FFmpeg: libswscale/ppc/swscale_altivec.c Source File

00001 /*
00002  * AltiVec-enhanced yuv2yuvX
00003  *
00004  * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
00005  * based on the equivalent C code in swscale.c
00006  *
00007  * This file is part of FFmpeg.
00008  *
00009  * FFmpeg is free software; you can redistribute it and/or
00010  * modify it under the terms of the GNU Lesser General Public
00011  * License as published by the Free Software Foundation; either
00012  * version 2.1 of the License, or (at your option) any later version.
00013  *
00014  * FFmpeg is distributed in the hope that it will be useful,
00015  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00016  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00017  * Lesser General Public License for more details.
00018  *
00019  * You should have received a copy of the GNU Lesser General Public
00020  * License along with FFmpeg; if not, write to the Free Software
00021  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00022  */
00023 
00024 #include <inttypes.h>
00025 
00026 #include "config.h"
00027 #include "libswscale/swscale.h"
00028 #include "libswscale/swscale_internal.h"
00029 #include "libavutil/attributes.h"
00030 #include "libavutil/cpu.h"
00031 #include "yuv2rgb_altivec.h"
00032 
00033 #define vzero vec_splat_s32(0)
00034 
00035 #define yuv2planeX_8(d1, d2, l1, src, x, perm, filter) do {     \
00036         vector signed short l2  = vec_ld(((x) << 1) + 16, src); \
00037         vector signed short ls  = vec_perm(l1, l2, perm);       \
00038         vector signed int   i1  = vec_mule(filter, ls);         \
00039         vector signed int   i2  = vec_mulo(filter, ls);         \
00040         vector signed int   vf1 = vec_mergeh(i1, i2);           \
00041         vector signed int   vf2 = vec_mergel(i1, i2);           \
00042         d1 = vec_add(d1, vf1);                                  \
00043         d2 = vec_add(d2, vf2);                                  \
00044         l1 = l2;                                                \
00045     } while (0)
00046 
00047 static void yuv2planeX_16_altivec(const int16_t *filter, int filterSize,
00048                                   const int16_t **src, uint8_t *dest,
00049                                   const uint8_t *dither, int offset, int x)
00050 {
00051     register int i, j;
00052     DECLARE_ALIGNED(16, int, val)[16];
00053     vector signed int vo1, vo2, vo3, vo4;
00054     vector unsigned short vs1, vs2;
00055     vector unsigned char vf;
00056     vector unsigned int altivec_vectorShiftInt19 =
00057         vec_add(vec_splat_u32(10), vec_splat_u32(9));
00058 
00059     for (i = 0; i < 16; i++)
00060         val[i] = dither[(x + i + offset) & 7] << 12;
00061 
00062     vo1 = vec_ld(0,  val);
00063     vo2 = vec_ld(16, val);
00064     vo3 = vec_ld(32, val);
00065     vo4 = vec_ld(48, val);
00066 
00067     for (j = 0; j < filterSize; j++) {
00068         vector signed short l1, vLumFilter = vec_ld(j << 1, filter);
00069         vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter);
00070         vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
00071         vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
00072 
00073         perm = vec_lvsl(x << 1, src[j]);
00074         l1   = vec_ld(x << 1, src[j]);
00075 
00076         yuv2planeX_8(vo1, vo2, l1, src[j], x,     perm, vLumFilter);
00077         yuv2planeX_8(vo3, vo4, l1, src[j], x + 8, perm, vLumFilter);
00078     }
00079 
00080     vo1 = vec_sra(vo1, altivec_vectorShiftInt19);
00081     vo2 = vec_sra(vo2, altivec_vectorShiftInt19);
00082     vo3 = vec_sra(vo3, altivec_vectorShiftInt19);
00083     vo4 = vec_sra(vo4, altivec_vectorShiftInt19);
00084     vs1 = vec_packsu(vo1, vo2);
00085     vs2 = vec_packsu(vo3, vo4);
00086     vf  = vec_packsu(vs1, vs2);
00087     vec_st(vf, 0, dest);
00088 }
00089 
00090 static inline void yuv2planeX_u(const int16_t *filter, int filterSize,
00091                                 const int16_t **src, uint8_t *dest, int dstW,
00092                                 const uint8_t *dither, int offset, int x)
00093 {
00094     int i, j;
00095 
00096     for (i = x; i < dstW; i++) {
00097         int t = dither[(i + offset) & 7] << 12;
00098         for (j = 0; j < filterSize; j++)
00099             t += src[j][i] * filter[j];
00100         dest[i] = av_clip_uint8(t >> 19);
00101     }
00102 }
00103 
00104 static void yuv2planeX_altivec(const int16_t *filter, int filterSize,
00105                                const int16_t **src, uint8_t *dest, int dstW,
00106                                const uint8_t *dither, int offset)
00107 {
00108     int dst_u = -(uintptr_t)dest & 15;
00109     int i;
00110 
00111     yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0);
00112 
00113     for (i = dst_u; i < dstW - 15; i += 16)
00114         yuv2planeX_16_altivec(filter, filterSize, src, dest + i, dither,
00115                               offset, i);
00116 
00117     yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i);
00118 }
00119 
00120 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
00121                                 const uint8_t *src, const int16_t *filter,
00122                                 const int32_t *filterPos, int filterSize)
00123 {
00124     register int i;
00125     DECLARE_ALIGNED(16, int, tempo)[4];
00126 
00127     if (filterSize % 4) {
00128         for (i = 0; i < dstW; i++) {
00129             register int j;
00130             register int srcPos = filterPos[i];
00131             register int val    = 0;
00132             for (j = 0; j < filterSize; j++)
00133                 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
00134             dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
00135         }
00136     } else
00137         switch (filterSize) {
00138         case 4:
00139             for (i = 0; i < dstW; i++) {
00140                 register int srcPos = filterPos[i];
00141 
00142                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00143                 vector unsigned char src_v1, src_vF;
00144                 vector signed short src_v, filter_v;
00145                 vector signed int val_vEven, val_s;
00146                 if ((((uintptr_t)src + srcPos) % 16) > 12) {
00147                     src_v1 = vec_ld(srcPos + 16, src);
00148                 }
00149                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00150 
00151                 src_v = // vec_unpackh sign-extends...
00152                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00153                 // now put our elements in the even slots
00154                 src_v = vec_mergeh(src_v, (vector signed short)vzero);
00155 
00156                 filter_v = vec_ld(i << 3, filter);
00157                 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
00158 
00159                 // The neat trick: We only care for half the elements,
00160                 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
00161                 // and we're going to use vec_mule, so we choose
00162                 // carefully how to "unpack" the elements into the even slots.
00163                 if ((i << 3) % 16)
00164                     filter_v = vec_mergel(filter_v, (vector signed short)vzero);
00165                 else
00166                     filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
00167 
00168                 val_vEven = vec_mule(src_v, filter_v);
00169                 val_s     = vec_sums(val_vEven, vzero);
00170                 vec_st(val_s, 0, tempo);
00171                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00172             }
00173         break;
00174 
00175         case 8:
00176             for (i = 0; i < dstW; i++) {
00177                 register int srcPos = filterPos[i];
00178 
00179                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00180                 vector unsigned char src_v1, src_vF;
00181                 vector signed short src_v, filter_v;
00182                 vector signed int val_v, val_s;
00183                 if ((((uintptr_t)src + srcPos) % 16) > 8) {
00184                     src_v1 = vec_ld(srcPos + 16, src);
00185                 }
00186                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00187 
00188                 src_v = // vec_unpackh sign-extends...
00189                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00190                 filter_v = vec_ld(i << 4, filter);
00191                 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
00192 
00193                 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
00194                 val_s = vec_sums(val_v, vzero);
00195                 vec_st(val_s, 0, tempo);
00196                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00197             }
00198         break;
00199 
00200         case 16:
00201             for (i = 0; i < dstW; i++) {
00202                 register int srcPos = filterPos[i];
00203 
00204                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00205                 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
00206                 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00207 
00208                 vector signed short src_vA = // vec_unpackh sign-extends...
00209                                              (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00210                 vector signed short src_vB = // vec_unpackh sign-extends...
00211                                              (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00212 
00213                 vector signed short filter_v0 = vec_ld(i << 5, filter);
00214                 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
00215                 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
00216 
00217                 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
00218                 vector signed int val_v   = vec_msums(src_vB, filter_v1, val_acc);
00219 
00220                 vector signed int val_s = vec_sums(val_v, vzero);
00221 
00222                 vec_st(val_s, 0, tempo);
00223                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00224             }
00225         break;
00226 
00227         default:
00228             for (i = 0; i < dstW; i++) {
00229                 register int j;
00230                 register int srcPos = filterPos[i];
00231 
00232                 vector signed int val_s, val_v = (vector signed int)vzero;
00233                 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
00234                 vector unsigned char permF     = vec_lvsl((i * 2 * filterSize), filter);
00235 
00236                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00237                 vector unsigned char permS  = vec_lvsl(srcPos, src);
00238 
00239                 for (j = 0; j < filterSize - 15; j += 16) {
00240                     vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
00241                     vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
00242 
00243                     vector signed short src_vA = // vec_unpackh sign-extends...
00244                                                  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00245                     vector signed short src_vB = // vec_unpackh sign-extends...
00246                                                  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00247 
00248                     vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00249                     vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
00250                     vector signed short filter_v0  = vec_perm(filter_v0R, filter_v1R, permF);
00251                     vector signed short filter_v1  = vec_perm(filter_v1R, filter_v2R, permF);
00252 
00253                     vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
00254                     val_v = vec_msums(src_vB, filter_v1, val_acc);
00255 
00256                     filter_v0R = filter_v2R;
00257                     src_v0     = src_v1;
00258                 }
00259 
00260                 if (j < filterSize - 7) {
00261                     // loading src_v0 is useless, it's already done above
00262                     // vector unsigned char src_v0 = vec_ld(srcPos + j, src);
00263                     vector unsigned char src_v1, src_vF;
00264                     vector signed short src_v, filter_v1R, filter_v;
00265                     if ((((uintptr_t)src + srcPos) % 16) > 8) {
00266                         src_v1 = vec_ld(srcPos + j + 16, src);
00267                     }
00268                     src_vF = vec_perm(src_v0, src_v1, permS);
00269 
00270                     src_v = // vec_unpackh sign-extends...
00271                             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00272                     // loading filter_v0R is useless, it's already done above
00273                     // vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
00274                     filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00275                     filter_v   = vec_perm(filter_v0R, filter_v1R, permF);
00276 
00277                     val_v = vec_msums(src_v, filter_v, val_v);
00278                 }
00279 
00280                 val_s = vec_sums(val_v, vzero);
00281 
00282                 vec_st(val_s, 0, tempo);
00283                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00284             }
00285         }
00286 }
00287 
00288 av_cold void ff_sws_init_swScale_altivec(SwsContext *c)
00289 {
00290     enum AVPixelFormat dstFormat = c->dstFormat;
00291 
00292     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
00293         return;
00294 
00295     if (c->srcBpc == 8 && c->dstBpc <= 14) {
00296         c->hyScale = c->hcScale = hScale_altivec_real;
00297     }
00298     if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
00299         dstFormat != AV_PIX_FMT_NV12 && dstFormat != AV_PIX_FMT_NV21 &&
00300         !c->alpPixBuf) {
00301         c->yuv2planeX = yuv2planeX_altivec;
00302     }
00303 
00304     /* The following list of supported dstFormat values should
00305      * match what's found in the body of ff_yuv2packedX_altivec() */
00306     if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
00307         switch (c->dstFormat) {
00308         case AV_PIX_FMT_ABGR:
00309             c->yuv2packedX = ff_yuv2abgr_X_altivec;
00310             break;
00311         case AV_PIX_FMT_BGRA:
00312             c->yuv2packedX = ff_yuv2bgra_X_altivec;
00313             break;
00314         case AV_PIX_FMT_ARGB:
00315             c->yuv2packedX = ff_yuv2argb_X_altivec;
00316             break;
00317         case AV_PIX_FMT_RGBA:
00318             c->yuv2packedX = ff_yuv2rgba_X_altivec;
00319             break;
00320         case AV_PIX_FMT_BGR24:
00321             c->yuv2packedX = ff_yuv2bgr24_X_altivec;
00322             break;
00323         case AV_PIX_FMT_RGB24:
00324             c->yuv2packedX = ff_yuv2rgb24_X_altivec;
00325             break;
00326         }
00327     }
00328 }