Cinelerra: hvirtual/mpeg2enc/fdctdata.c Source File

00001 
00002 //
00003 //  fdctam32.c - AP922 MMX(3D-Now) forward-DCT
00004 //  ----------
00005 //  Intel Application Note AP-922 - fast, precise implementation of DCT
00006 //        http://developer.intel.com/vtune/cbts/appnotes.htm
00007 //  ----------
00008 //  
00009 //       This routine uses a 3D-Now/MMX enhancement to increase the
00010 //  accuracy of the fdct_col_4 macro.  The dct_col function uses 3D-Now's
00011 //  PMHULHRW instead of MMX's PMHULHW(and POR).  The substitution improves
00012 //  accuracy very slightly with performance penalty.  If the target CPU
00013 //  does not support 3D-Now, then this function cannot be executed.  
00014 //  fdctmm32.c contains the standard MMX implementation of AP-922.
00015 //  
00016 //  For a fast, precise MMX implementation of inverse-DCT 
00017 //              visit http://www.elecard.com/peter
00018 //
00019 //  v1.0 07/22/2000 (initial release)
00020 //       Initial release of AP922 MMX(3D-Now) forward_DCT.
00021 //       This code was tested with Visual C++ 6.0Pro + service_pack4 + 
00022 //       processor_pack_beta!  If you have the processor_pack_beta, you can
00023 //       remove the #include for amd3dx.h, and substitute the 'normal'
00024 //       assembly lines for the macro'd versions.  Otherwise, this
00025 //       code should compile 'as is', under Visual C++ 6.0 Pro.
00026 //     
00027 //  liaor@iname.com  http://members.tripod.com/~liaor  
00029 
00030 #include <inttypes.h>
00031 
00033 //
00034 // constants for the forward DCT
00035 // -----------------------------
00036 //
00037 // Be sure to check that your compiler is aligning all constants to QWORD
00038 // (8-byte) memory boundaries!  Otherwise the unaligned memory access will
00039 // severely stall MMX execution.
00040 //
00042 
00043 #define BITS_FRW_ACC    3 //; 2 or 3 for accuracy
00044 #define SHIFT_FRW_COL   BITS_FRW_ACC
00045 #define SHIFT_FRW_ROW   (BITS_FRW_ACC + 17)
00046 //#define RND_FRW_ROW           (262144 * (BITS_FRW_ACC - 1)) //; 1 << (SHIFT_FRW_ROW-1)
00047 #define RND_FRW_ROW             (1 << (SHIFT_FRW_ROW-1))
00048 //#define RND_FRW_COL           (2 * (BITS_FRW_ACC - 1)) //; 1 << (SHIFT_FRW_COL-1)
00049 #define RND_FRW_COL             (1 << (SHIFT_FRW_COL-1))
00050 
00051 //concatenated table, for forward DCT transformation
00052 const int16_t fdct_tg_all_16[] = {
00053         13036, 13036, 13036, 13036,             // tg * (2<<16) + 0.5
00054         27146, 27146, 27146, 27146,             // tg * (2<<16) + 0.5
00055         -21746, -21746, -21746, -21746, // tg * (2<<16) + 0.5
00056         -19195, -19195, -19195, -19195, //cos * (2<<16) + 0.5
00057         23170, 23170, 23170, 23170 };   //cos * (2<<15) + 0.5
00058 const long long  fdct_one_corr = 0x0001000100010001LL;
00059 const long fdct_r_row[2] = {RND_FRW_ROW, RND_FRW_ROW };
00060 
00061 const int16_t tab_frw_01234567[] = {  // forward_dct coeff table
00062     //row0
00063     16384, 16384, 21407, -8867,     //    w09 w01 w08 w00
00064     16384, 16384, 8867, -21407,     //    w13 w05 w12 w04
00065     16384, -16384, 8867, 21407,     //    w11 w03 w10 w02
00066     -16384, 16384, -21407, -8867,   //    w15 w07 w14 w06
00067     22725, 12873, 19266, -22725,    //    w22 w20 w18 w16
00068     19266, 4520, -4520, -12873,     //    w23 w21 w19 w17
00069     12873, 4520, 4520, 19266,       //    w30 w28 w26 w24
00070     -22725, 19266, -12873, -22725,  //    w31 w29 w27 w25
00071 
00072     //row1
00073     22725, 22725, 29692, -12299,    //    w09 w01 w08 w00
00074     22725, 22725, 12299, -29692,    //    w13 w05 w12 w04
00075     22725, -22725, 12299, 29692,    //    w11 w03 w10 w02
00076     -22725, 22725, -29692, -12299,  //    w15 w07 w14 w06
00077     31521, 17855, 26722, -31521,    //    w22 w20 w18 w16
00078     26722, 6270, -6270, -17855,     //    w23 w21 w19 w17
00079     17855, 6270, 6270, 26722,       //    w30 w28 w26 w24
00080     -31521, 26722, -17855, -31521,  //    w31 w29 w27 w25
00081 
00082     //row2
00083     21407, 21407, 27969, -11585,    //    w09 w01 w08 w00
00084     21407, 21407, 11585, -27969,    //    w13 w05 w12 w04
00085     21407, -21407, 11585, 27969,    //    w11 w03 w10 w02
00086     -21407, 21407, -27969, -11585,  //    w15 w07 w14 w06
00087     29692, 16819, 25172, -29692,    //    w22 w20 w18 w16
00088     25172, 5906, -5906, -16819,     //    w23 w21 w19 w17
00089     16819, 5906, 5906, 25172,       //    w30 w28 w26 w24
00090     -29692, 25172, -16819, -29692,  //    w31 w29 w27 w25
00091 
00092     //row3
00093     19266, 19266, 25172, -10426,    //    w09 w01 w08 w00
00094     19266, 19266, 10426, -25172,    //    w13 w05 w12 w04
00095     19266, -19266, 10426, 25172,    //    w11 w03 w10 w02
00096     -19266, 19266, -25172, -10426,  //    w15 w07 w14 w06, 
00097     26722, 15137, 22654, -26722,    //    w22 w20 w18 w16
00098     22654, 5315, -5315, -15137,     //    w23 w21 w19 w17
00099     15137, 5315, 5315, 22654,       //    w30 w28 w26 w24
00100     -26722, 22654, -15137, -26722,  //    w31 w29 w27 w25, 
00101 
00102     //row4
00103     16384, 16384, 21407, -8867,     //    w09 w01 w08 w00
00104     16384, 16384, 8867, -21407,     //    w13 w05 w12 w04
00105     16384, -16384, 8867, 21407,     //    w11 w03 w10 w02
00106     -16384, 16384, -21407, -8867,   //    w15 w07 w14 w06
00107     22725, 12873, 19266, -22725,    //    w22 w20 w18 w16
00108     19266, 4520, -4520, -12873,     //    w23 w21 w19 w17
00109     12873, 4520, 4520, 19266,       //    w30 w28 w26 w24
00110     -22725, 19266, -12873, -22725,  //    w31 w29 w27 w25 
00111 
00112     //row5
00113     19266, 19266, 25172, -10426,    //    w09 w01 w08 w00
00114     19266, 19266, 10426, -25172,    //    w13 w05 w12 w04
00115     19266, -19266, 10426, 25172,    //    w11 w03 w10 w02
00116     -19266, 19266, -25172, -10426,  //    w15 w07 w14 w06
00117     26722, 15137, 22654, -26722,    //    w22 w20 w18 w16
00118     22654, 5315, -5315, -15137,     //    w23 w21 w19 w17
00119     15137, 5315, 5315, 22654,       //    w30 w28 w26 w24
00120     -26722, 22654, -15137, -26722,  //    w31 w29 w27 w25
00121 
00122     //row6
00123     21407, 21407, 27969, -11585,    //    w09 w01 w08 w00
00124     21407, 21407, 11585, -27969,    //    w13 w05 w12 w04
00125     21407, -21407, 11585, 27969,    //    w11 w03 w10 w02
00126     -21407, 21407, -27969, -11585,  //    w15 w07 w14 w06, 
00127     29692, 16819, 25172, -29692,    //    w22 w20 w18 w16
00128     25172, 5906, -5906, -16819,     //    w23 w21 w19 w17
00129     16819, 5906, 5906, 25172,       //    w30 w28 w26 w24
00130     -29692, 25172, -16819, -29692,  //    w31 w29 w27 w25, 
00131 
00132     //row7
00133     22725, 22725, 29692, -12299,    //    w09 w01 w08 w00
00134     22725, 22725, 12299, -29692,    //    w13 w05 w12 w04
00135     22725, -22725, 12299, 29692,    //    w11 w03 w10 w02
00136     -22725, 22725, -29692, -12299,  //    w15 w07 w14 w06, 
00137     31521, 17855, 26722, -31521,    //    w22 w20 w18 w16
00138     26722, 6270, -6270, -17855,     //    w23 w21 w19 w17
00139     17855, 6270, 6270, 26722,       //    w30 w28 w26 w24
00140     -31521, 26722, -17855, -31521   //    w31 w29 w27 w25
00141 };
00142 
00143