Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © 2014 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.  *
  24.  * Authors:
  25.  *    Zhao Yakui <yakui.zhao@intel.com>
  26.  *    Xiang Haihao <haihao.xiang@intel.com>
  27.  *    Li Zhong <zhong.li@intel.com>
  28.  *
  29.  */
  30.  
  31. #define SAVE_RET        add (1) RETURN_REG<1>:ud   ip:ud        32:ud
  32. #define RETURN          mov (1) ip:ud   RETURN_REG<0,1,0>:ud
  33.  
  34. /*
  35.  * __START
  36.  */
  37. __INTER_START:
  38. mov  (16) tmp_reg0.0<1>:UD      0x0:UD {align1};
  39. mov  (16) tmp_reg2.0<1>:UD      0x0:UD {align1};
  40. mov  (16) tmp_reg4.0<1>:UD      0x0:UD {align1} ;
  41. mov  (16) tmp_reg6.0<1>:UD      0x0:UD {align1} ;
  42.  
  43. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  44. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  45. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  46. mov  (1) read0_header.8<1>:UD   BLOCK_32X1 {align1};
  47. mov  (1) read0_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  48.  
  49. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  50. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  51. mov  (1) read1_header.8<1>:UD   BLOCK_4X16 {align1};
  52. mov  (1) read1_header.20<1>:UB  thread_id_ub {align1};                  /* dispatch id */
  53.        
  54. shl  (2) vme_m0.8<1>:UW         orig_xy_ub<2,2,1>:UB 4:UW {align1};    /* (x, y) * 16 */
  55. mov  (1) vme_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  56.  
  57. mul  (1) obw_m0.8<1>:UD         w_in_mb_uw<0,1,0>:UW orig_y_ub<0,1,0>:UB {align1};
  58. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD orig_x_ub<0,1,0>:UB {align1};
  59. mul  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 24:UD {align1};
  60. mov  (1) obw_m0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  61.        
  62. /*
  63.  * Media Read Message -- fetch Luma neighbor edge pixels
  64.  */
  65. /* ROW */
  66. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  67. send (8) msg_ind INEP_ROW<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 1 {align1};
  68.  
  69. /* COL */
  70. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  71. send (8) msg_ind INEP_COL0<1>:UB null read(BIND_IDX_INEP, 0, 0, 4) mlen 1 rlen 2 {align1};
  72.        
  73. /*
  74.  * Media Read Message -- fetch Chroma neighbor edge pixels
  75.  */
  76. /* ROW */
  77. shl  (2) read0_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16 , y * 8 */
  78. mul  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D  2:W {align1};
  79. add  (1) read0_header.0<1>:D    read0_header.0<0,1,0>:D -8:W {align1};     /* X offset */
  80. add  (1) read0_header.4<1>:D    read0_header.4<0,1,0>:D -1:W {align1};     /* Y offset */
  81. mov  (8) msg_reg0.0<1>:UD       read0_header.0<8,8,1>:UD {align1};        
  82. send (8) msg_ind CHROMA_ROW<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  83.  
  84. /* COL */
  85. shl  (2) read1_header.0<1>:D    orig_xy_ub<2,2,1>:UB 3:UW {align1};    /* x * 16, y * 8 */
  86. mul  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D  2:W {align1};
  87. add  (1) read1_header.0<1>:D    read1_header.0<0,1,0>:D -4:W {align1};     /* X offset */
  88. mov  (1) read1_header.8<1>:UD   BLOCK_8X4 {align1};
  89. mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};                
  90. send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
  91.  
  92. mov  (8) mb_mvp_ref.0<1>:ud     0:ud            {align1};
  93. mov  (8) mb_ref_win.0<1>:ud     0:ud            {align1};
  94. and.z.f0.0 (1)          null:uw mb_hwdep<0,1,0>:uw              0x04:uw   {align1};
  95. (f0.0) jmpi (1) __mb_hwdep_end;
  96. /* read back the data for MB A */
  97. /* the layout of MB result is: rx.0(Available). rx.4(MVa), rX.8(MVb), rX.16(Pred_L0 flag),
  98. *  rX.18 (Pred_L1 flag), rX.20(Forward reference ID), rX.22(Backwared reference ID)
  99. */
  100. mov  (8) mba_result.0<1>:ud     0x0:ud          {align1};
  101. mov  (8) mbb_result.0<1>:ud     0x0:ud          {align1};
  102. mov  (8) mbc_result.0<1>:ud     0x0:ud          {align1};
  103. mba_start:
  104. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  105. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  106. /* MB A doesn't exist. Zero MV. mba_flag is zero and ref ID = -1 */
  107. (f0.0)  mov  (2)        mba_result.20<1>:w      -1:w    {align1};
  108. (f0.0)  jmpi (1)        mbb_start;
  109. mov  (1) mba_result.0<1>:d      MB_AVAIL                {align1};      
  110. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  111. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     -1:w    {align1};
  112. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  113. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  114. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  115. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  116.  
  117. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  118. send (16)
  119.        mb_ind
  120.        mb_wb.0<1>:ud
  121.         NULL
  122.        data_port(
  123.                OBR_CACHE_TYPE,
  124.                OBR_MESSAGE_TYPE,
  125.                OBR_CONTROL_4,
  126.                OBR_BIND_IDX,
  127.                OBR_WRITE_COMMIT_CATEGORY,
  128.                OBR_HEADER_PRESENT
  129.        )
  130.        mlen 1
  131.        rlen 2
  132.        {align1};
  133.  
  134. /* TODO: RefID is required after multi-references are added */
  135. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  136. (f0.0)   mov (2)        mba_result.20<1>:w                      -1:w    {align1};
  137. (f0.0)   jmpi   (1)     mbb_start;
  138.  
  139. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  140. /* Read MV for MB A */
  141. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  142. send (16)
  143.        mb_ind
  144.        mb_mv0.0<1>:ud
  145.         NULL
  146.        data_port(
  147.                OBR_CACHE_TYPE,
  148.                OBR_MESSAGE_TYPE,
  149.                OBR_CONTROL_8,
  150.                OBR_BIND_IDX,
  151.                OBR_WRITE_COMMIT_CATEGORY,
  152.                OBR_HEADER_PRESENT
  153.        )
  154.        mlen 1
  155.        rlen 4
  156.        {align1};
  157. /* TODO: RefID is required after multi-references are added */
  158. /* MV */
  159. mov        (2)          mba_result.4<1>:ud              mb_mv1.8<2,2,1>:ud      {align1};
  160. mov        (1)          mba_result.16<1>:w              MB_PRED_FLAG            {align1};
  161.  
  162. mbb_start:
  163. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  164. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  165. /* MB B doesn't exist. Zero MV. mba_flag is zero */
  166. /* If MB B doesn't exist, neither MB C nor D exists */
  167. (f0.0)  mov  (2)        mbb_result.20<1>:w      -1:w            {align1};
  168. (f0.0)  mov  (2)        mbc_result.20<1>:w      -1:w            {align1};
  169. (f0.0)  jmpi (1)        mb_mvp_start;
  170. mov  (1) mbb_result.0<1>:d      MB_AVAIL                {align1};      
  171. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  172. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  173. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  174. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  175. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  176. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  177.  
  178. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  179. send (16)
  180.        mb_ind
  181.        mb_wb.0<1>:ud
  182.         NULL
  183.        data_port(
  184.                OBR_CACHE_TYPE,
  185.                OBR_MESSAGE_TYPE,
  186.                OBR_CONTROL_4,
  187.                OBR_BIND_IDX,
  188.                OBR_WRITE_COMMIT_CATEGORY,
  189.                OBR_HEADER_PRESENT
  190.        )
  191.        mlen 1
  192.        rlen 2
  193.        {align1};
  194.  
  195. /* TODO: RefID is required after multi-references are added */
  196. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  197. (f0.0)   mov (2)        mbb_result.20<1>:w                      -1:w    {align1};
  198. (f0.0)   jmpi   (1)     mbc_start;
  199. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  200. /* Read MV for MB B */
  201. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  202. send (16)
  203.        mb_ind
  204.        mb_mv0.0<1>:ud
  205.         NULL
  206.        data_port(
  207.                OBR_CACHE_TYPE,
  208.                OBR_MESSAGE_TYPE,
  209.                OBR_CONTROL_8,
  210.                OBR_BIND_IDX,
  211.                OBR_WRITE_COMMIT_CATEGORY,
  212.                OBR_HEADER_PRESENT
  213.        )
  214.        mlen 1
  215.        rlen 4
  216.        {align1};
  217. /* TODO: RefID is required after multi-references are added */
  218. mov        (2)          mbb_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  219. mov        (1)          mbb_result.16<1>:w              MB_PRED_FLAG            {align1};
  220.  
  221. mbc_start:
  222. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  223. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_C:uw   {align1};
  224. /* MB C doesn't exist. Zero MV. mba_flag is zero */
  225. /* Based on h264 spec the MB D will be replaced if MB C doesn't exist */
  226. (f0.0)  jmpi (1)        mbd_start;
  227. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  228. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  229. add  (1) tmp_reg0.2<1>:w        tmp_reg0.2<0,1,0>:w     -1:w    {align1};
  230. add  (1) tmp_reg0.0<1>:w        tmp_reg0.0<0,1,0>:w     1:w     {align1};
  231. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  232. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  233. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  234. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  235.  
  236. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  237. send (16)
  238.        mb_ind
  239.        mb_wb.0<1>:ud
  240.         NULL
  241.        data_port(
  242.                OBR_CACHE_TYPE,
  243.                OBR_MESSAGE_TYPE,
  244.                OBR_CONTROL_4,
  245.                OBR_BIND_IDX,
  246.                OBR_WRITE_COMMIT_CATEGORY,
  247.                OBR_HEADER_PRESENT
  248.        )
  249.        mlen 1
  250.        rlen 2
  251.        {align1};
  252.  
  253. /* TODO: RefID is required after multi-references are added */
  254. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  255. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  256. (f0.0)   jmpi   (1)     mb_mvp_start;
  257. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  258. /* Read MV for MB C */
  259. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  260. send (16)
  261.        mb_ind
  262.        mb_mv0.0<1>:ud
  263.         NULL
  264.        data_port(
  265.                OBR_CACHE_TYPE,
  266.                OBR_MESSAGE_TYPE,
  267.                OBR_CONTROL_8,
  268.                OBR_BIND_IDX,
  269.                OBR_WRITE_COMMIT_CATEGORY,
  270.                OBR_HEADER_PRESENT
  271.        )
  272.        mlen 1
  273.        rlen 4
  274.        {align1};
  275. /* TODO: RefID is required after multi-references are added */
  276. /* Forward MV */
  277. mov        (2)          mbc_result.4<1>:ud              mb_mv2.16<2,2,1>:ud     {align1};
  278. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  279.  
  280. jmpi   (1)    mb_mvp_start;
  281. mbd_start:
  282. mov  (8) mb_msg0.0<1>:ud        0:ud            {align1};
  283. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_D:uw   {align1};
  284. (f0.0)  mov (2) mbc_result.20<1>:w      -1:w    {align1};
  285. (f0.0)  jmpi (1)        mb_mvp_start;
  286. mov  (1) mbc_result.0<1>:d      MB_AVAIL                {align1};      
  287. mov  (2) tmp_reg0.0<1>:UW       orig_xy_ub<2,2,1>:UB    {align1};
  288. add  (2) tmp_reg0.0<1>:w        tmp_reg0.0<2,2,1>:w     -1:w    {align1};
  289. mul  (1) mb_msg0.8<1>:UD       w_in_mb_uw<0,1,0>:UW tmp_reg0.2<0,1,0>:UW {align1};
  290. add  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD   tmp_reg0.0<0,1,0>:uw {align1};
  291. mul  (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:UD 24:UD {align1};
  292. mov  (1) mb_msg0.20<1>:UB        thread_id_ub {align1};                  /* dispatch id */
  293.  
  294. /* bind index 3, read 4 oword (64bytes), msg type: 0(OWord Block Read) */
  295. send (16)
  296.        mb_ind
  297.        mb_wb.0<1>:ud
  298.         NULL
  299.        data_port(
  300.                OBR_CACHE_TYPE,
  301.                OBR_MESSAGE_TYPE,
  302.                OBR_CONTROL_4,
  303.                OBR_BIND_IDX,
  304.                OBR_WRITE_COMMIT_CATEGORY,
  305.                OBR_HEADER_PRESENT
  306.        )
  307.        mlen 1
  308.        rlen 2
  309.        {align1};
  310.  
  311. cmp.l.f0.0 (1)          null:w  mb_intra_wb.16<0,1,0>:uw        mb_inter_wb.8<0,1,0>:uw {align1};
  312. (f0.0)   mov (2)        mbc_result.20<1>:w                      -1:w    {align1};
  313. (f0.0)   jmpi   (1)     mb_mvp_start;
  314.  
  315. add   (1) mb_msg0.8<1>:UD       mb_msg0.8<0,1,0>:ud     3:ud {align1};
  316. /* Read MV for MB D */
  317. /* bind index 3, read 8 oword (128bytes), msg type: 0(OWord Block Read) */
  318. send (16)
  319.        mb_ind
  320.        mb_mv0.0<1>:ub
  321.         NULL
  322.        data_port(
  323.                OBR_CACHE_TYPE,
  324.                OBR_MESSAGE_TYPE,
  325.                OBR_CONTROL_8,
  326.                OBR_BIND_IDX,
  327.                OBR_WRITE_COMMIT_CATEGORY,
  328.                OBR_HEADER_PRESENT
  329.        )
  330.        mlen 1
  331.        rlen 4
  332.        {align1};
  333.  
  334. /* TODO: RefID is required after multi-references are added */
  335.  
  336. /* Forward MV */
  337. mov        (2)          mbc_result.4<1>:ud              mb_mv3.24<2,2,1>:ud     {align1};
  338. mov        (1)          mbc_result.16<1>:w              MB_PRED_FLAG            {align1};
  339.        
  340. mb_mvp_start:
  341. /*TODO: Add the skip prediction */
  342. /* Check whether both MB B and C are inavailable */
  343. add     (1)     tmp_reg0.0<1>:d         mbb_result.0<0,1,0>:d   mbc_result.0<0,1,0>:d   {align1};
  344. cmp.z.f0.0 (1)  null:d                  tmp_reg0.0<0,1,0>:d     0:d     {align1};
  345. (-f0.0) jmpi (1)        mb_median_start;
  346. cmp.nz.f0.0 (1) null:d  mba_result.0<0,1,0>:d           0:d             {align1};
  347. (f0.0)  mov     (1)     mbb_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  348. (f0.0)  mov     (1)     mbc_result.4<1>:ud              mba_result.4<0,1,0>:ud  {align1};      
  349. (f0.0)  mov     (1)     mbb_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  350. (f0.0)  mov     (1)     mbc_result.20<1>:uw             mba_result.20<0,1,0>:uw {align1};      
  351. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud              mba_result.4<0,1,0>:ud  {align1};
  352. (-f0.0) mov     (1)     mb_mvp_ref.0<1>:ud              0:ud                    {align1};
  353. jmpi    (1)     __mb_hwdep_end;
  354.        
  355. mb_median_start:
  356. /* check whether only one neighbour MB has the same ref ID with the current MB */
  357. mov (8) tmp_reg0.0<1>:ud                0:ud            {align1};
  358. cmp.z.f0.0      (1)     null:d  mba_result.20<0,1,0>:w  0:w     {align1};
  359. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  360. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mba_result.4<0,1,0>:ud  {align1};
  361. cmp.z.f0.0      (1)     null:d  mbb_result.20<0,1,0>:w  0:w     {align1};
  362. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  363. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbb_result.4<0,1,0>:ud  {align1};
  364. cmp.z.f0.0      (1)     null:d  mbc_result.20<0,1,0>:w  0:w     {align1};
  365. (f0.0)  add     (1)     tmp_reg0.0<1>:w         tmp_reg0.0<0,1,0>:w     1:w     {align1};
  366. (f0.0)  mov     (1)     tmp_reg0.4<1>:ud        mbc_result.4<0,1,0>:ud  {align1};
  367. cmp.e.f0.0      (1)     null:d  tmp_reg0.0<0,1,0>:w      1:w    {align1};
  368. (f0.0)  mov     (1)     mb_mvp_ref.0<1>:ud      tmp_reg0.4<0,1,0>:ud    {align1};
  369. (f0.0)  jmpi (1)  __mb_hwdep_end;
  370.  
  371. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.4<0,1,0>:w   {align1};
  372. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.4<0,1,0>:w   {align1};
  373. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.4<0,1,0>:w   {align1};
  374. SAVE_RET        {align1};
  375.  jmpi   (1)     word_imedian;
  376. mov     (1)     mb_mvp_ref.0<1>:w               RET_ARG<0,1,0>:w        {align1};
  377. mov     (1)     INPUT_ARG0.0<1>:w       mba_result.6<0,1,0>:w   {align1};
  378. mov     (1)     INPUT_ARG0.4<1>:w       mbb_result.6<0,1,0>:w   {align1};
  379. mov     (1)     INPUT_ARG0.8<1>:w       mbc_result.6<0,1,0>:w   {align1};
  380. SAVE_RET        {align1};
  381. jmpi    (1)     word_imedian;
  382. mov     (1)     mb_mvp_ref.2<1>:w               RET_ARG<0,1,0>:w        {align1};
  383.  
  384. __mb_hwdep_end:
  385. asr     (2)     mb_ref_win.0<1>:w       mb_mvp_ref.0<2,2,1>:w   2:w     {align1};
  386. add     (2)     mb_ref_win.8<1>:w       mb_ref_win.0<2,2,1>:w   3:w     {align1};
  387. and     (2)     mb_ref_win.16<1>:uw     mb_ref_win.8<2,2,1>:uw  0xFFFC:uw {align1};
  388. /* m2, get the MV/Mb cost passed from constant buffer when
  389. spawning thread by MEDIA_OBJECT */      
  390. mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
  391.  
  392. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  393.  
  394. /* m3 FWD/BWD cost center*/
  395. mov (8) vme_msg_3<1>:UD         0x0:UD {align1};               
  396.  
  397. /* m4 skip center*/
  398. mov (8) vme_msg_4<1>:UD         0x0:UD {align1};               
  399.  
  400. /* m5 */
  401. mov  (1) INEP_ROW.0<1>:UD       0x0:UD {align1};
  402. and  (1) INEP_ROW.4<1>:UD       INEP_ROW.4<0,1,0>:UD            0xFF000000:UD {align1};
  403. mov  (8) vme_msg_5<1>:UD         INEP_ROW.0<8,8,1>:UD {align1};
  404.  
  405.  
  406. /* Use the Luma mode */
  407. mov  (1) tmp_reg0.0<1>:UW       LUMA_INTRA_MODE:UW {align1};
  408. mov  (1) vme_msg_5.5<1>:UB      tmp_reg0.0<0,1,0>:UB {align1};
  409.  
  410. /* m6 */        
  411. mov  (8) vme_msg_6<1>:UD         0x0:UD {align1};
  412. mov (16) vme_msg_6.0<1>:UB       INEP_COL0.3<32,8,4>:UB {align1};
  413. mov  (1) vme_msg_6.16<1>:UD      INTRA_PREDICTORE_MODE {align1};
  414.  
  415. /* the penalty for Intra mode */
  416. mov  (1) vme_msg_6.28<1>:UD     0x010101:UD {align1};
  417. mov  (1) vme_msg_6.20<1>:UW      CHROMA_ROW.6<0,1,0>:UW {align1};
  418.  
  419.  
  420. /* m7 */
  421.  
  422. mov  (4) vme_msg_7.16<1>:UD      CHROMA_ROW.8<4,4,1>:UD {align1};
  423. mov  (8) vme_msg_7.0<1>:UW       CHROMA_COL.2<16,8,2>:UW {align1};
  424.  
  425. /*
  426. * SIC VME message
  427. */
  428.  
  429. /* m1 */
  430. mov  (1) intra_flag<1>:UW       0x0:UW {align1};
  431. mov  (1) intra_part_mask_ub<1>:UB  LUMA_INTRA_8x8_DISABLE {align1}; /* vp8 don't support intra_8x8 mode*/
  432.  
  433. /* assign MB intra struct from the thread payload*/
  434. mov (1) mb_intra_struct_ub<1>:UB input_mb_intra_ub<0,1,0>:UB {align1};
  435.  
  436. /* Disable DC HAAR component when calculating HARR SATD block */
  437. mov  (1) tmp_reg0.0<1>:UW       DC_HARR_DISABLE:UW              {align1};
  438. mov  (1) vme_m1.30<1>:UB        tmp_reg0.0<0,1,0>:UB  {align1};
  439. mov  (8) vme_msg_1<1>:UD        vme_m1.0<8,8,1>:UD {align1};
  440.  
  441. /* m0 */        
  442. mov  (1) vme_m0.12<1>:UD        INTRA_SAD_HAAR:UD {align1};    /* 16x16 Source, Intra_harr */
  443. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  444.  
  445. /* after verification it will be passed by using payload */
  446. send (8)
  447.         vme_msg_ind
  448.         vme_wb<1>:UD
  449.         null
  450.         cre(
  451.                 BIND_IDX_VME,
  452.                 VME_SIC_MESSAGE_TYPE
  453.         )
  454.         mlen sic_vme_msg_length
  455.         rlen vme_wb_length
  456.         {align1};
  457. /*
  458.  * Oword Block Write message
  459.  */
  460. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  461.        
  462. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  463. mov  (1) msg_reg1.4<1>:UD       vme_wb.16<0,1,0>:UD     {align1};
  464. mov  (1) msg_reg1.8<1>:UD       vme_wb.20<0,1,0>:UD     {align1};
  465. mov  (1) msg_reg1.12<1>:UD      vme_wb.24<0,1,0>:UD     {align1};
  466.  
  467. /* Distortion, Intra (17-16), */
  468. mov  (1) msg_reg1.16<1>:UW      vme_wb.12<0,1,0>:UW     {align1};
  469.  
  470. mov  (1) msg_reg1.20<1>:UD      vme_wb.8<0,1,0>:UD     {align1};
  471. /* VME clock counts */
  472. mov  (1) msg_reg1.24<1>:UD      vme_wb.28<0,1,0>:UD     {align1};
  473.  
  474. mov  (1) msg_reg1.28<1>:UD      obw_m0.8<0,1,0>:UD     {align1};
  475.  
  476. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  477. send (16)
  478.         msg_ind
  479.         obw_wb
  480.         null
  481.         data_port(
  482.                 OBW_CACHE_TYPE,
  483.                 OBW_MESSAGE_TYPE,
  484.                 OBW_CONTROL_2,
  485.                 OBW_BIND_IDX,
  486.                 OBW_WRITE_COMMIT_CATEGORY,
  487.                 OBW_HEADER_PRESENT
  488.         )
  489.         mlen 2
  490.         rlen obw_wb_length
  491.         {align1};
  492.  
  493. /* IME search */
  494. mov  (1) vme_m0.12<1>:UD        SEARCH_CTRL_SINGLE + VP8_INTER_PART_MASK + INTER_SAD_HAAR:UD {align1};    /* 16x16 Source, harr */
  495. mov  (1) vme_m0.22<1>:UW        REF_REGION_SIZE {align1};         /* Reference Width&Height, 48x40 */
  496.  
  497. mov  (1) vme_m0.0<1>:UD         vme_m0.8<0,1,0>:UD      {align1};
  498.  
  499. add  (1) vme_m0.0<1>:W          vme_m0.0<0,1,0>:W -16:W {align1};               /* Reference = (x-16,y-12)-(x+32,y+28) */
  500. add  (1) vme_m0.2<1>:W          vme_m0.2<0,1,0>:W -12:W {align1};
  501.  
  502. mov  (1) vme_m0.0<1>:W          -16:W                   {align1};
  503. mov  (1) vme_m0.2<1>:W          -12:W                   {align1};
  504.  
  505. mov  (1) vme_m0.4<1>:UD         vme_m0.0<0,1,0>:UD      {align1};
  506.  
  507. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
  508. (f0.0)  add     (1)     vme_m0.0<1>:w   vme_m0.0<0,1,0>:w       12:w    {align1};
  509. and.z.f0.0 (1)          null:uw input_mb_intra_ub<0,1,0>:ub     INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
  510. (f0.0)  add     (1)     vme_m0.2<1>:w   vme_m0.2<0,1,0>:w       8:w     {align1};
  511.  
  512. add  (2) vme_m0.0<1>:w          vme_m0.0<2,2,1>:w       mb_ref_win.16<2,2,1>:{align1};
  513. add  (2) vme_m0.4<1>:w          vme_m0.4<2,2,1>:w       mb_ref_win.16<2,2,1>:{align1};
  514. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
  515.  
  516. mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
  517. /* the Max MV number is passed by constant buffer */
  518. mov  (1) vme_m1.4<1>:UB         r4.28<0,1,0>:UB {align1};          
  519. mov  (1) vme_m1.8<1>:UD         START_CENTER + SEARCH_PATH_LEN:UD {align1};
  520. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD {align1};
  521.  
  522. /* Setup the Cost center */
  523. /* currently four 8x8 share the same cost center */
  524. mov  (4) vme_m3.0<2>:ud         mb_mvp_ref.0<0,1,0>:ud  {align1};
  525. mov  (4) vme_m3.4<2>:ud         mb_mvp_ref.0<0,1,0>:ud  {align1};
  526.  
  527. mov (8) vme_msg_3<1>:UD         vme_m3.0<8,8,1>:UD {align1};
  528. mov (8) vme_msg_2<1>:UD         vme_m2.0<8,8,1>:UD {align1};
  529.  
  530. /* M4/M5 search path */
  531. mov  (1) vme_msg_4.0<1>:UD      0x01010101:UD {align1};
  532. mov  (1) vme_msg_4.4<1>:UD      0x10010101:UD {align1};
  533. mov  (1) vme_msg_4.8<1>:UD      0x0F0F0F0F:UD {align1};
  534. mov  (1) vme_msg_4.12<1>:UD     0x100F0F0F:UD {align1};
  535. mov  (1) vme_msg_4.16<1>:UD     0x01010101:UD {align1};
  536. mov  (1) vme_msg_4.20<1>:UD     0x10010101:UD {align1};
  537. mov  (1) vme_msg_4.24<1>:UD     0x0F0F0F0F:UD {align1};
  538. mov  (1) vme_msg_4.28<1>:UD     0x100F0F0F:UD {align1};
  539.  
  540. mov  (1) vme_msg_5.0<1>:UD      0x01010101:UD {align1};
  541. mov  (1) vme_msg_5.4<1>:UD      0x10010101:UD {align1};
  542. mov  (1) vme_msg_5.8<1>:UD      0x0F0F0F0F:UD {align1};
  543. mov  (1) vme_msg_5.12<1>:UD     0x000F0F0F:UD {align1};
  544.  
  545. mov  (4) vme_msg_5.16<1>:UD     0x0:UD {align1};
  546.  
  547. send (8)
  548.         vme_msg_ind
  549.         vme_wb<1>:UD
  550.         null
  551.         vme(
  552.                 BIND_IDX_VME,
  553.                 0,
  554.                 0,
  555.                 VME_IME_MESSAGE_TYPE
  556.         )
  557.         mlen ime_vme_msg_length
  558.         rlen vme_wb_length {align1};
  559.  
  560. /* Set Macroblock-shape/mode for FBR */
  561.  
  562. mov  (1) vme_m2.20<1>:UD        0x0:UD {align1};
  563. mov  (1) vme_m2.21<1>:UB        vme_wb.25<0,1,0>:UB     {align1};
  564. mov  (1) vme_m2.22<1>:UB        vme_wb.26<0,1,0>:UB     {align1};
  565.  
  566. and  (1) tmp_reg0.0<1>:UW       vme_wb.0<0,1,0>:UW      0x03:UW {align1};
  567. mov  (1) vme_m2.20<1>:UB        tmp_reg0.0<0,1,0>:UB    {align1};
  568.  
  569. /* Send FBR message into CRE */
  570.  
  571. mov  (8) vme_msg_4.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  572. mov  (8) vme_msg_5.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  573. mov  (8) vme_msg_6.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  574. mov  (8) vme_msg_7.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  575.  
  576. 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 */
  577. mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD  {align1};
  578. mov  (8) vme_msg_1.0<1>:UD      vme_m1.0<8,8,1>:UD  {align1};
  579.  
  580. mov  (8) vme_msg_2.0<1>:UD              vme_m2.0<8,8,1>:UD      {align1};
  581. mov  (8) vme_msg_3.0<1>:UD              vme_m3.0<8,8,1>:UD      {align1};
  582.  
  583. /* after verification it will be passed by using payload */
  584. send (8)
  585.         vme_msg_ind
  586.         vme_wb<1>:UD
  587.         null
  588.         cre(
  589.                 BIND_IDX_VME,
  590.                 VME_FBR_MESSAGE_TYPE
  591.         )
  592.         mlen fbr_vme_msg_length
  593.         rlen vme_wb_length
  594.         {align1};
  595.  
  596. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x02:UD {align1};
  597. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  598. /* write FME info */
  599. mov  (1) msg_reg1.0<1>:UD       vme_wb.0<0,1,0>:UD      {align1};
  600.  
  601. mov  (1) msg_reg1.4<1>:UD       vme_wb.24<0,1,0>:UD     {align1};
  602. /* Inter distortion of FME */
  603. mov  (1) msg_reg1.8<1>:UD       vme_wb.8<0,1,0>:UD     {align1};
  604.  
  605. mov  (1) msg_reg1.12<1>:UD      vme_m2.20<0,1,0>:UD {align1};
  606.  
  607. /* bind index 3, write  oword (16bytes), msg type: 8(OWord Block Write) */
  608. send (16)
  609.         msg_ind
  610.         obw_wb
  611.         null
  612.         data_port(
  613.                 OBW_CACHE_TYPE,
  614.                 OBW_MESSAGE_TYPE,
  615.                 OBW_CONTROL_0,
  616.                 OBW_BIND_IDX,
  617.                 OBW_WRITE_COMMIT_CATEGORY,
  618.                 OBW_HEADER_PRESENT
  619.         )
  620.         mlen 2
  621.         rlen obw_wb_length
  622.         {align1};
  623.  
  624. /* Write FME/BME MV */
  625. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x01:UD {align1};
  626. mov  (8) msg_reg0.0<1>:UD       obw_m0.0<8,8,1>:UD {align1};
  627.  
  628.  
  629. mov  (8) msg_reg1.0<1>:UD       vme_wb1.0<8,8,1>:UD {align1};
  630. mov  (8) msg_reg2.0<1>:ud       vme_wb2.0<8,8,1>:ud {align1};
  631. mov  (8) msg_reg3.0<1>:ud       vme_wb3.0<8,8,1>:ud {align1};
  632. mov  (8) msg_reg4.0<1>:ud       vme_wb4.0<8,8,1>:ud {align1};                
  633. /* bind index 3, write  8 oword (128 bytes), msg type: 8(OWord Block Write) */
  634. send (16)
  635.         msg_ind
  636.         obw_wb
  637.         null
  638.         data_port(
  639.                 OBW_CACHE_TYPE,
  640.                 OBW_MESSAGE_TYPE,
  641.                 OBW_CONTROL_8,
  642.                 OBW_BIND_IDX,
  643.                 OBW_WRITE_COMMIT_CATEGORY,
  644.                 OBW_HEADER_PRESENT
  645.         )
  646.         mlen 5
  647.         rlen obw_wb_length
  648.         {align1};
  649.  
  650. /* Write FME/BME RefID */
  651. add  (1) obw_m0.8<1>:UD         obw_m0.8<0,1,0>:UD 0x08:UD {align1};
  652. mov  (8) msg_reg0.0<1>:UD       obw_m0<8,8,1>:UD {align1};
  653.  
  654. mov  (8) msg_reg1.0<1>:UD       vme_wb6.0<8,8,1>:UD {align1};
  655.  
  656. /* bind index 3, write 2 oword (32bytes), msg type: 8(OWord Block Write) */
  657. send (16)
  658.         msg_ind
  659.         obw_wb
  660.         null
  661.         data_port(
  662.                 OBW_CACHE_TYPE,
  663.                 OBW_MESSAGE_TYPE,
  664.                 OBW_CONTROL_2,
  665.                 OBW_BIND_IDX,
  666.                 OBW_WRITE_COMMIT_CATEGORY,
  667.                 OBW_HEADER_PRESENT
  668.         )
  669.         mlen 2
  670.         rlen obw_wb_length
  671.         {align1};
  672.  
  673. /* Issue message fence so that the previous write message is committed */
  674. send (16)
  675.         mb_ind
  676.         mb_wb.0<1>:ud
  677.         NULL
  678.         data_port(
  679.                 OBR_CACHE_TYPE,
  680.                 OBR_MESSAGE_FENCE,
  681.                 OBR_MF_COMMIT,
  682.                 OBR_BIND_IDX,
  683.                 OBR_WRITE_COMMIT_CATEGORY,
  684.                 OBR_HEADER_PRESENT
  685.         )
  686.         mlen 1
  687.         rlen 1
  688.         {align1};
  689.  
  690. __EXIT:
  691. /*
  692.  * kill thread
  693.  */        
  694. mov  (8) ts_msg_reg0<1>:UD         r0<8,8,1>:UD {align1};
  695. send (16) ts_msg_ind acc0<1>UW null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT};
  696.  
  697.  
  698.         nop             ;
  699.         nop             ;
  700. /* Compare three word data to get the min value */
  701. word_imin:
  702.         cmp.le.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  703.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  704.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  705.         cmp.le.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  706.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  707.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  708.         RETURN          {align1};      
  709.        
  710. /* Compare three word data to get the max value */
  711. word_imax:
  712.         cmp.ge.f0.0 (1)         null:w          INPUT_ARG0.0<0,1,0>:w   INPUT_ARG0.4<0,1,0>:w {align1};
  713.         (f0.0) mov  (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.0<0,1,0>:w                    {align1};
  714.         (-f0.0) mov (1)         TEMP_VAR0.0<1>:w INPUT_ARG0.4<0,1,0>:w                    {align1};
  715.         cmp.ge.f0.0 (1)         null:w          TEMP_VAR0.0<0,1,0>:w    INPUT_ARG0.8<0,1,0>:w {align1};
  716.         (f0.0) mov  (1)         RET_ARG<1>:w TEMP_VAR0.0<0,1,0>:w                         {align1};
  717.         (-f0.0) mov (1)         RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w                        {align1};
  718.         RETURN          {align1};      
  719.        
  720. word_imedian:
  721.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.4<0,1,0>:w {align1};
  722.         (f0.0)  jmpi (1) cmp_a_ge_b;
  723.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  724.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  725.         (f0.0) jmpi (1) cmp_end;
  726.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  727.         (f0.0) mov (1) RET_ARG<1>: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.         jmpi (1) cmp_end;
  730. cmp_a_ge_b:
  731.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.4<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  732.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.4<0,1,0>:w {align1};
  733.         (f0.0) jmpi (1) cmp_end;
  734.         cmp.ge.f0.0 (1) null:w INPUT_ARG0.0<0,1,0>:w INPUT_ARG0.8<0,1,0>:w {align1};
  735.         (f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.8<0,1,0>:w {align1};
  736.         (-f0.0) mov (1) RET_ARG<1>:w INPUT_ARG0.0<0,1,0>:w {align1};
  737. cmp_end:
  738.         RETURN  {align1};
  739.  
  740.