libavcodec/ppc/fdct_altivec.c
Go to the documentation of this file.
00001 /*
00002  * Copyright (C) 2003  James Klicman <james@klicman.org>
00003  *
00004  * This file is part of Libav.
00005  *
00006  * Libav 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 2.1 of the License, or (at your option) any later version.
00010  *
00011  * Libav 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 Libav; if not, write to the Free Software
00018  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00019  */
00020 
00021 #include "config.h"
00022 #if HAVE_ALTIVEC_H
00023 #include <altivec.h>
00024 #endif
00025 #include "libavutil/common.h"
00026 #include "libavcodec/dsputil.h"
00027 #include "dsputil_altivec.h"
00028 
00029 #define vs16(v) ((vector signed short)(v))
00030 #define vs32(v) ((vector signed int)(v))
00031 #define vu8(v)  ((vector unsigned char)(v))
00032 #define vu16(v) ((vector unsigned short)(v))
00033 #define vu32(v) ((vector unsigned int)(v))
00034 
00035 
00036 #define C1     0.98078525066375732421875000 /* cos(1*PI/16) */
00037 #define C2     0.92387950420379638671875000 /* cos(2*PI/16) */
00038 #define C3     0.83146959543228149414062500 /* cos(3*PI/16) */
00039 #define C4     0.70710676908493041992187500 /* cos(4*PI/16) */
00040 #define C5     0.55557024478912353515625000 /* cos(5*PI/16) */
00041 #define C6     0.38268342614173889160156250 /* cos(6*PI/16) */
00042 #define C7     0.19509032368659973144531250 /* cos(7*PI/16) */
00043 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2)      */
00044 
00045 
00046 #define W0 -(2 * C2)
00047 #define W1 (2 * C6)
00048 #define W2 (SQRT_2 * C6)
00049 #define W3 (SQRT_2 * C3)
00050 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
00051 #define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
00052 #define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
00053 #define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
00054 #define W8 (SQRT_2 * ( C7 - C3))
00055 #define W9 (SQRT_2 * (-C1 - C3))
00056 #define WA (SQRT_2 * (-C3 - C5))
00057 #define WB (SQRT_2 * ( C5 - C3))
00058 
00059 
00060 static vector float fdctconsts[3] = {
00061     { W0, W1, W2, W3 },
00062     { W4, W5, W6, W7 },
00063     { W8, W9, WA, WB }
00064 };
00065 
00066 #define LD_W0 vec_splat(cnsts0, 0)
00067 #define LD_W1 vec_splat(cnsts0, 1)
00068 #define LD_W2 vec_splat(cnsts0, 2)
00069 #define LD_W3 vec_splat(cnsts0, 3)
00070 #define LD_W4 vec_splat(cnsts1, 0)
00071 #define LD_W5 vec_splat(cnsts1, 1)
00072 #define LD_W6 vec_splat(cnsts1, 2)
00073 #define LD_W7 vec_splat(cnsts1, 3)
00074 #define LD_W8 vec_splat(cnsts2, 0)
00075 #define LD_W9 vec_splat(cnsts2, 1)
00076 #define LD_WA vec_splat(cnsts2, 2)
00077 #define LD_WB vec_splat(cnsts2, 3)
00078 
00079 
00080 #define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */                  \
00081     x0 = vec_add(b0, b7);               /* x0 = b0 + b7; */         \
00082     x7 = vec_sub(b0, b7);               /* x7 = b0 - b7; */         \
00083     x1 = vec_add(b1, b6);               /* x1 = b1 + b6; */         \
00084     x6 = vec_sub(b1, b6);               /* x6 = b1 - b6; */         \
00085     x2 = vec_add(b2, b5);               /* x2 = b2 + b5; */         \
00086     x5 = vec_sub(b2, b5);               /* x5 = b2 - b5; */         \
00087     x3 = vec_add(b3, b4);               /* x3 = b3 + b4; */         \
00088     x4 = vec_sub(b3, b4);               /* x4 = b3 - b4; */         \
00089                                                                     \
00090     b7 = vec_add(x0, x3);               /* b7 = x0 + x3; */         \
00091     b1 = vec_add(x1, x2);               /* b1 = x1 + x2; */         \
00092     b0 = vec_add(b7, b1);               /* b0 = b7 + b1; */         \
00093     b4 = vec_sub(b7, b1);               /* b4 = b7 - b1; */         \
00094                                                                     \
00095     b2 = vec_sub(x0, x3);               /* b2 = x0 - x3; */         \
00096     b6 = vec_sub(x1, x2);               /* b6 = x1 - x2; */         \
00097     b5 = vec_add(b6, b2);               /* b5 = b6 + b2; */         \
00098     cnst = LD_W2;                                                   \
00099     b5 = vec_madd(cnst, b5, mzero);     /* b5 = b5 * W2; */         \
00100     cnst = LD_W1;                                                   \
00101     b2 = vec_madd(cnst, b2, b5);        /* b2 = b5 + b2 * W1; */    \
00102     cnst = LD_W0;                                                   \
00103     b6 = vec_madd(cnst, b6, b5);        /* b6 = b5 + b6 * W0; */    \
00104                                                                     \
00105     x0 = vec_add(x4, x7);               /* x0 = x4 + x7; */         \
00106     x1 = vec_add(x5, x6);               /* x1 = x5 + x6; */         \
00107     x2 = vec_add(x4, x6);               /* x2 = x4 + x6; */         \
00108     x3 = vec_add(x5, x7);               /* x3 = x5 + x7; */         \
00109     x8 = vec_add(x2, x3);               /* x8 = x2 + x3; */         \
00110     cnst = LD_W3;                                                   \
00111     x8 = vec_madd(cnst, x8, mzero);     /* x8 = x8 * W3; */         \
00112                                                                     \
00113     cnst = LD_W8;                                                   \
00114     x0 = vec_madd(cnst, x0, mzero);     /* x0 *= W8; */             \
00115     cnst = LD_W9;                                                   \
00116     x1 = vec_madd(cnst, x1, mzero);     /* x1 *= W9; */             \
00117     cnst = LD_WA;                                                   \
00118     x2 = vec_madd(cnst, x2, x8);        /* x2 = x2 * WA + x8; */    \
00119     cnst = LD_WB;                                                   \
00120     x3 = vec_madd(cnst, x3, x8);        /* x3 = x3 * WB + x8; */    \
00121                                                                     \
00122     cnst = LD_W4;                                                   \
00123     b7 = vec_madd(cnst, x4, x0);        /* b7 = x4 * W4 + x0; */    \
00124     cnst = LD_W5;                                                   \
00125     b5 = vec_madd(cnst, x5, x1);        /* b5 = x5 * W5 + x1; */    \
00126     cnst = LD_W6;                                                   \
00127     b3 = vec_madd(cnst, x6, x1);        /* b3 = x6 * W6 + x1; */    \
00128     cnst = LD_W7;                                                   \
00129     b1 = vec_madd(cnst, x7, x0);        /* b1 = x7 * W7 + x0; */    \
00130                                                                     \
00131     b7 = vec_add(b7, x2);               /* b7 = b7 + x2; */         \
00132     b5 = vec_add(b5, x3);               /* b5 = b5 + x3; */         \
00133     b3 = vec_add(b3, x2);               /* b3 = b3 + x2; */         \
00134     b1 = vec_add(b1, x3);               /* b1 = b1 + x3; */         \
00135     /* }}} */
00136 
00137 #define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */                  \
00138     x0 = vec_add(b0, b7);               /* x0 = b0 + b7; */         \
00139     x7 = vec_sub(b0, b7);               /* x7 = b0 - b7; */         \
00140     x1 = vec_add(b1, b6);               /* x1 = b1 + b6; */         \
00141     x6 = vec_sub(b1, b6);               /* x6 = b1 - b6; */         \
00142     x2 = vec_add(b2, b5);               /* x2 = b2 + b5; */         \
00143     x5 = vec_sub(b2, b5);               /* x5 = b2 - b5; */         \
00144     x3 = vec_add(b3, b4);               /* x3 = b3 + b4; */         \
00145     x4 = vec_sub(b3, b4);               /* x4 = b3 - b4; */         \
00146                                                                     \
00147     b7 = vec_add(x0, x3);               /* b7 = x0 + x3; */         \
00148     b1 = vec_add(x1, x2);               /* b1 = x1 + x2; */         \
00149     b0 = vec_add(b7, b1);               /* b0 = b7 + b1; */         \
00150     b4 = vec_sub(b7, b1);               /* b4 = b7 - b1; */         \
00151                                                                     \
00152     b2 = vec_sub(x0, x3);               /* b2 = x0 - x3; */         \
00153     b6 = vec_sub(x1, x2);               /* b6 = x1 - x2; */         \
00154     b5 = vec_add(b6, b2);               /* b5 = b6 + b2; */         \
00155     cnst = LD_W2;                                                   \
00156     b5 = vec_madd(cnst, b5, mzero);     /* b5 = b5 * W2; */         \
00157     cnst = LD_W1;                                                   \
00158     b2 = vec_madd(cnst, b2, b5);        /* b2 = b5 + b2 * W1; */    \
00159     cnst = LD_W0;                                                   \
00160     b6 = vec_madd(cnst, b6, b5);        /* b6 = b5 + b6 * W0; */    \
00161                                                                     \
00162     x0 = vec_add(x4, x7);               /* x0 = x4 + x7; */         \
00163     x1 = vec_add(x5, x6);               /* x1 = x5 + x6; */         \
00164     x2 = vec_add(x4, x6);               /* x2 = x4 + x6; */         \
00165     x3 = vec_add(x5, x7);               /* x3 = x5 + x7; */         \
00166     x8 = vec_add(x2, x3);               /* x8 = x2 + x3; */         \
00167     cnst = LD_W3;                                                   \
00168     x8 = vec_madd(cnst, x8, mzero);     /* x8 = x8 * W3; */         \
00169                                                                     \
00170     cnst = LD_W8;                                                   \
00171     x0 = vec_madd(cnst, x0, mzero);     /* x0 *= W8; */             \
00172     cnst = LD_W9;                                                   \
00173     x1 = vec_madd(cnst, x1, mzero);     /* x1 *= W9; */             \
00174     cnst = LD_WA;                                                   \
00175     x2 = vec_madd(cnst, x2, x8);        /* x2 = x2 * WA + x8; */    \
00176     cnst = LD_WB;                                                   \
00177     x3 = vec_madd(cnst, x3, x8);        /* x3 = x3 * WB + x8; */    \
00178                                                                     \
00179     cnst = LD_W4;                                                   \
00180     b7 = vec_madd(cnst, x4, x0);        /* b7 = x4 * W4 + x0; */    \
00181     cnst = LD_W5;                                                   \
00182     b5 = vec_madd(cnst, x5, x1);        /* b5 = x5 * W5 + x1; */    \
00183     cnst = LD_W6;                                                   \
00184     b3 = vec_madd(cnst, x6, x1);        /* b3 = x6 * W6 + x1; */    \
00185     cnst = LD_W7;                                                   \
00186     b1 = vec_madd(cnst, x7, x0);        /* b1 = x7 * W7 + x0; */    \
00187                                                                     \
00188     b7 = vec_add(b7, x2);               /* b7 += x2; */             \
00189     b5 = vec_add(b5, x3);               /* b5 += x3; */             \
00190     b3 = vec_add(b3, x2);               /* b3 += x2; */             \
00191     b1 = vec_add(b1, x3);               /* b1 += x3; */             \
00192     /* }}} */
00193 
00194 
00195 
00196 /* two dimensional discrete cosine transform */
00197 
00198 void fdct_altivec(int16_t *block)
00199 {
00200     vector signed short *bp;
00201     vector float *cp;
00202     vector float b00, b10, b20, b30, b40, b50, b60, b70;
00203     vector float b01, b11, b21, b31, b41, b51, b61, b71;
00204     vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
00205     vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
00206 
00207     /* setup constants {{{ */
00208     /* mzero = -0.0 */
00209     mzero = ((vector float)vec_splat_u32(-1));
00210     mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero)));
00211     cp = fdctconsts;
00212     cnsts0 = vec_ld(0, cp); cp++;
00213     cnsts1 = vec_ld(0, cp); cp++;
00214     cnsts2 = vec_ld(0, cp);
00215     /* }}} */
00216 
00217 
00218     /* 8x8 matrix transpose (vector short[8]) {{{ */
00219 #define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
00220 
00221     bp = (vector signed short*)block;
00222     b00 = ((vector float)vec_ld(0,    bp));
00223     b40 = ((vector float)vec_ld(16*4, bp));
00224     b01 = ((vector float)MERGE_S16(h, b00, b40));
00225     b11 = ((vector float)MERGE_S16(l, b00, b40));
00226     bp++;
00227     b10 = ((vector float)vec_ld(0,    bp));
00228     b50 = ((vector float)vec_ld(16*4, bp));
00229     b21 = ((vector float)MERGE_S16(h, b10, b50));
00230     b31 = ((vector float)MERGE_S16(l, b10, b50));
00231     bp++;
00232     b20 = ((vector float)vec_ld(0,    bp));
00233     b60 = ((vector float)vec_ld(16*4, bp));
00234     b41 = ((vector float)MERGE_S16(h, b20, b60));
00235     b51 = ((vector float)MERGE_S16(l, b20, b60));
00236     bp++;
00237     b30 = ((vector float)vec_ld(0,    bp));
00238     b70 = ((vector float)vec_ld(16*4, bp));
00239     b61 = ((vector float)MERGE_S16(h, b30, b70));
00240     b71 = ((vector float)MERGE_S16(l, b30, b70));
00241 
00242     x0 = ((vector float)MERGE_S16(h, b01, b41));
00243     x1 = ((vector float)MERGE_S16(l, b01, b41));
00244     x2 = ((vector float)MERGE_S16(h, b11, b51));
00245     x3 = ((vector float)MERGE_S16(l, b11, b51));
00246     x4 = ((vector float)MERGE_S16(h, b21, b61));
00247     x5 = ((vector float)MERGE_S16(l, b21, b61));
00248     x6 = ((vector float)MERGE_S16(h, b31, b71));
00249     x7 = ((vector float)MERGE_S16(l, b31, b71));
00250 
00251     b00 = ((vector float)MERGE_S16(h, x0, x4));
00252     b10 = ((vector float)MERGE_S16(l, x0, x4));
00253     b20 = ((vector float)MERGE_S16(h, x1, x5));
00254     b30 = ((vector float)MERGE_S16(l, x1, x5));
00255     b40 = ((vector float)MERGE_S16(h, x2, x6));
00256     b50 = ((vector float)MERGE_S16(l, x2, x6));
00257     b60 = ((vector float)MERGE_S16(h, x3, x7));
00258     b70 = ((vector float)MERGE_S16(l, x3, x7));
00259 
00260 #undef MERGE_S16
00261     /* }}} */
00262 
00263 
00264 /* Some of the initial calculations can be done as vector short before
00265  * conversion to vector float.  The following code section takes advantage
00266  * of this.
00267  */
00268     /* fdct rows {{{ */
00269     x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
00270     x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));
00271     x1 = ((vector float)vec_add(vs16(b10), vs16(b60)));
00272     x6 = ((vector float)vec_sub(vs16(b10), vs16(b60)));
00273     x2 = ((vector float)vec_add(vs16(b20), vs16(b50)));
00274     x5 = ((vector float)vec_sub(vs16(b20), vs16(b50)));
00275     x3 = ((vector float)vec_add(vs16(b30), vs16(b40)));
00276     x4 = ((vector float)vec_sub(vs16(b30), vs16(b40)));
00277 
00278     b70 = ((vector float)vec_add(vs16(x0), vs16(x3)));
00279     b10 = ((vector float)vec_add(vs16(x1), vs16(x2)));
00280 
00281     b00 = ((vector float)vec_add(vs16(b70), vs16(b10)));
00282     b40 = ((vector float)vec_sub(vs16(b70), vs16(b10)));
00283 
00284 #define CTF0(n) \
00285     b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \
00286     b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \
00287     b##n##1 = vec_ctf(vs32(b##n##1), 0); \
00288     b##n##0 = vec_ctf(vs32(b##n##0), 0);
00289 
00290     CTF0(0);
00291     CTF0(4);
00292 
00293     b20 = ((vector float)vec_sub(vs16(x0), vs16(x3)));
00294     b60 = ((vector float)vec_sub(vs16(x1), vs16(x2)));
00295 
00296     CTF0(2);
00297     CTF0(6);
00298 
00299 #undef CTF0
00300 
00301     x0 = vec_add(b60, b20);
00302     x1 = vec_add(b61, b21);
00303 
00304     cnst = LD_W2;
00305     x0 = vec_madd(cnst, x0, mzero);
00306     x1 = vec_madd(cnst, x1, mzero);
00307     cnst = LD_W1;
00308     b20 = vec_madd(cnst, b20, x0);
00309     b21 = vec_madd(cnst, b21, x1);
00310     cnst = LD_W0;
00311     b60 = vec_madd(cnst, b60, x0);
00312     b61 = vec_madd(cnst, b61, x1);
00313 
00314 #define CTFX(x,b) \
00315     b##0 = ((vector float)vec_unpackh(vs16(x))); \
00316     b##1 = ((vector float)vec_unpackl(vs16(x))); \
00317     b##0 = vec_ctf(vs32(b##0), 0); \
00318     b##1 = vec_ctf(vs32(b##1), 0); \
00319 
00320     CTFX(x4, b7);
00321     CTFX(x5, b5);
00322     CTFX(x6, b3);
00323     CTFX(x7, b1);
00324 
00325 #undef CTFX
00326 
00327 
00328     x0 = vec_add(b70, b10);
00329     x1 = vec_add(b50, b30);
00330     x2 = vec_add(b70, b30);
00331     x3 = vec_add(b50, b10);
00332     x8 = vec_add(x2, x3);
00333     cnst = LD_W3;
00334     x8 = vec_madd(cnst, x8, mzero);
00335 
00336     cnst = LD_W8;
00337     x0 = vec_madd(cnst, x0, mzero);
00338     cnst = LD_W9;
00339     x1 = vec_madd(cnst, x1, mzero);
00340     cnst = LD_WA;
00341     x2 = vec_madd(cnst, x2, x8);
00342     cnst = LD_WB;
00343     x3 = vec_madd(cnst, x3, x8);
00344 
00345     cnst = LD_W4;
00346     b70 = vec_madd(cnst, b70, x0);
00347     cnst = LD_W5;
00348     b50 = vec_madd(cnst, b50, x1);
00349     cnst = LD_W6;
00350     b30 = vec_madd(cnst, b30, x1);
00351     cnst = LD_W7;
00352     b10 = vec_madd(cnst, b10, x0);
00353 
00354     b70 = vec_add(b70, x2);
00355     b50 = vec_add(b50, x3);
00356     b30 = vec_add(b30, x2);
00357     b10 = vec_add(b10, x3);
00358 
00359 
00360     x0 = vec_add(b71, b11);
00361     x1 = vec_add(b51, b31);
00362     x2 = vec_add(b71, b31);
00363     x3 = vec_add(b51, b11);
00364     x8 = vec_add(x2, x3);
00365     cnst = LD_W3;
00366     x8 = vec_madd(cnst, x8, mzero);
00367 
00368     cnst = LD_W8;
00369     x0 = vec_madd(cnst, x0, mzero);
00370     cnst = LD_W9;
00371     x1 = vec_madd(cnst, x1, mzero);
00372     cnst = LD_WA;
00373     x2 = vec_madd(cnst, x2, x8);
00374     cnst = LD_WB;
00375     x3 = vec_madd(cnst, x3, x8);
00376 
00377     cnst = LD_W4;
00378     b71 = vec_madd(cnst, b71, x0);
00379     cnst = LD_W5;
00380     b51 = vec_madd(cnst, b51, x1);
00381     cnst = LD_W6;
00382     b31 = vec_madd(cnst, b31, x1);
00383     cnst = LD_W7;
00384     b11 = vec_madd(cnst, b11, x0);
00385 
00386     b71 = vec_add(b71, x2);
00387     b51 = vec_add(b51, x3);
00388     b31 = vec_add(b31, x2);
00389     b11 = vec_add(b11, x3);
00390     /* }}} */
00391 
00392 
00393     /* 8x8 matrix transpose (vector float[8][2]) {{{ */
00394     x0 = vec_mergel(b00, b20);
00395     x1 = vec_mergeh(b00, b20);
00396     x2 = vec_mergel(b10, b30);
00397     x3 = vec_mergeh(b10, b30);
00398 
00399     b00 = vec_mergeh(x1, x3);
00400     b10 = vec_mergel(x1, x3);
00401     b20 = vec_mergeh(x0, x2);
00402     b30 = vec_mergel(x0, x2);
00403 
00404     x4 = vec_mergel(b41, b61);
00405     x5 = vec_mergeh(b41, b61);
00406     x6 = vec_mergel(b51, b71);
00407     x7 = vec_mergeh(b51, b71);
00408 
00409     b41 = vec_mergeh(x5, x7);
00410     b51 = vec_mergel(x5, x7);
00411     b61 = vec_mergeh(x4, x6);
00412     b71 = vec_mergel(x4, x6);
00413 
00414     x0 = vec_mergel(b01, b21);
00415     x1 = vec_mergeh(b01, b21);
00416     x2 = vec_mergel(b11, b31);
00417     x3 = vec_mergeh(b11, b31);
00418 
00419     x4 = vec_mergel(b40, b60);
00420     x5 = vec_mergeh(b40, b60);
00421     x6 = vec_mergel(b50, b70);
00422     x7 = vec_mergeh(b50, b70);
00423 
00424     b40 = vec_mergeh(x1, x3);
00425     b50 = vec_mergel(x1, x3);
00426     b60 = vec_mergeh(x0, x2);
00427     b70 = vec_mergel(x0, x2);
00428 
00429     b01 = vec_mergeh(x5, x7);
00430     b11 = vec_mergel(x5, x7);
00431     b21 = vec_mergeh(x4, x6);
00432     b31 = vec_mergel(x4, x6);
00433     /* }}} */
00434 
00435 
00436     FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
00437     FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
00438 
00439 
00440     /* round, convert back to short {{{ */
00441 #define CTS(n) \
00442     b##n##0 = vec_round(b##n##0); \
00443     b##n##1 = vec_round(b##n##1); \
00444     b##n##0 = ((vector float)vec_cts(b##n##0, 0)); \
00445     b##n##1 = ((vector float)vec_cts(b##n##1, 0)); \
00446     b##n##0 = ((vector float)vec_pack(vs32(b##n##0), vs32(b##n##1))); \
00447     vec_st(vs16(b##n##0), 0, bp);
00448 
00449     bp = (vector signed short*)block;
00450     CTS(0); bp++;
00451     CTS(1); bp++;
00452     CTS(2); bp++;
00453     CTS(3); bp++;
00454     CTS(4); bp++;
00455     CTS(5); bp++;
00456     CTS(6); bp++;
00457     CTS(7);
00458 
00459 #undef CTS
00460     /* }}} */
00461 }
00462 
00463 /* vim:set foldmethod=marker foldlevel=0: */