg0.2<0,1,0>:w     4:w     {align1};
  756. (f0.0)  jmpi (1)        vme_start;
  757.  
  758. jmpi (1) vme_done;
  759.  
  760. vme_start:
  761.         mov (8) tmp_vme_wb0.0<1>:ud     vme_wb0.0<8,8,1>:ud     {align1};
  762.         mov (8) tmp_vme_wb1.0<1>:ud     vme_wb1.0<8,8,1>:ud     {align1};
  763.  
  764. /* Calibrate the ref window for MPEG2 */
  765. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  766. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  767. mov  (4) INPUT_ARG0.0<1>:ud     vme_m0.0<4,4,1>:ud      {align1};
  768. add  (2) INPUT_ARG0.0<1>:w      INPUT_ARG0.0<2,2,1>:w   mb_ref_win.16<2,2,1>:w  {align1};
  769. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  770.  
  771. SAVE_RET        {align1};
  772. jmpi    (1)     ref_boundary_check;
  773. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  774.  
  775. /* IME search */
  776. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  777. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  778.  
  779. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  780.  
  781. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  782.  
  783. mov  (8) vme_m1.0<1>:ud         0x0:UD  {align1};
  784.  
  785. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  786. /* the Max MV number is passed by constant buffer */
  787. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  788. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  789. /* Set the MV cost center */
  790. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  791. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  792. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  793.  
  794. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  795. /* M3/M4 search path */
  796.  
  797. mov  (1) vme_msg_3.0<1>:UD      0x01010101:UD {align1};
  798. mov  (1) vme_msg_3.4<1>:UD      0x10010101:UD {align1};
  799. mov  (1) vme_msg_3.8<1>:UD      0x0F0F0F0F:UD {align1};
  800. mov  (1) vme_msg_3.12<1>:UD     0x100F0F0F:UD {align1};
  801. mov  (1) vme_msg_3.16<1>:UD     0x01010101:UD {align1};
  802. mov  (1) vme_msg_3.20<1>:UD     0x10010101:UD {align1};
  803. mov  (1) vme_msg_3.24<1>:UD     0x0F0F0F0F:UD {align1};
  804. mov  (1) vme_msg_3.28<1>:UD     0x100F0F0F:UD {align1};
  805.  
  806. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  807. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  808. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  809. mov  (1) vme_msg_4.12<1>:UD     0x000F0F0F:UD {align1};
  810.  
  811. mov  (4) vme_msg_4.16<1>:UD     0x0:UD {align1};
  812.  
  813. send (8)
  814.        vme_msg_ind
  815.        vme_wb<1>:UD
  816.        null
  817.        vme(
  818.                BIND_IDX_VME,
  819.                0,
  820.                0,
  821.                VME_IME_MESSAGE_TYPE
  822.        )
  823.        mlen ime_vme_msg_length
  824.        rlen vme_wb_length {align1};
  825.  
  826. /* Set Macroblock-shape/mode for FBR */
  827.  
  828. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  829. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  830. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  831.  
  832. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  833. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  834.  
  835. /* Send FBR message into CRE */
  836.  
  837. mov  (8) vme_msg_3.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  838. mov  (8) vme_msg_4.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  839. mov  (8) vme_msg_5.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  840. mov  (8) vme_msg_6.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  841.  
  842. 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 */
  843.  
  844. /* Bilinear filter */
  845. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  846. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  847.  
  848. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  849. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  850.  
  851. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  852.  
  853. /* after verification it will be passed by using payload */
  854. send (8)
  855.        vme_msg_ind
  856.        vme_wb<1>:UD
  857.        null
  858.        cre(
  859.                BIND_IDX_VME,
  860.                VME_FBR_MESSAGE_TYPE
  861.        )
  862.        mlen fbr_vme_msg_length
  863.        rlen vme_wb_length
  864.        {align1};
  865.  
  866. cmp.l.f0.0 (1)  null:uw vme_wb0.8<0,1,0>:uw     tmp_vme_wb0.8<0,1,0>:uw {align1};
  867. (f0.0)  jmpi (1) vme_done;
  868. mov     (8)     vme_wb0.0<1>:ud tmp_vme_wb0.0<8,8,1>:ud {align1};
  869. mov     (8)     vme_wb1.0<1>:ud tmp_vme_wb1.0<8,8,1>:ud {align1};
  870.  
  871. vme_done:
  872.       jmpi (1) vme_mv_output;
  873. nop;
  874. nop;
  875. nop;
  876.  
  877.