libmpcodecs/vf_halfpack.c

   1 #include <stdio.h>
   2 #include <stdlib.h>
   3 #include <string.h>
   4 #include <inttypes.h>
   5
   6 #include "config.h"
   7 #include "mp_msg.h"
   8 #include "cpudetect.h"
   9
  10 #include "img_format.h"
  11 #include "mp_image.h"
  12 #include "vf.h"
  13
  14 #include "libswscale/rgb2rgb.h"
  15
  16 struct vf_priv_s {
  17         int field;
  18 };
  19
  20 #if HAVE_MMX
  21 static void halfpack_MMX(unsigned char *dst, unsigned char *src[3],
  22                      int dststride, int srcstride[3],
  23                      int w, int h)
  24 {
  25         int j;
  26         unsigned char *y1, *y2, *u, *v;
  27         int dstinc, yinc, uinc, vinc;
  28
  29         y1 = src[0];
  30         y2 = src[0] + srcstride[0];
  31         u = src[1];
  32         v = src[2];
  33
  34         dstinc = dststride - 2*w;
  35         yinc = 2*srcstride[0] - w;
  36         uinc = srcstride[1] - w/2;
  37         vinc = srcstride[2] - w/2;
  38
  39         for (h/=2; h; h--) {
  40                 __asm__ (
  41                         "pxor %%mm0, %%mm0 \n\t"
  42                         ASMALIGN(4)
  43                         "1: \n\t"
  44                         "movq (%0), %%mm1 \n\t"
  45                         "movq (%0), %%mm2 \n\t"
  46                         "movq (%1), %%mm3 \n\t"
  47                         "movq (%1), %%mm4 \n\t"
  48                         "punpcklbw %%mm0, %%mm1 \n\t"
  49                         "punpckhbw %%mm0, %%mm2 \n\t"
  50                         "punpcklbw %%mm0, %%mm3 \n\t"
  51                         "punpckhbw %%mm0, %%mm4 \n\t"
  52                         "paddw %%mm3, %%mm1 \n\t"
  53                         "paddw %%mm4, %%mm2 \n\t"
  54                         "psrlw $1, %%mm1 \n\t"
  55                         "psrlw $1, %%mm2 \n\t"
  56
  57                         "movq (%2), %%mm3 \n\t"
  58                         "movq (%3), %%mm5 \n\t"
  59                         "punpcklbw %%mm0, %%mm3 \n\t"
  60                         "punpcklbw %%mm0, %%mm5 \n\t"
  61                         "movq %%mm3, %%mm4 \n\t"
  62                         "movq %%mm5, %%mm6 \n\t"
  63                         "punpcklwd %%mm0, %%mm3 \n\t"
  64                         "punpckhwd %%mm0, %%mm4 \n\t"
  65                         "punpcklwd %%mm0, %%mm5 \n\t"
  66                         "punpckhwd %%mm0, %%mm6 \n\t"
  67                         "pslld $8, %%mm3 \n\t"
  68                         "pslld $8, %%mm4 \n\t"
  69                         "pslld $24, %%mm5 \n\t"
  70                         "pslld $24, %%mm6 \n\t"
  71
  72                         "por %%mm3, %%mm1 \n\t"
  73                         "por %%mm4, %%mm2 \n\t"
  74                         "por %%mm5, %%mm1 \n\t"
  75                         "por %%mm6, %%mm2 \n\t"
  76
  77                         "add $8, %0 \n\t"
  78                         "add $8, %1 \n\t"
  79                         "add $4, %2 \n\t"
  80                         "add $4, %3 \n\t"
  81                         "movq %%mm1, (%8) \n\t"
  82                         "movq %%mm2, 8(%8) \n\t"
  83                         "add $16, %8 \n\t"
  84                         "decl %9 \n\t"
  85                         "jnz 1b \n\t"
  86                         : "=r" (y1), "=r" (y2), "=r" (u), "=r" (v)
  87                         : "0" (y1), "1" (y2), "2" (u), "3" (v), "r" (dst), "r" (w/8)
  88                         : "memory"
  89                 );
  90                 for (j = (w&7)/2; j; j--) {
  91                         *dst++ = (*y1++ + *y2++)/2;
  92                         *dst++ = *u++;
  93                         *dst++ = (*y1++ + *y2++)/2;
  94                         *dst++ = *v++;
  95                 }
  96                 y1 += yinc;
  97                 y2 += yinc;
  98                 u += uinc;
  99                 v += vinc;
 100                 dst += dstinc;
 101         }
 102         __asm__ volatile ( "emms \n\t" ::: "memory" );
 103 }
 104 #endif
 105
 106
 107
 108 static void halfpack_C(unsigned char *dst, unsigned char *src[3],
 109                      int dststride, int srcstride[3],
 110                      int w, int h)
 111 {
 112         int i, j;
 113         unsigned char *y1, *y2, *u, *v;
 114         int dstinc, yinc, uinc, vinc;
 115
 116         y1 = src[0];
 117         y2 = src[0] + srcstride[0];
 118         u = src[1];
 119         v = src[2];
 120
 121         dstinc = dststride - 2*w;
 122         yinc = 2*srcstride[0] - w;
 123         uinc = srcstride[1] - w/2;
 124         vinc = srcstride[2] - w/2;
 125
 126         for (i = h/2; i; i--) {
 127                 for (j = w/2; j; j--) {
 128                         *dst++ = (*y1++ + *y2++)>>1;
 129                         *dst++ = *u++;
 130                         *dst++ = (*y1++ + *y2++)>>1;
 131                         *dst++ = *v++;
 132                 }
 133                 y1 += yinc;
 134                 y2 += yinc;
 135                 u += uinc;
 136                 v += vinc;
 137                 dst += dstinc;
 138         }
 139 }
 140
 141 static void (*halfpack)(unsigned char *dst, unsigned char *src[3],
 142         int dststride, int srcstride[3], int w, int h);
 143
 144
 145 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
 146 {
 147         mp_image_t *dmpi;
 148
 149         // hope we'll get DR buffer:
 150         dmpi=vf_get_image(vf->next, IMGFMT_YUY2,
 151                           MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 152                           mpi->w, mpi->h/2);
 153
 154         switch(vf->priv->field) {
 155         case 0:
 156         case 1:
 157                 yuv422ptoyuy2(mpi->planes[0] + mpi->stride[0]*vf->priv->field,
 158                         mpi->planes[1], mpi->planes[2], dmpi->planes[0],
 159                         mpi->w, mpi->h/2, mpi->stride[0]*2, mpi->stride[1], dmpi->stride[0]);
 160                 break;
 161         default:
 162                 halfpack(dmpi->planes[0], mpi->planes, dmpi->stride[0],
 163                         mpi->stride, mpi->w, mpi->h);
 164         }
 165
 166         return vf_next_put_image(vf,dmpi, pts);
 167 }
 168
 169 static int config(struct vf_instance_s* vf,
 170                   int width, int height, int d_width, int d_height,
 171                   unsigned int flags, unsigned int outfmt)
 172 {
 173         /* FIXME - also support UYVY output? */
 174         return vf_next_config(vf, width, height/2, d_width, d_height, flags, IMGFMT_YUY2);
 175 }
 176
 177
 178 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
 179 {
 180         /* FIXME - really any YUV 4:2:0 input format should work */
 181         switch (fmt) {
 182         case IMGFMT_YV12:
 183         case IMGFMT_IYUV:
 184         case IMGFMT_I420:
 185                 return vf_next_query_format(vf,IMGFMT_YUY2);
 186         }
 187         return 0;
 188 }
 189
 190 static void uninit(struct vf_instance_s* vf)
 191 {
 192         free(vf->priv);
 193 }
 194
 195 static int open(vf_instance_t *vf, char* args)
 196 {
 197         vf->config=config;
 198         vf->query_format=query_format;
 199         vf->put_image=put_image;
 200         vf->uninit=uninit;
 201
 202         vf->priv = calloc(1, sizeof (struct vf_priv_s));
 203         vf->priv->field = 2;
 204         if (args) sscanf(args, "%d", &vf->priv->field);
 205
 206         halfpack = halfpack_C;
 207 #if HAVE_MMX
 208         if(gCpuCaps.hasMMX) halfpack = halfpack_MMX;
 209 #endif
 210         return 1;
 211 }
 212
 213 const vf_info_t vf_info_halfpack = {
 214         "yuv planar 4:2:0 -> packed 4:2:2, half height",
 215         "halfpack",
 216         "Richard Felker",
 217         "",
 218         open,
 219         NULL
 220 };
 221