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.  *
  8.  */
  9. // Modual name: Inter_frame_haswell.asm
  10. //
  11. // Make inter predition estimation for Inter-frame on Haswell
  12. //
  13.  
  14. //
  15. //  Now, begin source code....
  16. //
  17.  
  18. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  19. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  20.  
  21. /*
  22.  * __START
  23.  */
  24. __INTER_START:
  25. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  26. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  27. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  28. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  29.  
  30. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  31. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  32. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  33. mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
  34. mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  35.  
  36. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  37. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  38. mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
  39. mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  40.        
  41. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  42. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  43.  
  44. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  45. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  46. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
  47. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  48.        
  49. /*
  50.  * Media Read Message -- fetch Luma neighbor edge pixels
  51.  */
  52. /* ROW */
  53. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  54. send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
  55.  
  56. /* COL */
  57. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  58. send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
  59.        
  60. /*
  61.  * Media Read Message -- fetch Chroma neighbor edge pixels
  62.  */
  63. /* ROW */
  64. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
  65. mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
  66. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  67. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  68. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  69. send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  70.  
  71. /* COL */
  72. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
  73. mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
  74. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  75. mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
  76. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  77. send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  78.  
  79. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  80. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  81. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  82. (f0.0) jmpi (1) __mb_hwdep_end;
  83. /* read back the data for MB A */
  84. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  85. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  86. */
  87. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  88. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  89. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  90. mba_start:
  91. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  92. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  93. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  94. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  95. (f0.0)  jmpi (1)        mbb_start;
  96. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  97. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  98. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  99. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  100. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  101. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  102. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  103.  
  104. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  105. send (16)
  106.        mb_ind
  107.        mb_wb.0<1>:ud
  108.         NULL
  109.        data_port(
  110.                OBR_CACHE_TYPE,
  111.                OBR_MESSAGE_TYPE,
  112.                OBR_CONTROL_4,
  113.                OBR_BIND_IDX,
  114.                OBR_WRITE_COMMIT_CATEGORY,
  115.                OBR_HEADER_PRESENT
  116.        )
  117.        mlen 1
  118.        rlen 2
  119.        {align1};
  120.  
  121. /* TODO: RefID is required after multi-references are added */
  122. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  123. (f0.0)   mov (2)        mba_result.20<1>:w                      -1:w    {align1};
  124. (f0.0)   jmpi   (1)     mbb_start;
  125.  
  126. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  127. /* Read MV for MB A */
  128. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  129. send (16)
  130.        mb_ind
  131.        mb_mv0.0<1>:ud
  132.         NULL
  133.        data_port(
  134.                OBR_CACHE_TYPE,
  135.                OBR_MESSAGE_TYPE,
  136.                OBR_CONTROL_8,
  137.                OBR_BIND_IDX,
  138.                OBR_WRITE_COMMIT_CATEGORY,
  139.                OBR_HEADER_PRESENT
  140.        )
  141.        mlen 1
  142.        rlen 4
  143.        {align1};
  144. /* TODO: RefID is required after multi-references are added */
  145. /* MV */
  146. mov        (2)          mba_result.4<1>:ud              mb_mv1.8<2,2,1>:ud      {align1};
  147. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  148.  
  149. mbb_start:
  150. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  151. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  152. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  153. /* If MB B doesn't exist, neither MB C nor D exists */
  154. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  155. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  156. (f0.0)  jmpi (1)        mb_mvp_start;
  157. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  158. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  159. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  160. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  161. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  162. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  163. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  164.  
  165. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  166. send (16)
  167.        mb_ind
  168.        mb_wb.0<1>:ud
  169.         NULL
  170.        data_port(
  171.                OBR_CACHE_TYPE,
  172.                OBR_MESSAGE_TYPE,
  173.                OBR_CONTROL_4,
  174.                OBR_BIND_IDX,
  175.                OBR_WRITE_COMMIT_CATEGORY,
  176.                OBR_HEADER_PRESENT
  177.        )
  178.        mlen 1
  179.        rlen 2
  180.        {align1};
  181.  
  182. /* TODO: RefID is required after multi-references are added */
  183. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  184. (f0.0)   mov (2)        mbb_result.20<1>:w                      -1:w    {align1};
  185. (f0.0)   jmpi   (1)     mbc_start;
  186. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  187. /* Read MV for MB B */
  188. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  189. send (16)
  190.        mb_ind
  191.        mb_mv0.0<1>:ud
  192.         NULL
  193.        data_port(
  194.                OBR_CACHE_TYPE,
  195.                OBR_MESSAGE_TYPE,
  196.                OBR_CONTROL_8,
  197.                OBR_BIND_IDX,
  198.                OBR_WRITE_COMMIT_CATEGORY,
  199.                OBR_HEADER_PRESENT
  200.        )
  201.        mlen 1
  202.        rlen 4
  203.        {align1};
  204. /* TODO: RefID is required after multi-references are added */
  205. mov        (2)          mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  206. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  207.  
  208. mbc_start:
  209. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  210. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  211. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  212. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  213. (f0.0)  jmpi (1)        mbd_start;
  214. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  215. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  216. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  217. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  218. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  219. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  220. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  221. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  222.  
  223. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  224. send (16)
  225.        mb_ind
  226.        mb_wb.0<1>:ud
  227.         NULL
  228.        data_port(
  229.                OBR_CACHE_TYPE,
  230.                OBR_MESSAGE_TYPE,
  231.                OBR_CONTROL_4,
  232.                OBR_BIND_IDX,
  233.                OBR_WRITE_COMMIT_CATEGORY,
  234.                OBR_HEADER_PRESENT
  235.        )
  236.        mlen 1
  237.        rlen 2
  238.        {align1};
  239.  
  240. /* TODO: RefID is required after multi-references are added */
  241. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  242. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  243. (f0.0)   jmpi   (1)     mb_mvp_start;
  244. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  245. /* Read MV for MB C */
  246. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  247. send (16)
  248.        mb_ind
  249.        mb_mv0.0<1>:ud
  250.         NULL
  251.        data_port(
  252.                OBR_CACHE_TYPE,
  253.                OBR_MESSAGE_TYPE,
  254.                OBR_CONTROL_8,
  255.                OBR_BIND_IDX,
  256.                OBR_WRITE_COMMIT_CATEGORY,
  257.                OBR_HEADER_PRESENT
  258.        )
  259.        mlen 1
  260.        rlen 4
  261.        {align1};
  262. /* TODO: RefID is required after multi-references are added */
  263. /* Forward MV */
  264. mov        (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  265. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  266.  
  267. jmpi   (1)    mb_mvp_start;
  268. mbd_start:
  269. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  270. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  271. (f0.0)  jmpi (1)        mb_mvp_start;
  272. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  273. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  274. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  275. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  276. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  277. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  278. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  279.  
  280. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  281. send (16)
  282.        mb_ind
  283.        mb_wb.0<1>:ud
  284.         NULL
  285.        data_port(
  286.                OBR_CACHE_TYPE,
  287.                OBR_MESSAGE_TYPE,
  288.                OBR_CONTROL_4,
  289.                OBR_BIND_IDX,
  290.                OBR_WRITE_COMMIT_CATEGORY,
  291.                OBR_HEADER_PRESENT
  292.        )
  293.        mlen 1
  294.        rlen 2
  295.        {align1};
  296.  
  297. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  298. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  299. (f0.0)   jmpi   (1)     mb_mvp_start;
  300.  
  301. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  302. /* Read MV for MB D */
  303. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  304. send (16)
  305.        mb_ind
  306.        mb_mv0.0<1>:ub
  307.         NULL
  308.        data_port(
  309.                OBR_CACHE_TYPE,
  310.                OBR_MESSAGE_TYPE,
  311.                OBR_CONTROL_8,
  312.                OBR_BIND_IDX,
  313.                OBR_WRITE_COMMIT_CATEGORY,
  314.                OBR_HEADER_PRESENT
  315.        )
  316.        mlen 1
  317.        rlen 4
  318.        {align1};
  319.  
  320. /* TODO: RefID is required after multi-references are added */
  321.  
  322. /* Forward MV */
  323. mov        (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
  324. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  325.        
  326. mb_mvp_start:
  327. /*TODO: Add the skip prediction */
  328. /* Check whether both MB B and C are inavailable */
  329. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  330. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  331. (-f0.0) jmpi (1)        mb_median_start;
  332. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  333. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  334. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  335. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  336. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  337. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  338. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  339. jmpi    (1)     __mb_hwdep_end;
  340.        
  341. mb_median_start:
  342. /* check whether only one neighbour MB has the same ref ID with the current MB */
  343. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  344. cmp.z.f0.0      (1)     null:d  mba_result.20<1>:w      0:w     {align1};
  345. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  346. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  347. cmp.z.f0.0      (1)     null:d  mbb_result.20<1>:w      0:w     {align1};
  348. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  349. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  350. cmp.z.f0.0      (1)     null:d  mbc_result.20<1>:w      0:w     {align1};
  351. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  352. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  353. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  354. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  355. (f0.0)  jmpi (1)  __mb_hwdep_end;
  356.  
  357. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  358. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  359. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  360. SAVE_RET        {align1};
  361.  jmpi   (1)     word_imedian;
  362. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  363. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  364. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  365. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  366. SAVE_RET        {align1};
  367. jmpi    (1)     word_imedian;
  368. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  369.  
  370. __mb_hwdep_end:
  371. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  372. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  373. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  374.        
  375. /* m2, get the MV/Mb cost passed from constant buffer when
  376. spawning thread by MEDIA_OBJECT */      
  377. mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
  378.  
  379. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  380.  
  381. /* m3 */
  382. mov (8) vme_msg_3<1>:UD         0x0:UD {align1};               
  383.  
  384. /* m4 */
  385. mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
  386. and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
  387. mov  (8) vme_msg_4<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
  388.  
  389. /* m5 */        
  390. mov  (8) vme_msg_5<1>:UD         0x0:UD {align1};
  391. mov (16) vme_msg_5.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
  392. mov  (1) vme_msg_5.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
  393.  
  394. /* the penalty for Intra mode */
  395. mov  (1) vme_msg_5.28<1>:UD     0x010101:UD {align1};
  396. mov  (1) vme_msg_5.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
  397.  
  398.  
  399. /* m6 */
  400. mov  (4) vme_msg_6.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
  401. mov  (8) vme_msg_6.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
  402.  
  403. /*
  404. * SIC VME message
  405. */
  406. /* m0 */        
  407. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  408. mov  (1) tmp_reg0.0<1>:UW       LUMA_INTRA_MODE:UW {align1};
  409. /* Use the Luma mode */
  410. mov  (1) vme_msg_4.5<1>:UB      tmp_reg0.0<0,1,0>:UB {align1};
  411.  
  412. /* m1 */
  413. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  414. and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
  415. (f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1};
  416.  
  417. /* assign MB intra struct from the thread payload*/
  418. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  419.  
  420. /* Disable DC HAAR component when calculating HARR SATD block */
  421. mov  (1) tmp_reg0.0<1>:UW       DC_HARR_DISABLE:UW              {align1};
  422. mov  (1) vme_m1.30<1>:UB        tmp_reg0.0<0,1,0>:UB  {align1};
  423.  
  424. mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
  425. /* m0 */        
  426. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  427. mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
  428.  
  429. /* after verification it will be passed by using payload */
  430. send (8)
  431.        vme_msg_ind
  432.        vme_wb<1>:UD
  433.        null
  434.        cre(
  435.                BIND_IDX_VME,
  436.                VME_SIC_MESSAGE_TYPE
  437.        )
  438.        mlen sic_vme_msg_length
  439.        rlen vme_wb_length
  440.        {align1};
  441. /*
  442. * Oword Block Write message
  443. */
  444. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  445.        
  446. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  447. mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
  448. mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
  449. mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
  450.  
  451. /* Distortion, Intra (17-16), */
  452. mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
  453.  
  454. mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
  455. /* VME clock counts */
  456. mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
  457.  
  458. mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
  459.  
  460. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  461. send (16)
  462.        msg_ind
  463.        obw_wb
  464.        null
  465.        data_port(
  466.                OBW_CACHE_TYPE,
  467.                OBW_MESSAGE_TYPE,
  468.                OBW_CONTROL_2,
  469.                OBW_BIND_IDX,
  470.                OBW_WRITE_COMMIT_CATEGORY,
  471.                OBW_HEADER_PRESENT
  472.        )
  473.        mlen 2
  474.        rlen obw_wb_length
  475.        {align1};
  476.  
  477. /* IME search */
  478. cmp.z.f0.0 (1)          null<1>:uw      quality_level_ub<0,1,0>:ub              LOW_QUALITY_LEVEL:uw   {align1};
  479. (f0.0) jmpi (1) __low_quality_search;
  480.  
  481. __high_quality_search:
  482. /* M3/M4 search path */
  483. mov  (1) vme_msg_3.0<1>:UD      0x01010101:UD {align1};
  484. mov  (1) vme_msg_3.4<1>:UD      0x10010101:UD {align1};
  485. mov  (1) vme_msg_3.8<1>:UD      0x0F0F0F0F:UD {align1};
  486. mov  (1) vme_msg_3.12<1>:UD     0x100F0F0F:UD {align1};
  487. mov  (1) vme_msg_3.16<1>:UD     0x01010101:UD {align1};
  488. mov  (1) vme_msg_3.20<1>:UD     0x10010101:UD {align1};
  489. mov  (1) vme_msg_3.24<1>:UD     0x0F0F0F0F:UD {align1};
  490. mov  (1) vme_msg_3.28<1>:UD     0x100F0F0F:UD {align1};
  491. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  492. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  493. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  494. mov  (1) vme_msg_4.12<1>:UD     0x000F0F0F:UD {align1};
  495. mov  (4) vme_msg_4.16<1>:UD     0x0:UD {align1};
  496.  
  497. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  498. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  499. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  500. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  501.  
  502. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  503. (f0.0)  add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       12:w    {align1};
  504. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  505. (f0.0)  add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       8:w     {align1};
  506.  
  507. jmpi (1) __vme_msg;
  508.  
  509. __low_quality_search:
  510. /* M3/M4 search path */
  511. mov  (1) vme_msg_3.0<1>:UD      0x10010101:UD {align1};
  512. mov  (1) vme_msg_3.4<1>:UD      0x100F0F0F:UD {align1};
  513. mov  (1) vme_msg_3.8<1>:UD      0x10010101:UD {align1};
  514. mov  (1) vme_msg_3.12<1>:UD     0x000F0F0F:UD {align1};
  515. mov  (4) vme_msg_3.16<1>:UD     0x0:UD {align1};
  516. mov  (8) vme_msg_4.16<1>:UD     0x0:UD {align1};
  517.  
  518. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  519. mov  (1) vme_m0.22<1>:UW        MIN_REF_REGION_SIZE {align1};         /* Reference Width&Height, 32x32 */
  520. mov  (1) vme_m0.0<1>:W          -8:W                    {align1};
  521. mov  (1) vme_m0.2<1>:W          -8:W                    {align1};
  522.  
  523. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  524. (f0.0)  add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       4:w     {align1};
  525. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  526. (f0.0)  add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       4:w     {align1};
  527.  
  528. __vme_msg:
  529. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  530. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  531. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  532. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  533.  
  534. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  535. /* the Max MV number is passed by constant buffer */
  536. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  537. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  538. /* Set the MV cost center */
  539. mov  (1) vme_m1.16<1>:ud        mb_mvp_ref.0<0,1,0>:ud  {align1};
  540. mov  (1) vme_m1.20<1>:ud        mb_mvp_ref.0<0,1,0>:ud  {align1};
  541. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  542.  
  543. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  544.  
  545. send (8)
  546.        vme_msg_ind
  547.        vme_wb<1>:UD
  548.        null
  549.        vme(
  550.                BIND_IDX_VME,
  551.                0,
  552.                0,
  553.                VME_IME_MESSAGE_TYPE
  554.        )
  555.        mlen ime_vme_msg_length
  556.        rlen vme_wb_length {align1};
  557.  
  558. /* Set Macroblock-shape/mode for FBR */
  559.  
  560. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  561. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  562. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  563.  
  564. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  565. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  566.  
  567. /* Send FBR message into CRE */
  568.  
  569. mov  (8) vme_msg_3.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  570. mov  (8) vme_msg_4.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  571. mov  (8) vme_msg_5.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  572. mov  (8) vme_msg_6.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  573.  
  574. mov  (1) vme_m0.12<1>:UD        INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER + FBR_BME_DISABLE:UD {align1};    /* 16x16 Source, 1/4 pixel, harr, BME disable */
  575. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  576. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  577.  
  578. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  579.  
  580. /* after verification it will be passed by using payload */
  581. send (8)
  582.        vme_msg_ind
  583.        vme_wb<1>:UD
  584.        null
  585.        cre(
  586.                BIND_IDX_VME,
  587.                VME_FBR_MESSAGE_TYPE
  588.        )
  589.        mlen fbr_vme_msg_length
  590.        rlen vme_wb_length
  591.        {align1};
  592.  
  593. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
  594. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  595. /* write FME info */
  596. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  597.  
  598. mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  599. /* Inter distortion of FME */
  600. mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
  601.  
  602. mov  (1) msg_reg1.12<1>:UD      vme_m2.20<0,1,0>:UD {align1};
  603.  
  604. /* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
  605. send (16)
  606.        msg_ind
  607.        obw_wb
  608.        null
  609.        data_port(
  610.                OBW_CACHE_TYPE,
  611.                OBW_MESSAGE_TYPE,
  612.                OBW_CONTROL_0,
  613.                OBW_BIND_IDX,
  614.                OBW_WRITE_COMMIT_CATEGORY,
  615.                OBW_HEADER_PRESENT
  616.        )
  617.        mlen 2
  618.        rlen obw_wb_length
  619.        {align1};
  620.  
  621. /* Write FME/BME MV */
  622. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
  623. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  624.  
  625.  
  626. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  627. mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  628. mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  629. mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  630. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  631. send (16)
  632.        msg_ind
  633.        obw_wb
  634.        null
  635.        data_port(
  636.                OBW_CACHE_TYPE,
  637.                OBW_MESSAGE_TYPE,
  638.                OBW_CONTROL_8,
  639.                OBW_BIND_IDX,
  640.                OBW_WRITE_COMMIT_CATEGORY,
  641.                OBW_HEADER_PRESENT
  642.        )
  643.        mlen 5
  644.        rlen obw_wb_length
  645.        {align1};
  646.  
  647. /* Write FME/BME RefID */
  648. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
  649. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  650.  
  651. mov  (8) msg_reg1.0<1>:UD       vme_wb6.0<8,8,1>:UD {align1};
  652.  
  653. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  654. send (16)
  655.        msg_ind
  656.        obw_wb
  657.        null
  658.        data_port(
  659.                OBW_CACHE_TYPE,
  660.                OBW_MESSAGE_TYPE,
  661.                OBW_CONTROL_2,
  662.                OBW_BIND_IDX,
  663.                OBW_WRITE_COMMIT_CATEGORY,
  664.                OBW_HEADER_PRESENT
  665.        )
  666.        mlen 2
  667.        rlen obw_wb_length
  668.        {align1};
  669.  
  670. /* Issue message fence so that the previous write message is committed */
  671. send (16)
  672.        mb_ind
  673.        mb_wb.0<1>:ud
  674.         NULL
  675.        data_port(
  676.                OBR_CACHE_TYPE,
  677.                OBR_MESSAGE_FENCE,
  678.                OBR_MF_COMMIT,
  679.                OBR_BIND_IDX,
  680.                OBR_WRITE_COMMIT_CATEGORY,
  681.                OBR_HEADER_PRESENT
  682.        )
  683.        mlen 1
  684.        rlen 1
  685.        {align1};
  686.  
  687. __EXIT:
  688. /*
  689. * kill thread
  690. */        
  691. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  692. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  693.  
  694.         nop             ;
  695.         nop             ;
  696. /* Compare three word data to get the min value */
  697. word_imin:
  698.         cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  699.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  700.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  701.         cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  702.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  703.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  704.         RETURN          {align1};      
  705.        
  706. /* Compare three word data to get the max value */
  707. word_imax:
  708.         cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  709.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  710.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  711.         cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  712.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  713.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  714.         RETURN          {align1};      
  715.        
  716. word_imedian:
  717.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  718.         (f0.0)  jmpi (1) cmp_a_ge_b;
  719.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  720.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  721.         (f0.0) jmpi (1) cmp_end;
  722.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  723.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  724.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  725.         jmpi (1) cmp_end;
  726. cmp_a_ge_b:
  727.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  728.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  729.         (f0.0) jmpi (1) cmp_end;
  730.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  731.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  732.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  733. cmp_end:
  734.         RETURN  {align1};
  735.