Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /* GRF allocation:
  2.    g1~g30: constant buffer
  3.            g1~g2:intra IQ matrix
  4.            g3~g4:non intra IQ matrix
  5.            g5~g20:IDCT table
  6.    g31:    thread payload
  7.    g58~g81:reference data
  8.    g82:    thread payload backup
  9.    g83~g106:IDCT data
  10.    g115:   message descriptor for reading reference data   */
  11. mov (1) g32.8<1>UD 0x7000FUD {align1}; //8*16
  12. send (16) 0 g40.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
  13. send (16) 0 g46.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
  14. add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
  15. mov (1) g32.8<1>UD 0x0FUD {align1};
  16. send (16) 0 g44.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 1 {align1};//U
  17. send (16) 0 g50.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 1 {align1};//V
  18. //U
  19. add (8) g74.0<1>UW  g40.0<8,8,1>UB g40.1<8,8,1>UB {align1};
  20. add (8) g74.16<1>UW g41.0<8,8,1>UB g41.1<8,8,1>UB {align1};
  21. add (8) g75.0<1>UW  g42.0<8,8,1>UB g42.1<8,8,1>UB {align1};
  22. add (8) g75.16<1>UW g43.0<8,8,1>UB g43.1<8,8,1>UB {align1};
  23.  
  24. add (8) g74.0<1>UW  g74.0<8,8,1>UW  g41.0<8,8,1>UB {align1};
  25. add (8) g74.16<1>UW g74.16<8,8,1>UW g42.0<8,8,1>UB {align1};
  26. add (8) g75.0<1>UW  g75.0<8,8,1>UW  g43.0<8,8,1>UB {align1};
  27. add (8) g75.16<1>UW g75.16<8,8,1>UW g44.0<8,8,1>UB {align1};
  28.  
  29. add (8) g74.0<1>UW  g74.0<8,8,1>UW  g41.1<8,8,1>UB {align1};
  30. add (8) g74.16<1>UW g74.16<8,8,1>UW g42.1<8,8,1>UB {align1};
  31. add (8) g75.0<1>UW  g75.0<8,8,1>UW  g43.1<8,8,1>UB {align1};
  32. add (8) g75.16<1>UW g75.16<8,8,1>UW g44.1<8,8,1>UB {align1};
  33. //V
  34. add (8) g78.0<1>UW  g46.0<8,8,1>UB g46.1<8,8,1>UB {align1};
  35. add (8) g78.16<1>UW g47.0<8,8,1>UB g47.1<8,8,1>UB {align1};
  36. add (8) g79.0<1>UW  g48.0<8,8,1>UB g48.1<8,8,1>UB {align1};
  37. add (8) g79.16<1>UW g49.0<8,8,1>UB g49.1<8,8,1>UB {align1};
  38.  
  39. add (8) g78.0<1>UW  g78.0<8,8,1>UW  g47.0<8,8,1>UB {align1};
  40. add (8) g78.16<1>UW g78.16<8,8,1>UW g48.0<8,8,1>UB {align1};
  41. add (8) g79.0<1>UW  g79.0<8,8,1>UW  g49.0<8,8,1>UB {align1};
  42. add (8) g79.16<1>UW g79.16<8,8,1>UW g50.0<8,8,1>UB {align1};
  43.  
  44. add (8) g78.0<1>UW  g78.0<8,8,1>UW  g47.1<8,8,1>UB {align1};
  45. add (8) g78.16<1>UW g78.16<8,8,1>UW g48.1<8,8,1>UB {align1};
  46. add (8) g79.0<1>UW  g79.0<8,8,1>UW  g49.1<8,8,1>UB {align1};
  47. add (8) g79.16<1>UW g79.16<8,8,1>UW g50.1<8,8,1>UB {align1};
  48.  
  49. mov (1) g32.8<1>UD 0x7000FUD {align1}; //8*16
  50. send (16) 0 g40.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 4 {align1};//U
  51. send (16) 0 g46.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 4 {align1};//V
  52. add (1) g32.4<1>UD g32.4<1,1,1>UD 8UD {align1};
  53. mov (1) g32.8<1>UD 0x0FUD {align1};
  54. send (16) 0 g44.0<1>UW g32<8,8,1>UW read(surface_u, 2, 0, 2) mlen 1 rlen 1 {align1};//U
  55. send (16) 0 g50.0<1>UW g32<8,8,1>UW read(surface_v, 2, 0, 2) mlen 1 rlen 1 {align1};//V
  56. //U
  57. add (8) g76.0<1>UW  g40.0<8,8,1>UB g40.1<8,8,1>UB {align1};
  58. add (8) g76.16<1>UW g41.0<8,8,1>UB g41.1<8,8,1>UB {align1};
  59. add (8) g77.0<1>UW  g42.0<8,8,1>UB g42.1<8,8,1>UB {align1};
  60. add (8) g77.16<1>UW g43.0<8,8,1>UB g43.1<8,8,1>UB {align1};
  61.  
  62. add (8) g76.0<1>UW  g76.0<8,8,1>UW  g41.0<8,8,1>UB {align1};
  63. add (8) g76.16<1>UW g76.16<8,8,1>UW g42.0<8,8,1>UB {align1};
  64. add (8) g77.0<1>UW  g77.0<8,8,1>UW  g43.0<8,8,1>UB {align1};
  65. add (8) g77.16<1>UW g77.16<8,8,1>UW g44.0<8,8,1>UB {align1};
  66.  
  67. add (8) g76.0<1>UW  g76.0<8,8,1>UW  g41.1<8,8,1>UB {align1};
  68. add (8) g76.16<1>UW g76.16<8,8,1>UW g42.1<8,8,1>UB {align1};
  69. add (8) g77.0<1>UW  g77.0<8,8,1>UW  g43.1<8,8,1>UB {align1};
  70. add (8) g77.16<1>UW g77.16<8,8,1>UW g44.1<8,8,1>UB {align1};
  71. //V
  72. add (8) g80.0<1>UW  g46.0<8,8,1>UB g46.1<8,8,1>UB {align1};
  73. add (8) g80.16<1>UW g47.0<8,8,1>UB g47.1<8,8,1>UB {align1};
  74. add (8) g81.0<1>UW  g48.0<8,8,1>UB g48.1<8,8,1>UB {align1};
  75. add (8) g81.16<1>UW g49.0<8,8,1>UB g49.1<8,8,1>UB {align1};
  76.  
  77. add (8) g80.0<1>UW  g80.0<8,8,1>UW  g47.0<8,8,1>UB {align1};
  78. add (8) g80.16<1>UW g80.16<8,8,1>UW g48.0<8,8,1>UB {align1};
  79. add (8) g81.0<1>UW  g81.0<8,8,1>UW  g49.0<8,8,1>UB {align1};
  80. add (8) g81.16<1>UW g81.16<8,8,1>UW g50.0<8,8,1>UB {align1};
  81.  
  82. add (8) g80.0<1>UW  g80.0<8,8,1>UW  g47.1<8,8,1>UB {align1};
  83. add (8) g80.16<1>UW g80.16<8,8,1>UW g48.1<8,8,1>UB {align1};
  84. add (8) g81.0<1>UW  g81.0<8,8,1>UW  g49.1<8,8,1>UB {align1};
  85. add (8) g81.16<1>UW g81.16<8,8,1>UW g50.1<8,8,1>UB {align1};
  86.  
  87. shr (32) g74.0<1>UW g74.0<16,16,1>UW 2UW {align1 compr};
  88. shr (32) g76.0<1>UW g76.0<16,16,1>UW 2UW {align1 compr};
  89. shr (32) g78.0<1>UW g78.0<16,16,1>UW 2UW {align1 compr};
  90. shr (32) g80.0<1>UW g80.0<16,16,1>UW 2UW {align1 compr};
  91.  
  92.