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/cpu.h"
00030 #include "yuv2rgb_altivec.h"
00031 
00032 #define vzero vec_splat_s32(0)
00033 
00034 static inline void altivec_packIntArrayToCharArray(int *val, uint8_t *dest,
00035                                                    int dstW)
00036 {
00037     register int i;
00038     vector unsigned int altivec_vectorShiftInt19 =
00039         vec_add(vec_splat_u32(10), vec_splat_u32(9));
00040     if ((uintptr_t)dest % 16) {
00041         /* badly aligned store, we force store alignment */
00042         /* and will handle load misalignment on val w/ vec_perm */
00043         vector unsigned char perm1;
00044         vector signed int v1;
00045         for (i = 0; (i < dstW) &&
00046              (((uintptr_t)dest + i) % 16); i++) {
00047             int t = val[i] >> 19;
00048             dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00049         }
00050         perm1 = vec_lvsl(i << 2, val);
00051         v1    = vec_ld(i << 2, val);
00052         for (; i < (dstW - 15); i += 16) {
00053             int offset = i << 2;
00054             vector signed int v2  = vec_ld(offset + 16, val);
00055             vector signed int v3  = vec_ld(offset + 32, val);
00056             vector signed int v4  = vec_ld(offset + 48, val);
00057             vector signed int v5  = vec_ld(offset + 64, val);
00058             vector signed int v12 = vec_perm(v1, v2, perm1);
00059             vector signed int v23 = vec_perm(v2, v3, perm1);
00060             vector signed int v34 = vec_perm(v3, v4, perm1);
00061             vector signed int v45 = vec_perm(v4, v5, perm1);
00062 
00063             vector signed int vA      = vec_sra(v12, altivec_vectorShiftInt19);
00064             vector signed int vB      = vec_sra(v23, altivec_vectorShiftInt19);
00065             vector signed int vC      = vec_sra(v34, altivec_vectorShiftInt19);
00066             vector signed int vD      = vec_sra(v45, altivec_vectorShiftInt19);
00067             vector unsigned short vs1 = vec_packsu(vA, vB);
00068             vector unsigned short vs2 = vec_packsu(vC, vD);
00069             vector unsigned char vf   = vec_packsu(vs1, vs2);
00070             vec_st(vf, i, dest);
00071             v1 = v5;
00072         }
00073     } else { // dest is properly aligned, great
00074         for (i = 0; i < (dstW - 15); i += 16) {
00075             int offset = i << 2;
00076             vector signed int v1      = vec_ld(offset, val);
00077             vector signed int v2      = vec_ld(offset + 16, val);
00078             vector signed int v3      = vec_ld(offset + 32, val);
00079             vector signed int v4      = vec_ld(offset + 48, val);
00080             vector signed int v5      = vec_sra(v1, altivec_vectorShiftInt19);
00081             vector signed int v6      = vec_sra(v2, altivec_vectorShiftInt19);
00082             vector signed int v7      = vec_sra(v3, altivec_vectorShiftInt19);
00083             vector signed int v8      = vec_sra(v4, altivec_vectorShiftInt19);
00084             vector unsigned short vs1 = vec_packsu(v5, v6);
00085             vector unsigned short vs2 = vec_packsu(v7, v8);
00086             vector unsigned char vf   = vec_packsu(vs1, vs2);
00087             vec_st(vf, i, dest);
00088         }
00089     }
00090     for (; i < dstW; i++) {
00091         int t = val[i] >> 19;
00092         dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00093     }
00094 }
00095 
00096 // FIXME remove the usage of scratch buffers.
00097 static void 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     DECLARE_ALIGNED(16, int, val)[dstW];
00103 
00104     for (i = 0; i < dstW; i++)
00105         val[i] = dither[(i + offset) & 7] << 12;
00106 
00107     for (j = 0; j < filterSize; j++) {
00108         vector signed short l1, vLumFilter = vec_ld(j << 1, filter);
00109         vector unsigned char perm, perm0 = vec_lvsl(j << 1, filter);
00110         vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
00111         vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
00112 
00113         perm = vec_lvsl(0, src[j]);
00114         l1   = vec_ld(0, src[j]);
00115 
00116         for (i = 0; i < (dstW - 7); i += 8) {
00117             int offset = i << 2;
00118             vector signed short l2 = vec_ld((i << 1) + 16, src[j]);
00119 
00120             vector signed int v1 = vec_ld(offset, val);
00121             vector signed int v2 = vec_ld(offset + 16, val);
00122 
00123             vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
00124 
00125             vector signed int i1 = vec_mule(vLumFilter, ls);
00126             vector signed int i2 = vec_mulo(vLumFilter, ls);
00127 
00128             vector signed int vf1 = vec_mergeh(i1, i2);
00129             vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
00130 
00131             vector signed int vo1 = vec_add(v1, vf1);
00132             vector signed int vo2 = vec_add(v2, vf2);
00133 
00134             vec_st(vo1, offset, val);
00135             vec_st(vo2, offset + 16, val);
00136 
00137             l1 = l2;
00138         }
00139         for (; i < dstW; i++)
00140             val[i] += src[j][i] * filter[j];
00141     }
00142     altivec_packIntArrayToCharArray(val, dest, dstW);
00143 }
00144 
00145 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
00146                                 const uint8_t *src, const int16_t *filter,
00147                                 const int32_t *filterPos, int filterSize)
00148 {
00149     register int i;
00150     DECLARE_ALIGNED(16, int, tempo)[4];
00151 
00152     if (filterSize % 4) {
00153         for (i = 0; i < dstW; i++) {
00154             register int j;
00155             register int srcPos = filterPos[i];
00156             register int val    = 0;
00157             for (j = 0; j < filterSize; j++)
00158                 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
00159             dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
00160         }
00161     } else
00162         switch (filterSize) {
00163         case 4:
00164             for (i = 0; i < dstW; i++) {
00165                 register int srcPos = filterPos[i];
00166 
00167                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00168                 vector unsigned char src_v1, src_vF;
00169                 vector signed short src_v, filter_v;
00170                 vector signed int val_vEven, val_s;
00171                 if ((((uintptr_t)src + srcPos) % 16) > 12) {
00172                     src_v1 = vec_ld(srcPos + 16, src);
00173                 }
00174                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00175 
00176                 src_v = // vec_unpackh sign-extends...
00177                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00178                 // now put our elements in the even slots
00179                 src_v = vec_mergeh(src_v, (vector signed short)vzero);
00180 
00181                 filter_v = vec_ld(i << 3, filter);
00182                 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
00183 
00184                 // The neat trick: We only care for half the elements,
00185                 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
00186                 // and we're going to use vec_mule, so we choose
00187                 // carefully how to "unpack" the elements into the even slots.
00188                 if ((i << 3) % 16)
00189                     filter_v = vec_mergel(filter_v, (vector signed short)vzero);
00190                 else
00191                     filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
00192 
00193                 val_vEven = vec_mule(src_v, filter_v);
00194                 val_s     = vec_sums(val_vEven, vzero);
00195                 vec_st(val_s, 0, tempo);
00196                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00197             }
00198         break;
00199 
00200         case 8:
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, src_vF;
00206                 vector signed short src_v, filter_v;
00207                 vector signed int val_v, val_s;
00208                 if ((((uintptr_t)src + srcPos) % 16) > 8) {
00209                     src_v1 = vec_ld(srcPos + 16, src);
00210                 }
00211                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00212 
00213                 src_v = // vec_unpackh sign-extends...
00214                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00215                 filter_v = vec_ld(i << 4, filter);
00216                 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
00217 
00218                 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
00219                 val_s = vec_sums(val_v, vzero);
00220                 vec_st(val_s, 0, tempo);
00221                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00222             }
00223         break;
00224 
00225         case 16:
00226             for (i = 0; i < dstW; i++) {
00227                 register int srcPos = filterPos[i];
00228 
00229                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00230                 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
00231                 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00232 
00233                 vector signed short src_vA = // vec_unpackh sign-extends...
00234                                              (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00235                 vector signed short src_vB = // vec_unpackh sign-extends...
00236                                              (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00237 
00238                 vector signed short filter_v0 = vec_ld(i << 5, filter);
00239                 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
00240                 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
00241 
00242                 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
00243                 vector signed int val_v   = vec_msums(src_vB, filter_v1, val_acc);
00244 
00245                 vector signed int val_s = vec_sums(val_v, vzero);
00246 
00247                 vec_st(val_s, 0, tempo);
00248                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00249             }
00250         break;
00251 
00252         default:
00253             for (i = 0; i < dstW; i++) {
00254                 register int j;
00255                 register int srcPos = filterPos[i];
00256 
00257                 vector signed int val_s, val_v = (vector signed int)vzero;
00258                 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
00259                 vector unsigned char permF     = vec_lvsl((i * 2 * filterSize), filter);
00260 
00261                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00262                 vector unsigned char permS  = vec_lvsl(srcPos, src);
00263 
00264                 for (j = 0; j < filterSize - 15; j += 16) {
00265                     vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
00266                     vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
00267 
00268                     vector signed short src_vA = // vec_unpackh sign-extends...
00269                                                  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00270                     vector signed short src_vB = // vec_unpackh sign-extends...
00271                                                  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00272 
00273                     vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00274                     vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
00275                     vector signed short filter_v0  = vec_perm(filter_v0R, filter_v1R, permF);
00276                     vector signed short filter_v1  = vec_perm(filter_v1R, filter_v2R, permF);
00277 
00278                     vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
00279                     val_v = vec_msums(src_vB, filter_v1, val_acc);
00280 
00281                     filter_v0R = filter_v2R;
00282                     src_v0     = src_v1;
00283                 }
00284 
00285                 if (j < filterSize - 7) {
00286                     // loading src_v0 is useless, it's already done above
00287                     // vector unsigned char src_v0 = vec_ld(srcPos + j, src);
00288                     vector unsigned char src_v1, src_vF;
00289                     vector signed short src_v, filter_v1R, filter_v;
00290                     if ((((uintptr_t)src + srcPos) % 16) > 8) {
00291                         src_v1 = vec_ld(srcPos + j + 16, src);
00292                     }
00293                     src_vF = vec_perm(src_v0, src_v1, permS);
00294 
00295                     src_v = // vec_unpackh sign-extends...
00296                             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00297                     // loading filter_v0R is useless, it's already done above
00298                     // vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
00299                     filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00300                     filter_v   = vec_perm(filter_v0R, filter_v1R, permF);
00301 
00302                     val_v = vec_msums(src_v, filter_v, val_v);
00303                 }
00304 
00305                 val_s = vec_sums(val_v, vzero);
00306 
00307                 vec_st(val_s, 0, tempo);
00308                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00309             }
00310         }
00311 }
00312 
00313 void ff_sws_init_swScale_altivec(SwsContext *c)
00314 {
00315     enum PixelFormat dstFormat = c->dstFormat;
00316 
00317     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
00318         return;
00319 
00320     if (c->srcBpc == 8 && c->dstBpc <= 10) {
00321         c->hyScale = c->hcScale = hScale_altivec_real;
00322     }
00323     if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
00324         dstFormat != PIX_FMT_NV12 && dstFormat != PIX_FMT_NV21 &&
00325         !c->alpPixBuf) {
00326         c->yuv2planeX = yuv2planeX_altivec;
00327     }
00328 
00329     /* The following list of supported dstFormat values should
00330      * match what's found in the body of ff_yuv2packedX_altivec() */
00331     if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
00332         switch (c->dstFormat) {
00333         case PIX_FMT_ABGR:
00334             c->yuv2packedX = ff_yuv2abgr_X_altivec;
00335             break;
00336         case PIX_FMT_BGRA:
00337             c->yuv2packedX = ff_yuv2bgra_X_altivec;
00338             break;
00339         case PIX_FMT_ARGB:
00340             c->yuv2packedX = ff_yuv2argb_X_altivec;
00341             break;
00342         case PIX_FMT_RGBA:
00343             c->yuv2packedX = ff_yuv2rgba_X_altivec;
00344             break;
00345         case PIX_FMT_BGR24:
00346             c->yuv2packedX = ff_yuv2bgr24_X_altivec;
00347             break;
00348         case PIX_FMT_RGB24:
00349             c->yuv2packedX = ff_yuv2rgb24_X_altivec;
00350             break;
00351         }
00352     }
00353 }