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 static inline void altivec_packIntArrayToCharArray(int *val, uint8_t *dest,
00036                                                    int dstW)
00037 {
00038     register int i;
00039     vector unsigned int altivec_vectorShiftInt19 =
00040         vec_add(vec_splat_u32(10), vec_splat_u32(9));
00041     if ((uintptr_t)dest % 16) {
00042         /* badly aligned store, we force store alignment */
00043         /* and will handle load misalignment on val w/ vec_perm */
00044         vector unsigned char perm1;
00045         vector signed int v1;
00046         for (i = 0; (i < dstW) &&
00047              (((uintptr_t)dest + i) % 16); i++) {
00048             int t = val[i] >> 19;
00049             dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00050         }
00051         perm1 = vec_lvsl(i << 2, val);
00052         v1    = vec_ld(i << 2, val);
00053         for (; i < (dstW - 15); i += 16) {
00054             int offset = i << 2;
00055             vector signed int v2  = vec_ld(offset + 16, val);
00056             vector signed int v3  = vec_ld(offset + 32, val);
00057             vector signed int v4  = vec_ld(offset + 48, val);
00058             vector signed int v5  = vec_ld(offset + 64, val);
00059             vector signed int v12 = vec_perm(v1, v2, perm1);
00060             vector signed int v23 = vec_perm(v2, v3, perm1);
00061             vector signed int v34 = vec_perm(v3, v4, perm1);
00062             vector signed int v45 = vec_perm(v4, v5, perm1);
00063 
00064             vector signed int vA      = vec_sra(v12, altivec_vectorShiftInt19);
00065             vector signed int vB      = vec_sra(v23, altivec_vectorShiftInt19);
00066             vector signed int vC      = vec_sra(v34, altivec_vectorShiftInt19);
00067             vector signed int vD      = vec_sra(v45, altivec_vectorShiftInt19);
00068             vector unsigned short vs1 = vec_packsu(vA, vB);
00069             vector unsigned short vs2 = vec_packsu(vC, vD);
00070             vector unsigned char vf   = vec_packsu(vs1, vs2);
00071             vec_st(vf, i, dest);
00072             v1 = v5;
00073         }
00074     } else { // dest is properly aligned, great
00075         for (i = 0; i < (dstW - 15); i += 16) {
00076             int offset = i << 2;
00077             vector signed int v1      = vec_ld(offset, val);
00078             vector signed int v2      = vec_ld(offset + 16, val);
00079             vector signed int v3      = vec_ld(offset + 32, val);
00080             vector signed int v4      = vec_ld(offset + 48, val);
00081             vector signed int v5      = vec_sra(v1, altivec_vectorShiftInt19);
00082             vector signed int v6      = vec_sra(v2, altivec_vectorShiftInt19);
00083             vector signed int v7      = vec_sra(v3, altivec_vectorShiftInt19);
00084             vector signed int v8      = vec_sra(v4, altivec_vectorShiftInt19);
00085             vector unsigned short vs1 = vec_packsu(v5, v6);
00086             vector unsigned short vs2 = vec_packsu(v7, v8);
00087             vector unsigned char vf   = vec_packsu(vs1, vs2);
00088             vec_st(vf, i, dest);
00089         }
00090     }
00091     for (; i < dstW; i++) {
00092         int t = val[i] >> 19;
00093         dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
00094     }
00095 }
00096 
00097 // FIXME remove the usage of scratch buffers.
00098 static void yuv2planeX_altivec(const int16_t *filter, int filterSize,
00099                                const int16_t **src, uint8_t *dest, int dstW,
00100                                const uint8_t *dither, int offset)
00101 {
00102     register int i, j;
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     altivec_packIntArrayToCharArray(val, dest, dstW);
00144 }
00145 
00146 static void hScale_altivec_real(SwsContext *c, int16_t *dst, int dstW,
00147                                 const uint8_t *src, const int16_t *filter,
00148                                 const int32_t *filterPos, int filterSize)
00149 {
00150     register int i;
00151     DECLARE_ALIGNED(16, int, tempo)[4];
00152 
00153     if (filterSize % 4) {
00154         for (i = 0; i < dstW; i++) {
00155             register int j;
00156             register int srcPos = filterPos[i];
00157             register int val    = 0;
00158             for (j = 0; j < filterSize; j++)
00159                 val += ((int)src[srcPos + j]) * filter[filterSize * i + j];
00160             dst[i] = FFMIN(val >> 7, (1 << 15) - 1);
00161         }
00162     } else
00163         switch (filterSize) {
00164         case 4:
00165             for (i = 0; i < dstW; i++) {
00166                 register int srcPos = filterPos[i];
00167 
00168                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00169                 vector unsigned char src_v1, src_vF;
00170                 vector signed short src_v, filter_v;
00171                 vector signed int val_vEven, val_s;
00172                 if ((((uintptr_t)src + srcPos) % 16) > 12) {
00173                     src_v1 = vec_ld(srcPos + 16, src);
00174                 }
00175                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00176 
00177                 src_v = // vec_unpackh sign-extends...
00178                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00179                 // now put our elements in the even slots
00180                 src_v = vec_mergeh(src_v, (vector signed short)vzero);
00181 
00182                 filter_v = vec_ld(i << 3, filter);
00183                 // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
00184 
00185                 // The neat trick: We only care for half the elements,
00186                 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
00187                 // and we're going to use vec_mule, so we choose
00188                 // carefully how to "unpack" the elements into the even slots.
00189                 if ((i << 3) % 16)
00190                     filter_v = vec_mergel(filter_v, (vector signed short)vzero);
00191                 else
00192                     filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
00193 
00194                 val_vEven = vec_mule(src_v, filter_v);
00195                 val_s     = vec_sums(val_vEven, vzero);
00196                 vec_st(val_s, 0, tempo);
00197                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00198             }
00199         break;
00200 
00201         case 8:
00202             for (i = 0; i < dstW; i++) {
00203                 register int srcPos = filterPos[i];
00204 
00205                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00206                 vector unsigned char src_v1, src_vF;
00207                 vector signed short src_v, filter_v;
00208                 vector signed int val_v, val_s;
00209                 if ((((uintptr_t)src + srcPos) % 16) > 8) {
00210                     src_v1 = vec_ld(srcPos + 16, src);
00211                 }
00212                 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00213 
00214                 src_v = // vec_unpackh sign-extends...
00215                         (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00216                 filter_v = vec_ld(i << 4, filter);
00217                 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
00218 
00219                 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
00220                 val_s = vec_sums(val_v, vzero);
00221                 vec_st(val_s, 0, tempo);
00222                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00223             }
00224         break;
00225 
00226         case 16:
00227             for (i = 0; i < dstW; i++) {
00228                 register int srcPos = filterPos[i];
00229 
00230                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00231                 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
00232                 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
00233 
00234                 vector signed short src_vA = // vec_unpackh sign-extends...
00235                                              (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00236                 vector signed short src_vB = // vec_unpackh sign-extends...
00237                                              (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00238 
00239                 vector signed short filter_v0 = vec_ld(i << 5, filter);
00240                 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
00241                 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
00242 
00243                 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
00244                 vector signed int val_v   = vec_msums(src_vB, filter_v1, val_acc);
00245 
00246                 vector signed int val_s = vec_sums(val_v, vzero);
00247 
00248                 vec_st(val_s, 0, tempo);
00249                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00250             }
00251         break;
00252 
00253         default:
00254             for (i = 0; i < dstW; i++) {
00255                 register int j;
00256                 register int srcPos = filterPos[i];
00257 
00258                 vector signed int val_s, val_v = (vector signed int)vzero;
00259                 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
00260                 vector unsigned char permF     = vec_lvsl((i * 2 * filterSize), filter);
00261 
00262                 vector unsigned char src_v0 = vec_ld(srcPos, src);
00263                 vector unsigned char permS  = vec_lvsl(srcPos, src);
00264 
00265                 for (j = 0; j < filterSize - 15; j += 16) {
00266                     vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
00267                     vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
00268 
00269                     vector signed short src_vA = // vec_unpackh sign-extends...
00270                                                  (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00271                     vector signed short src_vB = // vec_unpackh sign-extends...
00272                                                  (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
00273 
00274                     vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00275                     vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
00276                     vector signed short filter_v0  = vec_perm(filter_v0R, filter_v1R, permF);
00277                     vector signed short filter_v1  = vec_perm(filter_v1R, filter_v2R, permF);
00278 
00279                     vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
00280                     val_v = vec_msums(src_vB, filter_v1, val_acc);
00281 
00282                     filter_v0R = filter_v2R;
00283                     src_v0     = src_v1;
00284                 }
00285 
00286                 if (j < filterSize - 7) {
00287                     // loading src_v0 is useless, it's already done above
00288                     // vector unsigned char src_v0 = vec_ld(srcPos + j, src);
00289                     vector unsigned char src_v1, src_vF;
00290                     vector signed short src_v, filter_v1R, filter_v;
00291                     if ((((uintptr_t)src + srcPos) % 16) > 8) {
00292                         src_v1 = vec_ld(srcPos + j + 16, src);
00293                     }
00294                     src_vF = vec_perm(src_v0, src_v1, permS);
00295 
00296                     src_v = // vec_unpackh sign-extends...
00297                             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
00298                     // loading filter_v0R is useless, it's already done above
00299                     // vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
00300                     filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
00301                     filter_v   = vec_perm(filter_v0R, filter_v1R, permF);
00302 
00303                     val_v = vec_msums(src_v, filter_v, val_v);
00304                 }
00305 
00306                 val_s = vec_sums(val_v, vzero);
00307 
00308                 vec_st(val_s, 0, tempo);
00309                 dst[i] = FFMIN(tempo[3] >> 7, (1 << 15) - 1);
00310             }
00311         }
00312 }
00313 
00314 av_cold void ff_sws_init_swScale_altivec(SwsContext *c)
00315 {
00316     enum PixelFormat dstFormat = c->dstFormat;
00317 
00318     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
00319         return;
00320 
00321     if (c->srcBpc == 8 && c->dstBpc <= 14) {
00322         c->hyScale = c->hcScale = hScale_altivec_real;
00323     }
00324     if (!is16BPS(dstFormat) && !is9_OR_10BPS(dstFormat) &&
00325         dstFormat != PIX_FMT_NV12 && dstFormat != PIX_FMT_NV21 &&
00326         !c->alpPixBuf) {
00327         c->yuv2planeX = yuv2planeX_altivec;
00328     }
00329 
00330     /* The following list of supported dstFormat values should
00331      * match what's found in the body of ff_yuv2packedX_altivec() */
00332     if (!(c->flags & (SWS_BITEXACT | SWS_FULL_CHR_H_INT)) && !c->alpPixBuf) {
00333         switch (c->dstFormat) {
00334         case PIX_FMT_ABGR:
00335             c->yuv2packedX = ff_yuv2abgr_X_altivec;
00336             break;
00337         case PIX_FMT_BGRA:
00338             c->yuv2packedX = ff_yuv2bgra_X_altivec;
00339             break;
00340         case PIX_FMT_ARGB:
00341             c->yuv2packedX = ff_yuv2argb_X_altivec;
00342             break;
00343         case PIX_FMT_RGBA:
00344             c->yuv2packedX = ff_yuv2rgba_X_altivec;
00345             break;
00346         case PIX_FMT_BGR24:
00347             c->yuv2packedX = ff_yuv2bgr24_X_altivec;
00348             break;
00349         case PIX_FMT_RGB24:
00350             c->yuv2packedX = ff_yuv2rgb24_X_altivec;
00351             break;
00352         }
00353     }
00354 }