src/shaders/vme/batchbuffer.asm

   1 /*
   2  * Copyright © 2012 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 "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Xiang Haihao <haihao.xiang@intel.com>
  25  */
  26
  27 /*
  28  * __START
  29  */
  30 __START:
  31         mov             (16)    tmp_reg0<1>:ud                  0x0:ud {align1} ;
  32         mov             (16)    tmp_reg2<1>:ud                  0x0:ud {align1} ;
  33         mov             (1)     obw_header.20<1>:ub             thread_id_ub {align1};                  /* dispatch id */
  34
  35         mov             (8)     media_object_ud<1>:ud           0x0:ud {align1} ;
  36         mov             (1)     media_object0_ud<1>:ud          CMD_MEDIA_OBJECT {align1} ;
  37         mov             (1)     media_object1_ud<1>:ud          mtype_ub<0,1,0>ub {align1};
  38         mov             (1)     media_object6_width<1>:uw       width_in_mb<0,1,0>:uw {align1};
  39         mov             (1)     media_object7_flag<1>:uw        transform_8x8_ub<0,1,0>ub {align1};
  40         mov             (1)     media_object7_num_mbs<1>:uw     NUM_MACROBLOCKS_PER_COMMAND:uw {align1} ;
  41
  42         mov             (1)     width_per_row<1>:ud             width_in_mb<0,1,0>:uw {align1} ;
  43         and.z.f0.1      (1)     remainder_cmds<1>:ud            total_mbs<0,1,0>:ud     (NUM_MACROBLOCKS_PER_COMMAND - 1):ud {align1} ;
  44         and.z.f0.0      (1)     total_mbs<1>:ud                 total_mbs<0,1,0>:ud     -NUM_MACROBLOCKS_PER_COMMAND:ud {align1} ;
  45
  46         (f0.0)jmpi      (1)     __REMAINDER ;
  47
  48 __CMD_LOOP:
  49         mov             (8)     msg_reg0.0<1>:ud                obw_header<8,8,1>:ud {align1};
  50         mov             (8)     msg_reg1<1>:ud                  media_object_ud<8,8,1>:ud {align1};
  51
  52 /* bind index 5, write 2 oword, msg type: 8(OWord Block Write) */
  53 send (16)
  54         msg_ind
  55         obw_wb
  56         null
  57         data_port(
  58                 OBW_CACHE_TYPE,
  59                 OBW_MESSAGE_TYPE,
  60                 OBW_CONTROL_2,
  61                 OBW_BIND_IDX,
  62                 OBW_WRITE_COMMIT_CATEGORY,
  63                 OBW_HEADER_PRESENT
  64         )
  65         mlen 2
  66         rlen obw_wb_length
  67         {align1};
  68
  69         /* (x, y) of the first macroblock */
  70         add             (1)     count<1>:ud                     count<0,1,0>:ud         NUM_MACROBLOCKS_PER_COMMAND:uw {align1} ;
  71         math            (1)     quotient<1>:ud                  count<0,1,0>:ud         width_per_row<0,1,0>:ud intdivmod {align1} ;
  72         shl             (1)     quotient<1>:ud                  quotient<0,1,0>:ud      8:uw {align1} ;
  73         add             (1)     quotient<1>:ud                  quotient<0,1,0>:ud      remainder<0,1,0>:ud {align1} ;
  74         mov             (1)     media_object6_xy<1>:uw          quotient<0,1,0>:uw {align1} ;
  75
  76         /* the new offset */
  77         add             (1)     obw_header.8<1>:ud              obw_header.8<0,1,0>:ud  2:uw {align1} ;
  78
  79         add.z.f0.0      (1)     total_mbs<1>:w                  total_mbs<0,1,0>:w      -NUM_MACROBLOCKS_PER_COMMAND:w {align1} ;
  80         (-f0.0)jmpi     (1)     __CMD_LOOP ;
  81
  82 __REMAINDER:
  83         (f0.1)jmpi      (1)     __DONE ;
  84
  85         mov             (1)     media_object7_num_mbs<1>:uw     remainder_cmds<0,1,0>:uw {align1} ;
  86         mov             (8)     msg_reg0.0<1>:ud                obw_header<8,8,1>:ud {align1};
  87         mov             (8)     msg_reg1<1>:ud                  media_object_ud<8,8,1>:ud {align1};
  88
  89 send (16)
  90         msg_ind
  91         obw_wb
  92         null
  93         data_port(
  94                 OBW_CACHE_TYPE,
  95                 OBW_MESSAGE_TYPE,
  96                 OBW_CONTROL_2,
  97                 OBW_BIND_IDX,
  98                 OBW_WRITE_COMMIT_CATEGORY,
  99                 OBW_HEADER_PRESENT
 100         )
 101         mlen 2
 102         rlen obw_wb_length
 103         {align1};
 104
 105         /* the new offset */
 106         add             (1)     obw_header.8<1>:ud              obw_header.8<0,1,0>:ud  2:uw {align1} ;
 107
 108 __DONE:
 109
 110 /* bind index 5, write 1 oword, msg type: 8(OWord Block Write) */
 111         mov             (8)     msg_reg0.0<1>:ud                obw_header<8,8,1>:ud {align1} ;
 112         mov             (4)     msg_reg1.0<1>:ud                0x0:ud {align1} ;
 113         mov             (1)     msg_reg1.4<1>:ud                MI_BATCH_BUFFER_END {align1} ;
 114
 115 send (16)
 116         msg_ind
 117         obw_wb
 118         null
 119         data_port(
 120                 OBW_CACHE_TYPE,
 121                 OBW_MESSAGE_TYPE,
 122                 OBW_CONTROL_0,
 123                 OBW_BIND_IDX,
 124                 OBW_WRITE_COMMIT_CATEGORY,
 125                 OBW_HEADER_PRESENT
 126         )
 127         mlen 2
 128         rlen obw_wb_length
 129         {align1};
 130
 131 __EXIT:
 132         mov             (8)     msg_reg0<1>:ud                  r0<8,8,1>:ud {align1} ;
 133         send            (16)    msg_ind acc0<1>ud null thread_spawner(0, 0, 1) mlen 1 rlen 0 {align1 EOT} ;