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. // Modual name: Inter_bframe_haswell.asm
  26. //
  27. // Make inter predition estimation for Inter frame for B-frame
  28. //
  29.  
  30. //
  31. //  Now, begin source code....
  32. //
  33.  
  34. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  35. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  36.  
  37. /*
  38.  * __START
  39.  */
  40. __INTER_START:
  41. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  42. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  43. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  44. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  45.  
  46. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  47. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  48. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  49. mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
  50. mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  51.  
  52. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  53. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  54. mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
  55. mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  56.        
  57. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  58. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  59.  
  60. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  61. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  62. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
  63. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  64.        
  65. /*
  66.  * Media Read Message -- fetch Luma neighbor edge pixels
  67.  */
  68. /* ROW */
  69. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  70. send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
  71.  
  72. /* COL */
  73. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  74. send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
  75.        
  76. /*
  77.  * Media Read Message -- fetch Chroma neighbor edge pixels
  78.  */
  79. /* ROW */
  80. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
  81. mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
  82. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  83. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  84. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  85. send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  86.  
  87. /* COL */
  88. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
  89. mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
  90. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  91. mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
  92. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  93. send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  94.  
  95. mov  (8) vme_m1.0<1>:ud         0:ud            {align1};
  96. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  97. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  98. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  99. (f0.0) jmpi (1) __mb_hwdep_end;
  100.  
  101. /* read back the data for MB A */
  102. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  103. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  104. */
  105. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  106. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  107. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  108. mba_start:
  109. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  110. and.z.f0.0 (2)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  111. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  112. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  113. (f0.0)  jmpi (1)        mbb_start;
  114. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  115. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  116. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  117. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  118. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  119. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  120. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  121.  
  122. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  123. send (16)
  124.        mb_ind
  125.        mb_wb.0<1>:ud
  126.         NULL
  127.        data_port(
  128.                OBR_CACHE_TYPE,
  129.                OBR_MESSAGE_TYPE,
  130.                OBR_CONTROL_4,
  131.                OBR_BIND_IDX,
  132.                OBR_WRITE_COMMIT_CATEGORY,
  133.                OBR_HEADER_PRESENT
  134.        )
  135.        mlen 1
  136.        rlen 2
  137.        {align1};
  138.  
  139. /* TODO: RefID is required after multi-references are added */
  140. cmp.l.f0.0 (2)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  141. (f0.0)   mov (2)        mba_result.20<1>:w                      -1:w    {align1};
  142. (f0.0)   jmpi   (1)     mbb_start;
  143.  
  144. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  145. /* Read MV for MB A */
  146. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  147. send (16)
  148.        mb_ind
  149.        mb_mv0.0<1>:ud
  150.         NULL
  151.        data_port(
  152.                OBR_CACHE_TYPE,
  153.                OBR_MESSAGE_TYPE,
  154.                OBR_CONTROL_8,
  155.                OBR_BIND_IDX,
  156.                OBR_WRITE_COMMIT_CATEGORY,
  157.                OBR_HEADER_PRESENT
  158.        )
  159.        mlen 1
  160.        rlen 4
  161.        {align1};
  162. /* TODO: RefID is required after multi-references are added */
  163. /* MV */
  164. mov     (2)     mba_result.20<1>:w              -1:w            {align1};
  165. mov     (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
  166. mov     (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
  167. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK1:ud {align1};
  168. SAVE_RET        {align1};
  169. jmpi    (1)     mb_pred_func;
  170. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  171. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  172. (f0.0)  mov   (1)       mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
  173. (f0.0)  mov   (1)       mba_result.20<1>:w              0:w             {align1};
  174. (f0.0)  mov   (1)       mba_result.4<1>:ud              mb_mv1.8<0,1,0>:ud      {align1};
  175. (f0.0)  jmpi    (1) mbb_start;
  176. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  177. (f0.0)  mov   (1)       mba_result.18<1>:uw             MB_PRED_FLAG            {align1};
  178. (f0.0)  mov   (1)       mba_result.22<1>:w              0:w             {align1};
  179. (f0.0)  mov   (1)       mba_result.8<1>:ud              mb_mv1.12<0,1,0>:ud     {align1};
  180. (f0.0)  jmpi    (1) mbb_start;
  181. mov   (2)       mba_result.4<1>:ud              mb_mv1.8<2,2,1>:ud      {align1};
  182. mov   (2)       mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
  183. mov   (2)       mba_result.20<1>:w              0:w             {align1};
  184.  
  185. mbb_start:
  186. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  187. and.z.f0.0 (2)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  188. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  189. /* If MB B doesn't exist, neither MB C nor D exists */
  190. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  191. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  192. (f0.0)  jmpi (1)        mb_mvp_start;
  193. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  194. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  195. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  196. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  197. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  198. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  199. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  200.  
  201. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  202. send (16)
  203.        mb_ind
  204.        mb_wb.0<1>:ud
  205.         NULL
  206.        data_port(
  207.                OBR_CACHE_TYPE,
  208.                OBR_MESSAGE_TYPE,
  209.                OBR_CONTROL_4,
  210.                OBR_BIND_IDX,
  211.                OBR_WRITE_COMMIT_CATEGORY,
  212.                OBR_HEADER_PRESENT
  213.        )
  214.        mlen 1
  215.        rlen 2
  216.        {align1};
  217.  
  218. /* TODO: RefID is required after multi-references are added */
  219. cmp.l.f0.0 (2)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  220. (f0.0)   mov (2)        mbb_result.20<1>:w                      -1:w    {align1};
  221. (f0.0)   jmpi   (1)     mbc_start;
  222. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  223. /* Read MV for MB B */
  224. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  225. send (16)
  226.        mb_ind
  227.        mb_mv0.0<1>:ud
  228.         NULL
  229.        data_port(
  230.                OBR_CACHE_TYPE,
  231.                OBR_MESSAGE_TYPE,
  232.                OBR_CONTROL_8,
  233.                OBR_BIND_IDX,
  234.                OBR_WRITE_COMMIT_CATEGORY,
  235.                OBR_HEADER_PRESENT
  236.        )
  237.        mlen 1
  238.        rlen 4
  239.        {align1};
  240. /* TODO: RefID is required after multi-references are added */
  241. mov        (2)          mbb_result.20<1>:w              -1:w    {align1};
  242. mov     (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
  243. mov     (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
  244. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
  245. SAVE_RET        {align1};
  246. jmpi    (1)     mb_pred_func;
  247. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  248. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  249. (f0.0)  mov   (1)       mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
  250. (f0.0)  mov   (1)       mbb_result.20<1>:w              0:w             {align1};
  251. (f0.0)  mov   (1)       mbb_result.4<1>:ud              mb_mv2.16<0,1,0>:ud     {align1};
  252. (f0.0)  jmpi    (1) mbc_start;
  253. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  254. (f0.0)  mov   (1)       mbb_result.18<1>:uw             MB_PRED_FLAG            {align1};
  255. (f0.0)  mov   (1)       mbb_result.22<1>:w              0:w             {align1};
  256. (f0.0)  mov   (1)       mbb_result.8<1>:ud              mb_mv2.20<0,1,0>:ud     {align1};
  257. (f0.0)  jmpi    (1) mbc_start;
  258. mov   (2)       mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
  259. mov   (2)       mbb_result.20<1>:w              0:w             {align1};
  260. mov        (2)          mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  261.  
  262. mbc_start:
  263. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  264. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  265. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  266. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  267. (f0.0)  jmpi (1)        mbd_start;
  268. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  269. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  270. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  271. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  272. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  273. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  274. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  275. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  276.  
  277. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  278. send (16)
  279.        mb_ind
  280.        mb_wb.0<1>:ud
  281.         NULL
  282.        data_port(
  283.                OBR_CACHE_TYPE,
  284.                OBR_MESSAGE_TYPE,
  285.                OBR_CONTROL_4,
  286.                OBR_BIND_IDX,
  287.                OBR_WRITE_COMMIT_CATEGORY,
  288.                OBR_HEADER_PRESENT
  289.        )
  290.        mlen 1
  291.        rlen 2
  292.        {align1};
  293.  
  294. /* TODO: RefID is required after multi-references are added */
  295. cmp.l.f0.0 (2)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  296. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  297. (f0.0)   jmpi   (1)     mb_mvp_start;
  298. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  299. /* Read MV for MB C */
  300. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  301. send (16)
  302.        mb_ind
  303.        mb_mv0.0<1>:ud
  304.         NULL
  305.        data_port(
  306.                OBR_CACHE_TYPE,
  307.                OBR_MESSAGE_TYPE,
  308.                OBR_CONTROL_8,
  309.                OBR_BIND_IDX,
  310.                OBR_WRITE_COMMIT_CATEGORY,
  311.                OBR_HEADER_PRESENT
  312.        )
  313.        mlen 1
  314.        rlen 4
  315.        {align1};
  316. /* TODO: RefID is required after multi-references are added */
  317. /* Forward MV */
  318. mov        (2)          mbc_result.20<1>:w              -1:w    {align1};
  319. mov     (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
  320. mov     (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
  321. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
  322. SAVE_RET        {align1};
  323. jmpi    (1)     mb_pred_func;
  324. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  325. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  326. (f0.0)  mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
  327. (f0.0)  mov    (1) mbc_result.20<1>:w           0:w             {align1};
  328. (f0.0)  mov    (1)  mbc_result.4<1>:ud          mb_mv2.16<0,1,0>:ud     {align1};
  329. (f0.0)  jmpi   (1) mb_mvp_start;
  330. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  331. (f0.0)  mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
  332. (f0.0)  mov    (1) mbc_result.22<1>:w           0:w             {align1};
  333. (f0.0)  mov    (1)  mbc_result.8<1>:ud          mb_mv2.20<0,1,0>:ud     {align1};
  334. (f0.0)  jmpi  (1) mb_mvp_start;
  335. mov  (2)  mbc_result.16<1>:uw           MB_PRED_FLAG            {align1};
  336. mov  (2)  mbc_result.20<1>:w            0:w     {align1};
  337. mov        (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  338.  
  339. jmpi   (1)    mb_mvp_start;
  340. mbd_start:
  341. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  342. and.z.f0.0 (2)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  343. (f0.0)  mov  (2)        mbc_result.20<1>:w              -1:w    {align1};
  344. (f0.0)  jmpi (1)        mb_mvp_start;
  345. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  346. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  347. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  348. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  349. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  350. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  351. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  352.  
  353. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  354. send (16)
  355.        mb_ind
  356.        mb_wb.0<1>:ud
  357.         NULL
  358.        data_port(
  359.                OBR_CACHE_TYPE,
  360.                OBR_MESSAGE_TYPE,
  361.                OBR_CONTROL_4,
  362.                OBR_BIND_IDX,
  363.                OBR_WRITE_COMMIT_CATEGORY,
  364.                OBR_HEADER_PRESENT
  365.        )
  366.        mlen 1
  367.        rlen 2
  368.        {align1};
  369.  
  370. cmp.l.f0.0 (2)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  371. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  372. (f0.0)   jmpi   (1)     mb_mvp_start;
  373.  
  374. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  375. /* Read MV for MB D */
  376. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  377. send (16)
  378.        mb_ind
  379.        mb_mv0.0<1>:ub
  380.         NULL
  381.        data_port(
  382.                OBR_CACHE_TYPE,
  383.                OBR_MESSAGE_TYPE,
  384.                OBR_CONTROL_8,
  385.                OBR_BIND_IDX,
  386.                OBR_WRITE_COMMIT_CATEGORY,
  387.                OBR_HEADER_PRESENT
  388.        )
  389.        mlen 1
  390.        rlen 4
  391.        {align1};
  392.  
  393. /* TODO: RefID is required after multi-references are added */
  394.  
  395. /* Forward MV */
  396. mov        (2)          mbc_result.20<1>:w              -1:w    {align1};
  397. mov     (1)     INPUT_ARG0.0<1>:ud      mb_inter_wb.4<0,1,0>:ud {align1};
  398. mov     (1)     INPUT_ARG0.4<1>:ud      mb_inter_wb.0<0,1,0>:ud {align1};
  399. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK3:ud {align1};
  400. SAVE_RET        {align1};
  401. jmpi    (1)     mb_pred_func;
  402. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  403. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  404. (f0.0)  mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
  405. (f0.0)  mov    (1) mbc_result.20<1>:w           0:w             {align1};
  406. (f0.0)  mov    (1) mbc_result.4<1>:ud           mb_mv3.24<0,1,0>:ud     {align1};
  407. (f0.0)  jmpi    (1) mb_mvp_start;
  408. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  409. (f0.0)  mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
  410. (f0.0)  mov    (1) mbc_result.22<1>:w           0:w             {align1};
  411. (f0.0)  mov    (1) mbc_result.8<1>:ud           mb_mv3.28<0,1,0>:ud     {align1};
  412. (f0.0)  jmpi  (1) mb_mvp_start;
  413. mov  (2)  mbc_result.16<1>:uw           MB_PRED_FLAG            {align1};
  414. mov  (2)  mbc_result.20<1>:w            0:w     {align1};
  415. mov        (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
  416.        
  417. mb_mvp_start:
  418. /*TODO: Add the skip prediction */
  419. /* Check whether both MB B and C are inavailable */
  420. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  421. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  422. (-f0.0) jmpi (1)        mb_median_start;
  423. cmp.nz.f0.0 (2) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  424. (f0.0)  mov     (2)     mbb_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};      
  425. (f0.0)  mov     (2)     mbc_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};      
  426. (f0.0)  mov     (2)     mbb_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};      
  427. (f0.0)  mov     (2)     mbc_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};      
  428. (f0.0)  mov     (2)     mb_mvp_ref.0<1>:ud              mba_result.4<2,2,1>:ud  {align1};
  429. (-f0.0) mov     (2)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  430. jmpi    (1)     __mb_hwdep_end;
  431.        
  432. mb_median_start:
  433. /* forward_MVP */
  434. /* check whether only one neighbour MB has the same ref ID with the current MB */
  435. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  436. cmp.z.f0.0      (1)     null:d  mba_result.20<0,1,0>:w  0:w     {align1};
  437. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  438. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  439. cmp.z.f0.0      (1)     null:d  mbb_result.20<0,1,0>:w  0:w     {align1};
  440. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  441. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  442. cmp.z.f0.0      (1)     null:d  mbc_result.20<0,1,0>:w  0:w     {align1};
  443. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  444. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  445. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
  446. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  447. (f0.0)  jmpi (1)  mvp_backward;
  448.  
  449. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  450. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  451. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  452. SAVE_RET        {align1};
  453.  jmpi   (1)     word_imedian;
  454. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  455. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  456. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  457. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  458. SAVE_RET        {align1};
  459. jmpi    (1)     word_imedian;
  460. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  461.  
  462.  
  463. mvp_backward:
  464. /* check whether only one neighbour MB has the same ref ID with the current MB */
  465. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  466. cmp.z.f0.0      (1)     null:d  mba_result.22<0,1,0>:w  0:w     {align1};
  467. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  468. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.8<0,1,0>:ud  {align1};
  469. cmp.z.f0.0      (1)     null:d  mbb_result.22<0,1,0>:w  0:w     {align1};
  470. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  471. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.8<0,1,0>:ud  {align1};
  472. cmp.z.f0.0      (1)     null:d  mbc_result.22<0,1,0>:w  0:w     {align1};
  473. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  474. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.8<0,1,0>:ud  {align1};
  475. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
  476. (f0.0)  mov     (1)     mb_mvp_ref.4<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  477. (f0.0)  jmpi (1) __mb_hwdep_end;
  478.  
  479. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.8<0,1,0>:w   {align1};
  480. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.8<0,1,0>:w   {align1};
  481. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.8<0,1,0>:w   {align1};
  482. SAVE_RET        {align1};
  483.  jmpi   (1)     word_imedian;
  484. mov     (1)     mb_mvp_ref.4<1>:w               RET_ARG<0,1,0>:w        {align1};
  485. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.10<0,1,0>:w  {align1};
  486. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.10<0,1,0>:w  {align1};
  487. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.10<0,1,0>:w  {align1};
  488. SAVE_RET        {align1};
  489. jmpi    (1)     word_imedian;
  490. mov     (1)     mb_mvp_ref.6<1>:w               RET_ARG<0,1,0>:w        {align1};
  491.  
  492. __mb_hwdep_end:
  493. asr     (4)     mb_ref_win.0<1>:w       mb_mvp_ref.0<4,4,1>:w   2:w     {align1};
  494. add     (4)     mb_ref_win.8<1>:w       mb_ref_win.0<4,4,1>:w   3:w     {align1};
  495. and     (4)     mb_ref_win.16<1>:uw     mb_ref_win.8<4,4,1>:uw  0xFFFC:uw {align1};
  496. /* m2, get the MV/Mb cost passed from constant buffer when
  497. spawning thread by MEDIA_OBJECT */      
  498. mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
  499.  
  500. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  501. /* m3 cost center */
  502. mov (8) vme_m3.0<1>:ud          0x0:ud  {align1};
  503. mov (8) vme_msg_3<1>:UD         vme_m3.0<8,8,1>:UD {align1};           
  504.  
  505. /* m4. skip center */
  506. mov (8) vme_msg_4<1>:ud         0x0:ud  {align1};
  507.  
  508. /* m5 */
  509. mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
  510. and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
  511. mov  (8) vme_msg_5<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
  512. /* Use the Luma mode */
  513. mov  (1) tmp_reg0.0<1>:UW       LUMA_INTRA_MODE:UW {align1};
  514. mov  (1) vme_msg_5.5<1>:UB      tmp_reg0.0<0,1,0>:UB {align1};
  515.  
  516. /* m6 */        
  517. mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
  518. mov (16) vme_msg_6.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
  519. mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
  520.  
  521. /* the penalty for Intra mode */
  522. mov  (1) vme_msg_6.28<1>:UD     0x010101:UD {align1};
  523. mov  (1) vme_msg_6.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
  524.  
  525.  
  526. /* m7 */
  527.  
  528. mov  (4) vme_msg_7.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
  529. mov  (8) vme_msg_7.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
  530.  
  531. /*
  532. * SIC VME message
  533. */
  534. /* m1 */
  535. mov  (1) intra_flag<1>:UW       0x0:UW {align1}                     ;
  536. and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
  537. (f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1};
  538.  
  539. /* assign MB intra struct from the thread payload*/
  540. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  541.  
  542. /* Disable DC HAAR component when calculating HARR SATD block */
  543. mov  (1) tmp_reg0.0<1>:UW       DC_HARR_DISABLE:UW              {align1};
  544. mov  (1) vme_m1.30<1>:UB        tmp_reg0.0<0,1,0>:UB  {align1};
  545. mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
  546.  
  547. /* m0 */        
  548. mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
  549. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  550.  
  551. /* after verification it will be passed by using payload */
  552. send (8)
  553.        vme_msg_ind
  554.        vme_wb<1>:UD
  555.        null
  556.        cre(
  557.                BIND_IDX_VME,
  558.                VME_SIC_MESSAGE_TYPE
  559.        )
  560.        mlen sic_vme_msg_length
  561.        rlen vme_wb_length
  562.        {align1};
  563. /*
  564. * Oword Block Write message
  565. */
  566. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  567.        
  568. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  569. mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
  570. mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
  571. mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
  572.  
  573. /* Distortion, Intra (17-16), */
  574. mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
  575.  
  576. mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
  577. /* VME clock counts */
  578. mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
  579.  
  580. mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
  581.  
  582. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  583. send (16)
  584.        msg_ind
  585.        obw_wb
  586.        null
  587.        data_port(
  588.                OBW_CACHE_TYPE,
  589.                OBW_MESSAGE_TYPE,
  590.                OBW_CONTROL_2,
  591.                OBW_BIND_IDX,
  592.                OBW_WRITE_COMMIT_CATEGORY,
  593.                OBW_HEADER_PRESENT
  594.        )
  595.        mlen 2
  596.        rlen obw_wb_length
  597.        {align1};
  598.  
  599. /* IME search */
  600. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_DUAL_REFERENCE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  601. mov  (1) vme_m0.22<1>:UW        DREF_REGION_SIZE {align1};        
  602. /* Dual Reference Width&Height,32x32 */
  603.  
  604. mov  (1) vme_m0.0<1>:UD         vme_m0.8<0,1,0>:UD      {align1};
  605.  
  606. /* Reference = (x-8,y-8)-(x+8,y+8) */
  607. add  (1) vme_m0.0<1>:W          vme_m0.0<0,1,0>:W -8:W {align1};               
  608. add  (1) vme_m0.2<1>:W          vme_m0.2<0,1,0>:W -8:W {align1};
  609.  
  610. mov  (1) vme_m0.0<1>:W          -8:W                    {align1};
  611. mov  (1) vme_m0.2<1>:W          -8:W                    {align1};
  612.  
  613. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  614. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  615. (f0.0)  add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       4:w     {align1};
  616. (f0.0)  add     (1)     vme_m0.4<1>:w   vme_m0.4<0,1,0>:w       4:w     {align1};
  617. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  618. (f0.0)  add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       4:w     {align1};
  619. (f0.0)  add     (1)     vme_m0.6<1>:w   vme_m0.6<0,1,0>:w       4:w     {align1};
  620.  
  621. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  622. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.20<2,2,1>:w  {align1};
  623.  
  624. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  625.  
  626. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  627. /* the Max MV number is passed by constant buffer */
  628. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  629. mov  (1) vme_m1.8<1>:UD         DSTART_CENTER + DSEARCH_PATH_LEN:UD {align1};
  630. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  631.  
  632. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  633.  
  634.  
  635. /* Setup the Cost center */
  636. /* currently four 8x8 share the same cost center */
  637. mov  (4) vme_m3.0<2>:ud         mb_mvp_ref.0<0,1,0>:ud  {align1};
  638. mov  (4) vme_m3.4<2>:ud         mb_mvp_ref.4<0,1,0>:ud  {align1};
  639.  
  640. /* M4/M5 search path */
  641.  
  642. mov  (1) vme_msg_4.0<1>:UD      0x10010101:UD {align1};
  643. mov  (1) vme_msg_4.4<1>:UD      0x100F0F0F:UD {align1};
  644. mov  (1) vme_msg_4.8<1>:UD      0x10010101:UD {align1};
  645. mov  (1) vme_msg_4.12<1>:UD     0x000F0F0F:UD {align1};
  646.  
  647. mov  (4) vme_msg_4.16<1>:UD     0x0:UD {align1};
  648. mov  (8) vme_msg_5.16<1>:UD     0x0:UD {align1};
  649.  
  650. send (8)
  651.        vme_msg_ind
  652.        vme_wb<1>:UD
  653.        null
  654.        vme(
  655.                BIND_IDX_VME,
  656.                0,
  657.                0,
  658.                VME_IME_MESSAGE_TYPE
  659.        )
  660.        mlen ime_vme_msg_length
  661.        rlen vme_wb_length {align1};
  662.  
  663. /* Set Macroblock-shape/mode for FBR */
  664.  
  665. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  666. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  667. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  668.  
  669. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  670. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  671.  
  672. /* Send FBR message into CRE */
  673.  
  674. mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  675. mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  676. mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  677. mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  678.  
  679. /* 16x16 Source, 1/4 pixel, harr, BME ENABLE */
  680. mov  (1) vme_m0.12<1>:UD        INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER + FBR_BME_ENABLE:UD {align1};
  681.  
  682. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  683.  
  684. mov  (1) tmp_reg0.0<1>:uw       BI_WEIGHT       {align1};
  685. mov  (1) vme_m1.6<1>:UB         tmp_reg0.0<0,1,0>:ub    {align1};
  686. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  687.  
  688. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  689. mov  (8) vme_msg_3.0<1>:UD              vme_m3.0<8,8,1>:UD      {align1};
  690.  
  691. /* after verification it will be passed by using payload */
  692. send (8)
  693.        vme_msg_ind
  694.        vme_wb<1>:UD
  695.        null
  696.        cre(
  697.                BIND_IDX_VME,
  698.                VME_FBR_MESSAGE_TYPE
  699.        )
  700.        mlen fbr_vme_msg_length
  701.        rlen vme_wb_length
  702.        {align1};
  703.  
  704. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
  705. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  706. /* write FME info */
  707. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  708.  
  709. mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  710. /* Inter distortion of FME */
  711. mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
  712.  
  713. mov  (1) msg_reg1.12<1>:UD      vme_m2.20<0,1,0>:UD {align1};
  714.  
  715. /* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
  716. send (16)
  717.        msg_ind
  718.        obw_wb
  719.        null
  720.        data_port(
  721.                OBW_CACHE_TYPE,
  722.                OBW_MESSAGE_TYPE,
  723.                OBW_CONTROL_0,
  724.                OBW_BIND_IDX,
  725.                OBW_WRITE_COMMIT_CATEGORY,
  726.                OBW_HEADER_PRESENT
  727.        )
  728.        mlen 2
  729.        rlen obw_wb_length
  730.        {align1};
  731.  
  732. /* Write FME/BME MV */
  733. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
  734. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  735.  
  736.  
  737. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  738. mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  739. mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  740. mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  741. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  742. send (16)
  743.        msg_ind
  744.        obw_wb
  745.        null
  746.        data_port(
  747.                OBW_CACHE_TYPE,
  748.                OBW_MESSAGE_TYPE,
  749.                OBW_CONTROL_8,
  750.                OBW_BIND_IDX,
  751.                OBW_WRITE_COMMIT_CATEGORY,
  752.                OBW_HEADER_PRESENT
  753.        )
  754.        mlen 5
  755.        rlen obw_wb_length
  756.        {align1};
  757.  
  758. /* Write FME/BME RefID */
  759. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
  760. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  761.  
  762. mov  (8) msg_reg1.0<1>:UD       vme_wb6.0<8,8,1>:UD {align1};
  763.  
  764. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  765. send (16)
  766.        msg_ind
  767.        obw_wb
  768.        null
  769.        data_port(
  770.                OBW_CACHE_TYPE,
  771.                OBW_MESSAGE_TYPE,
  772.                OBW_CONTROL_2,
  773.                OBW_BIND_IDX,
  774.                OBW_WRITE_COMMIT_CATEGORY,
  775.                OBW_HEADER_PRESENT
  776.        )
  777.        mlen 2
  778.        rlen obw_wb_length
  779.        {align1};
  780.  
  781.  
  782. /* Issue message fence so that the previous write message is committed */
  783. send (16)
  784.        mb_ind
  785.        mb_wb.0<1>:ud
  786.         NULL
  787.        data_port(
  788.                OBR_CACHE_TYPE,
  789.                OBR_MESSAGE_FENCE,
  790.                OBR_MF_COMMIT,
  791.                OBR_BIND_IDX,
  792.                OBR_WRITE_COMMIT_CATEGORY,
  793.                OBR_HEADER_PRESENT
  794.        )
  795.        mlen 1
  796.        rlen 1
  797.        {align1};
  798.  
  799. __EXIT:
  800. /*
  801. * kill thread
  802. */        
  803. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  804. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  805.  
  806.  
  807.         nop             ;
  808.         nop             ;
  809. /* Compare three word data to get the min value */
  810. word_imin:
  811.         cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  812.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  813.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  814.         cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  815.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  816.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  817.         RETURN          {align1};      
  818.        
  819. /* Compare three word data to get the max value */
  820. word_imax:
  821.         cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  822.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  823.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  824.         cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  825.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  826.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  827.         RETURN          {align1};      
  828.        
  829. word_imedian:
  830.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  831.         (f0.0)  jmpi (1) cmp_a_ge_b;
  832.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  833.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  834.         (f0.0) jmpi (1) cmp_end;
  835.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  836.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  837.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  838.         jmpi (1) cmp_end;
  839. cmp_a_ge_b:
  840.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  841.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  842.         (f0.0) jmpi (1) cmp_end;
  843.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  844.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  845.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  846. cmp_end:
  847.         RETURN  {align1};
  848.  
  849. mb_pred_func:
  850.         mov     (8)     TEMP_VAR0.0<1>:ud       0:ud    {align1};
  851.         mov     (1)     TEMP_VAR0.0<1>:ub       INPUT_ARG0.2<0,1,0>:ub  {align1};
  852.         and     (1)   TEMP_VAR0.4<1>:uw   INPUT_ARG0.4<0,1,0>:uw        INTER_MASK:uw   {align1};
  853.         /* INTER16x16 mode. The bit1-0 is the prediction mode */
  854.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_16X16MODE:uw      {align1};
  855.         (f0.0)  and (1) RET_ARG<1>:uw   TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  856.         (f0.0)  jmpi (1) end_mb_pred;
  857.         /* Check whether it is INTER8x8 mode. */
  858.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_8X8MODE:uw        {align1};
  859.         (f0.0)  jmpi (1) mb_pred_func_8;
  860.        
  861.         /* Check whether it is INTER16x8 mode. */
  862.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<0,1,0>:uw   INTER_16X8MODE:uw       {align1};
  863.         (f0.0)  jmpi (1) mb_pred_func_168;
  864. mb_pred_func_816:      
  865.         /* Block 0/2 uses the bit1-0. Block 1/3 uses the bit3-2 */
  866.         mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
  867.         and.z.f0.0 (1)  null:uw TEMP_VAR0.8<0,1,0>:uw   INTER_BLOCK1:uw {align1};
  868.         (f0.0)  and     (1) RET_ARG<1>:uw       TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  869.         (f0.0)  jmpi (1) end_mb_pred;
  870.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
  871.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};   
  872.         jmpi    (1) end_mb_pred;
  873.  
  874. mb_pred_func_168:
  875.         /* Block 0/1 uses the bit1-0. Block 2/3 uses the bit3-2 */
  876.         mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
  877.         cmp.l.f0.0 (1)  null:uw TEMP_VAR0.8<0,1,0>:uw   INTER_BLOCK2:uw {align1};
  878.         (f0.0)  and  (1) RET_ARG<1>:uw  TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  879.         (f0.0)  jmpi (1) end_mb_pred;
  880.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
  881.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};   
  882.         jmpi    (1) end_mb_pred;
  883.  
  884. mb_pred_func_8:
  885.         /* 8X8 mode. Every block uses two bits as the prediction mode. */
  886.         mul     (1)  TEMP_VAR0.8<1>:uw  INPUT_ARG0.8<0,1,0>:uw  2:uw {align1};
  887.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   TEMP_VAR0.8<0,1,0>:uw {align1};
  888.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<0,1,0>:uw  PRED_MASK  {align1};   
  889. end_mb_pred:
  890.         RETURN  {align1};
  891.        
  892.