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 #include "config.h"
00026 #include "libswscale/swscale.h"
00027 #include "libswscale/swscale_internal.h"
00028 #include "libavutil/cpu.h"
00029 #include "yuv2rgb_altivec.h"
00030 
00031 #define vzero vec_splat_s32(0)
00032 
00033 static inline void
00034 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW)
00035 {
00036     register int i;
00037     vector unsigned int altivec_vectorShiftInt19 =
00038         vec_add(vec_splat_u32(10), vec_splat_u32(9));
00039     if ((uintptr_t)dest % 16) {
00040         /* badly aligned store, we force store alignment */
00041         /* and will handle load misalignment on val w/ vec_perm */
00042         vector unsigned char perm1;
00043         vector signed int v1;
00044         for (i = 0 ; (i < dstW) &&
00045             (((uintptr_t)dest + i) % 16) ; i++) {
00046                 int t = val[i] >> 19;
00047                 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00048         }
00049         perm1 = vec_lvsl(i << 2, val);
00050         v1 = vec_ld(i << 2, val);
00051         for ( ; i < (dstW - 15); i+=16) {
00052             int offset = i << 2;
00053             vector signed int v2 = vec_ld(offset + 16, val);
00054             vector signed int v3 = vec_ld(offset + 32, val);
00055             vector signed int v4 = vec_ld(offset + 48, val);
00056             vector signed int v5 = vec_ld(offset + 64, val);
00057             vector signed int v12 = vec_perm(v1, v2, perm1);
00058             vector signed int v23 = vec_perm(v2, v3, perm1);
00059             vector signed int v34 = vec_perm(v3, v4, perm1);
00060             vector signed int v45 = vec_perm(v4, v5, perm1);
00061 
00062             vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
00063             vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
00064             vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
00065             vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
00066             vector unsigned short vs1 = vec_packsu(vA, vB);
00067             vector unsigned short vs2 = vec_packsu(vC, vD);
00068             vector unsigned char vf = vec_packsu(vs1, vs2);
00069             vec_st(vf, i, dest);
00070             v1 = v5;
00071         }
00072     } else { // dest is properly aligned, great
00073         for (i = 0; i < (dstW - 15); i+=16) {
00074             int offset = i << 2;
00075             vector signed int v1 = vec_ld(offset, val);
00076             vector signed int v2 = vec_ld(offset + 16, val);
00077             vector signed int v3 = vec_ld(offset + 32, val);
00078             vector signed int v4 = vec_ld(offset + 48, val);
00079             vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
00080             vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
00081             vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
00082             vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
00083             vector unsigned short vs1 = vec_packsu(v5, v6);
00084             vector unsigned short vs2 = vec_packsu(v7, v8);
00085             vector unsigned char vf = vec_packsu(vs1, vs2);
00086             vec_st(vf, i, dest);
00087         }
00088     }
00089     for ( ; i < dstW ; i++) {
00090         int t = val[i] >> 19;
00091         dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00092     }
00093 }
00094 
00095 //FIXME remove the usage of scratch buffers.
00096 static void
00097 yuv2planeX_altivec(const int16_t *filter, int filterSize,
00098                    const int16_t **src, uint8_t *dest, int dstW,
00099                    const uint8_t *dither, int offset)
00100 {
00101     register int i, j;
00102     {
00103         DECLARE_ALIGNED(16, int, val)[dstW];
00104 
00105         for (i=0; i<dstW; i++)
00106             val[i] = dither[(i + offset) & 7] << 12;
00107 
00108         for (j = 0; j < filterSize; j++) {
00109             vector signed short l1, vLumFilter = vec_ld(j << 1, filter);
00110             vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter);
00111             vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
00112             vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
00113 
00114             perm = vec_lvsl(0, src[j]);
00115             l1 = vec_ld(0, src[j]);
00116 
00117             for (i = 0; i < (dstW - 7); i+=8) {
00118                 int offset = i << 2;
00119                 vector signed short l2 = vec_ld((i << 1) + 16, src[j]);
00120 
00121                 vector signed int v1 = vec_ld(offset, val);
00122                 vector signed int v2 = vec_ld(offset + 16, val);
00123 
00124                 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
00125 
00126                 vector signed int i1 = vec_mule(vLumFilter, ls);
00127                 vector signed int i2 = vec_mulo(vLumFilter, ls);
00128 
00129                 vector signed int vf1 = vec_mergeh(i1, i2);
00130                 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
00131 
00132                 vector signed int vo1 = vec_add(v1, vf1);
00133                 vector signed int vo2 = vec_add(v2, vf2);
00134 
00135                 vec_st(vo1, offset, val);
00136                 vec_st(vo2, offset + 16, val);
00137 
00138                 l1 = l2;
00139             }
00140             for ( ; i < dstW; i++) {
00141                 val[i] += src[j][i] * filter[j];
00142             }
00143         }
00144         altivec_packIntArrayToCharArray(val, dest, dstW);
00145     }
00146 }
00147 
00148 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
00149                                 const uint8_t *src, const int16_t *filter,
00150                                 const int16_t *filterPos, int filterSize)
00151 {
00152     register int i;
00153     DECLARE_ALIGNED(16, int, tempo)[4];
00154 
00155     if (filterSize % 4) {
00156         for (i=0; i<dstW; i++) {
00157             register int j;
00158             register int srcPos = filterPos[i];
00159             register int val = 0;
00160             for (j=0; j<filterSize; j++) {
00161                 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
00162             }
00163             dst[i] = FFMIN(val>>7, (1<<15)-1);
00164         }
00165     }
00166     else
00167     switch (filterSize) {
00168     case 4:
00169     {
00170     for (i=0; i<dstW; i++) {
00171         register int srcPos = filterPos[i];
00172 
00173         vector unsigned char src_v0 = vec_ld(srcPos, src);
00174         vector unsigned char src_v1, src_vF;
00175         vector signed short src_v, filter_v;
00176         vector signed int val_vEven, val_s;
00177         if ((((uintptr_t)src + srcPos) % 16) > 12) {
00178             src_v1 = vec_ld(srcPos + 16, src);
00179         }
00180         src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00181 
00182         src_v = // vec_unpackh sign-extends...
00183             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00184         // now put our elements in the even slots
00185         src_v = vec_mergeh(src_v, (vector signed short)vzero);
00186 
00187         filter_v = vec_ld(i << 3, filter);
00188         // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
00189 
00190         // The neat trick: We only care for half the elements,
00191         // high or low depending on (i<<3)%16 (it's 0 or 8 here),
00192         // and we're going to use vec_mule, so we choose
00193         // carefully how to "unpack" the elements into the even slots.
00194         if ((i << 3) % 16)
00195             filter_v = vec_mergel(filter_v, (vector signed short)vzero);
00196         else
00197             filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
00198 
00199         val_vEven = vec_mule(src_v, filter_v);
00200         val_s = vec_sums(val_vEven, vzero);
00201         vec_st(val_s, 0, tempo);
00202         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
00203     }
00204     }
00205     break;
00206 
00207     case 8:
00208     {
00209     for (i=0; i<dstW; i++) {
00210         register int srcPos = filterPos[i];
00211 
00212         vector unsigned char src_v0 = vec_ld(srcPos, src);
00213         vector unsigned char src_v1, src_vF;
00214         vector signed short src_v, filter_v;
00215         vector signed int val_v, val_s;
00216         if ((((uintptr_t)src + srcPos) % 16) > 8) {
00217             src_v1 = vec_ld(srcPos + 16, src);
00218         }
00219         src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00220 
00221         src_v = // vec_unpackh sign-extends...
00222             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00223         filter_v = vec_ld(i << 4, filter);
00224         // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
00225 
00226         val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
00227         val_s = vec_sums(val_v, vzero);
00228         vec_st(val_s, 0, tempo);
00229         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
00230     }
00231     }
00232     break;
00233 
00234     case 16:
00235     {
00236         for (i=0; i<dstW; i++) {
00237             register int srcPos = filterPos[i];
00238 
00239             vector unsigned char src_v0 = vec_ld(srcPos, src);
00240             vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
00241             vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
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_v0 = vec_ld(i << 5, filter);
00249             vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
00250             // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
00251 
00252             vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
00253             vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
00254 
00255             vector signed int val_s = vec_sums(val_v, vzero);
00256 
00257             vec_st(val_s, 0, tempo);
00258             dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
00259         }
00260     }
00261     break;
00262 
00263     default:
00264     {
00265     for (i=0; i<dstW; i++) {
00266         register int j;
00267         register int srcPos = filterPos[i];
00268 
00269         vector signed int val_s, val_v = (vector signed int)vzero;
00270         vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
00271         vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
00272 
00273         vector unsigned char src_v0 = vec_ld(srcPos, src);
00274         vector unsigned char permS = vec_lvsl(srcPos, src);
00275 
00276         for (j = 0 ; j < filterSize - 15; j += 16) {
00277             vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
00278             vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
00279 
00280             vector signed short src_vA = // vec_unpackh sign-extends...
00281                 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00282             vector signed short src_vB = // vec_unpackh sign-extends...
00283                 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00284 
00285             vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00286             vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
00287             vector signed short filter_v0  = vec_perm(filter_v0R, filter_v1R, permF);
00288             vector signed short filter_v1  = vec_perm(filter_v1R, filter_v2R, permF);
00289 
00290             vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
00291             val_v = vec_msums(src_vB, filter_v1, val_acc);
00292 
00293             filter_v0R = filter_v2R;
00294             src_v0 = src_v1;
00295         }
00296 
00297         if (j < filterSize-7) {
00298             // loading src_v0 is useless, it's already done above
00299             //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
00300             vector unsigned char src_v1, src_vF;
00301             vector signed short src_v, filter_v1R, filter_v;
00302             if ((((uintptr_t)src + srcPos) % 16) > 8) {
00303                 src_v1 = vec_ld(srcPos + j + 16, src);
00304             }
00305             src_vF = vec_perm(src_v0, src_v1, permS);
00306 
00307             src_v = // vec_unpackh sign-extends...
00308                 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00309             // loading filter_v0R is useless, it's already done above
00310             //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
00311             filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00312             filter_v = vec_perm(filter_v0R, filter_v1R, permF);
00313 
00314             val_v = vec_msums(src_v, filter_v, val_v);
00315         }
00316 
00317         val_s = vec_sums(val_v, vzero);
00318 
00319         vec_st(val_s, 0, tempo);
00320         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
00321     }
00322 
00323     }
00324     }
00325 }
00326 
00327 void ff_sws_init_swScale_altivec(SwsContext *c)
00328 {
00329     enum PixelFormat dstFormat = c->dstFormat;
00330 
00331     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
00332         return;
00333 
00334     if (c->srcBpc == 8 && c->dstBpc <= 10) {
00335         c->hyScale = c->hcScale = hScale_altivec_real;
00336     }
00337     if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
00338         dstFormat != PIX_FMT_NV12 && dstFormat != PIX_FMT_NV21 &&
00339         !c->alpPixBuf) {
00340         c->yuv2planeX = yuv2planeX_altivec;
00341     }
00342 
00343     /* The following list of supported dstFormat values should
00344      * match what's found in the body of ff_yuv2packedX_altivec() */
00345     if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
00346         switch (c->dstFormat) {
00347         case PIX_FMT_ABGR:  c->yuv2packedX = ff_yuv2abgr_X_altivec;  break;
00348         case PIX_FMT_BGRA:  c->yuv2packedX = ff_yuv2bgra_X_altivec;  break;
00349         case PIX_FMT_ARGB:  c->yuv2packedX = ff_yuv2argb_X_altivec;  break;
00350         case PIX_FMT_RGBA:  c->yuv2packedX = ff_yuv2rgba_X_altivec;  break;
00351         case PIX_FMT_BGR24: c->yuv2packedX = ff_yuv2bgr24_X_altivec; break;
00352         case PIX_FMT_RGB24: c->yuv2packedX = ff_yuv2rgb24_X_altivec; break;
00353         }
00354     }
00355 }