Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © <2010>, Intel Corporation.
  3.  *
  4.  * Permission is hereby granted, free of charge, to any person obtaining a
  5.  * copy of this software and associated documentation files (the
  6.  * "Software"), to deal in the Software without restriction, including
  7.  * without limitation the rights to use, copy, modify, merge, publish,
  8.  * distribute, sub license, and/or sell copies of the Software, and to
  9.  * permit persons to whom the Software is furnished to do so, subject to
  10.  * the following conditions:
  11.  *
  12.  * The above copyright notice and this permission notice (including the
  13.  * next paragraph) shall be included in all copies or substantial portions
  14.  * of the Software.
  15.  *
  16.  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  17.  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  18.  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  19.  * IN NO EVENT SHALL PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR
  20.  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  21.  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  22.  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23.  * Authors: Zhao Yakui <yakui.zhao@intel.com>
  24.  *
  25.  */
  26. // Modual name: InterFrame_ivy.asm
  27. //
  28. // Make intra predition estimation for Inter-B frame
  29. //
  30.  
  31. //
  32. //  Now, begin source code....
  33. //
  34.  
  35. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  36. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  37.  
  38. /*
  39.  * __START
  40.  */
  41. __INTER_START:
  42. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  43. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  44. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  45. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  46.  
  47. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  48. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  49. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  50. mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
  51. mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  52.  
  53. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  54. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  55. mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
  56. mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  57.        
  58. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  59. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  60.  
  61. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  62. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  63. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  64. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  65.        
  66. /*
  67.  * Media Read Message -- fetch Luma neighbor edge pixels
  68.  */
  69. /* ROW */
  70. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  71. send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
  72.  
  73. /* COL */
  74. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  75. send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
  76.        
  77. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  78. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  79. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  80. (f0.0) jmpi (1) __mb_hwdep_end;
  81. /* read back the data for MB A */
  82. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  83. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  84. */
  85. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  86. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  87. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  88. mba_start:
  89. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  90. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  91. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  92. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  93. (f0.0)  jmpi (1)        mbb_start;
  94. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  95. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  96. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  97. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  98. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  99. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  100. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  101. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  102.  
  103. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  104. /* bind index 3, read 1 oword (16bytes), 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_0,
  113.                OBR_BIND_IDX,
  114.                OBR_WRITE_COMMIT_CATEGORY,
  115.                OBR_HEADER_PRESENT
  116.        )
  117.        mlen 1
  118.        rlen 1
  119.        {align1};
  120.  
  121. /* TODO: RefID is required after multi-references are added */
  122. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0 <0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  123. (-f0.0)   mov (2)       mba_result.20<1>:w                      -1:w    {align1};
  124. (-f0.0)   jmpi  (1)     mbb_start;
  125.  
  126. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>: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     (2)     mba_result.20<1>:w              -1:w            {align1};
  148. mov     (1)     INPUT_ARG0.0<1>:ud      mb_mode_wb.4<0,1,0>:ud  {align1};
  149. mov     (1)     INPUT_ARG0.4<1>:ud      mb_mode_wb.0<0,1,0>:ud  {align1};
  150. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK1:ud {align1};
  151. SAVE_RET        {align1};
  152. jmpi    (1)     mb_pred_func;
  153. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  154. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  155. (f0.0)  mov   (1)       mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
  156. (f0.0)  mov   (1)       mba_result.20<1>:w              0:w             {align1};
  157. (f0.0)  jmpi    (1) mbb_start;
  158. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  159. (f0.0)  mov   (1)       mba_result.18<1>:uw             MB_PRED_FLAG            {align1};
  160. (f0.0)  mov   (1)       mba_result.22<1>:w              0:w             {align1};
  161. (f0.0)  jmpi    (1) mbb_start;
  162. mov   (2)       mba_result.16<1>:uw             MB_PRED_FLAG            {align1};
  163. mov   (2)       mba_result.20<1>:w              0:w             {align1};
  164.  
  165. mbb_start:
  166. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  167. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  168. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  169. /* If MB B doesn't exist, neither MB C nor D exists */
  170. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  171. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  172. (f0.0)  jmpi (1)        mb_mvp_start;
  173. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  174. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  175. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  176. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  177. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  178. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  179. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  180. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  181.  
  182. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  183.  
  184. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  185. send (16)
  186.        mb_ind
  187.        mb_wb.0<1>:ud
  188.         NULL
  189.        data_port(
  190.                OBR_CACHE_TYPE,
  191.                OBR_MESSAGE_TYPE,
  192.                OBR_CONTROL_0,
  193.                OBR_BIND_IDX,
  194.                OBR_WRITE_COMMIT_CATEGORY,
  195.                OBR_HEADER_PRESENT
  196.        )
  197.        mlen 1
  198.        rlen 1
  199.        {align1};
  200.  
  201. /* TODO: RefID is required after multi-references are added */
  202. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0 <0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  203. (-f0.0)   mov (2)       mbb_result.20<1>:w                      -1:w    {align1};
  204. (-f0.0)   jmpi  (1)     mbc_start;
  205.  
  206. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  207. /* Read MV for MB B */
  208. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  209. send (16)
  210.        mb_ind
  211.        mb_mv0.0<1>:ud
  212.         NULL
  213.        data_port(
  214.                OBR_CACHE_TYPE,
  215.                OBR_MESSAGE_TYPE,
  216.                OBR_CONTROL_8,
  217.                OBR_BIND_IDX,
  218.                OBR_WRITE_COMMIT_CATEGORY,
  219.                OBR_HEADER_PRESENT
  220.        )
  221.        mlen 1
  222.        rlen 4
  223.        {align1};
  224. /* TODO: RefID is required after multi-references are added */
  225.  
  226. mov     (2)     mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  227. mov     (2)     mbb_result.20<1>:w              -1:w            {align1};
  228. mov     (1)     INPUT_ARG0.0<1>:ud      mb_mode_wb.4<0,1,0>:ud  {align1};
  229. mov     (1)     INPUT_ARG0.4<1>:ud      mb_mode_wb.0<0,1,0>:ud  {align1};
  230. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
  231. SAVE_RET        {align1};
  232. jmpi    (1)     mb_pred_func;
  233. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  234. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  235. (f0.0)  mov   (1)       mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
  236. (f0.0)  mov   (1)       mbb_result.20<1>:w              0:w             {align1};
  237. (f0.0)  jmpi    (1) mbc_start;
  238. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  239. (f0.0)  mov   (1)       mbb_result.18<1>:uw             MB_PRED_FLAG            {align1};
  240. (f0.0)  mov   (1)       mbb_result.22<1>:w              0:w             {align1};
  241. (f0.0)  jmpi    (1) mbc_start;
  242. mov   (2)       mbb_result.16<1>:uw             MB_PRED_FLAG            {align1};
  243.  
  244. mbc_start:
  245. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  246. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  247. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  248. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  249. (f0.0)  jmpi (1)        mbd_start;
  250. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  251. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  252. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  253. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  254. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  255. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  256. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  257. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  258.  
  259. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  260.  
  261. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  262. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  263. send (16)
  264.        mb_ind
  265.        mb_wb.0<1>:ud
  266.         NULL
  267.        data_port(
  268.                OBR_CACHE_TYPE,
  269.                OBR_MESSAGE_TYPE,
  270.                OBR_CONTROL_0,
  271.                OBR_BIND_IDX,
  272.                OBR_WRITE_COMMIT_CATEGORY,
  273.                OBR_HEADER_PRESENT
  274.        )
  275.        mlen 1
  276.        rlen 1
  277.        {align1};
  278.  
  279. /* TODO: RefID is required after multi-references are added */
  280. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0 <0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  281. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  282. (-f0.0)   jmpi  (1)     mb_mvp_start;
  283. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud {align1};
  284. /* Read MV for MB C */
  285. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  286. send (16)
  287.        mb_ind
  288.        mb_mv0.0<1>:ud
  289.         NULL
  290.        data_port(
  291.                OBR_CACHE_TYPE,
  292.                OBR_MESSAGE_TYPE,
  293.                OBR_CONTROL_8,
  294.                OBR_BIND_IDX,
  295.                OBR_WRITE_COMMIT_CATEGORY,
  296.                OBR_HEADER_PRESENT
  297.        )
  298.        mlen 1
  299.        rlen 4
  300.        {align1};
  301. /* TODO: RefID is required after multi-references are added */
  302. /* Forward MV */
  303. mov        (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  304. mov        (2)          mbc_result.20<1>:w              -1:w    {align1};
  305. mov     (1)     INPUT_ARG0.0<1>:ud      mb_mode_wb.4<0,1,0>:ud  {align1};
  306. mov     (1)     INPUT_ARG0.4<1>:ud      mb_mode_wb.0<0,1,0>:ud  {align1};
  307. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK2:ud {align1};
  308. SAVE_RET        {align1};
  309. jmpi    (1)     mb_pred_func;
  310. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  311. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  312. (f0.0)  mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
  313. (f0.0)  mov    (1) mbc_result.20<1>:w           0:w             {align1};
  314. (f0.0)  jmpi   (1) mb_mvp_start;
  315. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  316. (f0.0)  mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
  317. (f0.0)  mov    (1) mbc_result.22<1>:w           0:w             {align1};
  318. (f0.0)  jmpi  (1) mb_mvp_start;
  319. mov  (2)  mbc_result.16<1>:uw           MB_PRED_FLAG            {align1};
  320. mov  (2)  mbc_result.20<1>:w            0:w     {align1};
  321.  
  322. jmpi   (1)    mb_mvp_start;
  323. mbd_start:
  324. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  325. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  326. (f0.0)  jmpi (1)        mb_mvp_start;
  327. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  328. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  329. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  330. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  331. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  332.  
  333. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD INTER_VME_OUTPUT_IN_OWS:UD {align1};
  334. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  335. mov  (1) mb_msg_tmp.8<1>:ud     mb_msg0.8<0,1,0>:ud     {align1};
  336.  
  337. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  338. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  339. send (16)
  340.        mb_ind
  341.        mb_wb.0<1>:ud
  342.         NULL
  343.        data_port(
  344.                OBR_CACHE_TYPE,
  345.                OBR_MESSAGE_TYPE,
  346.                OBR_CONTROL_0,
  347.                OBR_BIND_IDX,
  348.                OBR_WRITE_COMMIT_CATEGORY,
  349.                OBR_HEADER_PRESENT
  350.        )
  351.        mlen 1
  352.        rlen 1
  353.        {align1};
  354.  
  355. and.z.f0.0      (1)     null<1>:ud        mb_mode_wb.0 <0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  356. (-f0.0)   mov (2)       mbc_result.20<1>:w                      -1:w    {align1};
  357. (-f0.0)   jmpi  (1)     mb_mvp_start;
  358.  
  359. mov   (1) mb_msg0.8<1>:UD       mb_msg_tmp.8<0,1,0>:ud  {align1};
  360. /* Read MV for MB D */
  361. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  362. send (16)
  363.        mb_ind
  364.        mb_mv0.0<1>:ub
  365.         NULL
  366.        data_port(
  367.                OBR_CACHE_TYPE,
  368.                OBR_MESSAGE_TYPE,
  369.                OBR_CONTROL_8,
  370.                OBR_BIND_IDX,
  371.                OBR_WRITE_COMMIT_CATEGORY,
  372.                OBR_HEADER_PRESENT
  373.        )
  374.        mlen 1
  375.        rlen 4
  376.        {align1};
  377.  
  378. /* TODO: RefID is required after multi-references are added */
  379.  
  380. /* Forward MV */
  381. mov        (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
  382. mov        (2)          mbc_result.20<1>:w              -1:w    {align1};
  383. mov     (1)     INPUT_ARG0.0<1>:ud      mb_mode_wb.4<0,1,0>:ud  {align1};
  384. mov     (1)     INPUT_ARG0.4<1>:ud      mb_mode_wb.0<0,1,0>:ud  {align1};
  385. mov     (1)     INPUT_ARG0.8<1>:ud      INTER_BLOCK3:ud {align1};
  386. SAVE_RET        {align1};
  387. jmpi    (1)     mb_pred_func;
  388. mov     (1)     mb_pred_mode.0<1>:uw    RET_ARG<0,1,0>:uw       {align1};
  389. mov        (1)  mbc_result.18<1>:w              MB_PRED_FLAG            {align1};
  390. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L0 {align1};
  391. (f0.0)  mov    (1) mbc_result.16<1>:uw          MB_PRED_FLAG            {align1};
  392. (f0.0)  mov    (1) mbc_result.20<1>:w           0:w             {align1};
  393. (f0.0)  jmpi    (1) mb_mvp_start;
  394. cmp.e.f0.0 (1)  null:uw         mb_pred_mode.0<0,1,0>:uw        PRED_L1 {align1};
  395. (f0.0)  mov   (1)       mbc_result.18<1>:uw             MB_PRED_FLAG            {align1};
  396. (f0.0)  mov    (1) mbc_result.22<1>:w           0:w             {align1};
  397. (f0.0)  jmpi  (1) mb_mvp_start;
  398. mov  (2)  mbc_result.16<1>:uw           MB_PRED_FLAG            {align1};
  399. mov  (2)  mbc_result.20<1>:w            0:w     {align1};
  400.        
  401. mb_mvp_start:
  402. /*TODO: Add the skip prediction */
  403. /* Check whether both MB B and C are invailable */
  404. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  405. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  406. (-f0.0) jmpi (1)        mb_median_start;
  407. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  408. (f0.0)  mov     (2)     mbb_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};      
  409. (f0.0)  mov     (2)     mbc_result.4<1>:ud              mba_result.4<2,2,1>:ud  {align1};      
  410. (f0.0)  mov     (2)     mbb_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};      
  411. (f0.0)  mov     (2)     mbc_result.20<1>:uw             mba_result.20<2,2,1>:uw {align1};      
  412. (f0.0)  mov     (2)     mb_mvp_ref.0<1>:ud              mba_result.4<2,2,1>:ud  {align1};
  413. (-f0.0) mov     (2)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  414. jmpi    (1)     __mb_hwdep_end;
  415.        
  416. mb_median_start:
  417. /* forward_MVP */
  418. /* check whether only one neighbour MB has the same ref ID with the current MB */
  419. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  420. cmp.z.f0.0      (1)     null:d  mba_result.20<1>:w      0:w     {align1};
  421. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  422. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  423. cmp.z.f0.0      (1)     null:d  mbb_result.20<1>:w      0:w     {align1};
  424. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  425. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  426. cmp.z.f0.0      (1)     null:d  mbc_result.20<1>:w      0:w     {align1};
  427. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  428. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  429. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  430. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  431. (f0.0)  jmpi (1)  mvp_backward;
  432.  
  433. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  434. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  435. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  436. SAVE_RET        {align1};
  437.  jmpi   (1)     word_imedian;
  438. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  439. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  440. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  441. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  442. SAVE_RET        {align1};
  443. jmpi    (1)     word_imedian;
  444. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  445.  
  446.  
  447. mvp_backward:
  448. /* check whether only one neighbour MB has the same ref ID with the current MB */
  449. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  450. cmp.z.f0.0      (1)     null:d  mba_result.22<1>:w      0:w     {align1};
  451. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  452. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.8<0,1,0>:ud  {align1};
  453. cmp.z.f0.0      (1)     null:d  mbb_result.22<1>:w      0:w     {align1};
  454. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  455. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.8<0,1,0>:ud  {align1};
  456. cmp.z.f0.0      (1)     null:d  mbc_result.22<1>:w      0:w     {align1};
  457. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<1>:w 1:w     {align1};
  458. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.8<0,1,0>:ud  {align1};
  459. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<1>:w  1:w    {align1};
  460. (f0.0)  mov     (1)     mb_mvp_ref.4<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  461. (f0.0)  jmpi (1) __mb_hwdep_end;
  462.  
  463. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.8<0,1,0>:w   {align1};
  464. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.8<0,1,0>:w   {align1};
  465. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.8<0,1,0>:w   {align1};
  466. SAVE_RET        {align1};
  467.  jmpi   (1)     word_imedian;
  468. mov     (1)     mb_mvp_ref.4<1>:w               RET_ARG<0,1,0>:w        {align1};
  469. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.10<0,1,0>:w  {align1};
  470. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.10<0,1,0>:w  {align1};
  471. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.10<0,1,0>:w  {align1};
  472. SAVE_RET        {align1};
  473. jmpi    (1)     word_imedian;
  474. mov     (1)     mb_mvp_ref.6<1>:w               RET_ARG<0,1,0>:w        {align1};
  475.  
  476. __mb_hwdep_end:
  477. asr     (4)     mb_ref_win.0<1>:w       mb_mvp_ref.0<4,4,1>:w   2:w     {align1};
  478. add     (4)     mb_ref_win.8<1>:w       mb_ref_win.0<4,4,1>:w   3:w     {align1};
  479. and     (4)     mb_ref_win.16<1>:uw     mb_ref_win.8<4,4,1>:uw  0xFFFC:uw {align1};
  480.  
  481. /* m2 */        
  482. mov  (8) vme_msg_2<1>:UD        0x0:UD {align1};
  483.  
  484. /* m3 */
  485. mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
  486. and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
  487. mov  (8) vme_msg_3<1>:UD        INEP_ROW.0<8,8,1>:UD {align1};        
  488.  
  489. /* m4 */
  490. mov  (8) vme_msg_4<1>:UD        0x0 {align1};
  491. mov (16) vme_msg_4.0<1>:UB      INEP_COL0.3<32,8,4>:UB {align1};
  492. mov  (1) vme_msg_4.16<1>:UD     INTRA_PREDICTORE_MODE {align1};
  493.  
  494.  
  495. /* m1 */
  496. mov  (8) vme_m1.0<1>:ud         0x0:ud  {align1};
  497. and.z.f0.0 (1) null<1>:UW transform_8x8_ub<0,1,0>:UB 1:UW {align1};
  498. (f0.0) mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE:uw {align1};
  499.  
  500. /* assign MB intra struct from the thread payload*/
  501. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  502.  
  503.  
  504. /* M0 */
  505. /* IME search */
  506. mov  (1) vme_m0.12<1>:UD   SEARCH_CTRL_DUAL_REFERENCE + INTER_PART_MASK + INTER_SAD_HAAR + SUB_PEL_MODE_QUARTER:UD {align1};  
  507. /* 16x16 Source, 1/4 pixel, harr */
  508. mov  (1) vme_m0.22<1>:UW        DREF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  509.  
  510. mov  (1) vme_m0.0<1>:UD         vme_m0.8<0,1,0>:UD      {align1};
  511.  
  512. mov  (1) vme_m0.0<1>:W          -8:W                    {align1};
  513. mov  (1) vme_m0.2<1>:W          -8:W                    {align1};
  514.  
  515. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  516. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  517. (f0.0)  add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       4:w     {align1};
  518. (f0.0)  add     (1)     vme_m0.4<1>:w   vme_m0.4<0,1,0>:w       4:w     {align1};
  519. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  520. (f0.0)  add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       4:w     {align1};
  521. (f0.0)  add     (1)     vme_m0.6<1>:w   vme_m0.6<0,1,0>:w       4:w     {align1};
  522.  
  523. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  524.  
  525. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:w  {align1};
  526. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.20<2,2,1>:w  {align1};
  527. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  528.  
  529. /* m1 */
  530.  
  531. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  532. /* MV num is passed by constant buffer. R4.28 */
  533. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};
  534. add  (1) vme_m1.4<1>:UD         vme_m1.4<0,1,0>:UD      FB_PRUNING_ENABLE:UD {align1};
  535. add  (1) vme_m1.6<1>:uw         vme_m1.6<0,1,0>:uw      BI_WEIGHT       {align1};
  536.  
  537.  
  538. mov  (1) vme_m1.8<1>:UD         DSTART_CENTER + DSEARCH_PATH_LEN:UD {align1};
  539.  
  540. /* Set the MV cost center */
  541. mov  (1) vme_m1.16<1>:ud        mb_mvp_ref.0<0,1,0>:ud  {align1};
  542. mov  (1) vme_m1.20<1>:ud        mb_mvp_ref.4<0,1,0>:ud  {align1};
  543. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  544.  
  545.  
  546. send (8)
  547.        vme_msg_ind
  548.        vme_wb
  549.        null
  550.        vme(
  551.                BIND_IDX_VME,
  552.                0,
  553.                0,
  554.                VME_MESSAGE_TYPE_MIXED
  555.        )
  556.        mlen vme_msg_length
  557.        rlen vme_inter_wb_length
  558.        {align1};
  559.  
  560. and.z.f0.0      (1)     null<1>:ud              vme_wb0.0<0,1,0>:ud     INTRAMBFLAG_MASK:ud {align1} ;
  561.  
  562. (-f0.0)jmpi     (1)     __INTRA_INFO ;
  563.  
  564. __INTER_INFO:
  565. /* Write MV pairs */   
  566. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  567.  
  568. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD   {align1};
  569.  
  570. mov  (8) msg_reg2.0<1>:UD       vme_wb2.0<8,8,1>:UD   {align1};
  571.  
  572. mov  (8) msg_reg3.0<1>:UD       vme_wb3.0<8,8,1>:UD   {align1};
  573.  
  574. mov  (8) msg_reg4.0<1>:UD       vme_wb4.0<8,8,1>:UD   {align1};                
  575. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  576. send (16)
  577.        msg_ind
  578.        obw_wb
  579.        null
  580.        data_port(
  581.                OBW_CACHE_TYPE,
  582.                OBW_MESSAGE_TYPE,
  583.                OBW_CONTROL_8,
  584.                OBW_BIND_IDX,
  585.                OBW_WRITE_COMMIT_CATEGORY,
  586.                OBW_HEADER_PRESENT
  587.        )
  588.        mlen 5
  589.        rlen obw_wb_length
  590.        {align1};
  591.  
  592. mov             (1)     tmp_uw1<1>:uw           0:uw {align1} ;
  593. mov             (1)     tmp_ud1<1>:ud           0:ud {align1} ;
  594. and             (1)     tmp_uw1<1>:uw           vme_wb0.2<0,1,0>:uw     MV32_BIT_MASK:uw {align1} ;
  595. shr             (1)     tmp_uw1<1>:uw           tmp_uw1<1>:uw           MV32_BIT_SHIFT:uw {align1} ;
  596. mul       (1)     tmp_ud1<1>:ud           tmp_uw1<0,1,0>:uw       96:uw {align1} ;
  597. add       (1)     tmp_ud1<1>:ud           tmp_ud1<0,1,0>:ud       32:uw {align1} ;
  598. shl       (1)     tmp_uw1<1>:uw           tmp_uw1<0,1,0>:uw       MFC_MV32_BIT_SHIFT:uw {align1} ;
  599. add       (1)     tmp_uw1<1>:uw           tmp_uw1<0,1,0>:uw       MVSIZE_UW_BASE:uw {align1} ;
  600. add             (1)     tmp_uw1<1>:uw           tmp_uw1<0,1,0>:uw       CBP_DC_YUV_UW:uw {align1} ;
  601.  
  602. mov             (1)     msg_reg1.0<1>:uw        vme_wb0.0<0,1,0>:uw     {align1} ;
  603. mov             (1)     msg_reg1.2<1>:uw        tmp_uw1<0,1,0>:uw       {align1} ;
  604. mov             (1)     msg_reg1.4<1>:UD        vme_wb0.28<0,1,0>:UD    {align1};
  605. mov             (1)     msg_reg1.8<1>:ud        tmp_ud1<0,1,0>:ud       {align1} ;
  606. mov             (1)     msg_reg1.12<1>:ud        vme_wb0.0<0,1,0>:ud     {align1} ;
  607. mov             (1)     msg_reg1.16<1>:ud        0x25:ud     {align1} ;
  608. jmpi            (1)     __OUTPUT_INFO;
  609.  
  610. __INTRA_INFO:
  611. mov             (1)     msg_reg1.0<1>:UD        vme_wb.0<0,1,0>:UD      {align1};
  612. mov             (1)     msg_reg1.4<1>:UD        vme_wb.16<0,1,0>:UD     {align1};
  613. mov             (1)     msg_reg1.8<1>:UD        vme_wb.20<0,1,0>:UD     {align1};
  614. mov             (1)     msg_reg1.12<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  615. mov             (1)     msg_reg1.16<1>:ud        0x35:ud     {align1} ;
  616.  
  617. __OUTPUT_INFO:
  618.  
  619. mov     (1)     msg_reg1.20<1>:ud       obw_m0.8<0,1,0>:ud      {align1};
  620. add     (1)     obw_m0.8<1>:UD       obw_m0.8<0,1,0>:UD      INTER_VME_OUTPUT_MV_IN_OWS:UD {align1};
  621. mov     (8)     msg_reg0.0<1>:ud        obw_m0.0<8,8,1>:ud      {align1};
  622.  
  623.  
  624. /* bind index 3, write 1 oword, msg type: 8(OWord Block Write) */
  625. send (16)
  626.        msg_ind
  627.        obw_wb
  628.        null
  629.        data_port(
  630.                OBW_CACHE_TYPE,
  631.                OBW_MESSAGE_TYPE,
  632.                OBW_CONTROL_2,
  633.                OBW_BIND_IDX,
  634.                OBW_WRITE_COMMIT_CATEGORY,
  635.                OBW_HEADER_PRESENT
  636.        )
  637.        mlen 2
  638.        rlen obw_wb_length
  639.        {align1};
  640.  
  641. NOP;
  642.  
  643. /* Issue message fence so that the previous write message is committed */
  644. send (16)
  645.        mb_ind
  646.        mb_wb.0<1>:ud
  647.         NULL
  648.        data_port(
  649.                OBR_CACHE_TYPE,
  650.                OBR_MESSAGE_FENCE,
  651.                OBR_MF_COMMIT,
  652.                OBR_BIND_IDX,
  653.                OBR_WRITE_COMMIT_CATEGORY,
  654.                OBR_HEADER_PRESENT
  655.        )
  656.        mlen 1
  657.        rlen 1
  658.        {align1};
  659.  
  660. __EXIT:
  661. /*
  662. * kill thread
  663. */        
  664. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  665. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  666.  
  667.  
  668.         nop             ;
  669.         nop             ;
  670. /* Compare three word data to get the min value */
  671. word_imin:
  672.         cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  673.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  674.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  675.         cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  676.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  677.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  678.         RETURN          {align1};      
  679.        
  680. /* Compare three word data to get the max value */
  681. word_imax:
  682.         cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  683.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  684.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  685.         cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  686.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  687.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  688.         RETURN          {align1};      
  689.        
  690. word_imedian:
  691.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  692.         (f0.0)  jmpi (1) cmp_a_ge_b;
  693.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  694.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  695.         (f0.0) jmpi (1) cmp_end;
  696.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  697.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  698.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  699.         jmpi (1) cmp_end;
  700. cmp_a_ge_b:
  701.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  702.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  703.         (f0.0) jmpi (1) cmp_end;
  704.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  705.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  706.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  707. cmp_end:
  708.         RETURN  {align1};
  709.  
  710. /* This is to get the prediction mode for the correspongding partions/sub-mb
  711. * Parameter description:
  712.         ARG0.0: INTER_SHAPE/PRED_MODE
  713.         ARG0.4: INTER_MODE(16x16, 16x8, 8x16 or 8X8)
  714.         ARG0.8: sub-mb block number(block 0/1/2/3)
  715. */
  716.  
  717. mb_pred_func:
  718.         mov     (8)     TEMP_VAR0.0<1>:ud       0:ud    {align1};
  719.         mov     (1)     TEMP_VAR0.0<1>:ub       INPUT_ARG0.1<0,1,0>:ub  {align1};
  720.         and     (1)   TEMP_VAR0.4<1>:uw   INPUT_ARG0.4<0,1,0>:uw        INTER_MASK:uw   {align1};
  721.         /* INTER16x16 mode. The bit1-0 is the prediction mode */
  722.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<1>:uw       INTER_16X16MODE:uw      {align1};
  723.         (f0.0)  and (1) RET_ARG<1>:uw   TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  724.         (f0.0)  jmpi (1) end_mb_pred;
  725.         /* Check whether it is INTER8x8 mode. */
  726.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<1>:uw       INTER_8X8MODE:uw        {align1};
  727.         (f0.0)  jmpi (1) mb_pred_func_8;
  728.        
  729.         /* Check whether it is INTER16x8 mode. */
  730.         cmp.e.f0.0 (1) null:uw  TEMP_VAR0.4<1>:uw       INTER_16X8MODE:uw       {align1};
  731.         (f0.0)  jmpi (1) mb_pred_func_168;
  732. mb_pred_func_816:      
  733.         /* Block 0/2 uses the bit1-0. Block 1/3 uses the bit3-2 */
  734.         mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
  735.         and.z.f0.0 (1)  null:uw TEMP_VAR0.8<1>:uw       INTER_BLOCK1:uw {align1};
  736.         (f0.0)  and     (1) RET_ARG<1>:uw       TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  737.         (f0.0)  jmpi (1) end_mb_pred;
  738.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
  739.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<1>:uw      PRED_MASK  {align1};   
  740.         jmpi    (1) end_mb_pred;
  741.  
  742. mb_pred_func_168:
  743.         /* Block 0/1 uses the bit1-0. Block 2/3 uses the bit3-2 */
  744.         mov     (1) TEMP_VAR0.8<1>:uw INPUT_ARG0.8<0,1,0>:uw    {align1};
  745.         cmp.l.f0.0 (1)  null:uw TEMP_VAR0.8<1>:uw       INTER_BLOCK2:uw {align1};
  746.         (f0.0)  and  (1) RET_ARG<1>:uw  TEMP_VAR0.0<0,1,0>:uw   PRED_MASK {align1};
  747.         (f0.0)  jmpi (1) end_mb_pred;
  748.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   2:uw {align1};
  749.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<1>:uw      PRED_MASK  {align1};   
  750.         jmpi    (1) end_mb_pred;
  751.  
  752. mb_pred_func_8:
  753.         /* 8X8 mode. Every block uses two bits as the prediction mode. */
  754.         mul     (1)  TEMP_VAR0.8<1>:uw  INPUT_ARG0.8<0,1,0>:uw  2:uw {align1};
  755.         shr     (1)  TEMP_VAR0.16<1>:uw TEMP_VAR0.0<0,1,0>:uw   TEMP_VAR0.8<0,1,0>:uw {align1};
  756.         and     (1)  RET_ARG<1>:uw      TEMP_VAR0.16<1>:uw      PRED_MASK  {align1};   
  757. end_mb_pred:
  758.         RETURN  {align1};
  759.        
  760.