Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © <2010>, Intel Corporation.
  3.  *
  4.  * This program is licensed under the terms and conditions of the
  5.  * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at
  6.  * http://www.opensource.org/licenses/eclipse-1.0.php.
  7.  * Authors: Zhao Yakui <yakui.zhao@intel.com>
  8.  *
  9.  */
  10. // Modual name: mpeg2_inter_ivb.asm
  11. //
  12. // Make inter predition estimation for Mpeg2 Inter frame on Ivy
  13. //
  14.  
  15. //
  16. //  Now, begin source code....
  17. //
  18.  
  19. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  20. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  21.  
  22. /*
  23.  * __START
  24.  */
  25. __INTER_START:
  26. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  27. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  28. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  29. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  30.  
  31. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  32. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  33.  
  34. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  35. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  36. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  37. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  38.        
  39. shl     (2) pic_ref.0<1>:uw     r4.24<2,2,1>:uw 4:uw    {align1};
  40. mov     (2) pic_ref.16<1>:uw    r4.20<2,2,1>:uw {align1};
  41.  
  42. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  43. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  44. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  45. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  46. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  47. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  48. (f0.0) jmpi (1) __mb_hwdep_end;
  49. /* read back the data for MB A */
  50. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  51. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  52. */
  53. mba_start:
  54. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  55. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  56. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  57. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  58. (f0.0)  jmpi (1)        mbb_start;
  59. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  60. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  61. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  62. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  63. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  64. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  65. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  66. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  67.  
  68. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  69. /* bind index 3, read 1 oword (16bytes), msg type: 0(OWord Block Read) */
  70. send (16)
  71.        mb_ind
  72.        mb_wb.0<1>:ud
  73.         NULL
  74.        data_port(
  75.                OBR_CACHE_TYPE,
  76.                OBR_MESSAGE_TYPE,
  77.                OBR_CONTROL_0,
  78.                OBR_BIND_IDX,
  79.                OBR_WRITE_COMMIT_CATEGORY,
  80.                OBR_HEADER_PRESENT
  81.        )
  82.        mlen 1
  83.        rlen 1
  84.        {align1};
  85.  
  86. /* TODO: RefID is required after multi-references are added */
  87. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  88. (-f0.0)   mov (2)       mba_result.20<1>:w                      -1:w    {align1};
  89. (-f0.0)   jmpi  (1)     mbb_start;
  90.  
  91. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  92. /* Read MV for MB A */
  93. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  94. send (16)
  95.        mb_ind
  96.        mb_mv0.0<1>:ud
  97.         NULL
  98.        data_port(
  99.                OBR_CACHE_TYPE,
  100.                OBR_MESSAGE_TYPE,
  101.                OBR_CONTROL_2,
  102.                OBR_BIND_IDX,
  103.                OBR_WRITE_COMMIT_CATEGORY,
  104.                OBR_HEADER_PRESENT
  105.        )
  106.        mlen 1
  107.        rlen 1
  108.        {align1};
  109. /* TODO: RefID is required after multi-references are added */
  110. /* MV */
  111. mov        (2)          mba_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  112. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  113.  
  114. mbb_start:
  115. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  116. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  117. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  118. /* If MB B doesn't exist, neither MB C nor D exists */
  119. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  120. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  121. (f0.0)  jmpi (1)        mb_mvp_start;
  122. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  123. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  124. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  125. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  126. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  127. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  128. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  129. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  130.  
  131. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  132.  
  133. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  134. send (16)
  135.        mb_ind
  136.        mb_wb.0<1>:ud
  137.         NULL
  138.        data_port(
  139.                OBR_CACHE_TYPE,
  140.                OBR_MESSAGE_TYPE,
  141.                OBR_CONTROL_0,
  142.                OBR_BIND_IDX,
  143.                OBR_WRITE_COMMIT_CATEGORY,
  144.                OBR_HEADER_PRESENT
  145.        )
  146.        mlen 1
  147.        rlen 1
  148.        {align1};
  149.  
  150. /* TODO: RefID is required after multi-references are added */
  151. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  152. (-f0.0)   mov (2)       mbb_result.20<1>:w                      -1:w    {align1};
  153. (-f0.0)   jmpi  (1)     mbc_start;
  154.  
  155. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  156. /* Read MV for MB B */
  157. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  158. send (16)
  159.        mb_ind
  160.        mb_mv0.0<1>:ud
  161.         NULL
  162.        data_port(
  163.                OBR_CACHE_TYPE,
  164.                OBR_MESSAGE_TYPE,
  165.                OBR_CONTROL_2,
  166.                OBR_BIND_IDX,
  167.                OBR_WRITE_COMMIT_CATEGORY,
  168.                OBR_HEADER_PRESENT
  169.        )
  170.        mlen 1
  171.        rlen 1
  172.        {align1};
  173. /* TODO: RefID is required after multi-references are added */
  174. mov        (2)          mbb_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  175. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  176.  
  177. mbc_start:
  178. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  179. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  180. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  181. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  182. (f0.0)  jmpi (1)        mbd_start;
  183. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  184. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  185. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  186. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  187. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  188. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  189. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  190. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  191.  
  192. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  193.  
  194. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  195. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  196. send (16)
  197.        mb_ind
  198.        mb_wb.0<1>:ud
  199.         NULL
  200.        data_port(
  201.                OBR_CACHE_TYPE,
  202.                OBR_MESSAGE_TYPE,
  203.                OBR_CONTROL_0,
  204.                OBR_BIND_IDX,
  205.                OBR_WRITE_COMMIT_CATEGORY,
  206.                OBR_HEADER_PRESENT
  207.        )
  208.        mlen 1
  209.        rlen 1
  210.        {align1};
  211.  
  212. /* TODO: RefID is required after multi-references are added */
  213. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  214. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  215. (-f0.0)   jmpi  (1)     mb_mvp_start;
  216. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud {align1};
  217. /* Read MV for MB C */
  218. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  219. send (16)
  220.        mb_ind
  221.        mb_mv0.0<1>:ud
  222.         NULL
  223.        data_port(
  224.                OBR_CACHE_TYPE,
  225.                OBR_MESSAGE_TYPE,
  226.                OBR_CONTROL_2,
  227.                OBR_BIND_IDX,
  228.                OBR_WRITE_COMMIT_CATEGORY,
  229.                OBR_HEADER_PRESENT
  230.        )
  231.        mlen 1
  232.        rlen 1
  233.        {align1};
  234. /* TODO: RefID is required after multi-references are added */
  235. /* Forward MV */
  236. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  237. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  238.  
  239. jmpi   (1)    mb_mvp_start;
  240. mbd_start:
  241. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  242. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  243. (f0.0)  jmpi (1)        mb_mvp_start;
  244. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  245. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  246. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  247. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  248. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  249.  
  250. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  251. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  252. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  253.  
  254. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  255. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  256. send (16)
  257.        mb_ind
  258.        mb_wb.0<1>:ud
  259.         NULL
  260.        data_port(
  261.                OBR_CACHE_TYPE,
  262.                OBR_MESSAGE_TYPE,
  263.                OBR_CONTROL_0,
  264.                OBR_BIND_IDX,
  265.                OBR_WRITE_COMMIT_CATEGORY,
  266.                OBR_HEADER_PRESENT
  267.        )
  268.        mlen 1
  269.        rlen 1
  270.        {align1};
  271.  
  272. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  273. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  274. (-f0.0)   jmpi  (1)     mb_mvp_start;
  275.  
  276. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  277. /* Read MV for MB D */
  278. /* bind index 3, read 2 oword (16bytes), msg type: 0(OWord Block Read) */
  279. send (16)
  280.        mb_ind
  281.        mb_mv0.0<1>:ub
  282.         NULL
  283.        data_port(
  284.                OBR_CACHE_TYPE,
  285.                OBR_MESSAGE_TYPE,
  286.                OBR_CONTROL_2,
  287.                OBR_BIND_IDX,
  288.                OBR_WRITE_COMMIT_CATEGORY,
  289.                OBR_HEADER_PRESENT
  290.        )
  291.        mlen 1
  292.        rlen 1
  293.        {align1};
  294.  
  295. /* TODO: RefID is required after multi-references are added */
  296.  
  297. /* Forward MV */
  298. mov        (2)          mbc_result.4<1>:ud              mb_mv0.0<2,2,1>:ud      {align1};
  299. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  300.        
  301. mb_mvp_start:
  302. /*TODO: Add the skip prediction */
  303. /* Check whether both MB B and C are invailable */
  304. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  305. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  306. (-f0.0) jmpi (1)        mb_median_start;
  307. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  308. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  309. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  310. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  311. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  312. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  313. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  314. jmpi    (1)     __mb_hwdep_end;
  315.        
  316. mb_median_start:
  317. /* check whether only one neighbour MB has the same ref ID with the current MB */
  318. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  319. cmp.z.f0.0      (1)     null:d  mba_result.20<1>:w      0:w     {align1};
  320. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  321. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  322. cmp.z.f0.0      (1)     null:d  mbb_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        mbb_result.4<0,1,0>:ud  {align1};
  325. cmp.z.f0.0      (1)     null:d  mbc_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        mbc_result.4<0,1,0>:ud  {align1};
  328. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  329. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  330. (f0.0)  jmpi (1)  __mb_hwdep_end;
  331.  
  332. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  333. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  334. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  335. SAVE_RET        {align1};
  336.  jmpi   (1)     word_imedian;
  337. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  338. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  339. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  340. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  341. SAVE_RET        {align1};
  342. jmpi    (1)     word_imedian;
  343. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  344.  
  345. __mb_hwdep_end:
  346.  
  347. mov     (2)     mv_cc_ref.0<1>:w        mba_result.4<2,2,1>:w   {align1};
  348.  
  349. /* Calibrate the ref window for MPEG2 */
  350. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  351. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  352. mov  (1) INPUT_ARG0.0<1>:ud     vme_m0.0<0,1,0>:ud      {align1};
  353. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  354. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  355.  
  356. SAVE_RET        {align1};
  357. jmpi    (1)     ref_boundary_check;
  358. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  359.  
  360. /* m2 */        
  361. mov  (8) vme_msg_2<1>:UD        0x0:UD {align1};
  362.  
  363. /* m3 */
  364. mov  (8) vme_msg_3<1>:UD        0x0:UD {align1};        
  365. /* Use the Luma mode */
  366. mov  (1) tmp_reg0.0<1>:UW       INTRA16_DC_PRED:UW {align1};
  367. mov  (1) vme_msg_3.4<1>:ub      tmp_reg0.0<0,1,0>:UB {align1};
  368.  
  369. /* m4 */
  370. mov  (8) vme_msg_4<1>:UD        0x0:UD {align1};
  371. mov  (1) vme_msg_4.16<1>:UD     INTRA_PREDICTORE_MODE {align1};
  372.  
  373.  
  374. /* m1 */
  375. mov  (8) vme_m1.0<1>:ud         0x0:ud  {align1};
  376. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  377. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  378. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  379. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  380. /* m1 */
  381. /* assign MB intra struct from the thread payload*/
  382. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  383.  
  384.  
  385. /* M0 */
  386. /* IME search */
  387. mov  (1) vme_m0.12<1>:UD   SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_HALF:UD {align1};  
  388. /* 16x16 Source, 1/2 pixel, harr */
  389. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  390.  
  391. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  392. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  393. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  394. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  395.  
  396. /* m1 */
  397.  
  398. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  399. /* MV num is passed by constant buffer. R4.28 */
  400. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};
  401. add  (1) vme_m1.4<1>:UD         vme_m1.4<0,1,0>:UD      FB_PRUNING_DISABLE:UD {align1};
  402. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  403.  
  404. /* Bilinear filter */
  405. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  406. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  407.  
  408. /* Set the MV cost center */
  409. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  410. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  411. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  412.  
  413.  
  414. send (8)
  415.        vme_msg_ind
  416.        vme_wb
  417.        null
  418.        vme(
  419.                BIND_IDX_VME,
  420.                0,
  421.                0,
  422.                VME_MESSAGE_TYPE_MIXED
  423.        )
  424.        mlen vme_msg_length
  425.        rlen vme_inter_wb_length
  426.        {align1};
  427.  
  428. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  429. (-f0.0) jmpi (1) vme_run_again;
  430.  
  431. vme_mv_output:
  432.  
  433. and.z.f0.0      (1)     null<1>:ud              vme_wb0.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  434.  
  435. (-f0.0)jmpi     (1)     __INTRA_INFO ;
  436.  
  437. __INTER_INFO:
  438. /* Write MV pairs */   
  439. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  440.  
  441. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD   {align1};
  442.  
  443. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  444. send (16)
  445.        msg_ind
  446.        obw_wb
  447.        null
  448.        data_port(
  449.                OBW_CACHE_TYPE,
  450.                OBW_MESSAGE_TYPE,
  451.                OBW_CONTROL_2,
  452.                OBW_BIND_IDX,
  453.                OBW_WRITE_COMMIT_CATEGORY,
  454.                OBW_HEADER_PRESENT
  455.        )
  456.        mlen 2
  457.        rlen obw_wb_length
  458.        {align1};
  459.  
  460.  
  461. mov             (1)     msg_reg1.0<1>:ud        vme_wb0.0<0,1,0>:ud     {align1} ;
  462. mov             (1)     msg_reg1.4<1>:UD        vme_wb0.28<0,1,0>:UD    {align1};
  463. mov             (1)     msg_reg1.8<1>:ud        tmp_ud1<0,1,0>:ud       {align1} ;
  464. mov             (1)     msg_reg1.12<1>:ud        vme_wb0.0<0,1,0>:ud     {align1} ;
  465. mov             (1)     msg_reg1.16<1>:ud        0x25:ud     {align1} ;
  466. jmpi            (1)     __OUTPUT_INFO;
  467.  
  468. __INTRA_INFO:
  469. mov             (1)     msg_reg1.0<1>:UD        vme_wb.0<0,1,0>:UD      {align1};
  470. mov             (1)     msg_reg1.4<1>:UD        vme_wb.16<0,1,0>:UD     {align1};
  471. mov             (1)     msg_reg1.8<1>:UD        vme_wb.20<0,1,0>:UD     {align1};
  472. mov             (1)     msg_reg1.12<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  473. mov             (1)     msg_reg1.16<1>:ud        0x35:ud     {align1} ;
  474.  
  475. __OUTPUT_INFO:
  476.  
  477. mov     (1)     msg_reg1.20<1>:ud       obw_m0.8<0,1,0>:ud      {align1};
  478. add     (1)     obw_m0.8<1>:UD       obw_m0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  479. mov     (8)     msg_reg0.0<1>:ud        obw_m0.0<8,8,1>:ud      {align1};
  480.  
  481.  
  482. /* bind index 3, write 1 oword, msg type: 8(OWord Block Write) */
  483. send (16)
  484.        msg_ind
  485.        obw_wb
  486.        null
  487.        data_port(
  488.                OBW_CACHE_TYPE,
  489.                OBW_MESSAGE_TYPE,
  490.                OBW_CONTROL_2,
  491.                OBW_BIND_IDX,
  492.                OBW_WRITE_COMMIT_CATEGORY,
  493.                OBW_HEADER_PRESENT
  494.        )
  495.        mlen 2
  496.        rlen obw_wb_length
  497.        {align1};
  498.  
  499. /* Issue message fence so that the previous write message is committed */
  500. send (16)
  501.        mb_ind
  502.        mb_wb.0<1>:ud
  503.         NULL
  504.        data_port(
  505.                OBR_CACHE_TYPE,
  506.                OBR_MESSAGE_FENCE,
  507.                OBR_MF_COMMIT,
  508.                OBR_BIND_IDX,
  509.                OBR_WRITE_COMMIT_CATEGORY,
  510.                OBR_HEADER_PRESENT
  511.        )
  512.        mlen 1
  513.        rlen 1
  514.         {align1};
  515.  
  516. __EXIT:
  517. /*
  518. * kill thread
  519. */        
  520. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  521. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  522.  
  523.  
  524.         nop             ;
  525.         nop             ;
  526.        
  527. word_imedian:
  528.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  529.         (f0.0)  jmpi (1) cmp_a_ge_b;
  530.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  531.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  532.         (f0.0) jmpi (1) cmp_end;
  533.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  534.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  535.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  536.         jmpi (1) cmp_end;
  537. cmp_a_ge_b:
  538.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  539.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  540.         (f0.0) jmpi (1) cmp_end;
  541.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  542.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  543.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  544. cmp_end:
  545.         RETURN  {align1};
  546.  
  547. nop;
  548. nop;
  549.  
  550. ref_boundary_check:
  551.  
  552. /* The left/up coordinate of reference window */
  553. add  (2) TEMP_VAR0.0<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG0.0<2,2,1>:w   {align1};
  554. /* The right/bottom coordinate of reference window */
  555. add  (1) TEMP_VAR0.16<1>:w      TEMP_VAR0.0<0,1,0>:w    48:w                    {align1};
  556. add  (1) TEMP_VAR0.18<1>:w      TEMP_VAR0.2<0,1,0>:w    40:w                    {align1};
  557.  
  558. /* Firstly the MV range is checked */
  559. mul  (2) TEMP_VAR1.16<1>:w      INPUT_ARG1.16<2,2,1>:w  -1:w    {align1};
  560. add  (2) TEMP_VAR1.0<1>:w       INPUT_ARG0.8<2,2,1>:w   TEMP_VAR1.16<2,2,1>:w   {align1};
  561. add  (2) TEMP_VAR1.4<1>:w       INPUT_ARG0.8<2,2,1>:w   INPUT_ARG1.16<2,2,1>:w  {align1};
  562.  
  563. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.0<0,1,0>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  564. (f0.0)  mov     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.0<0,1,0>:w    {align1};
  565. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   TEMP_VAR1.4<0,1,0>:w    {align1};
  566. (f0.0)  add     (1) TEMP_VAR0.0<1>:w    TEMP_VAR1.4<0,1,0>:w    -48:w   {align1};
  567. cmp.l.f0.0 (1)  null:w  TEMP_VAR0.2<0,1,0>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  568. (f0.0)  mov     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.2<0,1,0>:w    {align1};
  569. cmp.g.f0.0 (1)  null:w  TEMP_VAR0.18<0,1,0>:w   TEMP_VAR1.6<0,1,0>:w    {align1};
  570. (f0.0)  add     (1) TEMP_VAR0.2<1>:w    TEMP_VAR1.6<0,1,0>:w    -40:w   {align1};
  571.  
  572.  
  573. x_left_cmp:
  574.         cmp.l.f0.0 (1)          null:w          TEMP_VAR0.0<0,1,0>:w    0:w     {align1};
  575.         (-f0.0) jmpi    (1)     x_right_cmp;
  576.         (f0.0)  mov     (1)     TEMP_VAR0.0<1>:w        0:w             {align1};
  577.         jmpi    (1)     y_top_cmp;
  578. x_right_cmp:
  579.         cmp.g.f0.0 (1)  null:w  TEMP_VAR0.16<0,1,0>:w   INPUT_ARG1.0<0,1,0>:w   {align1};
  580.         (-f0.0) jmpi    (1)     y_top_cmp;     
  581.         (f0.0)  add     (1)     TEMP_VAR0.0<1>:w        INPUT_ARG1.0<0,1,0>:w   -48:w   {align1};
  582. y_top_cmp:
  583.         cmp.l.f0.0  (1)         null:w  TEMP_VAR0.2<0,1,0>:w    0:w             {align1};
  584.         (-f0.0) jmpi    (1)     y_bottom_cmp;
  585.         (f0.0)  mov     (1)     TEMP_VAR0.2<1>:w        0:w             {align1};
  586.         jmpi    (1)     y_bottom_end;
  587. y_bottom_cmp:
  588.         cmp.g.f0.0  (1)         null:w  TEMP_VAR0.18<0,1,0>:w   INPUT_ARG1.2<0,1,0>:w {align1};
  589.         (f0.0)  add     (1)     TEMP_VAR0.2<1>:w                INPUT_ARG1.2<0,1,0>:w   -40:w   {align1};
  590.  
  591. y_bottom_end:
  592. mul     (2)     TEMP_VAR1.0<1>:w        INPUT_ARG0.8<2,2,1>:w   -1:w    {align1};
  593. add     (2)     RET_ARG<1>:w            TEMP_VAR0.0<2,2,1>:w    TEMP_VAR1.0<2,2,1>:w    {align1};      
  594.         RETURN  {align1};
  595. nop;
  596. nop;
  597.  
  598. vme_run_again:
  599.  
  600. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  601. mov     (2)     tmp_reg0.0<1>:w         mb_ref_win.0<2,2,1>:w           {align1};
  602. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  603. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  604.  
  605. cmp.l.f0.0      (1) null:w      tmp_reg0.0<0,1,0>:w     0:w     {align1};
  606. (f0.0)  mul     (1) tmp_reg0.0<1>:w     tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  607. cmp.l.f0.0      (1) null:w      tmp_reg0.2<0,1,0>:w     0:w     {align1};
  608. (f0.0)  mul     (1) tmp_reg0.2<1>:w     tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  609.  
  610. cmp.ge.f0.0     (1) null:w      tmp_reg0.0<0,1,0>:w     4:w     {align1};
  611. (f0.0)  jmpi (1)        vme_start;
  612. cmp.ge.f0.0     (1) null:w      tmp_reg0.2<0,1,0>:w     4:w     {align1};
  613. (f0.0)  jmpi (1)        vme_start;
  614.  
  615. jmpi (1) vme_done;
  616.  
  617. vme_start:
  618.         mov (8) tmp_vme_wb0.0<1>:ud     vme_wb0.0<8,8,1>:ud     {align1};
  619.         mov (8) tmp_vme_wb1.0<1>:ud     vme_wb1.0<8,8,1>:ud     {align1};
  620.  
  621. /* Calibrate the ref window for MPEG2 */
  622. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  623. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  624. mov  (1) INPUT_ARG0.8<1>:ud     vme_m0.8<0,1,0>:ud      {align1};
  625. add  (2) INPUT_ARG0.0<1>:w      vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  626. mov  (8) INPUT_ARG1.0<1>:ud     pic_ref.0<8,8,1>:ud     {align1};
  627.  
  628. SAVE_RET        {align1};
  629. jmpi    (1)     ref_boundary_check;
  630. mov  (2) vme_m0.0<1>:w          RET_ARG<2,2,1>:w        {align1};
  631.  
  632. /* m2 */        
  633. mov  (8) vme_msg_2<1>:UD        0x0:UD {align1};
  634.  
  635. /* m3 */
  636. mov  (8) vme_msg_3<1>:UD        0x0:UD {align1};        
  637.  
  638. /* m4 */
  639. mov  (8) vme_msg_4<1>:UD        0x0:UD {align1};
  640.  
  641.  
  642. /* m1 */
  643. mov  (8) vme_m1.0<1>:ud         0x0:ud  {align1};
  644. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  645. mov   (1) tmp_reg0.0<1>:uw      LUMA_INTRA_8x8_DISABLE:uw               {align1};
  646. add   (1) tmp_reg0.0<1>:uw      tmp_reg0.0<0,1,0>:uw    LUMA_INTRA_4x4_DISABLE:uw {align1};
  647. mov  (1) intra_part_mask_ub<1>:UB  tmp_reg0.0<0,1,0>:ub {align1};
  648. /* m1 */
  649. /* assign MB intra struct from the thread payload*/
  650. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  651.  
  652.  
  653. /* M0 */
  654. /* IME search */
  655. mov  (1) vme_m0.12<1>:UD   SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_HALF:UD {align1};  
  656. /* 16x16 Source, 1/2 pixel, harr */
  657. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  658.  
  659. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  660. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  661.  
  662. /* m1 */
  663.  
  664. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  665. /* MV num is passed by constant buffer. R4.28 */
  666. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};
  667. add  (1) vme_m1.4<1>:UD         vme_m1.4<0,1,0>:UD      FB_PRUNING_DISABLE:UD {align1};
  668. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  669.  
  670. /* Bilinear filter */
  671. mov  (1) tmp_reg0.0<1>:uw       0x04:uw {align1};
  672. add  (1) vme_m1.30<1>:ub        vme_m1.30<0,1,0>:ub     tmp_reg0.0<0,1,0>:ub    {align1};
  673.  
  674. /* Set the MV cost center */
  675. mov  (1) vme_m1.16<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  676. mov  (1) vme_m1.20<1>:ud        mv_cc_ref.0<0,1,0>:ud   {align1};
  677. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  678.  
  679.  
  680. send (8)
  681.        vme_msg_ind
  682.        vme_wb
  683.        null
  684.        vme(
  685.                BIND_IDX_VME,
  686.                0,
  687.                0,
  688.                VME_MESSAGE_TYPE_INTER
  689.        )
  690.        mlen vme_msg_length
  691.        rlen vme_inter_wb_length
  692.        {align1};
  693.  
  694.  
  695. cmp.l.f0.0 (1)  null:uw vme_wb0.6<0,1,0>:uw     tmp_vme_wb0.6<0,1,0>:uw {align1};
  696. (f0.0)  jmpi (1) vme_done;
  697. mov     (8)     vme_wb0.0<1>:ud tmp_vme_wb0.0<8,8,1>:ud {align1};
  698. mov     (8)     vme_wb1.0<1>:ud tmp_vme_wb1.0<8,8,1>:ud {align1};
  699.  
  700. vme_done:
  701.       jmpi (1) vme_mv_output;
  702. nop;
  703. nop;
  704. nop;
  705.  
  706.