i965_drv_video/shaders/h264/mc/weightedPred.asm

   1 /*\r
   2  * Weighted prediction of luminance and chrominance\r
   3  * Copyright © <2010>, Intel Corporation.\r
   4  *\r
   5  * This program is licensed under the terms and conditions of the\r
   6  * Eclipse Public License (EPL), version 1.0.  The full text of the EPL is at\r
   7  * http://www.opensource.org/licenses/eclipse-1.0.php.\r
   8  *\r
   9  */\r
  10 // Kernel name: WeightedPred.asm\r
  11 //\r
  12 // Weighted prediction of luminance and chrominance\r
  13 //\r
  14 \r
  15 \r
  16 //#if !defined(__WeightedPred__)                // Make sure this is only included once\r
  17 //#define __WeightedPred__\r
  18 \r
  19 \r
  20         and.z.f0.0 (1) gWEIGHTFLAG:w    gWPREDFLAG:ub                                   nWBIDIR_MASK:w\r
  21         cmp.e.f0.1 (1) null:w                   gPREDFLAG:w                                             2:w\r
  22         (-f0.0) jmpi INTERLABEL(WeightedPred)\r
  23         (f0.1) jmpi INTERLABEL(DefaultWeightedPred_BiPred)\r
  24         \r
  25 INTERLABEL(DefaultWeightedPred_UniPred):\r
  26 \r
  27         cmp.e.f0.0 (1) null:w                   gPREDFLAG:w                                             0:w\r
  28         (f0.0) jmpi INTERLABEL(Return_WeightedPred)\r
  29 \r
  30         // luma\r
  31         mov (32)        gubYPRED(0)<2>          gubINTPY1(0)    {Compr}\r
  32         mov (32)        gubYPRED(2)<2>          gubINTPY1(2)    {Compr}\r
  33 \r
  34 #ifndef MONO\r
  35         // chroma       \r
  36         mov (32)        gubCPRED(0)<2>          gubINTPC1(0)    {Compr}\r
  37 #endif\r
  38 \r
  39         jmpi INTERLABEL(Return_WeightedPred)\r
  40         \r
  41 INTERLABEL(DefaultWeightedPred_BiPred):\r
  42 \r
  43         // luma\r
  44         avg.sat (32) gubYPRED(0)<2>             gubINTPY0(0)                                    gubINTPY1(0)    {Compr}\r
  45         avg.sat (32) gubYPRED(2)<2>             gubINTPY0(2)                                    gubINTPY1(2)    {Compr}\r
  46         \r
  47 #ifndef MONO\r
  48         // chroma\r
  49         avg.sat (32) gubCPRED(0)<2>             gubINTPC0(0)                                    gubINTPC1(0)    {Compr}\r
  50 #endif\r
  51         \r
  52         jmpi INTERLABEL(Return_WeightedPred)\r
  53         \r
  54 INTERLABEL(WeightedPred):\r
  55         cmp.e.f0.1 (1) null:w                   gWEIGHTFLAG:w                                   0x80:w\r
  56         (-f0.1) jmpi INTERLABEL(WeightedPred_Explicit)\r
  57         \r
  58         cmp.e.f0.0 (1) null:w                   gPREDFLAG:w                                             2:w\r
  59         (-f0.0) jmpi INTERLABEL(DefaultWeightedPred_UniPred)\r
  60 \r
  61         mov (2)         gYADD<1>:w                      32:w                                                            {NoDDClr}       \r
  62         mov (2)         gYSHIFT<1>:w            6:w                                                                     {NoDDChk}\r
  63         mov (4)         gOFFSET<1>:w            0:w\r
  64         mov (8)         gWT0<2>:w                       r[pWGT,0]<0;2,1>:w\r
  65         \r
  66         jmpi INTERLABEL(WeightedPred_LOOP)\r
  67         \r
  68         // Explicit Prediction\r
  69 INTERLABEL(WeightedPred_Explicit):\r
  70         \r
  71         // WA for weighted prediction - 2007/09/06      \r
  72 #ifdef  SW_W_128                // CTG SW WA\r
  73         cmp.e.f0.1 (8) null:ud                  r[pWGT,0]<8;8,1>:uw                             gudW128(0)<0;1,0>\r
  74 #else                                   // ILK HW solution\r
  75         and.ne.f0.1 (8) null:uw                 r[pWGT,12]<0;1,0>:ub                    0x88848421:v    // Expand W=128 flag to all components. 2 MSB are don't care\r
  76 #endif  \r
  77         asr.nz.f0.0 (2) gBIPRED<1>:w    gPREDFLAG<0;1,0>:w                              1:w\r
  78         asr (1)         gWEIGHTFLAG:w           gWEIGHTFLAG:w                                   6:w     \r
  79         (-f0.0) mov (2) gPREDFLAG1<1>:w gPREDFLAG<0;1,0>:w                                                              \r
  80         (f0.0) mov (2)  gPREDFLAG0<1>:ud 0x00010001:ud\r
  81         (-f0.0) add (2) gPREDFLAG0<1>:w -gPREDFLAG1<2;2,1>:w                    1:w\r
  82         \r
  83         // WA for weighted prediction - 2007/09/06      \r
  84         (f0.1) mov (8)  gWT0<1>:ud              0x00000080:ud\r
  85         (-f0.1) mov (8) gWT0<2>:w               r[pWGT,0]<16;8,2>:b\r
  86         (-f0.1) mov (8) gO0<2>:w                r[pWGT,1]<16;8,2>:b\r
  87         mul (16)                gWT0<1>:w               gWT0<16;16,1>:w                                 gPREDFLAG0<0;4,1>:w\r
  88 \r
  89         // Compute addition\r
  90         cmp.e.f0.1 (2) null<1>:w                gYWDENOM<2;2,1>:ub                              0:w\r
  91         (-f0.1) shl (2) gW0<1>:w                gWEIGHTFLAG<0;1,0>:w                    gYWDENOM<2;2,1>:ub\r
  92         (f0.1) mov (2) gW0<1>:w                 0:w\r
  93         (-f0.1) asr (2) gW0<1>:w                gW0<2;2,1>:w                                    1:w\r
  94         shl (2)         gYADD<1>:w                      gW0<2;2,1>:w                                    gBIPRED<0;1,0>:w\r
  95         (f0.1) add (2)  gYADD<1>:w              gYADD<2;2,1>:w                                  gBIPRED<0;1,0>:w\r
  96         \r
  97         // Compute shift\r
  98         add (2)         gYSHIFT<1>:w            gYWDENOM<2;2,1>:ub                              gBIPRED<0;1,0>:w\r
  99         \r
 100         // Compute offset\r
 101         add (4)         acc0<1>:w                       gO0<16;4,4>:w                                   gO1<16;4,4>:w\r
 102         add (4)         acc0<1>:w                       acc0<4;4,1>:w                                   gBIPRED<0;1,0>:w\r
 103         asr (4)         gOFFSET<1>:w            acc0<4;4,1>:w                                   gBIPRED<0;1,0>:w\r
 104 \r
 105 INTERLABEL(WeightedPred_LOOP):  \r
 106         // luma\r
 107         $for(0;<4;2) {  \r
 108         mul (16)        acc0<1>:w                       gubINTPY0(%1)                                   gWT0<0;1,0>:w\r
 109         mul (16)        acc1<1>:w                       gubINTPY0(%1+1)                                 gWT0<0;1,0>:w\r
 110         mac (16)        acc0<1>:w                       gubINTPY1(%1)                                   gWT1<0;1,0>:w\r
 111         mac (16)        acc1<1>:w                       gubINTPY1(%1+1)                                 gWT1<0;1,0>:w\r
 112         add (16)        acc0<1>:w                       acc0<16;16,1>:w                                 gYADD:w\r
 113         add (16)        acc1<1>:w                       acc1<16;16,1>:w                                 gYADD:w\r
 114         // Accumulator cannot be used as destination for ASR\r
 115         asr (16)        gwINTERIM_BUF3(0)<1> acc0<16;16,1>:w                            gYSHIFT:w\r
 116         asr (16)        gwINTERIM_BUF3(1)<1> acc1<16;16,1>:w                            gYSHIFT:w\r
 117         add.sat (16) gubYPRED(%1)<2>    gwINTERIM_BUF3(0)                               gOFFSET:w\r
 118         add.sat (16) gubYPRED(%1+1)<2>  gwINTERIM_BUF3(1)                               gOFFSET:w\r
 119         }       \r
 120 \r
 121 #ifndef MONO\r
 122         // chroma\r
 123         mul (16)        acc0<1>:w                       gubINTPC0(0)                                    gUW0<0;2,4>:w\r
 124         mul (16)        acc1<1>:w                       gubINTPC0(1)                                    gUW0<0;2,4>:w\r
 125         mac (16)        acc0<1>:w                       gubINTPC1(0)                                    gUW1<0;2,4>:w\r
 126         mac (16)        acc1<1>:w                       gubINTPC1(1)                                    gUW1<0;2,4>:w\r
 127         add (16)        acc0<1>:w                       acc0<16;16,1>:w                                 gCADD:w\r
 128         add (16)        acc1<1>:w                       acc1<16;16,1>:w                                 gCADD:w\r
 129         // Accumulator cannot be used as destination for ASR\r
 130         asr (16)        gwINTERIM_BUF3(0)<1> acc0<16;16,1>:w                            gCSHIFT:w\r
 131         asr (16)        gwINTERIM_BUF3(1)<1> acc1<16;16,1>:w                            gCSHIFT:w\r
 132         add.sat (16) gubCPRED(0)<2>             gwINTERIM_BUF3(0)                               gUOFFSET<0;2,1>:w\r
 133         add.sat (16) gubCPRED(1)<2>             gwINTERIM_BUF3(1)                               gUOFFSET<0;2,1>:w\r
 134 #endif\r
 135 \r
 136 \r
 137 INTERLABEL(Return_WeightedPred):\r
 138 \r
 139         \r
 140 //#endif        // !defined(__WeightedPred__)\r