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.  * Authors: Zhao Yakui <yakui.zhao@intel.com>
  24.  *
  25.  */
  26. // Modual name: mpeg2_inter_ivb.asm
  27. //
  28. // Make inter predition estimation for Mpeg2 Inter frame on Ivy
  29. //
  30.  
  31. //
  32. //  Now, begin source code....
  33. //
  34.  
  35. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  36. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  37.  
  38. /*
  39.  * __START
  40.  */
  41. __INTER_START:
  42. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  43. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  44. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  45. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  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 INTER_VME_OUTPUT_IN_OWS:UD {align1};
  53. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  54.        
  55. shl     (2) pic_ref.0<1>:uw     r4.24<2,2,1>:uw 4:uw    {align1};
  56. mov     (2) pic_ref.16<1>:uw    r4.20<2,2,1>:uw {align1};
  57.  
  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. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  64. (f0.0) jmpi (1) __mb_hwdep_end;
  65. /* read back the data for MB A */
  66. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  67. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  68. */
  69. mba_start:
  70. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  71. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  72. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  73. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  74. (f0.0)  jmpi (1)        mbb_start;
  75. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  76. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  77. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  78. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  79. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  80. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  81. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  82. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  83.  
  84. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  85. /* bind index 3, read 1 oword (16bytes), msg type: 0(OWord Block Read) */
  86. send (16)
  87.        mb_ind
  88.        mb_wb.0<1>:ud
  89.         NULL
  90.        data_port(
  91.                OBR_CACHE_TYPE,
  92.                OBR_MESSAGE_TYPE,
  93.                OBR_CONTROL_0,
  94.                OBR_BIND_IDX,
  95.                OBR_WRITE_COMMIT_CATEGORY,
  96.                OBR_HEADER_PRESENT
  97.        )
  98.        mlen 1
  99.        rlen 1
  100.        {align1};
  101.  
  102. /* TODO: RefID is required after multi-references are added */
  103. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  104. (-f0.0)   mov (2)       mba_result.20<1>:w                      -1:w    {align1};
  105. (-f0.0)   jmpi  (1)     mbb_start;
  106.  
  107. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  108. /* Read MV for MB A */
  109. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  110. send (16)
  111.        mb_ind
  112.        mb_mv0.0<1>:ud
  113.         NULL
  114.        data_port(
  115.                OBR_CACHE_TYPE,
  116.                OBR_MESSAGE_TYPE,
  117.                OBR_CONTROL_2,
  118.                OBR_BIND_IDX,
  119.                OBR_WRITE_COMMIT_CATEGORY,
  120.                OBR_HEADER_PRESENT
  121.        )
  122.        mlen 1
  123.        rlen 1
  124.        {align1};
  125. /* TODO: RefID is required after multi-references are added */
  126. /* MV */
  127. mov        (2)          mba_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  128. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  129.  
  130. mbb_start:
  131. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  132. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  133. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  134. /* If MB B doesn't exist, neither MB C nor D exists */
  135. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  136. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  137. (f0.0)  jmpi (1)        mb_mvp_start;
  138. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  139. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  140. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  141. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  142. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  143. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  144. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  145. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  146.  
  147. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  148.  
  149. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  150. send (16)
  151.        mb_ind
  152.        mb_wb.0<1>:ud
  153.         NULL
  154.        data_port(
  155.                OBR_CACHE_TYPE,
  156.                OBR_MESSAGE_TYPE,
  157.                OBR_CONTROL_0,
  158.                OBR_BIND_IDX,
  159.                OBR_WRITE_COMMIT_CATEGORY,
  160.                OBR_HEADER_PRESENT
  161.        )
  162.        mlen 1
  163.        rlen 1
  164.        {align1};
  165.  
  166. /* TODO: RefID is required after multi-references are added */
  167. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  168. (-f0.0)   mov (2)       mbb_result.20<1>:w                      -1:w    {align1};
  169. (-f0.0)   jmpi  (1)     mbc_start;
  170.  
  171. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  172. /* Read MV for MB B */
  173. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  174. send (16)
  175.        mb_ind
  176.        mb_mv0.0<1>:ud
  177.         NULL
  178.        data_port(
  179.                OBR_CACHE_TYPE,
  180.                OBR_MESSAGE_TYPE,
  181.                OBR_CONTROL_2,
  182.                OBR_BIND_IDX,
  183.                OBR_WRITE_COMMIT_CATEGORY,
  184.                OBR_HEADER_PRESENT
  185.        )
  186.        mlen 1
  187.        rlen 1
  188.        {align1};
  189. /* TODO: RefID is required after multi-references are added */
  190. mov        (2)          mbb_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  191. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  192.  
  193. mbc_start:
  194. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  195. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  196. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  197. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  198. (f0.0)  jmpi (1)        mbd_start;
  199. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  200. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  201. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  202. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  203. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  204. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  205. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  206. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  207.  
  208. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  209.  
  210. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  211. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  212. send (16)
  213.        mb_ind
  214.        mb_wb.0<1>:ud
  215.         NULL
  216.        data_port(
  217.                OBR_CACHE_TYPE,
  218.                OBR_MESSAGE_TYPE,
  219.                OBR_CONTROL_0,
  220.                OBR_BIND_IDX,
  221.                OBR_WRITE_COMMIT_CATEGORY,
  222.                OBR_HEADER_PRESENT
  223.        )
  224.        mlen 1
  225.        rlen 1
  226.        {align1};
  227.  
  228. /* TODO: RefID is required after multi-references are added */
  229. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  230. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  231. (-f0.0)   jmpi  (1)     mb_mvp_start;
  232. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud {align1};
  233. /* Read MV for MB C */
  234. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  235. send (16)
  236.        mb_ind
  237.        mb_mv0.0<1>:ud
  238.         NULL
  239.        data_port(
  240.                OBR_CACHE_TYPE,
  241.                OBR_MESSAGE_TYPE,
  242.                OBR_CONTROL_2,
  243.                OBR_BIND_IDX,
  244.                OBR_WRITE_COMMIT_CATEGORY,
  245.                OBR_HEADER_PRESENT
  246.        )
  247.        mlen 1
  248.        rlen 1
  249.        {align1};
  250. /* TODO: RefID is required after multi-references are added */
  251. /* Forward MV */
  252. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  253. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  254.  
  255. jmpi   (1)    mb_mvp_start;
  256. mbd_start:
  257. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  258. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  259. (f0.0)  jmpi (1)        mb_mvp_start;
  260. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  261. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  262. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  263. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  264. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  265.  
  266. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  267. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  268. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  269.  
  270. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  271. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  272. send (16)
  273.        mb_ind
  274.        mb_wb.0<1>:ud
  275.         NULL
  276.        data_port(
  277.                OBR_CACHE_TYPE,
  278.                OBR_MESSAGE_TYPE,
  279.                OBR_CONTROL_0,
  280.                OBR_BIND_IDX,
  281.                OBR_WRITE_COMMIT_CATEGORY,
  282.                OBR_HEADER_PRESENT
  283.        )
  284.        mlen 1
  285.        rlen 1
  286.        {align1};
  287.  
  288. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  289. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  290. (-f0.0)   jmpi  (1)     mb_mvp_start;
  291.  
  292. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  293. /* Read MV for MB D */
  294. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  295. send (16)
  296.        mb_ind
  297.        mb_mv0.0<1>:ub
  298.         NULL
  299.        data_port(
  300.                OBR_CACHE_TYPE,
  301.                OBR_MESSAGE_TYPE,
  302.                OBR_CONTROL_2,
  303.                OBR_BIND_IDX,
  304.                OBR_WRITE_COMMIT_CATEGORY,
  305.                OBR_HEADER_PRESENT
  306.        )
  307.        mlen 1
  308.        rlen 1
  309.        {align1};
  310.  
  311. /* TODO: RefID is required after multi-references are added */
  312.  
  313. /* Forward MV */
  314. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  315. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  316.        
  317. mb_mvp_start:
  318. /*TODO: Add the skip prediction */
  319. /* Check whether both MB B and C are invailable */
  320. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  321. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  322. (-f0.0) jmpi (1)        mb_median_start;
  323. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  324. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  325. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  326. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  327. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  328. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  329. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  330. jmpi    (1)     __mb_hwdep_end;
  331.        
  332. mb_median_start:
  333. /* check whether only one neighbour MB has the same ref ID with the current MB */
  334. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  335. cmp.z.f0.0      (1)     null:d  mba_result.20<1>:w      0:w     {align1};
  336. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  337. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  338. cmp.z.f0.0      (1)     null:d  mbb_result.20<1>:w      0:w     {align1};
  339. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  340. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  341. cmp.z.f0.0      (1)     null:d  mbc_result.20<1>:w      0:w     {align1};
  342. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  343. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  344. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  345. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  346. (f0.0)  jmpi (1)  __mb_hwdep_end;
  347.  
  348. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  349. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  350. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  351. SAVE_RET        {align1};
  352.  jmpi   (1)     word_imedian;
  353. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  354. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  355. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  356. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  357. SAVE_RET        {align1};
  358. jmpi    (1)     word_imedian;
  359. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  360.  
  361. __mb_hwdep_end:
  362.  
  363. mov     (2)     mv_cc_ref.0<1>:w        mba_result.4<2,2,1>:w   {align1};
  364.  
  365. /* Calibrate the ref window for MPEG2 */
  366. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  367. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  368. mov  (1) INPUT_ARG0.0<1>:ud     vme_m0.0<0,1,0>:ud      {align1};
  369. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  370. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  371.  
  372. SAVE_RET        {align1};
  373. jmpi    (1)     ref_boundary_check;
  374. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  375.  
  376. /* m2 */        
  377. mov  (8) vme_msg_2<1>:UD        0x0:UD {align1};
  378.  
  379. /* m3 */
  380. mov  (8) vme_msg_3<1>:UD        0x0:UD {align1};        
  381. /* Use the Luma mode */
  382. mov  (1) tmp_reg0.0<1>:UW       INTRA16_DC_PRED:UW {align1};
  383. mov  (1) vme_msg_3.4<1>:ub      tmp_reg0.0<0,1,0>:UB {align1};
  384.  
  385. /* m4 */
  386. mov  (8) vme_msg_4<1>:UD        0x0:UD {align1};
  387. mov  (1) vme_msg_4.16<1>:UD     INTRA_PREDICTORE_MODE {align1};
  388.  
  389.  
  390. /* m1 */
  391. mov  (8) vme_m1.0<1>:ud         0x0:ud  {align1};
  392. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  393. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  394. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  395. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  396. /* m1 */
  397. /* assign MB intra struct from the thread payload*/
  398. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  399.  
  400.  
  401. /* M0 */
  402. /* IME search */
  403. mov  (1) vme_m0.12<1>:UD   SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_HALF:UD {align1};  
  404. /* 16x16 Source, 1/2 pixel, harr */
  405. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  406.  
  407. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  408. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  409. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  410. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  411.  
  412. /* m1 */
  413.  
  414. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  415. /* MV num is passed by constant buffer. R4.28 */
  416. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};
  417. add  (1) vme_m1.4<1>:UD         vme_m1.4<0,1,0>:UD      FB_PRUNING_DISABLE:UD {align1};
  418. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  419.  
  420. /* Bilinear filter */
  421. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  422. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  423.  
  424. /* Set the MV cost center */
  425. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  426. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  427. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  428.  
  429.  
  430. send (8)
  431.        vme_msg_ind
  432.        vme_wb
  433.        null
  434.        vme(
  435.                BIND_IDX_VME,
  436.                0,
  437.                0,
  438.                VME_MESSAGE_TYPE_MIXED
  439.        )
  440.        mlen vme_msg_length
  441.        rlen vme_inter_wb_length
  442.        {align1};
  443.  
  444. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  445. (-f0.0) jmpi (1) vme_run_again;
  446.  
  447. vme_mv_output:
  448.  
  449. and.z.f0.0      (1)     null<1>:ud              vme_wb0.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  450.  
  451. (-f0.0)jmpi     (1)     __INTRA_INFO ;
  452.  
  453. __INTER_INFO:
  454. /* Write MV pairs */   
  455. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  456.  
  457. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD   {align1};
  458.  
  459. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  460. send (16)
  461.        msg_ind
  462.        obw_wb
  463.        null
  464.        data_port(
  465.                OBW_CACHE_TYPE,
  466.                OBW_MESSAGE_TYPE,
  467.                OBW_CONTROL_2,
  468.                OBW_BIND_IDX,
  469.                OBW_WRITE_COMMIT_CATEGORY,
  470.                OBW_HEADER_PRESENT
  471.        )
  472.        mlen 2
  473.        rlen obw_wb_length
  474.        {align1};
  475.  
  476.  
  477. mov             (1)     msg_reg1.0<1>:ud        vme_wb0.0<0,1,0>:ud     {align1} ;
  478. mov             (1)     msg_reg1.4<1>:UD        vme_wb0.28<0,1,0>:UD    {align1};
  479. mov             (1)     msg_reg1.8<1>:ud        tmp_ud1<0,1,0>:ud       {align1} ;
  480. mov             (1)     msg_reg1.12<1>:ud        vme_wb0.0<0,1,0>:ud     {align1} ;
  481. mov             (1)     msg_reg1.16<1>:ud        0x25:ud     {align1} ;
  482. jmpi            (1)     __OUTPUT_INFO;
  483.  
  484. __INTRA_INFO:
  485. mov             (1)     msg_reg1.0<1>:UD        vme_wb.0<0,1,0>:UD      {align1};
  486. mov             (1)     msg_reg1.4<1>:UD        vme_wb.16<0,1,0>:UD     {align1};
  487. mov             (1)     msg_reg1.8<1>:UD        vme_wb.20<0,1,0>:UD     {align1};
  488. mov             (1)     msg_reg1.12<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  489. mov             (1)     msg_reg1.16<1>:ud        0x35:ud     {align1} ;
  490.  
  491. __OUTPUT_INFO:
  492.  
  493. mov     (1)     msg_reg1.20<1>:ud       obw_m0.8<0,1,0>:ud      {align1};
  494. add     (1)     obw_m0.8<1>:UD       obw_m0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  495. mov     (8)     msg_reg0.0<1>:ud        obw_m0.0<8,8,1>:ud      {align1};
  496.  
  497.  
  498. /* bind index 3, write 1 oword, msg type: 8(OWord Block Write) */
  499. send (16)
  500.        msg_ind
  501.        obw_wb
  502.        null
  503.        data_port(
  504.                OBW_CACHE_TYPE,
  505.                OBW_MESSAGE_TYPE,
  506.                OBW_CONTROL_2,
  507.                OBW_BIND_IDX,
  508.                OBW_WRITE_COMMIT_CATEGORY,
  509.                OBW_HEADER_PRESENT
  510.        )
  511.        mlen 2
  512.        rlen obw_wb_length
  513.        {align1};
  514.  
  515. /* Issue message fence so that the previous write message is committed */
  516. send (16)
  517.        mb_ind
  518.        mb_wb.0<1>:ud
  519.         NULL
  520.        data_port(
  521.                OBR_CACHE_TYPE,
  522.                OBR_MESSAGE_FENCE,
  523.                OBR_MF_COMMIT,
  524.                OBR_BIND_IDX,
  525.                OBR_WRITE_COMMIT_CATEGORY,
  526.                OBR_HEADER_PRESENT
  527.        )
  528.        mlen 1
  529.        rlen 1
  530.         {align1};
  531.  
  532. __EXIT:
  533. /*
  534. * kill thread
  535. */        
  536. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  537. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  538.  
  539.  
  540.         nop             ;
  541.         nop             ;
  542.        
  543. word_imedian:
  544.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  545.         (f0.0)  jmpi (1) cmp_a_ge_b;
  546.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  547.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  548.         (f0.0) jmpi (1) cmp_end;
  549.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  550.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  551.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  552.         jmpi (1) cmp_end;
  553. cmp_a_ge_b:
  554.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  555.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  556.         (f0.0) jmpi (1) cmp_end;
  557.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  558.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  559.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  560. cmp_end:
  561.         RETURN  {align1};
  562.  
  563. nop;
  564. nop;
  565.  
  566. ref_boundary_check:
  567.  
  568. /* The left/up coordinate of reference window */
  569. add  (2) TEMP_VAR0.0<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG0.0<2,2,1>:w   {align1};
  570. /* The right/bottom coordinate of reference window */
  571. add  (1) TEMP_VAR0.16<1>:w      TEMP_VAR0.0<0,1,0>:w    48:w                    {align1};
  572. add  (1) TEMP_VAR0.18<1>:w      TEMP_VAR0.2<0,1,0>:w    40:w                    {align1};
  573.  
  574. /* Firstly the MV range is checked */
  575. mul  (2) TEMP_VAR1.16<1>:w      INPUT_ARG1.16<2,2,1>:w  -1:w    {align1};
  576. add  (2) TEMP_VAR1.0<1>:w       INPUT_ARG0.8<2,2,1>:w   TEMP_VAR1.16<2,2,1>:w   {align1};
  577. add  (2) TEMP_VAR1.4<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG1.16<2,2,1>:w  {align1};
  578.  
  579. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.0<0,1,0>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  580. (f0.0)  mov     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  581. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   TEMP_VAR1.4<0,1,0>:w    {align1};
  582. (f0.0)  add     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.4<0,1,0>:w    -48:w   {align1};
  583. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.2<0,1,0>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  584. (f0.0)  mov     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  585. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.18<0,1,0>:w   TEMP_VAR1.6<0,1,0>:w    {align1};
  586. (f0.0)  add     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.6<0,1,0>:w    -40:w   {align1};
  587.  
  588.  
  589. x_left_cmp:
  590.         cmp.l.f0.0 (1)          null:w          TEMP_VAR0.0<0,1,0>:w    0:w     {align1};
  591.         (-f0.0) jmpi    (1)     x_right_cmp;
  592.         (f0.0)  mov     (1)     TEMP_VAR0.0<1>:w        0:w             {align1};
  593.         jmpi    (1)     y_top_cmp;
  594. x_right_cmp:
  595.         cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   INPUT_ARG1.0<0,1,0>:w   {align1};
  596.         (-f0.0) jmpi    (1)     y_top_cmp;     
  597.         (f0.0)  add     (1)     TEMP_VAR0.0<1>:w        INPUT_ARG1.0<0,1,0>:w   -48:w   {align1};
  598. y_top_cmp:
  599.         cmp.l.f0.0  (1)         null:w  TEMP_VAR0.2<0,1,0>:w    0:w             {align1};
  600.         (-f0.0) jmpi    (1)     y_bottom_cmp;
  601.         (f0.0)  mov     (1)     TEMP_VAR0.2<1>:w        0:w             {align1};
  602.         jmpi    (1)     y_bottom_end;
  603. y_bottom_cmp:
  604.         cmp.g.f0.0  (1)         null:w  TEMP_VAR0.18<0,1,0>:w   INPUT_ARG1.2<0,1,0>:w {align1};
  605.         (f0.0)  add     (1)     TEMP_VAR0.2<1>:w                INPUT_ARG1.2<0,1,0>:w   -40:w   {align1};
  606.  
  607. y_bottom_end:
  608. mul     (2)     TEMP_VAR1.0<1>:w        INPUT_ARG0.8<2,2,1>:w   -1:w    {align1};
  609. add     (2)     RET_ARG<1>:w            TEMP_VAR0.0<2,2,1>:w    TEMP_VAR1.0<2,2,1>:w    {align1};      
  610.         RETURN  {align1};
  611. nop;
  612. nop;
  613.  
  614. vme_run_again:
  615.  
  616. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  617. mov     (2)     tmp_reg0.0<1>:w         mb_ref_win.0<2,2,1>:w           {align1};
  618. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  619. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  620.  
  621. cmp.l.f0.0      (1) null:w      tmp_reg0.0<0,1,0>:w     0:w     {align1};
  622. (f0.0)  mul     (1) tmp_reg0.0<1>:w     tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  623. cmp.l.f0.0      (1) null:w      tmp_reg0.2<0,1,0>:w     0:w     {align1};
  624. (f0.0)  mul     (1) tmp_reg0.2<1>:w     tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  625.  
  626. cmp.ge.f0.0     (1) null:w      tmp_reg0.0<0,1,0>:w     4:w     {align1};
  627. (f0.0)  jmpi (1)        vme_start;
  628. cmp.ge.f0.0     (1) null:w      tmp_reg0.2<0,1,0>:w     4:w     {align1};
  629. (f0.0)  jmpi (1)        vme_start;
  630.  
  631. jmpi (1) vme_done;
  632.  
  633. vme_start:
  634.         mov (8) tmp_vme_wb0.0<1>:ud     vme_wb0.0<8,8,1>:ud     {align1};
  635.         mov (8) tmp_vme_wb1.0<1>:ud     vme_wb1.0<8,8,1>:ud     {align1};
  636.  
  637. /* Calibrate the ref window for MPEG2 */
  638. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  639. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  640. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  641. add  (2) INPUT_ARG0.0<1>:w      vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  642. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  643.  
  644. SAVE_RET        {align1};
  645. jmpi    (1)     ref_boundary_check;
  646. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  647.  
  648. /* m2 */        
  649. mov  (8) vme_msg_2<1>:UD        0x0:UD {align1};
  650.  
  651. /* m3 */
  652. mov  (8) vme_msg_3<1>:UD        0x0:UD {align1};        
  653.  
  654. /* m4 */
  655. mov  (8) vme_msg_4<1>:UD        0x0:UD {align1};
  656.  
  657.  
  658. /* m1 */
  659. mov  (8) vme_m1.0<1>:ud         0x0:ud  {align1};
  660. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  661. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  662. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  663. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  664. /* m1 */
  665. /* assign MB intra struct from the thread payload*/
  666. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  667.  
  668.  
  669. /* M0 */
  670. /* IME search */
  671. mov  (1) vme_m0.12<1>:UD   SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_HALF:UD {align1};  
  672. /* 16x16 Source, 1/2 pixel, harr */
  673. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  674.  
  675. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  676. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  677.  
  678. /* m1 */
  679.  
  680. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  681. /* MV num is passed by constant buffer. R4.28 */
  682. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};
  683. add  (1) vme_m1.4<1>:UD         vme_m1.4<0,1,0>:UD      FB_PRUNING_DISABLE:UD {align1};
  684. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  685.  
  686. /* Bilinear filter */
  687. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  688. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  689.  
  690. /* Set the MV cost center */
  691. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  692. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  693. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  694.  
  695.  
  696. send (8)
  697.        vme_msg_ind
  698.        vme_wb
  699.        null
  700.        vme(
  701.                BIND_IDX_VME,
  702.                0,
  703.                0,
  704.                VME_MESSAGE_TYPE_INTER
  705.        )
  706.        mlen vme_msg_length
  707.        rlen vme_inter_wb_length
  708.        {align1};
  709.  
  710.  
  711. cmp.l.f0.0 (1)  null:uw vme_wb0.6<0,1,0>:uw     tmp_vme_wb0.6<0,1,0>:uw {align1};
  712. (f0.0)  jmpi (1) vme_done;
  713. mov     (8)     vme_wb0.0<1>:ud tmp_vme_wb0.0<8,8,1>:ud {align1};
  714. mov     (8)     vme_wb1.0<1>:ud tmp_vme_wb1.0<8,8,1>:ud {align1};
  715.  
  716. vme_done:
  717.       jmpi (1) vme_mv_output;
  718. nop;
  719. nop;
  720. nop;
  721.  
  722.