Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright © 2009 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 FROM,
  20.  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21.  * SOFTWARE.
  22.  *
  23.  * Author:
  24.  *    Zou Nan hai <nanhai.zou@intel.com>
  25.  *    Yan Li <li.l.yan@intel.com>
  26.  *    Liu Xi bin<xibin.liu@intel.com>
  27.  */
  28. /* GRF allocation:
  29.    g1~g30: constant buffer
  30.            g1~g2:intra IQ matrix in UB format
  31.            g3~g4:non intra IQ matrix in UB format
  32.            g5~g20:IDCT table
  33.    g32~g55:DCT data before IQ
  34.    g56~g79:DCT data after IQ
  35.    g83~g106: IDCT data after idct
  36.    g82:    thread payload backup
  37.    g125:   ip before jump
  38. */
  39. include(`iq_non_intra.g4i')
  40.  
  41. define(`ROW_SHIFT', `11UD') //define for idct
  42. define(`ROW_ADD', `0x400UD')            
  43. define(`COL_SHIFT', `20UD')
  44. define(`COL_ADD', `0x80000UD')          
  45.  
  46. mov (1) a0.0<1>UD 0x06F006E0UD {align1};//0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16,the start of DCT data)
  47.  
  48. //Y0
  49. and.nz (1) null g82.8<1,1,1>UW 0x800UW {align1};
  50. (f0) jmpi do_idct_y0;
  51.  
  52. mov (8) g83.0<1>UW 0UW {align1};
  53. mov (8) g84.0<1>UW 0UW {align1};
  54. mov (8) g85.0<1>UW 0UW {align1};
  55. mov (8) g86.0<1>UW 0UW {align1};
  56. mov (8) g87.0<1>UW 0UW {align1};
  57. mov (8) g88.0<1>UW 0UW {align1};
  58. mov (8) g89.0<1>UW 0UW {align1};
  59. mov (8) g90.0<1>UW 0UW {align1};
  60. add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
  61. jmpi block_y1;
  62. do_idct_y0:
  63. mov (1) g125.0<1>UD ip {align1};
  64. jmpi IDCT_START;
  65. mov (8) g83.0<1>W g32.0<16,8,2>W {align1};
  66. mov (8) g84.0<1>W g33.0<16,8,2>W {align1};
  67. mov (8) g85.0<1>W g34.0<16,8,2>W {align1};
  68. mov (8) g86.0<1>W g35.0<16,8,2>W {align1};
  69. mov (8) g87.0<1>W g36.0<16,8,2>W {align1};
  70. mov (8) g88.0<1>W g37.0<16,8,2>W {align1};
  71. mov (8) g89.0<1>W g38.0<16,8,2>W {align1};
  72. mov (8) g90.0<1>W g39.0<16,8,2>W {align1};
  73.  
  74. //Y1
  75. block_y1:
  76. and.nz (1) null g82.8<1,1,1>UW 0x400UW {align1};
  77. (f0) jmpi do_idct_y1;
  78. mov (8) g83.16<1>UW 0UW {align1};
  79. mov (8) g84.16<1>UW 0UW {align1};
  80. mov (8) g85.16<1>UW 0UW {align1};
  81. mov (8) g86.16<1>UW 0UW {align1};
  82. mov (8) g87.16<1>UW 0UW {align1};
  83. mov (8) g88.16<1>UW 0UW {align1};
  84. mov (8) g89.16<1>UW 0UW {align1};
  85. mov (8) g90.16<1>UW 0UW {align1};
  86. add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
  87. jmpi block_y2;
  88. do_idct_y1:
  89. mov (1) g125.0<1>UD ip {align1};
  90. jmpi IDCT_START;
  91. mov (8) g83.16<1>W g32.0<16,8,2>W {align1};
  92. mov (8) g84.16<1>W g33.0<16,8,2>W {align1};
  93. mov (8) g85.16<1>W g34.0<16,8,2>W {align1};
  94. mov (8) g86.16<1>W g35.0<16,8,2>W {align1};
  95. mov (8) g87.16<1>W g36.0<16,8,2>W {align1};
  96. mov (8) g88.16<1>W g37.0<16,8,2>W {align1};
  97. mov (8) g89.16<1>W g38.0<16,8,2>W {align1};
  98. mov (8) g90.16<1>W g39.0<16,8,2>W {align1};
  99.  
  100. //Y2
  101. block_y2:
  102. and.nz (1) null g82.8<1,1,1>UW 0x200UW {align1};
  103. (f0) jmpi do_idct_y2;
  104. mov (8) g91.0<1>UW 0UW {align1};
  105. mov (8) g92.0<1>UW 0UW {align1};
  106. mov (8) g93.0<1>UW 0UW {align1};
  107. mov (8) g94.0<1>UW 0UW {align1};
  108. mov (8) g95.0<1>UW 0UW {align1};
  109. mov (8) g96.0<1>UW 0UW {align1};
  110. mov (8) g97.0<1>UW 0UW {align1};
  111. mov (8) g98.0<1>UW 0UW {align1};
  112. add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
  113. jmpi block_y3;
  114. do_idct_y2:
  115. mov (1) g125.0<1>UD ip {align1};
  116. jmpi IDCT_START;
  117. mov (8) g91.0<1>W g32.0<16,8,2>W {align1};
  118. mov (8) g92.0<1>W g33.0<16,8,2>W {align1};
  119. mov (8) g93.0<1>W g34.0<16,8,2>W {align1};
  120. mov (8) g94.0<1>W g35.0<16,8,2>W {align1};
  121. mov (8) g95.0<1>W g36.0<16,8,2>W {align1};
  122. mov (8) g96.0<1>W g37.0<16,8,2>W {align1};
  123. mov (8) g97.0<1>W g38.0<16,8,2>W {align1};
  124. mov (8) g98.0<1>W g39.0<16,8,2>W {align1};
  125.  
  126. //Y3
  127. block_y3:
  128. and.nz (1) null g82.8<1,1,1>UW 0x100UW {align1};
  129. (f0) jmpi do_idct_y3;
  130. mov (8) g91.16<1>UW 0UW {align1};
  131. mov (8) g92.16<1>UW 0UW {align1};
  132. mov (8) g93.16<1>UW 0UW {align1};
  133. mov (8) g94.16<1>UW 0UW {align1};
  134. mov (8) g95.16<1>UW 0UW {align1};
  135. mov (8) g96.16<1>UW 0UW {align1};
  136. mov (8) g97.16<1>UW 0UW {align1};
  137. mov (8) g98.16<1>UW 0UW {align1};
  138. add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
  139. jmpi block_u;
  140. do_idct_y3:
  141. mov (1) g125.0<1>UD ip {align1};
  142. jmpi IDCT_START;
  143. mov (8) g91.16<1>W g32.0<16,8,2>W {align1};
  144. mov (8) g92.16<1>W g33.0<16,8,2>W {align1};
  145. mov (8) g93.16<1>W g34.0<16,8,2>W {align1};
  146. mov (8) g94.16<1>W g35.0<16,8,2>W {align1};
  147. mov (8) g95.16<1>W g36.0<16,8,2>W {align1};
  148. mov (8) g96.16<1>W g37.0<16,8,2>W {align1};
  149. mov (8) g97.16<1>W g38.0<16,8,2>W {align1};
  150. mov (8) g98.16<1>W g39.0<16,8,2>W {align1};
  151.  
  152. //U
  153. block_u:
  154. and.nz (1) null g82.8<1,1,1>UW 0x80UW {align1};
  155. (f0) jmpi do_idct_u;
  156. mov (16) g99.0<1>UW 0UW {align1};
  157. mov (16) g100.0<1>UW 0UW {align1};
  158. mov (16) g101.0<1>UW 0UW {align1};
  159. mov (16) g102.0<1>UW 0UW {align1};
  160. add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00800080UD {align1};
  161. jmpi block_v;
  162. do_idct_u:
  163. mov (1) g125.0<1>UD ip {align1};
  164. jmpi IDCT_START;
  165. mov (16) g99.0<1>W g32.0<16,8,2>W {align1};
  166. mov (16) g100.0<1>W g34.0<16,8,2>W {align1};
  167. mov (16) g101.0<1>W g36.0<16,8,2>W {align1};
  168. mov (16) g102.0<1>W g38.0<16,8,2>W {align1};
  169.  
  170. //V
  171. block_v:
  172. and.nz (1) null g82.8<1,1,1>UW 0x40UW {align1};
  173. (f0) jmpi do_idct_v;
  174. mov (16) g103.0<1>UW 0UW {align1};
  175. mov (16) g104.0<1>UW 0UW {align1};
  176. mov (16) g105.0<1>UW 0UW {align1};
  177. mov (16) g106.0<1>UW 0UW {align1};  
  178. jmpi block_end;
  179. do_idct_v:
  180. mov (1) g125.0<1>UD ip {align1};
  181. jmpi IDCT_START;
  182. mov (16) g103.0<1>W g32.0<16,8,2>W {align1};
  183. mov (16) g104.0<1>W g34.0<16,8,2>W {align1};
  184. mov (16) g105.0<1>W g36.0<16,8,2>W {align1};
  185. mov (16) g106.0<1>W g38.0<16,8,2>W {align1};
  186. block_end:
  187.  
  188. add (1) ip g126.8<1,1,1>UD 0x20UD {align1};            //jump back
  189. include(`do_iq_non_intra.g4i')
  190. include(`idct.g4i')
  191.