| 1 | /* |
| 2 | * XVID MPEG-4 VIDEO CODEC |
| 3 | * - MMX and XMM forward discrete cosine transform - |
| 4 | * |
| 5 | * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
| 6 | * |
| 7 | * Originally provided by Intel at AP-922 |
| 8 | * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
| 9 | * (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
| 10 | * but in a limited edition. |
| 11 | * New macro implements a column part for precise iDCT |
| 12 | * The routine precision now satisfies IEEE standard 1180-1990. |
| 13 | * |
| 14 | * Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
| 15 | * Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> |
| 16 | * |
| 17 | * http://www.elecard.com/peter/idct.html |
| 18 | * http://www.linuxvideo.org/mpeg2dec/ |
| 19 | * |
| 20 | * These examples contain code fragments for first stage iDCT 8x8 |
| 21 | * (for rows) and first stage DCT 8x8 (for columns) |
| 22 | * |
| 23 | * conversion to gcc syntax by Michael Niedermayer |
| 24 | * |
| 25 | * This file is part of FFmpeg. |
| 26 | * |
| 27 | * FFmpeg is free software; you can redistribute it and/or |
| 28 | * modify it under the terms of the GNU Lesser General Public |
| 29 | * License as published by the Free Software Foundation; either |
| 30 | * version 2.1 of the License, or (at your option) any later version. |
| 31 | * |
| 32 | * FFmpeg is distributed in the hope that it will be useful, |
| 33 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 34 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 35 | * Lesser General Public License for more details. |
| 36 | * |
| 37 | * You should have received a copy of the GNU Lesser General Public License |
| 38 | * along with FFmpeg; if not, write to the Free Software Foundation, |
| 39 | * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 40 | */ |
| 41 | |
| 42 | #include <inttypes.h> |
| 43 | |
| 44 | #include "config.h" |
| 45 | |
| 46 | #include "libavutil/mem.h" |
| 47 | |
| 48 | #include "libavcodec/avcodec.h" |
| 49 | #include "libavcodec/idctdsp.h" |
| 50 | |
| 51 | #include "idctdsp.h" |
| 52 | #include "xvididct.h" |
| 53 | |
| 54 | #if HAVE_MMX_INLINE |
| 55 | |
| 56 | // ----------------------------------------------------------------------------- |
| 57 | // Various memory constants (trigonometric values or rounding values) |
| 58 | // ----------------------------------------------------------------------------- |
| 59 | |
| 60 | DECLARE_ALIGNED(8, static const int16_t, tg_1_16)[4 * 4] = { |
| 61 | 13036, 13036, 13036, 13036, // tg * (2 << 16) + 0.5 |
| 62 | 27146, 27146, 27146, 27146, // tg * (2 << 16) + 0.5 |
| 63 | -21746, -21746, -21746, -21746, // tg * (2 << 16) + 0.5 |
| 64 | 23170, 23170, 23170, 23170 |
| 65 | }; // cos * (2 << 15) + 0.5 |
| 66 | |
| 67 | DECLARE_ALIGNED(8, static const int32_t, rounder_0)[2 * 8] = { |
| 68 | 65536, 65536, |
| 69 | 3597, 3597, |
| 70 | 2260, 2260, |
| 71 | 1203, 1203, |
| 72 | 0, 0, |
| 73 | 120, 120, |
| 74 | 512, 512, |
| 75 | 512, 512 |
| 76 | }; |
| 77 | |
| 78 | // ----------------------------------------------------------------------------- |
| 79 | // |
| 80 | // The first stage iDCT 8x8 - inverse DCTs of rows |
| 81 | // |
| 82 | // ----------------------------------------------------------------------------- |
| 83 | // The 8-point inverse DCT direct algorithm |
| 84 | // ----------------------------------------------------------------------------- |
| 85 | // |
| 86 | // static const short w[32] = { |
| 87 | // FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16), |
| 88 | // FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16), |
| 89 | // FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16), |
| 90 | // FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16), |
| 91 | // FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16), |
| 92 | // FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16), |
| 93 | // FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16), |
| 94 | // FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) }; |
| 95 | // |
| 96 | // #define DCT_8_INV_ROW(x, y) |
| 97 | // { |
| 98 | // int a0, a1, a2, a3, b0, b1, b2, b3; |
| 99 | // |
| 100 | // a0 = x[0] * w[0] + x[2] * w[1] + x[4] * w[2] + x[6] * w[3]; |
| 101 | // a1 = x[0] * w[4] + x[2] * w[5] + x[4] * w[6] + x[6] * w[7]; |
| 102 | // a2 = x[0] * w[8] + x[2] * w[9] + x[4] * w[10] + x[6] * w[11]; |
| 103 | // a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15]; |
| 104 | // b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19]; |
| 105 | // b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23]; |
| 106 | // b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27]; |
| 107 | // b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31]; |
| 108 | // |
| 109 | // y[0] = SHIFT_ROUND(a0 + b0); |
| 110 | // y[1] = SHIFT_ROUND(a1 + b1); |
| 111 | // y[2] = SHIFT_ROUND(a2 + b2); |
| 112 | // y[3] = SHIFT_ROUND(a3 + b3); |
| 113 | // y[4] = SHIFT_ROUND(a3 - b3); |
| 114 | // y[5] = SHIFT_ROUND(a2 - b2); |
| 115 | // y[6] = SHIFT_ROUND(a1 - b1); |
| 116 | // y[7] = SHIFT_ROUND(a0 - b0); |
| 117 | // } |
| 118 | // |
| 119 | // ----------------------------------------------------------------------------- |
| 120 | // |
| 121 | // In this implementation the outputs of the iDCT-1D are multiplied |
| 122 | // for rows 0,4 - by cos_4_16, |
| 123 | // for rows 1,7 - by cos_1_16, |
| 124 | // for rows 2,6 - by cos_2_16, |
| 125 | // for rows 3,5 - by cos_3_16 |
| 126 | // and are shifted to the left for better accuracy. |
| 127 | // |
| 128 | // For the constants used, |
| 129 | // FIX(float_const) = (short) (float_const * (1 << 15) + 0.5) |
| 130 | // |
| 131 | // ----------------------------------------------------------------------------- |
| 132 | |
| 133 | // ----------------------------------------------------------------------------- |
| 134 | // Tables for mmx processors |
| 135 | // ----------------------------------------------------------------------------- |
| 136 | |
| 137 | // Table for rows 0,4 - constants are multiplied by cos_4_16 |
| 138 | DECLARE_ALIGNED(8, static const int16_t, tab_i_04_mmx)[32 * 4] = { |
| 139 | 16384, 16384, 16384, -16384, // movq-> w06 w04 w02 w00 |
| 140 | 21407, 8867, 8867, -21407, // w07 w05 w03 w01 |
| 141 | 16384, -16384, 16384, 16384, // w14 w12 w10 w08 |
| 142 | -8867, 21407, -21407, -8867, // w15 w13 w11 w09 |
| 143 | 22725, 12873, 19266, -22725, // w22 w20 w18 w16 |
| 144 | 19266, 4520, -4520, -12873, // w23 w21 w19 w17 |
| 145 | 12873, 4520, 4520, 19266, // w30 w28 w26 w24 |
| 146 | -22725, 19266, -12873, -22725, // w31 w29 w27 w25 |
| 147 | // Table for rows 1,7 - constants are multiplied by cos_1_16 |
| 148 | 22725, 22725, 22725, -22725, // movq-> w06 w04 w02 w00 |
| 149 | 29692, 12299, 12299, -29692, // w07 w05 w03 w01 |
| 150 | 22725, -22725, 22725, 22725, // w14 w12 w10 w08 |
| 151 | -12299, 29692, -29692, -12299, // w15 w13 w11 w09 |
| 152 | 31521, 17855, 26722, -31521, // w22 w20 w18 w16 |
| 153 | 26722, 6270, -6270, -17855, // w23 w21 w19 w17 |
| 154 | 17855, 6270, 6270, 26722, // w30 w28 w26 w24 |
| 155 | -31521, 26722, -17855, -31521, // w31 w29 w27 w25 |
| 156 | // Table for rows 2,6 - constants are multiplied by cos_2_16 |
| 157 | 21407, 21407, 21407, -21407, // movq-> w06 w04 w02 w00 |
| 158 | 27969, 11585, 11585, -27969, // w07 w05 w03 w01 |
| 159 | 21407, -21407, 21407, 21407, // w14 w12 w10 w08 |
| 160 | -11585, 27969, -27969, -11585, // w15 w13 w11 w09 |
| 161 | 29692, 16819, 25172, -29692, // w22 w20 w18 w16 |
| 162 | 25172, 5906, -5906, -16819, // w23 w21 w19 w17 |
| 163 | 16819, 5906, 5906, 25172, // w30 w28 w26 w24 |
| 164 | -29692, 25172, -16819, -29692, // w31 w29 w27 w25 |
| 165 | // Table for rows 3,5 - constants are multiplied by cos_3_16 |
| 166 | 19266, 19266, 19266, -19266, // movq-> w06 w04 w02 w00 |
| 167 | 25172, 10426, 10426, -25172, // w07 w05 w03 w01 |
| 168 | 19266, -19266, 19266, 19266, // w14 w12 w10 w08 |
| 169 | -10426, 25172, -25172, -10426, // w15 w13 w11 w09 |
| 170 | 26722, 15137, 22654, -26722, // w22 w20 w18 w16 |
| 171 | 22654, 5315, -5315, -15137, // w23 w21 w19 w17 |
| 172 | 15137, 5315, 5315, 22654, // w30 w28 w26 w24 |
| 173 | -26722, 22654, -15137, -26722, // w31 w29 w27 w25 |
| 174 | }; |
| 175 | // ----------------------------------------------------------------------------- |
| 176 | // Tables for xmm processors |
| 177 | // ----------------------------------------------------------------------------- |
| 178 | |
| 179 | // %3 for rows 0,4 - constants are multiplied by cos_4_16 |
| 180 | DECLARE_ALIGNED(8, static const int16_t, tab_i_04_xmm)[32 * 4] = { |
| 181 | 16384, 21407, 16384, 8867, // movq-> w05 w04 w01 w00 |
| 182 | 16384, 8867, -16384, -21407, // w07 w06 w03 w02 |
| 183 | 16384, -8867, 16384, -21407, // w13 w12 w09 w08 |
| 184 | -16384, 21407, 16384, -8867, // w15 w14 w11 w10 |
| 185 | 22725, 19266, 19266, -4520, // w21 w20 w17 w16 |
| 186 | 12873, 4520, -22725, -12873, // w23 w22 w19 w18 |
| 187 | 12873, -22725, 4520, -12873, // w29 w28 w25 w24 |
| 188 | 4520, 19266, 19266, -22725, // w31 w30 w27 w26 |
| 189 | // %3 for rows 1,7 - constants are multiplied by cos_1_16 |
| 190 | 22725, 29692, 22725, 12299, // movq-> w05 w04 w01 w00 |
| 191 | 22725, 12299, -22725, -29692, // w07 w06 w03 w02 |
| 192 | 22725, -12299, 22725, -29692, // w13 w12 w09 w08 |
| 193 | -22725, 29692, 22725, -12299, // w15 w14 w11 w10 |
| 194 | 31521, 26722, 26722, -6270, // w21 w20 w17 w16 |
| 195 | 17855, 6270, -31521, -17855, // w23 w22 w19 w18 |
| 196 | 17855, -31521, 6270, -17855, // w29 w28 w25 w24 |
| 197 | 6270, 26722, 26722, -31521, // w31 w30 w27 w26 |
| 198 | // %3 for rows 2,6 - constants are multiplied by cos_2_16 |
| 199 | 21407, 27969, 21407, 11585, // movq-> w05 w04 w01 w00 |
| 200 | 21407, 11585, -21407, -27969, // w07 w06 w03 w02 |
| 201 | 21407, -11585, 21407, -27969, // w13 w12 w09 w08 |
| 202 | -21407, 27969, 21407, -11585, // w15 w14 w11 w10 |
| 203 | 29692, 25172, 25172, -5906, // w21 w20 w17 w16 |
| 204 | 16819, 5906, -29692, -16819, // w23 w22 w19 w18 |
| 205 | 16819, -29692, 5906, -16819, // w29 w28 w25 w24 |
| 206 | 5906, 25172, 25172, -29692, // w31 w30 w27 w26 |
| 207 | // %3 for rows 3,5 - constants are multiplied by cos_3_16 |
| 208 | 19266, 25172, 19266, 10426, // movq-> w05 w04 w01 w00 |
| 209 | 19266, 10426, -19266, -25172, // w07 w06 w03 w02 |
| 210 | 19266, -10426, 19266, -25172, // w13 w12 w09 w08 |
| 211 | -19266, 25172, 19266, -10426, // w15 w14 w11 w10 |
| 212 | 26722, 22654, 22654, -5315, // w21 w20 w17 w16 |
| 213 | 15137, 5315, -26722, -15137, // w23 w22 w19 w18 |
| 214 | 15137, -26722, 5315, -15137, // w29 w28 w25 w24 |
| 215 | 5315, 22654, 22654, -26722, // w31 w30 w27 w26 |
| 216 | }; |
| 217 | // ============================================================================= |
| 218 | // Helper macros for the code |
| 219 | // ============================================================================= |
| 220 | |
| 221 | // ----------------------------------------------------------------------------- |
| 222 | // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER |
| 223 | // ----------------------------------------------------------------------------- |
| 224 | |
| 225 | #define DCT_8_INV_ROW_MMX(A1, A2, A3, A4) \ |
| 226 | "movq "#A1", %%mm0 \n\t" /* 0 ; x3 x2 x1 x0 */ \ |
| 227 | "movq 8+"#A1", %%mm1 \n\t" /* 1 ; x7 x6 x5 x4 */ \ |
| 228 | "movq %%mm0, %%mm2 \n\t" /* 2 ; x3 x2 x1 x0 */ \ |
| 229 | "movq "#A3", %%mm3 \n\t" /* 3 ; w06 w04 w02 w00 */ \ |
| 230 | "punpcklwd %%mm1, %%mm0 \n\t" /* x5 x1 x4 x0 */ \ |
| 231 | "movq %%mm0, %%mm5 \n\t" /* 5 ; x5 x1 x4 x0 */ \ |
| 232 | "punpckldq %%mm0, %%mm0 \n\t" /* x4 x0 x4 x0 */ \ |
| 233 | "movq 8+"#A3", %%mm4 \n\t" /* 4 ; w07 w05 w03 w01 */ \ |
| 234 | "punpckhwd %%mm1, %%mm2 \n\t" /* 1 ; x7 x3 x6 x2 */ \ |
| 235 | "pmaddwd %%mm0, %%mm3 \n\t" /* x4*w06+x0*w04 x4*w02+x0*w00 */ \ |
| 236 | "movq %%mm2, %%mm6 \n\t" /* 6 ; x7 x3 x6 x2 */ \ |
| 237 | "movq 32+"#A3", %%mm1 \n\t" /* 1 ; w22 w20 w18 w16 */ \ |
| 238 | "punpckldq %%mm2, %%mm2 \n\t" /* x6 x2 x6 x2 */ \ |
| 239 | "pmaddwd %%mm2, %%mm4 \n\t" /* x6*w07+x2*w05 x6*w03+x2*w01 */ \ |
| 240 | "punpckhdq %%mm5, %%mm5 \n\t" /* x5 x1 x5 x1 */ \ |
| 241 | "pmaddwd 16+"#A3", %%mm0 \n\t" /* x4*w14+x0*w12 x4*w10+x0*w08 */ \ |
| 242 | "punpckhdq %%mm6, %%mm6 \n\t" /* x7 x3 x7 x3 */ \ |
| 243 | "movq 40+ "#A3", %%mm7 \n\t" /* 7 ; w23 w21 w19 w17 */ \ |
| 244 | "pmaddwd %%mm5, %%mm1 \n\t" /* x5*w22+x1*w20 x5*w18+x1*w16 */ \ |
| 245 | "paddd "#A4", %%mm3 \n\t" /* +%4 */ \ |
| 246 | "pmaddwd %%mm6, %%mm7 \n\t" /* x7*w23+x3*w21 x7*w19+x3*w17 */ \ |
| 247 | "pmaddwd 24+"#A3", %%mm2 \n\t" /* x6*w15+x2*w13 x6*w11+x2*w09 */ \ |
| 248 | "paddd %%mm4, %%mm3 \n\t" /* 4 ; a1=sum(even1) a0=sum(even0) */ \ |
| 249 | "pmaddwd 48+"#A3", %%mm5 \n\t" /* x5*w30+x1*w28 x5*w26+x1*w24 */ \ |
| 250 | "movq %%mm3, %%mm4 \n\t" /* 4 ; a1 a0 */ \ |
| 251 | "pmaddwd 56+"#A3", %%mm6 \n\t" /* x7*w31+x3*w29 x7*w27+x3*w25 */ \ |
| 252 | "paddd %%mm7, %%mm1 \n\t" /* 7 ; b1=sum(odd1) b0=sum(odd0) */ \ |
| 253 | "paddd "#A4", %%mm0 \n\t" /* +%4 */ \ |
| 254 | "psubd %%mm1, %%mm3 \n\t" /* a1-b1 a0-b0 */ \ |
| 255 | "psrad $11, %%mm3 \n\t" /* y6=a1-b1 y7=a0-b0 */ \ |
| 256 | "paddd %%mm4, %%mm1 \n\t" /* 4 ; a1+b1 a0+b0 */ \ |
| 257 | "paddd %%mm2, %%mm0 \n\t" /* 2 ; a3=sum(even3) a2=sum(even2) */ \ |
| 258 | "psrad $11, %%mm1 \n\t" /* y1=a1+b1 y0=a0+b0 */ \ |
| 259 | "paddd %%mm6, %%mm5 \n\t" /* 6 ; b3=sum(odd3) b2=sum(odd2) */ \ |
| 260 | "movq %%mm0, %%mm4 \n\t" /* 4 ; a3 a2 */ \ |
| 261 | "paddd %%mm5, %%mm0 \n\t" /* a3+b3 a2+b2 */ \ |
| 262 | "psubd %%mm5, %%mm4 \n\t" /* 5 ; a3-b3 a2-b2 */ \ |
| 263 | "psrad $11, %%mm0 \n\t" /* y3=a3+b3 y2=a2+b2 */ \ |
| 264 | "psrad $11, %%mm4 \n\t" /* y4=a3-b3 y5=a2-b2 */ \ |
| 265 | "packssdw %%mm0, %%mm1 \n\t" /* 0 ; y3 y2 y1 y0 */ \ |
| 266 | "packssdw %%mm3, %%mm4 \n\t" /* 3 ; y6 y7 y4 y5 */ \ |
| 267 | "movq %%mm4, %%mm7 \n\t" /* 7 ; y6 y7 y4 y5 */ \ |
| 268 | "psrld $16, %%mm4 \n\t" /* 0 y6 0 y4 */ \ |
| 269 | "pslld $16, %%mm7 \n\t" /* y7 0 y5 0 */ \ |
| 270 | "movq %%mm1, "#A2" \n\t" /* 1 ; save y3 y2 y1 y0 */ \ |
| 271 | "por %%mm4, %%mm7 \n\t" /* 4 ; y7 y6 y5 y4 */ \ |
| 272 | "movq %%mm7, 8+"#A2" \n\t" /* 7 ; save y7 y6 y5 y4 */ \ |
| 273 | |
| 274 | |
| 275 | // ----------------------------------------------------------------------------- |
| 276 | // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER |
| 277 | // ----------------------------------------------------------------------------- |
| 278 | |
| 279 | #define DCT_8_INV_ROW_XMM(A1, A2, A3, A4) \ |
| 280 | "movq "#A1", %%mm0 \n\t" /* 0 ; x3 x2 x1 x0 */ \ |
| 281 | "movq 8+"#A1", %%mm1 \n\t" /* 1 ; x7 x6 x5 x4 */ \ |
| 282 | "movq %%mm0, %%mm2 \n\t" /* 2 ; x3 x2 x1 x0 */ \ |
| 283 | "movq "#A3", %%mm3 \n\t" /* 3 ; w05 w04 w01 w00 */ \ |
| 284 | "pshufw $0x88, %%mm0, %%mm0 \n\t" /* x2 x0 x2 x0 */ \ |
| 285 | "movq 8+"#A3", %%mm4 \n\t" /* 4 ; w07 w06 w03 w02 */ \ |
| 286 | "movq %%mm1, %%mm5 \n\t" /* 5 ; x7 x6 x5 x4 */ \ |
| 287 | "pmaddwd %%mm0, %%mm3 \n\t" /* x2*w05+x0*w04 x2*w01+x0*w00 */ \ |
| 288 | "movq 32+"#A3", %%mm6 \n\t" /* 6 ; w21 w20 w17 w16 */ \ |
| 289 | "pshufw $0x88, %%mm1, %%mm1 \n\t" /* x6 x4 x6 x4 */ \ |
| 290 | "pmaddwd %%mm1, %%mm4 \n\t" /* x6*w07+x4*w06 x6*w03+x4*w02 */ \ |
| 291 | "movq 40+"#A3", %%mm7 \n\t" /* 7; w23 w22 w19 w18 */ \ |
| 292 | "pshufw $0xdd, %%mm2, %%mm2 \n\t" /* x3 x1 x3 x1 */ \ |
| 293 | "pmaddwd %%mm2, %%mm6 \n\t" /* x3*w21+x1*w20 x3*w17+x1*w16 */ \ |
| 294 | "pshufw $0xdd, %%mm5, %%mm5 \n\t" /* x7 x5 x7 x5 */ \ |
| 295 | "pmaddwd %%mm5, %%mm7 \n\t" /* x7*w23+x5*w22 x7*w19+x5*w18 */ \ |
| 296 | "paddd "#A4", %%mm3 \n\t" /* +%4 */ \ |
| 297 | "pmaddwd 16+"#A3", %%mm0 \n\t" /* x2*w13+x0*w12 x2*w09+x0*w08 */ \ |
| 298 | "paddd %%mm4, %%mm3 \n\t" /* 4 ; a1=sum(even1) a0=sum(even0) */ \ |
| 299 | "pmaddwd 24+"#A3", %%mm1 \n\t" /* x6*w15+x4*w14 x6*w11+x4*w10 */ \ |
| 300 | "movq %%mm3, %%mm4 \n\t" /* 4 ; a1 a0 */ \ |
| 301 | "pmaddwd 48+"#A3", %%mm2 \n\t" /* x3*w29+x1*w28 x3*w25+x1*w24 */ \ |
| 302 | "paddd %%mm7, %%mm6 \n\t" /* 7 ; b1=sum(odd1) b0=sum(odd0) */ \ |
| 303 | "pmaddwd 56+"#A3", %%mm5 \n\t" /* x7*w31+x5*w30 x7*w27+x5*w26 */ \ |
| 304 | "paddd %%mm6, %%mm3 \n\t" /* a1+b1 a0+b0 */ \ |
| 305 | "paddd "#A4", %%mm0 \n\t" /* +%4 */ \ |
| 306 | "psrad $11, %%mm3 \n\t" /* y1=a1+b1 y0=a0+b0 */ \ |
| 307 | "paddd %%mm1, %%mm0 \n\t" /* 1 ; a3=sum(even3) a2=sum(even2) */ \ |
| 308 | "psubd %%mm6, %%mm4 \n\t" /* 6 ; a1-b1 a0-b0 */ \ |
| 309 | "movq %%mm0, %%mm7 \n\t" /* 7 ; a3 a2 */ \ |
| 310 | "paddd %%mm5, %%mm2 \n\t" /* 5 ; b3=sum(odd3) b2=sum(odd2) */ \ |
| 311 | "paddd %%mm2, %%mm0 \n\t" /* a3+b3 a2+b2 */ \ |
| 312 | "psrad $11, %%mm4 \n\t" /* y6=a1-b1 y7=a0-b0 */ \ |
| 313 | "psubd %%mm2, %%mm7 \n\t" /* 2 ; a3-b3 a2-b2 */ \ |
| 314 | "psrad $11, %%mm0 \n\t" /* y3=a3+b3 y2=a2+b2 */ \ |
| 315 | "psrad $11, %%mm7 \n\t" /* y4=a3-b3 y5=a2-b2 */ \ |
| 316 | "packssdw %%mm0, %%mm3 \n\t" /* 0 ; y3 y2 y1 y0 */ \ |
| 317 | "packssdw %%mm4, %%mm7 \n\t" /* 4 ; y6 y7 y4 y5 */ \ |
| 318 | "movq %%mm3, "#A2" \n\t" /* 3 ; save y3 y2 y1 y0 */ \ |
| 319 | "pshufw $0xb1, %%mm7, %%mm7 \n\t" /* y7 y6 y5 y4 */ \ |
| 320 | "movq %%mm7, 8+"#A2" \n\t" /* 7 ; save y7 y6 y5 y4 */ \ |
| 321 | |
| 322 | |
| 323 | // ----------------------------------------------------------------------------- |
| 324 | // |
| 325 | // The first stage DCT 8x8 - forward DCTs of columns |
| 326 | // |
| 327 | // The %2puts are multiplied |
| 328 | // for rows 0,4 - on cos_4_16, |
| 329 | // for rows 1,7 - on cos_1_16, |
| 330 | // for rows 2,6 - on cos_2_16, |
| 331 | // for rows 3,5 - on cos_3_16 |
| 332 | // and are shifted to the left for rise of accuracy |
| 333 | // |
| 334 | // ----------------------------------------------------------------------------- |
| 335 | // |
| 336 | // The 8-point scaled forward DCT algorithm (26a8m) |
| 337 | // |
| 338 | // ----------------------------------------------------------------------------- |
| 339 | // |
| 340 | //#define DCT_8_FRW_COL(x, y) |
| 341 | // { |
| 342 | // short t0, t1, t2, t3, t4, t5, t6, t7; |
| 343 | // short tp03, tm03, tp12, tm12, tp65, tm65; |
| 344 | // short tp465, tm465, tp765, tm765; |
| 345 | // |
| 346 | // t0 = LEFT_SHIFT(x[0] + x[7]); |
| 347 | // t1 = LEFT_SHIFT(x[1] + x[6]); |
| 348 | // t2 = LEFT_SHIFT(x[2] + x[5]); |
| 349 | // t3 = LEFT_SHIFT(x[3] + x[4]); |
| 350 | // t4 = LEFT_SHIFT(x[3] - x[4]); |
| 351 | // t5 = LEFT_SHIFT(x[2] - x[5]); |
| 352 | // t6 = LEFT_SHIFT(x[1] - x[6]); |
| 353 | // t7 = LEFT_SHIFT(x[0] - x[7]); |
| 354 | // |
| 355 | // tp03 = t0 + t3; |
| 356 | // tm03 = t0 - t3; |
| 357 | // tp12 = t1 + t2; |
| 358 | // tm12 = t1 - t2; |
| 359 | // |
| 360 | // y[0] = tp03 + tp12; |
| 361 | // y[4] = tp03 - tp12; |
| 362 | // |
| 363 | // y[2] = tm03 + tm12 * tg_2_16; |
| 364 | // y[6] = tm03 * tg_2_16 - tm12; |
| 365 | // |
| 366 | // tp65 = (t6 + t5) * cos_4_16; |
| 367 | // tm65 = (t6 - t5) * cos_4_16; |
| 368 | // |
| 369 | // tp765 = t7 + tp65; |
| 370 | // tm765 = t7 - tp65; |
| 371 | // tp465 = t4 + tm65; |
| 372 | // tm465 = t4 - tm65; |
| 373 | // |
| 374 | // y[1] = tp765 + tp465 * tg_1_16; |
| 375 | // y[7] = tp765 * tg_1_16 - tp465; |
| 376 | // y[5] = tm765 * tg_3_16 + tm465; |
| 377 | // y[3] = tm765 - tm465 * tg_3_16; |
| 378 | // } |
| 379 | // |
| 380 | // ----------------------------------------------------------------------------- |
| 381 | |
| 382 | // ----------------------------------------------------------------------------- |
| 383 | // DCT_8_INV_COL_4 INP,OUT |
| 384 | // ----------------------------------------------------------------------------- |
| 385 | |
| 386 | #define DCT_8_INV_COL(A1, A2) \ |
| 387 | "movq 2*8(%3), %%mm0 \n\t" \ |
| 388 | "movq 16*3+"#A1", %%mm3 \n\t" \ |
| 389 | "movq %%mm0, %%mm1 \n\t" /* tg_3_16 */ \ |
| 390 | "movq 16*5+"#A1", %%mm5 \n\t" \ |
| 391 | "pmulhw %%mm3, %%mm0 \n\t" /* x3*(tg_3_16-1) */ \ |
| 392 | "movq (%3), %%mm4 \n\t" \ |
| 393 | "pmulhw %%mm5, %%mm1 \n\t" /* x5*(tg_3_16-1) */ \ |
| 394 | "movq 16*7+"#A1", %%mm7 \n\t" \ |
| 395 | "movq %%mm4, %%mm2 \n\t" /* tg_1_16 */ \ |
| 396 | "movq 16*1+"#A1", %%mm6 \n\t" \ |
| 397 | "pmulhw %%mm7, %%mm4 \n\t" /* x7*tg_1_16 */ \ |
| 398 | "paddsw %%mm3, %%mm0 \n\t" /* x3*tg_3_16 */ \ |
| 399 | "pmulhw %%mm6, %%mm2 \n\t" /* x1*tg_1_16 */ \ |
| 400 | "paddsw %%mm3, %%mm1 \n\t" /* x3+x5*(tg_3_16-1) */ \ |
| 401 | "psubsw %%mm5, %%mm0 \n\t" /* x3*tg_3_16-x5 = tm35 */ \ |
| 402 | "movq 3*8(%3), %%mm3 \n\t" \ |
| 403 | "paddsw %%mm5, %%mm1 \n\t" /* x3+x5*tg_3_16 = tp35 */ \ |
| 404 | "paddsw %%mm6, %%mm4 \n\t" /* x1+tg_1_16*x7 = tp17 */ \ |
| 405 | "psubsw %%mm7, %%mm2 \n\t" /* x1*tg_1_16-x7 = tm17 */ \ |
| 406 | "movq %%mm4, %%mm5 \n\t" /* tp17 */ \ |
| 407 | "movq %%mm2, %%mm6 \n\t" /* tm17 */ \ |
| 408 | "paddsw %%mm1, %%mm5 \n\t" /* tp17+tp35 = b0 */ \ |
| 409 | "psubsw %%mm0, %%mm6 \n\t" /* tm17-tm35 = b3 */ \ |
| 410 | "psubsw %%mm1, %%mm4 \n\t" /* tp17-tp35 = t1 */ \ |
| 411 | "paddsw %%mm0, %%mm2 \n\t" /* tm17+tm35 = t2 */ \ |
| 412 | "movq 1*8(%3), %%mm7 \n\t" \ |
| 413 | "movq %%mm4, %%mm1 \n\t" /* t1 */ \ |
| 414 | "movq %%mm5, 3*16+"#A2" \n\t" /* save b0 */ \ |
| 415 | "paddsw %%mm2, %%mm1 \n\t" /* t1+t2 */ \ |
| 416 | "movq %%mm6, 5*16+"#A2" \n\t" /* save b3 */ \ |
| 417 | "psubsw %%mm2, %%mm4 \n\t" /* t1-t2 */ \ |
| 418 | "movq 2*16+"#A1", %%mm5 \n\t" \ |
| 419 | "movq %%mm7, %%mm0 \n\t" /* tg_2_16 */ \ |
| 420 | "movq 6*16+"#A1", %%mm6 \n\t" \ |
| 421 | "pmulhw %%mm5, %%mm0 \n\t" /* x2*tg_2_16 */ \ |
| 422 | "pmulhw %%mm6, %%mm7 \n\t" /* x6*tg_2_16 */ \ |
| 423 | "pmulhw %%mm3, %%mm1 \n\t" /* ocos_4_16*(t1+t2) = b1/2 */ \ |
| 424 | "movq 0*16+"#A1", %%mm2 \n\t" \ |
| 425 | "pmulhw %%mm3, %%mm4 \n\t" /* ocos_4_16*(t1-t2) = b2/2 */ \ |
| 426 | "psubsw %%mm6, %%mm0 \n\t" /* t2*tg_2_16-x6 = tm26 */ \ |
| 427 | "movq %%mm2, %%mm3 \n\t" /* x0 */ \ |
| 428 | "movq 4*16+"#A1", %%mm6 \n\t" \ |
| 429 | "paddsw %%mm5, %%mm7 \n\t" /* x2+x6*tg_2_16 = tp26 */ \ |
| 430 | "paddsw %%mm6, %%mm2 \n\t" /* x0+x4 = tp04 */ \ |
| 431 | "psubsw %%mm6, %%mm3 \n\t" /* x0-x4 = tm04 */ \ |
| 432 | "movq %%mm2, %%mm5 \n\t" /* tp04 */ \ |
| 433 | "movq %%mm3, %%mm6 \n\t" /* tm04 */ \ |
| 434 | "psubsw %%mm7, %%mm2 \n\t" /* tp04-tp26 = a3 */ \ |
| 435 | "paddsw %%mm0, %%mm3 \n\t" /* tm04+tm26 = a1 */ \ |
| 436 | "paddsw %%mm1, %%mm1 \n\t" /* b1 */ \ |
| 437 | "paddsw %%mm4, %%mm4 \n\t" /* b2 */ \ |
| 438 | "paddsw %%mm7, %%mm5 \n\t" /* tp04+tp26 = a0 */ \ |
| 439 | "psubsw %%mm0, %%mm6 \n\t" /* tm04-tm26 = a2 */ \ |
| 440 | "movq %%mm3, %%mm7 \n\t" /* a1 */ \ |
| 441 | "movq %%mm6, %%mm0 \n\t" /* a2 */ \ |
| 442 | "paddsw %%mm1, %%mm3 \n\t" /* a1+b1 */ \ |
| 443 | "paddsw %%mm4, %%mm6 \n\t" /* a2+b2 */ \ |
| 444 | "psraw $6, %%mm3 \n\t" /* dst1 */ \ |
| 445 | "psubsw %%mm1, %%mm7 \n\t" /* a1-b1 */ \ |
| 446 | "psraw $6, %%mm6 \n\t" /* dst2 */ \ |
| 447 | "psubsw %%mm4, %%mm0 \n\t" /* a2-b2 */ \ |
| 448 | "movq 3*16+"#A2", %%mm1 \n\t" /* load b0 */ \ |
| 449 | "psraw $6, %%mm7 \n\t" /* dst6 */ \ |
| 450 | "movq %%mm5, %%mm4 \n\t" /* a0 */ \ |
| 451 | "psraw $6, %%mm0 \n\t" /* dst5 */ \ |
| 452 | "movq %%mm3, 1*16+"#A2" \n\t" \ |
| 453 | "paddsw %%mm1, %%mm5 \n\t" /* a0+b0 */ \ |
| 454 | "movq %%mm6, 2*16+"#A2" \n\t" \ |
| 455 | "psubsw %%mm1, %%mm4 \n\t" /* a0-b0 */ \ |
| 456 | "movq 5*16+"#A2", %%mm3 \n\t" /* load b3 */ \ |
| 457 | "psraw $6, %%mm5 \n\t" /* dst0 */ \ |
| 458 | "movq %%mm2, %%mm6 \n\t" /* a3 */ \ |
| 459 | "psraw $6, %%mm4 \n\t" /* dst7 */ \ |
| 460 | "movq %%mm0, 5*16+"#A2" \n\t" \ |
| 461 | "paddsw %%mm3, %%mm2 \n\t" /* a3+b3 */ \ |
| 462 | "movq %%mm7, 6*16+"#A2" \n\t" \ |
| 463 | "psubsw %%mm3, %%mm6 \n\t" /* a3-b3 */ \ |
| 464 | "movq %%mm5, 0*16+"#A2" \n\t" \ |
| 465 | "psraw $6, %%mm2 \n\t" /* dst3 */ \ |
| 466 | "movq %%mm4, 7*16+"#A2" \n\t" \ |
| 467 | "psraw $6, %%mm6 \n\t" /* dst4 */ \ |
| 468 | "movq %%mm2, 3*16+"#A2" \n\t" \ |
| 469 | "movq %%mm6, 4*16+"#A2" \n\t" \ |
| 470 | |
| 471 | // ============================================================================= |
| 472 | // Code |
| 473 | // ============================================================================= |
| 474 | |
| 475 | // ----------------------------------------------------------------------------- |
| 476 | // void idct_mmx(uint16_t block[64]); |
| 477 | // ----------------------------------------------------------------------------- |
| 478 | |
| 479 | void ff_xvid_idct_mmx(short *block) |
| 480 | { |
| 481 | __asm__ volatile ( |
| 482 | // # Process each row |
| 483 | DCT_8_INV_ROW_MMX(0 * 16(%0), 0 * 16(%0), 64 * 0(%2), 8 * 0(%1)) |
| 484 | DCT_8_INV_ROW_MMX(1 * 16(%0), 1 * 16(%0), 64 * 1(%2), 8 * 1(%1)) |
| 485 | DCT_8_INV_ROW_MMX(2 * 16(%0), 2 * 16(%0), 64 * 2(%2), 8 * 2(%1)) |
| 486 | DCT_8_INV_ROW_MMX(3 * 16(%0), 3 * 16(%0), 64 * 3(%2), 8 * 3(%1)) |
| 487 | DCT_8_INV_ROW_MMX(4 * 16(%0), 4 * 16(%0), 64 * 0(%2), 8 * 4(%1)) |
| 488 | DCT_8_INV_ROW_MMX(5 * 16(%0), 5 * 16(%0), 64 * 3(%2), 8 * 5(%1)) |
| 489 | DCT_8_INV_ROW_MMX(6 * 16(%0), 6 * 16(%0), 64 * 2(%2), 8 * 6(%1)) |
| 490 | DCT_8_INV_ROW_MMX(7 * 16(%0), 7 * 16(%0), 64 * 1(%2), 8 * 7(%1)) |
| 491 | |
| 492 | // # Process the columns (4 at a time) |
| 493 | DCT_8_INV_COL(0(%0), 0(%0)) |
| 494 | DCT_8_INV_COL(8(%0), 8(%0)) |
| 495 | :: "r" (block), "r" (rounder_0), "r" (tab_i_04_mmx), "r" (tg_1_16)); |
| 496 | } |
| 497 | |
| 498 | void ff_xvid_idct_mmx_put(uint8_t *dest, int line_size, int16_t *block) |
| 499 | { |
| 500 | ff_xvid_idct_mmx(block); |
| 501 | ff_put_pixels_clamped(block, dest, line_size); |
| 502 | } |
| 503 | |
| 504 | void ff_xvid_idct_mmx_add(uint8_t *dest, int line_size, int16_t *block) |
| 505 | { |
| 506 | ff_xvid_idct_mmx(block); |
| 507 | ff_add_pixels_clamped(block, dest, line_size); |
| 508 | } |
| 509 | |
| 510 | #endif /* HAVE_MMX_INLINE */ |
| 511 | |
| 512 | #if HAVE_MMXEXT_INLINE |
| 513 | |
| 514 | // ----------------------------------------------------------------------------- |
| 515 | // void idct_xmm(uint16_t block[64]); |
| 516 | // ----------------------------------------------------------------------------- |
| 517 | |
| 518 | void ff_xvid_idct_mmxext(short *block) |
| 519 | { |
| 520 | __asm__ volatile ( |
| 521 | // # Process each row |
| 522 | DCT_8_INV_ROW_XMM(0 * 16(%0), 0 * 16(%0), 64 * 0(%2), 8 * 0(%1)) |
| 523 | DCT_8_INV_ROW_XMM(1 * 16(%0), 1 * 16(%0), 64 * 1(%2), 8 * 1(%1)) |
| 524 | DCT_8_INV_ROW_XMM(2 * 16(%0), 2 * 16(%0), 64 * 2(%2), 8 * 2(%1)) |
| 525 | DCT_8_INV_ROW_XMM(3 * 16(%0), 3 * 16(%0), 64 * 3(%2), 8 * 3(%1)) |
| 526 | DCT_8_INV_ROW_XMM(4 * 16(%0), 4 * 16(%0), 64 * 0(%2), 8 * 4(%1)) |
| 527 | DCT_8_INV_ROW_XMM(5 * 16(%0), 5 * 16(%0), 64 * 3(%2), 8 * 5(%1)) |
| 528 | DCT_8_INV_ROW_XMM(6 * 16(%0), 6 * 16(%0), 64 * 2(%2), 8 * 6(%1)) |
| 529 | DCT_8_INV_ROW_XMM(7 * 16(%0), 7 * 16(%0), 64 * 1(%2), 8 * 7(%1)) |
| 530 | |
| 531 | // # Process the columns (4 at a time) |
| 532 | DCT_8_INV_COL(0(%0), 0(%0)) |
| 533 | DCT_8_INV_COL(8(%0), 8(%0)) |
| 534 | :: "r" (block), "r" (rounder_0), "r" (tab_i_04_xmm), "r" (tg_1_16)); |
| 535 | } |
| 536 | |
| 537 | void ff_xvid_idct_mmxext_put(uint8_t *dest, int line_size, int16_t *block) |
| 538 | { |
| 539 | ff_xvid_idct_mmxext(block); |
| 540 | ff_put_pixels_clamped(block, dest, line_size); |
| 541 | } |
| 542 | |
| 543 | void ff_xvid_idct_mmxext_add(uint8_t *dest, int line_size, int16_t *block) |
| 544 | { |
| 545 | ff_xvid_idct_mmxext(block); |
| 546 | ff_add_pixels_clamped(block, dest, line_size); |
| 547 | } |
| 548 | |
| 549 | #endif /* HAVE_MMXEXT_INLINE */ |