Cinelerra: hvirtual/quicktime/ffmpeg/libavcodec/i386/dsputil_mmx

00001 /*
00002  * DSP utils mmx functions are compiled twice for rnd/no_rnd
00003  * Copyright (c) 2000, 2001 Fabrice Bellard.
00004  * Copyright (c) 2003-2004 Michael Niedermayer <michaelni@gmx.at>
00005  *
00006  * This library is free software; you can redistribute it and/or
00007  * modify it under the terms of the GNU Lesser General Public
00008  * License as published by the Free Software Foundation; either
00009  * version 2 of the License, or (at your option) any later version.
00010  *
00011  * This library is distributed in the hope that it will be useful,
00012  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00013  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00014  * Lesser General Public License for more details.
00015  *
00016  * You should have received a copy of the GNU Lesser General Public
00017  * License along with this library; if not, write to the Free Software
00018  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
00019  *
00020  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
00021  * mostly rewritten by Michael Niedermayer <michaelni@gmx.at>
00022  * and improved by Zdenek Kabelac <kabi@users.sf.net>
00023  */
00024 
00025 // put_pixels
00026 static void DEF(put, pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00027 {
00028     MOVQ_BFE(mm6);
00029     __asm __volatile(
00030         "lea    (%3, %3), %%"REG_a"     \n\t"
00031         ".balign 8                      \n\t"
00032         "1:                             \n\t"
00033         "movq   (%1), %%mm0             \n\t"
00034         "movq   1(%1), %%mm1            \n\t"
00035         "movq   (%1, %3), %%mm2         \n\t"
00036         "movq   1(%1, %3), %%mm3        \n\t"
00037         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00038         "movq   %%mm4, (%2)             \n\t"
00039         "movq   %%mm5, (%2, %3)         \n\t"
00040         "add    %%"REG_a", %1           \n\t"
00041         "add    %%"REG_a", %2           \n\t"
00042         "movq   (%1), %%mm0             \n\t"
00043         "movq   1(%1), %%mm1            \n\t"
00044         "movq   (%1, %3), %%mm2         \n\t"
00045         "movq   1(%1, %3), %%mm3        \n\t"
00046         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00047         "movq   %%mm4, (%2)             \n\t"
00048         "movq   %%mm5, (%2, %3)         \n\t"
00049         "add    %%"REG_a", %1           \n\t"
00050         "add    %%"REG_a", %2           \n\t"
00051         "subl   $4, %0                  \n\t"
00052         "jnz    1b                      \n\t"
00053         :"+g"(h), "+S"(pixels), "+D"(block)
00054         :"r"((long)line_size)
00055         :REG_a, "memory");
00056 }
00057 
00058 static void DEF(put, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
00059 {
00060     MOVQ_BFE(mm6);
00061     __asm __volatile(
00062         "testl $1, %0                   \n\t"
00063         " jz 1f                         \n\t"
00064         "movq   (%1), %%mm0             \n\t"
00065         "movq   (%2), %%mm1             \n\t"
00066         "add    %4, %1                  \n\t"
00067         "add    $8, %2                  \n\t"
00068         PAVGB(%%mm0, %%mm1, %%mm4, %%mm6)
00069         "movq   %%mm4, (%3)             \n\t"
00070         "add    %5, %3                  \n\t"
00071         "decl   %0                      \n\t"
00072         ".balign 8                      \n\t"
00073         "1:                             \n\t"
00074         "movq   (%1), %%mm0             \n\t"
00075         "movq   (%2), %%mm1             \n\t"
00076         "add    %4, %1                  \n\t"
00077         "movq   (%1), %%mm2             \n\t"
00078         "movq   8(%2), %%mm3            \n\t"
00079         "add    %4, %1                  \n\t"
00080         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00081         "movq   %%mm4, (%3)             \n\t"
00082         "add    %5, %3                  \n\t"
00083         "movq   %%mm5, (%3)             \n\t"
00084         "add    %5, %3                  \n\t"
00085         "movq   (%1), %%mm0             \n\t"
00086         "movq   16(%2), %%mm1           \n\t"
00087         "add    %4, %1                  \n\t"
00088         "movq   (%1), %%mm2             \n\t"
00089         "movq   24(%2), %%mm3           \n\t"
00090         "add    %4, %1                  \n\t"
00091         "add    $32, %2                 \n\t"
00092         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00093         "movq   %%mm4, (%3)             \n\t"
00094         "add    %5, %3                  \n\t"
00095         "movq   %%mm5, (%3)             \n\t"
00096         "add    %5, %3                  \n\t"
00097         "subl   $4, %0                  \n\t"
00098         "jnz    1b                      \n\t"
00099 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
00100         :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
00101 #else
00102         :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
00103 #endif
00104         :"S"((long)src1Stride), "D"((long)dstStride)
00105         :"memory");
00106 }
00107 
00108 static void DEF(put, pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00109 {
00110     MOVQ_BFE(mm6);
00111     __asm __volatile(
00112         "lea    (%3, %3), %%"REG_a"     \n\t"
00113         ".balign 8                      \n\t"
00114         "1:                             \n\t"
00115         "movq   (%1), %%mm0             \n\t"
00116         "movq   1(%1), %%mm1            \n\t"
00117         "movq   (%1, %3), %%mm2         \n\t"
00118         "movq   1(%1, %3), %%mm3        \n\t"
00119         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00120         "movq   %%mm4, (%2)             \n\t"
00121         "movq   %%mm5, (%2, %3)         \n\t"
00122         "movq   8(%1), %%mm0            \n\t"
00123         "movq   9(%1), %%mm1            \n\t"
00124         "movq   8(%1, %3), %%mm2        \n\t"
00125         "movq   9(%1, %3), %%mm3        \n\t"
00126         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00127         "movq   %%mm4, 8(%2)            \n\t"
00128         "movq   %%mm5, 8(%2, %3)        \n\t"
00129         "add    %%"REG_a", %1           \n\t"
00130         "add    %%"REG_a", %2           \n\t"
00131         "movq   (%1), %%mm0             \n\t"
00132         "movq   1(%1), %%mm1            \n\t"
00133         "movq   (%1, %3), %%mm2         \n\t"
00134         "movq   1(%1, %3), %%mm3        \n\t"
00135         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00136         "movq   %%mm4, (%2)             \n\t"
00137         "movq   %%mm5, (%2, %3)         \n\t"
00138         "movq   8(%1), %%mm0            \n\t"
00139         "movq   9(%1), %%mm1            \n\t"
00140         "movq   8(%1, %3), %%mm2        \n\t"
00141         "movq   9(%1, %3), %%mm3        \n\t"
00142         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00143         "movq   %%mm4, 8(%2)            \n\t"
00144         "movq   %%mm5, 8(%2, %3)        \n\t"
00145         "add    %%"REG_a", %1           \n\t"
00146         "add    %%"REG_a", %2           \n\t"
00147         "subl   $4, %0                  \n\t"
00148         "jnz    1b                      \n\t"
00149         :"+g"(h), "+S"(pixels), "+D"(block)
00150         :"r"((long)line_size)
00151         :REG_a, "memory");
00152 }
00153 
00154 static void DEF(put, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
00155 {
00156     MOVQ_BFE(mm6);
00157     __asm __volatile(
00158         "testl $1, %0                   \n\t"
00159         " jz 1f                         \n\t"
00160         "movq   (%1), %%mm0             \n\t"
00161         "movq   (%2), %%mm1             \n\t"
00162         "movq   8(%1), %%mm2            \n\t"
00163         "movq   8(%2), %%mm3            \n\t"
00164         "add    %4, %1                  \n\t"
00165         "add    $16, %2                 \n\t"
00166         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00167         "movq   %%mm4, (%3)             \n\t"
00168         "movq   %%mm5, 8(%3)            \n\t"
00169         "add    %5, %3                  \n\t"
00170         "decl   %0                      \n\t"
00171         ".balign 8                      \n\t"
00172         "1:                             \n\t"
00173         "movq   (%1), %%mm0             \n\t"
00174         "movq   (%2), %%mm1             \n\t"
00175         "movq   8(%1), %%mm2            \n\t"
00176         "movq   8(%2), %%mm3            \n\t"
00177         "add    %4, %1                  \n\t"
00178         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00179         "movq   %%mm4, (%3)             \n\t"
00180         "movq   %%mm5, 8(%3)            \n\t"
00181         "add    %5, %3                  \n\t"
00182         "movq   (%1), %%mm0             \n\t"
00183         "movq   16(%2), %%mm1           \n\t"
00184         "movq   8(%1), %%mm2            \n\t"
00185         "movq   24(%2), %%mm3           \n\t"
00186         "add    %4, %1                  \n\t"
00187         PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
00188         "movq   %%mm4, (%3)             \n\t"
00189         "movq   %%mm5, 8(%3)            \n\t"
00190         "add    %5, %3                  \n\t"
00191         "add    $32, %2                 \n\t"
00192         "subl   $2, %0                  \n\t"
00193         "jnz    1b                      \n\t"
00194 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
00195         :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
00196 #else
00197         :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
00198 #endif
00199         :"S"((long)src1Stride), "D"((long)dstStride)
00200         :"memory"); 
00201 }
00202 
00203 static void DEF(put, pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00204 {
00205     MOVQ_BFE(mm6);
00206     __asm __volatile(
00207         "lea (%3, %3), %%"REG_a"        \n\t"
00208         "movq (%1), %%mm0               \n\t"
00209         ".balign 8                      \n\t"
00210         "1:                             \n\t"
00211         "movq   (%1, %3), %%mm1         \n\t"
00212         "movq   (%1, %%"REG_a"),%%mm2   \n\t"
00213         PAVGBP(%%mm1, %%mm0, %%mm4,   %%mm2, %%mm1, %%mm5)
00214         "movq   %%mm4, (%2)             \n\t"
00215         "movq   %%mm5, (%2, %3)         \n\t"
00216         "add    %%"REG_a", %1           \n\t"
00217         "add    %%"REG_a", %2           \n\t"
00218         "movq   (%1, %3), %%mm1         \n\t"
00219         "movq   (%1, %%"REG_a"),%%mm0   \n\t"
00220         PAVGBP(%%mm1, %%mm2, %%mm4,   %%mm0, %%mm1, %%mm5)
00221         "movq   %%mm4, (%2)             \n\t"
00222         "movq   %%mm5, (%2, %3)         \n\t"
00223         "add    %%"REG_a", %1           \n\t"
00224         "add    %%"REG_a", %2           \n\t"
00225         "subl   $4, %0                  \n\t"
00226         "jnz    1b                      \n\t"
00227         :"+g"(h), "+S"(pixels), "+D"(block)
00228         :"r"((long)line_size)
00229         :REG_a, "memory");
00230 }
00231 
00232 static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00233 {
00234     MOVQ_ZERO(mm7);
00235     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
00236     __asm __volatile(
00237         "movq   (%1), %%mm0             \n\t"
00238         "movq   1(%1), %%mm4            \n\t"
00239         "movq   %%mm0, %%mm1            \n\t"
00240         "movq   %%mm4, %%mm5            \n\t"
00241         "punpcklbw %%mm7, %%mm0         \n\t"
00242         "punpcklbw %%mm7, %%mm4         \n\t"
00243         "punpckhbw %%mm7, %%mm1         \n\t"
00244         "punpckhbw %%mm7, %%mm5         \n\t"
00245         "paddusw %%mm0, %%mm4           \n\t"
00246         "paddusw %%mm1, %%mm5           \n\t"
00247         "xor    %%"REG_a", %%"REG_a"    \n\t"
00248         "add    %3, %1                  \n\t"
00249         ".balign 8                      \n\t"
00250         "1:                             \n\t"
00251         "movq   (%1, %%"REG_a"), %%mm0  \n\t"
00252         "movq   1(%1, %%"REG_a"), %%mm2 \n\t"
00253         "movq   %%mm0, %%mm1            \n\t"
00254         "movq   %%mm2, %%mm3            \n\t"
00255         "punpcklbw %%mm7, %%mm0         \n\t"
00256         "punpcklbw %%mm7, %%mm2         \n\t"
00257         "punpckhbw %%mm7, %%mm1         \n\t"
00258         "punpckhbw %%mm7, %%mm3         \n\t"
00259         "paddusw %%mm2, %%mm0           \n\t"
00260         "paddusw %%mm3, %%mm1           \n\t"
00261         "paddusw %%mm6, %%mm4           \n\t"
00262         "paddusw %%mm6, %%mm5           \n\t"
00263         "paddusw %%mm0, %%mm4           \n\t"
00264         "paddusw %%mm1, %%mm5           \n\t"
00265         "psrlw  $2, %%mm4               \n\t"
00266         "psrlw  $2, %%mm5               \n\t"
00267         "packuswb  %%mm5, %%mm4         \n\t"
00268         "movq   %%mm4, (%2, %%"REG_a")  \n\t"
00269         "add    %3, %%"REG_a"           \n\t"
00270 
00271         "movq   (%1, %%"REG_a"), %%mm2  \n\t" // 0 <-> 2   1 <-> 3
00272         "movq   1(%1, %%"REG_a"), %%mm4 \n\t"
00273         "movq   %%mm2, %%mm3            \n\t"
00274         "movq   %%mm4, %%mm5            \n\t"
00275         "punpcklbw %%mm7, %%mm2         \n\t"
00276         "punpcklbw %%mm7, %%mm4         \n\t"
00277         "punpckhbw %%mm7, %%mm3         \n\t"
00278         "punpckhbw %%mm7, %%mm5         \n\t"
00279         "paddusw %%mm2, %%mm4           \n\t"
00280         "paddusw %%mm3, %%mm5           \n\t"
00281         "paddusw %%mm6, %%mm0           \n\t"
00282         "paddusw %%mm6, %%mm1           \n\t"
00283         "paddusw %%mm4, %%mm0           \n\t"
00284         "paddusw %%mm5, %%mm1           \n\t"
00285         "psrlw  $2, %%mm0               \n\t"
00286         "psrlw  $2, %%mm1               \n\t"
00287         "packuswb  %%mm1, %%mm0         \n\t"
00288         "movq   %%mm0, (%2, %%"REG_a")  \n\t"
00289         "add    %3, %%"REG_a"           \n\t"
00290 
00291         "subl   $2, %0                  \n\t"
00292         "jnz    1b                      \n\t"
00293         :"+g"(h), "+S"(pixels)
00294         :"D"(block), "r"((long)line_size)
00295         :REG_a, "memory");
00296 }
00297 
00298 // avg_pixels
00299 static void DEF(avg, pixels4)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00300 {
00301     MOVQ_BFE(mm6);
00302     JUMPALIGN();
00303     do {
00304         __asm __volatile(
00305              "movd  %0, %%mm0           \n\t"
00306              "movd  %1, %%mm1           \n\t"
00307              PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00308              "movd  %%mm2, %0           \n\t"
00309              :"+m"(*block)
00310              :"m"(*pixels)
00311              :"memory");
00312         pixels += line_size;
00313         block += line_size;
00314     }
00315     while (--h);
00316 }
00317 
00318 // in case more speed is needed - unroling would certainly help
00319 static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00320 {
00321     MOVQ_BFE(mm6);
00322     JUMPALIGN();
00323     do {
00324         __asm __volatile(
00325              "movq  %0, %%mm0           \n\t"
00326              "movq  %1, %%mm1           \n\t"
00327              PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00328              "movq  %%mm2, %0           \n\t"
00329              :"+m"(*block)
00330              :"m"(*pixels)
00331              :"memory");
00332         pixels += line_size;
00333         block += line_size;
00334     }
00335     while (--h);
00336 }
00337 
00338 static void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00339 {
00340     MOVQ_BFE(mm6);
00341     JUMPALIGN();
00342     do {
00343         __asm __volatile(
00344              "movq  %0, %%mm0           \n\t"
00345              "movq  %1, %%mm1           \n\t"
00346              PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00347              "movq  %%mm2, %0           \n\t"
00348              "movq  8%0, %%mm0          \n\t"
00349              "movq  8%1, %%mm1          \n\t"
00350              PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00351              "movq  %%mm2, 8%0          \n\t"
00352              :"+m"(*block)
00353              :"m"(*pixels)
00354              :"memory");
00355         pixels += line_size;
00356         block += line_size;
00357     }
00358     while (--h);
00359 }
00360 
00361 static void DEF(avg, pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00362 {
00363     MOVQ_BFE(mm6);
00364     JUMPALIGN();
00365     do {
00366         __asm __volatile(
00367             "movq  %1, %%mm0            \n\t"
00368             "movq  1%1, %%mm1           \n\t"
00369             "movq  %0, %%mm3            \n\t"
00370             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00371             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00372             "movq  %%mm0, %0            \n\t"
00373             :"+m"(*block)
00374             :"m"(*pixels)
00375             :"memory");
00376         pixels += line_size;
00377         block += line_size;
00378     } while (--h);
00379 }
00380 
00381 static __attribute__((unused)) void DEF(avg, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
00382 {
00383     MOVQ_BFE(mm6);
00384     JUMPALIGN();
00385     do {
00386         __asm __volatile(
00387             "movq  %1, %%mm0            \n\t"
00388             "movq  %2, %%mm1            \n\t"
00389             "movq  %0, %%mm3            \n\t"
00390             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00391             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00392             "movq  %%mm0, %0            \n\t"
00393             :"+m"(*dst)
00394             :"m"(*src1), "m"(*src2)
00395             :"memory");
00396         dst += dstStride;
00397         src1 += src1Stride;
00398         src2 += 8;
00399     } while (--h);
00400 }
00401 
00402 static void DEF(avg, pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00403 {
00404     MOVQ_BFE(mm6);
00405     JUMPALIGN();
00406     do {
00407         __asm __volatile(
00408             "movq  %1, %%mm0            \n\t"
00409             "movq  1%1, %%mm1           \n\t"
00410             "movq  %0, %%mm3            \n\t"
00411             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00412             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00413             "movq  %%mm0, %0            \n\t"
00414             "movq  8%1, %%mm0           \n\t"
00415             "movq  9%1, %%mm1           \n\t"
00416             "movq  8%0, %%mm3           \n\t"
00417             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00418             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00419             "movq  %%mm0, 8%0           \n\t"
00420             :"+m"(*block)
00421             :"m"(*pixels)
00422             :"memory");
00423         pixels += line_size;
00424         block += line_size;
00425     } while (--h);
00426 }
00427 
00428 static __attribute__((unused)) void DEF(avg, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
00429 {
00430     MOVQ_BFE(mm6);
00431     JUMPALIGN();
00432     do {
00433         __asm __volatile(
00434             "movq  %1, %%mm0            \n\t"
00435             "movq  %2, %%mm1            \n\t"
00436             "movq  %0, %%mm3            \n\t"
00437             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00438             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00439             "movq  %%mm0, %0            \n\t"
00440             "movq  8%1, %%mm0           \n\t"
00441             "movq  8%2, %%mm1           \n\t"
00442             "movq  8%0, %%mm3           \n\t"
00443             PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
00444             PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
00445             "movq  %%mm0, 8%0           \n\t"
00446             :"+m"(*dst)
00447             :"m"(*src1), "m"(*src2)
00448             :"memory");
00449         dst += dstStride;
00450         src1 += src1Stride;
00451         src2 += 16;
00452     } while (--h);
00453 }
00454 
00455 static void DEF(avg, pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00456 {
00457     MOVQ_BFE(mm6);
00458     __asm __volatile(
00459         "lea    (%3, %3), %%"REG_a"     \n\t"
00460         "movq   (%1), %%mm0             \n\t"
00461         ".balign 8                      \n\t"
00462         "1:                             \n\t"
00463         "movq   (%1, %3), %%mm1         \n\t"
00464         "movq   (%1, %%"REG_a"), %%mm2  \n\t"
00465         PAVGBP(%%mm1, %%mm0, %%mm4,   %%mm2, %%mm1, %%mm5)
00466         "movq   (%2), %%mm3             \n\t"
00467         PAVGB(%%mm3, %%mm4, %%mm0, %%mm6)
00468         "movq   (%2, %3), %%mm3         \n\t"
00469         PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
00470         "movq   %%mm0, (%2)             \n\t"
00471         "movq   %%mm1, (%2, %3)         \n\t"
00472         "add    %%"REG_a", %1           \n\t"
00473         "add    %%"REG_a", %2           \n\t"
00474 
00475         "movq   (%1, %3), %%mm1         \n\t"
00476         "movq   (%1, %%"REG_a"), %%mm0  \n\t"
00477         PAVGBP(%%mm1, %%mm2, %%mm4,   %%mm0, %%mm1, %%mm5)
00478         "movq   (%2), %%mm3             \n\t"
00479         PAVGB(%%mm3, %%mm4, %%mm2, %%mm6)
00480         "movq   (%2, %3), %%mm3         \n\t"
00481         PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
00482         "movq   %%mm2, (%2)             \n\t"
00483         "movq   %%mm1, (%2, %3)         \n\t"
00484         "add    %%"REG_a", %1           \n\t"
00485         "add    %%"REG_a", %2           \n\t"
00486 
00487         "subl   $4, %0                  \n\t"
00488         "jnz    1b                      \n\t"
00489         :"+g"(h), "+S"(pixels), "+D"(block)
00490         :"r"((long)line_size)
00491         :REG_a, "memory");
00492 }
00493 
00494 // this routine is 'slightly' suboptimal but mostly unused
00495 static void DEF(avg, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
00496 {
00497     MOVQ_ZERO(mm7);
00498     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
00499     __asm __volatile(
00500         "movq   (%1), %%mm0             \n\t"
00501         "movq   1(%1), %%mm4            \n\t"
00502         "movq   %%mm0, %%mm1            \n\t"
00503         "movq   %%mm4, %%mm5            \n\t"
00504         "punpcklbw %%mm7, %%mm0         \n\t"
00505         "punpcklbw %%mm7, %%mm4         \n\t"
00506         "punpckhbw %%mm7, %%mm1         \n\t"
00507         "punpckhbw %%mm7, %%mm5         \n\t"
00508         "paddusw %%mm0, %%mm4           \n\t"
00509         "paddusw %%mm1, %%mm5           \n\t"
00510         "xor    %%"REG_a", %%"REG_a"    \n\t"
00511         "add    %3, %1                  \n\t"
00512         ".balign 8                      \n\t"
00513         "1:                             \n\t"
00514         "movq   (%1, %%"REG_a"), %%mm0  \n\t"
00515         "movq   1(%1, %%"REG_a"), %%mm2 \n\t"
00516         "movq   %%mm0, %%mm1            \n\t"
00517         "movq   %%mm2, %%mm3            \n\t"
00518         "punpcklbw %%mm7, %%mm0         \n\t"
00519         "punpcklbw %%mm7, %%mm2         \n\t"
00520         "punpckhbw %%mm7, %%mm1         \n\t"
00521         "punpckhbw %%mm7, %%mm3         \n\t"
00522         "paddusw %%mm2, %%mm0           \n\t"
00523         "paddusw %%mm3, %%mm1           \n\t"
00524         "paddusw %%mm6, %%mm4           \n\t"
00525         "paddusw %%mm6, %%mm5           \n\t"
00526         "paddusw %%mm0, %%mm4           \n\t"
00527         "paddusw %%mm1, %%mm5           \n\t"
00528         "psrlw  $2, %%mm4               \n\t"
00529         "psrlw  $2, %%mm5               \n\t"
00530                 "movq   (%2, %%"REG_a"), %%mm3  \n\t"
00531         "packuswb  %%mm5, %%mm4         \n\t"
00532                 "pcmpeqd %%mm2, %%mm2   \n\t"
00533                 "paddb %%mm2, %%mm2     \n\t"
00534                 PAVGB(%%mm3, %%mm4, %%mm5, %%mm2)
00535                 "movq   %%mm5, (%2, %%"REG_a")  \n\t"
00536         "add    %3, %%"REG_a"           \n\t"
00537 
00538         "movq   (%1, %%"REG_a"), %%mm2  \n\t" // 0 <-> 2   1 <-> 3
00539         "movq   1(%1, %%"REG_a"), %%mm4 \n\t"
00540         "movq   %%mm2, %%mm3            \n\t"
00541         "movq   %%mm4, %%mm5            \n\t"
00542         "punpcklbw %%mm7, %%mm2         \n\t"
00543         "punpcklbw %%mm7, %%mm4         \n\t"
00544         "punpckhbw %%mm7, %%mm3         \n\t"
00545         "punpckhbw %%mm7, %%mm5         \n\t"
00546         "paddusw %%mm2, %%mm4           \n\t"
00547         "paddusw %%mm3, %%mm5           \n\t"
00548         "paddusw %%mm6, %%mm0           \n\t"
00549         "paddusw %%mm6, %%mm1           \n\t"
00550         "paddusw %%mm4, %%mm0           \n\t"
00551         "paddusw %%mm5, %%mm1           \n\t"
00552         "psrlw  $2, %%mm0               \n\t"
00553         "psrlw  $2, %%mm1               \n\t"
00554                 "movq   (%2, %%"REG_a"), %%mm3  \n\t"
00555         "packuswb  %%mm1, %%mm0         \n\t"
00556                 "pcmpeqd %%mm2, %%mm2   \n\t"
00557                 "paddb %%mm2, %%mm2     \n\t"
00558                 PAVGB(%%mm3, %%mm0, %%mm1, %%mm2)
00559                 "movq   %%mm1, (%2, %%"REG_a")  \n\t"
00560         "add    %3, %%"REG_a"           \n\t"
00561 
00562         "subl   $2, %0                  \n\t"
00563         "jnz    1b                      \n\t"
00564         :"+g"(h), "+S"(pixels)
00565         :"D"(block), "r"((long)line_size)
00566         :REG_a, "memory");
00567 }
00568 
00569 //FIXME optimize
00570 static void DEF(put, pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
00571     DEF(put, pixels8_y2)(block  , pixels  , line_size, h);
00572     DEF(put, pixels8_y2)(block+8, pixels+8, line_size, h);
00573 }
00574 
00575 static void DEF(put, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
00576     DEF(put, pixels8_xy2)(block  , pixels  , line_size, h);
00577     DEF(put, pixels8_xy2)(block+8, pixels+8, line_size, h);
00578 }
00579 
00580 static void DEF(avg, pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
00581     DEF(avg, pixels8_y2)(block  , pixels  , line_size, h);
00582     DEF(avg, pixels8_y2)(block+8, pixels+8, line_size, h);
00583 }
00584 
00585 static void DEF(avg, pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){
00586     DEF(avg, pixels8_xy2)(block  , pixels  , line_size, h);
00587     DEF(avg, pixels8_xy2)(block+8, pixels+8, line_size, h);
00588 }
00589 
00590
hvirtual/quicktime/ffmpeg/libavcodec/i386/dsputil_mmx_rnd.h
