faandct.c (7058B)
1 /* 2 * Floating point AAN DCT 3 * this implementation is based upon the IJG integer AAN DCT (see jfdctfst.c) 4 * 5 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at> 6 * Copyright (c) 2003 Roman Shaposhnik 7 * 8 * Permission to use, copy, modify, and/or distribute this software for any 9 * purpose with or without fee is hereby granted, provided that the above 10 * copyright notice and this permission notice appear in all copies. 11 * 12 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 13 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 14 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 15 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 16 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 17 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 18 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 19 */ 20 21 /** 22 * @file 23 * @brief 24 * Floating point AAN DCT 25 * @author Michael Niedermayer <michaelni@gmx.at> 26 */ 27 28 #include "faandct.h" 29 #include "libavutil/emms.h" 30 #include "libavutil/internal.h" 31 #include "libavutil/libm.h" 32 33 typedef float FLOAT; 34 35 /* numbers generated by arbitrary precision arithmetic followed by truncation 36 to 36 fractional digits (enough for a 128-bit IEEE quad, see /usr/include/math.h 37 for this approach). Unfortunately, long double is not always available correctly, 38 e.g ppc has issues. 39 TODO: add L suffixes when ppc and toolchains sort out their stuff. 40 */ 41 #define B0 1.000000000000000000000000000000000000 42 #define B1 0.720959822006947913789091890943021267 // (cos(pi*1/16)sqrt(2))^-1 43 #define B2 0.765366864730179543456919968060797734 // (cos(pi*2/16)sqrt(2))^-1 44 #define B3 0.850430094767256448766702844371412325 // (cos(pi*3/16)sqrt(2))^-1 45 #define B4 1.000000000000000000000000000000000000 // (cos(pi*4/16)sqrt(2))^-1 46 #define B5 1.272758580572833938461007018281767032 // (cos(pi*5/16)sqrt(2))^-1 47 #define B6 1.847759065022573512256366378793576574 // (cos(pi*6/16)sqrt(2))^-1 48 #define B7 3.624509785411551372409941227504289587 // (cos(pi*7/16)sqrt(2))^-1 49 50 #define A1 M_SQRT1_2 // cos(pi*4/16) 51 #define A2 0.54119610014619698435 // cos(pi*6/16)sqrt(2) 52 #define A5 0.38268343236508977170 // cos(pi*6/16) 53 #define A4 1.30656296487637652774 // cos(pi*2/16)sqrt(2) 54 55 static const FLOAT postscale[64]={ 56 B0*B0, B0*B1, B0*B2, B0*B3, B0*B4, B0*B5, B0*B6, B0*B7, 57 B1*B0, B1*B1, B1*B2, B1*B3, B1*B4, B1*B5, B1*B6, B1*B7, 58 B2*B0, B2*B1, B2*B2, B2*B3, B2*B4, B2*B5, B2*B6, B2*B7, 59 B3*B0, B3*B1, B3*B2, B3*B3, B3*B4, B3*B5, B3*B6, B3*B7, 60 B4*B0, B4*B1, B4*B2, B4*B3, B4*B4, B4*B5, B4*B6, B4*B7, 61 B5*B0, B5*B1, B5*B2, B5*B3, B5*B4, B5*B5, B5*B6, B5*B7, 62 B6*B0, B6*B1, B6*B2, B6*B3, B6*B4, B6*B5, B6*B6, B6*B7, 63 B7*B0, B7*B1, B7*B2, B7*B3, B7*B4, B7*B5, B7*B6, B7*B7, 64 }; 65 66 static av_always_inline void row_fdct(FLOAT temp[64], int16_t *data) 67 { 68 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 69 FLOAT tmp10, tmp11, tmp12, tmp13; 70 FLOAT z2, z4, z11, z13; 71 int i; 72 73 for (i=0; i<8*8; i+=8) { 74 tmp0= data[0 + i] + data[7 + i]; 75 tmp7= data[0 + i] - data[7 + i]; 76 tmp1= data[1 + i] + data[6 + i]; 77 tmp6= data[1 + i] - data[6 + i]; 78 tmp2= data[2 + i] + data[5 + i]; 79 tmp5= data[2 + i] - data[5 + i]; 80 tmp3= data[3 + i] + data[4 + i]; 81 tmp4= data[3 + i] - data[4 + i]; 82 83 tmp10= tmp0 + tmp3; 84 tmp13= tmp0 - tmp3; 85 tmp11= tmp1 + tmp2; 86 tmp12= tmp1 - tmp2; 87 88 temp[0 + i]= tmp10 + tmp11; 89 temp[4 + i]= tmp10 - tmp11; 90 91 tmp12 += tmp13; 92 tmp12 *= A1; 93 temp[2 + i]= tmp13 + tmp12; 94 temp[6 + i]= tmp13 - tmp12; 95 96 tmp4 += tmp5; 97 tmp5 += tmp6; 98 tmp6 += tmp7; 99 100 z2= tmp4*(A2+A5) - tmp6*A5; 101 z4= tmp6*(A4-A5) + tmp4*A5; 102 103 tmp5*=A1; 104 105 z11= tmp7 + tmp5; 106 z13= tmp7 - tmp5; 107 108 temp[5 + i]= z13 + z2; 109 temp[3 + i]= z13 - z2; 110 temp[1 + i]= z11 + z4; 111 temp[7 + i]= z11 - z4; 112 } 113 } 114 115 void ff_faandct(int16_t *data) 116 { 117 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 118 FLOAT tmp10, tmp11, tmp12, tmp13; 119 FLOAT z2, z4, z11, z13; 120 FLOAT temp[64]; 121 int i; 122 123 emms_c(); 124 125 row_fdct(temp, data); 126 127 for (i=0; i<8; i++) { 128 tmp0= temp[8*0 + i] + temp[8*7 + i]; 129 tmp7= temp[8*0 + i] - temp[8*7 + i]; 130 tmp1= temp[8*1 + i] + temp[8*6 + i]; 131 tmp6= temp[8*1 + i] - temp[8*6 + i]; 132 tmp2= temp[8*2 + i] + temp[8*5 + i]; 133 tmp5= temp[8*2 + i] - temp[8*5 + i]; 134 tmp3= temp[8*3 + i] + temp[8*4 + i]; 135 tmp4= temp[8*3 + i] - temp[8*4 + i]; 136 137 tmp10= tmp0 + tmp3; 138 tmp13= tmp0 - tmp3; 139 tmp11= tmp1 + tmp2; 140 tmp12= tmp1 - tmp2; 141 142 data[8*0 + i]= lrintf(postscale[8*0 + i] * (tmp10 + tmp11)); 143 data[8*4 + i]= lrintf(postscale[8*4 + i] * (tmp10 - tmp11)); 144 145 tmp12 += tmp13; 146 tmp12 *= A1; 147 data[8*2 + i]= lrintf(postscale[8*2 + i] * (tmp13 + tmp12)); 148 data[8*6 + i]= lrintf(postscale[8*6 + i] * (tmp13 - tmp12)); 149 150 tmp4 += tmp5; 151 tmp5 += tmp6; 152 tmp6 += tmp7; 153 154 z2= tmp4*(A2+A5) - tmp6*A5; 155 z4= tmp6*(A4-A5) + tmp4*A5; 156 157 tmp5*=A1; 158 159 z11= tmp7 + tmp5; 160 z13= tmp7 - tmp5; 161 162 data[8*5 + i]= lrintf(postscale[8*5 + i] * (z13 + z2)); 163 data[8*3 + i]= lrintf(postscale[8*3 + i] * (z13 - z2)); 164 data[8*1 + i]= lrintf(postscale[8*1 + i] * (z11 + z4)); 165 data[8*7 + i]= lrintf(postscale[8*7 + i] * (z11 - z4)); 166 } 167 } 168 169 void ff_faandct248(int16_t *data) 170 { 171 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 172 FLOAT tmp10, tmp11, tmp12, tmp13; 173 FLOAT temp[64]; 174 int i; 175 176 emms_c(); 177 178 row_fdct(temp, data); 179 180 for (i=0; i<8; i++) { 181 tmp0 = temp[8*0 + i] + temp[8*1 + i]; 182 tmp1 = temp[8*2 + i] + temp[8*3 + i]; 183 tmp2 = temp[8*4 + i] + temp[8*5 + i]; 184 tmp3 = temp[8*6 + i] + temp[8*7 + i]; 185 tmp4 = temp[8*0 + i] - temp[8*1 + i]; 186 tmp5 = temp[8*2 + i] - temp[8*3 + i]; 187 tmp6 = temp[8*4 + i] - temp[8*5 + i]; 188 tmp7 = temp[8*6 + i] - temp[8*7 + i]; 189 190 tmp10 = tmp0 + tmp3; 191 tmp11 = tmp1 + tmp2; 192 tmp12 = tmp1 - tmp2; 193 tmp13 = tmp0 - tmp3; 194 195 data[8*0 + i] = lrintf(postscale[8*0 + i] * (tmp10 + tmp11)); 196 data[8*4 + i] = lrintf(postscale[8*4 + i] * (tmp10 - tmp11)); 197 198 tmp12 += tmp13; 199 tmp12 *= A1; 200 data[8*2 + i] = lrintf(postscale[8*2 + i] * (tmp13 + tmp12)); 201 data[8*6 + i] = lrintf(postscale[8*6 + i] * (tmp13 - tmp12)); 202 203 tmp10 = tmp4 + tmp7; 204 tmp11 = tmp5 + tmp6; 205 tmp12 = tmp5 - tmp6; 206 tmp13 = tmp4 - tmp7; 207 208 data[8*1 + i] = lrintf(postscale[8*0 + i] * (tmp10 + tmp11)); 209 data[8*5 + i] = lrintf(postscale[8*4 + i] * (tmp10 - tmp11)); 210 211 tmp12 += tmp13; 212 tmp12 *= A1; 213 data[8*3 + i] = lrintf(postscale[8*2 + i] * (tmp13 + tmp12)); 214 data[8*7 + i] = lrintf(postscale[8*6 + i] * (tmp13 - tmp12)); 215 } 216 }