libavcodec/ppc/gmc_altivec.c

   1 /*
   2  * GMC (Global Motion Compensation)
   3  * AltiVec-enabled
   4  * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
   5  *
   6  * This file is part of Libav.
   7  *
   8  * Libav is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * Libav is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with Libav; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "libavutil/mem.h"
  24 #include "libavutil/ppc/types_altivec.h"
  25 #include "libavutil/ppc/util_altivec.h"
  26 #include "dsputil_altivec.h"
  27
  28 /*
  29   altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
  30   to preserve proper dst alignment.
  31 */
  32 void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
  33 {
  34     const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
  35     const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] =
  36         {
  37             (16-x16)*(16-y16), /* A */
  38             (   x16)*(16-y16), /* B */
  39             (16-x16)*(   y16), /* C */
  40             (   x16)*(   y16), /* D */
  41             0, 0, 0, 0         /* padding */
  42         };
  43     register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
  44     register const vector unsigned short vcsr8 = (const vector unsigned short)vec_splat_u16(8);
  45     register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
  46     register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
  47     int i;
  48     unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
  49     unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
  50
  51     tempA = vec_ld(0, (const unsigned short*)ABCD);
  52     Av = vec_splat(tempA, 0);
  53     Bv = vec_splat(tempA, 1);
  54     Cv = vec_splat(tempA, 2);
  55     Dv = vec_splat(tempA, 3);
  56
  57     rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);
  58
  59     // we'll be able to pick-up our 9 char elements
  60     // at src from those 32 bytes
  61     // we load the first batch here, as inside the loop
  62     // we can re-use 'src+stride' from one iteration
  63     // as the 'src' of the next.
  64     src_0 = vec_ld(0, src);
  65     src_1 = vec_ld(16, src);
  66     srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));
  67
  68     if (src_really_odd != 0x0000000F) {
  69         // if src & 0xF == 0xF, then (src+1) is properly aligned
  70         // on the second vector.
  71         srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
  72     } else {
  73         srcvB = src_1;
  74     }
  75     srcvA = vec_mergeh(vczero, srcvA);
  76     srcvB = vec_mergeh(vczero, srcvB);
  77
  78     for(i=0; i<h; i++) {
  79         dst_odd = (unsigned long)dst & 0x0000000F;
  80         src_really_odd = (((unsigned long)src) + stride) & 0x0000000F;
  81
  82         dstv = vec_ld(0, dst);
  83
  84         // we we'll be able to pick-up our 9 char elements
  85         // at src + stride from those 32 bytes
  86         // then reuse the resulting 2 vectors srvcC and srcvD
  87         // as the next srcvA and srcvB
  88         src_0 = vec_ld(stride + 0, src);
  89         src_1 = vec_ld(stride + 16, src);
  90         srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
  91
  92         if (src_really_odd != 0x0000000F) {
  93             // if src & 0xF == 0xF, then (src+1) is properly aligned
  94             // on the second vector.
  95             srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
  96         } else {
  97             srcvD = src_1;
  98         }
  99
 100         srcvC = vec_mergeh(vczero, srcvC);
 101         srcvD = vec_mergeh(vczero, srcvD);
 102
 103
 104         // OK, now we (finally) do the math :-)
 105         // those four instructions replaces 32 int muls & 32 int adds.
 106         // isn't AltiVec nice ?
 107         tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
 108         tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
 109         tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);
 110         tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC);
 111
 112         srcvA = srcvC;
 113         srcvB = srcvD;
 114
 115         tempD = vec_sr(tempD, vcsr8);
 116
 117         dstv2 = vec_pack(tempD, (vector unsigned short)vczero);
 118
 119         if (dst_odd) {
 120             dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1));
 121         } else {
 122             dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3));
 123         }
 124
 125         vec_st(dstv2, 0, dst);
 126
 127         dst += stride;
 128         src += stride;
 129     }
 130 }