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. /*
  29.    GRF allocation:
  30.    g1~g30: constant buffer
  31.            g1~g2:intra IQ matrix
  32.            g3~g4:non intra IQ matrix
  33.            g5~g20:IDCT tab
  34.    g31:    read and write message descriptor
  35.    g32~g55:DCT data
  36.    g58~g81:reference data
  37.    g82:    thread payload
  38.    g83~g106:IDCT data
  39. */
  40. mov (8) g82.0<1>UD g31.0<8,8,1>UD {align1};
  41. mov (2) g31.0<1>UD g82.12<2,2,1>UW {align1};
  42.  
  43. include(`iq_intra.g4i')
  44.  
  45. //defined for idct
  46. define(`ROW_SHIFT', `11UD')
  47. define(`ROW_ADD', `0x400UD')          
  48. define(`COL_SHIFT', `20UD')
  49. define(`COL_ADD', `0x80000UD')          
  50.  
  51. mov (1) a0.0<1>UD 0x06F006E0UD {align1};      //0x06F006E0UD+0x00200020UD=0x07100700UD (g56.0 and g56.16)
  52.  
  53. //Y0
  54. mov (1) g125.0<1>UD ip {align1};
  55. jmpi IDCT_START;
  56. add (8) g83.0<1>W g32.0<16,8,2>W 128UW {align1};
  57. add (8) g84.0<1>W g33.0<16,8,2>W 128UW {align1};
  58. add (8) g85.0<1>W g34.0<16,8,2>W 128UW {align1};
  59. add (8) g86.0<1>W g35.0<16,8,2>W 128UW {align1};
  60. add (8) g87.0<1>W g36.0<16,8,2>W 128UW {align1};
  61. add (8) g88.0<1>W g37.0<16,8,2>W 128UW {align1};
  62. add (8) g89.0<1>W g38.0<16,8,2>W 128UW {align1};
  63. add (8) g90.0<1>W g39.0<16,8,2>W 128UW {align1};
  64.  
  65. //Y1
  66. mov (1) g125.0<1>UD ip {align1};
  67. jmpi IDCT_START;
  68. add (8) g83.16<1>W g32.0<16,8,2>W 128UW {align1};
  69. add (8) g84.16<1>W g33.0<16,8,2>W 128UW {align1};
  70. add (8) g85.16<1>W g34.0<16,8,2>W 128UW {align1};
  71. add (8) g86.16<1>W g35.0<16,8,2>W 128UW {align1};
  72. add (8) g87.16<1>W g36.0<16,8,2>W 128UW {align1};
  73. add (8) g88.16<1>W g37.0<16,8,2>W 128UW {align1};
  74. add (8) g89.16<1>W g38.0<16,8,2>W 128UW {align1};
  75. add (8) g90.16<1>W g39.0<16,8,2>W 128UW {align1};
  76.  
  77. //Y2
  78. mov (1) g125.0<1>UD ip {align1};
  79. jmpi IDCT_START;
  80. add (8) g91.0<1>W g32.0<16,8,2>W 128UW {align1};
  81. add (8) g92.0<1>W g33.0<16,8,2>W 128UW {align1};
  82. add (8) g93.0<1>W g34.0<16,8,2>W 128UW {align1};
  83. add (8) g94.0<1>W g35.0<16,8,2>W 128UW {align1};
  84. add (8) g95.0<1>W g36.0<16,8,2>W 128UW {align1};
  85. add (8) g96.0<1>W g37.0<16,8,2>W 128UW {align1};
  86. add (8) g97.0<1>W g38.0<16,8,2>W 128UW {align1};
  87. add (8) g98.0<1>W g39.0<16,8,2>W 128UW {align1};
  88.  
  89. //Y3
  90. mov (1) g125.0<1>UD ip {align1};
  91. jmpi IDCT_START;
  92. add (8) g91.16<1>W g32.0<16,8,2>W 128UW {align1};
  93. add (8) g92.16<1>W g33.0<16,8,2>W 128UW {align1};
  94. add (8) g93.16<1>W g34.0<16,8,2>W 128UW {align1};
  95. add (8) g94.16<1>W g35.0<16,8,2>W 128UW {align1};
  96. add (8) g95.16<1>W g36.0<16,8,2>W 128UW {align1};
  97. add (8) g96.16<1>W g37.0<16,8,2>W 128UW {align1};
  98. add (8) g97.16<1>W g38.0<16,8,2>W 128UW {align1};
  99. add (8) g98.16<1>W g39.0<16,8,2>W 128UW {align1};
  100.  
  101. //U
  102. mov (1) g125.0<1>UD ip {align1};
  103. jmpi IDCT_START;
  104. add (16) g99.0<1>W g32.0<16,8,2>W 128UW {align1};
  105. add (16) g100.0<1>W g34.0<16,8,2>W 128UW {align1};
  106. add (16) g101.0<1>W g36.0<16,8,2>W 128UW {align1};
  107. add (16) g102.0<1>W g38.0<16,8,2>W 128UW {align1};
  108.  
  109. //V
  110. mov (1) g125.0<1>UD ip {align1};
  111. jmpi IDCT_START;
  112. add (16) g103.0<1>W g32.0<16,8,2>W 128UW {align1};
  113. add (16) g104.0<1>W g34.0<16,8,2>W 128UW {align1};
  114. add (16) g105.0<1>W g36.0<16,8,2>W 128UW {align1};
  115. add (16) g106.0<1>W g38.0<16,8,2>W 128UW {align1};
  116.  
  117. //send msg
  118. mov (1) g31.8<1>UD 0x00F000FUD {align1};
  119. mov.sat (16) g83.0<2>UB g83.0<16,16,1>W {align1};
  120. mov.sat (16) g84.0<2>UB g84.0<16,16,1>W {align1};
  121. mov.sat (16) g85.0<2>UB g85.0<16,16,1>W {align1};
  122. mov.sat (16) g86.0<2>UB g86.0<16,16,1>W {align1};
  123. mov.sat (16) g87.0<2>UB g87.0<16,16,1>W {align1};
  124. mov.sat (16) g88.0<2>UB g88.0<16,16,1>W {align1};
  125. mov.sat (16) g89.0<2>UB g89.0<16,16,1>W {align1};
  126. mov.sat (16) g90.0<2>UB g90.0<16,16,1>W {align1};
  127. mov.sat (16) g91.0<2>UB g91.0<16,16,1>W {align1};
  128. mov.sat (16) g92.0<2>UB g92.0<16,16,1>W {align1};
  129. mov.sat (16) g93.0<2>UB g93.0<16,16,1>W {align1};
  130. mov.sat (16) g94.0<2>UB g94.0<16,16,1>W {align1};
  131. mov.sat (16) g95.0<2>UB g95.0<16,16,1>W {align1};
  132. mov.sat (16) g96.0<2>UB g96.0<16,16,1>W {align1};
  133. mov.sat (16) g97.0<2>UB g97.0<16,16,1>W {align1};
  134. mov.sat (16) g98.0<2>UB g98.0<16,16,1>W {align1};
  135.  
  136. and.nz (1) null g82.2<1,1,1>UW 0x20UW{align1};
  137. (f0) jmpi field_dct;
  138.  
  139. mov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
  140. mov (16) m1.16<1>UB g84.0<16,16,2>UB {align1};
  141. mov (16) m2.0<1>UB  g85.0<16,16,2>UB {align1};
  142. mov (16) m2.16<1>UB g86.0<16,16,2>UB {align1};
  143. mov (16) m3.0<1>UB  g87.0<16,16,2>UB {align1};
  144. mov (16) m3.16<1>UB g88.0<16,16,2>UB {align1};
  145. mov (16) m4.0<1>UB  g89.0<16,16,2>UB {align1};
  146. mov (16) m4.16<1>UB g90.0<16,16,2>UB {align1};
  147. mov (16) m5.0<1>UB  g91.0<16,16,2>UB {align1};
  148. mov (16) m5.16<1>UB g92.0<16,16,2>UB {align1};
  149. mov (16) m6.0<1>UB  g93.0<16,16,2>UB {align1};
  150. mov (16) m6.16<1>UB g94.0<16,16,2>UB {align1};
  151. mov (16) m7.0<1>UB  g95.0<16,16,2>UB {align1};
  152. mov (16) m7.16<1>UB g96.0<16,16,2>UB {align1};
  153. mov (16) m8.0<1>UB  g97.0<16,16,2>UB {align1};
  154. mov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
  155. jmpi write_back;
  156.  
  157. field_dct:
  158. mov (16) m1.0<1>UB  g83.0<16,16,2>UB {align1};
  159. mov (16) m1.16<1>UB g91.0<16,16,2>UB {align1};
  160. mov (16) m2.0<1>UB  g84.0<16,16,2>UB {align1};
  161. mov (16) m2.16<1>UB g92.0<16,16,2>UB {align1};
  162. mov (16) m3.0<1>UB  g85.0<16,16,2>UB {align1};
  163. mov (16) m3.16<1>UB g93.0<16,16,2>UB {align1};
  164. mov (16) m4.0<1>UB  g86.0<16,16,2>UB {align1};
  165. mov (16) m4.16<1>UB g94.0<16,16,2>UB {align1};
  166. mov (16) m5.0<1>UB  g87.0<16,16,2>UB {align1};
  167. mov (16) m5.16<1>UB g95.0<16,16,2>UB {align1};
  168. mov (16) m6.0<1>UB  g88.0<16,16,2>UB {align1};
  169. mov (16) m6.16<1>UB g96.0<16,16,2>UB {align1};
  170. mov (16) m7.0<1>UB  g89.0<16,16,2>UB {align1};
  171. mov (16) m7.16<1>UB g97.0<16,16,2>UB {align1};
  172. mov (16) m8.0<1>UB  g90.0<16,16,2>UB {align1};
  173. mov (16) m8.16<1>UB g98.0<16,16,2>UB {align1};
  174.  
  175. write_back:
  176. send (16) 0 acc0<1>UW g31<8,8,1>UW write(0,0,2,0) mlen 9 rlen 0 {align1};
  177.  
  178. //U
  179. mov (1) g31.8<1>UD 0x0070007UD  { align1 };
  180. shr (2) g31.0<1>UD g82.12<2,2,1>UW 1D {align1};
  181. mov.sat (16) g99.0<2>UB g99.0<16,16,1>W {align1};
  182. mov.sat (16) g100.0<2>UB g100.0<16,16,1>W {align1};
  183. mov.sat (16) g101.0<2>UB g101.0<16,16,1>W {align1};
  184. mov.sat (16) g102.0<2>UB g102.0<16,16,1>W {align1};
  185.  
  186. mov (16) m1.0<1>UB  g99.0<16,16,2>UB {align1};
  187. mov (16) m1.16<1>UB g100.0<16,16,2>UB {align1};
  188. mov (16) m2.0<1>UB  g101.0<16,16,2>UB {align1};
  189. mov (16) m2.16<1>UB g102.0<16,16,2>UB {align1};
  190. send (16) 0 acc0<1>UW g31<8,8,1>UW write(1, 0, 2, 0) mlen 3 rlen 0 { align1 };
  191.  
  192. //V
  193. mov.sat (16) g103.0<2>UB g103.0<16,16,1>W {align1};
  194. mov.sat (16) g104.0<2>UB g104.0<16,16,1>W {align1};
  195. mov.sat (16) g105.0<2>UB g105.0<16,16,1>W {align1};
  196. mov.sat (16) g106.0<2>UB g106.0<16,16,1>W {align1};
  197.  
  198. mov (16) m1.0<1>UB  g103.0<16,16,2>UB {align1};
  199. mov (16) m1.16<1>UB g104.0<16,16,2>UB {align1};
  200. mov (16) m2.0<1>UB  g105.0<16,16,2>UB {align1};
  201. mov (16) m2.16<1>UB g106.0<16,16,2>UB {align1};
  202. send (16) 0 acc0<1>UW g31<8,8,1>UW write(2, 0, 2, 0) mlen 3 rlen 0 { align1 };
  203.  
  204. OUT:
  205. send (16) 0 acc0<1>UW g0<8,8,1>UW
  206.         thread_spawner(0, 0, 0) mlen 1 rlen 0 { align1 EOT};
  207.  
  208. include(`do_iq_intra.g4i')
  209. include(`idct.g4i')
  210.