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 "../libvo/fastmemcpy.h"
  15 #include "../postproc/rgb2rgb.h"
  16
  17 struct vf_priv_s {
  18         int field;
  19 };
  20
  21 #ifdef HAVE_MMX
  22 static void halfpack_MMX(unsigned char *dst, unsigned char *src[3],
  23                      unsigned int dststride, unsigned int srcstride[3],
  24                      int w, int h)
  25 {
  26         int j;
  27         unsigned char *y1, *y2, *u, *v;
  28         unsigned int dstinc, yinc, uinc, vinc;
  29
  30         y1 = src[0];
  31         y2 = src[0] + srcstride[0];
  32         u = src[1];
  33         v = src[2];
  34
  35         dstinc = dststride - 2*w;
  36         yinc = 2*srcstride[0] - w;
  37         uinc = srcstride[1] - w/2;
  38         vinc = srcstride[2] - w/2;
  39
  40         for (h/=2; h; h--) {
  41                 asm (
  42                         "pxor %%mm0, %%mm0 \n\t"
  43                         ".balign 16 \n\t"
  44                         "1: \n\t"
  45                         "movq (%0), %%mm1 \n\t"
  46                         "movq (%0), %%mm2 \n\t"
  47                         "movq (%1), %%mm3 \n\t"
  48                         "movq (%1), %%mm4 \n\t"
  49                         "punpcklbw %%mm0, %%mm1 \n\t"
  50                         "punpckhbw %%mm0, %%mm2 \n\t"
  51                         "punpcklbw %%mm0, %%mm3 \n\t"
  52                         "punpckhbw %%mm0, %%mm4 \n\t"
  53                         "paddw %%mm3, %%mm1 \n\t"
  54                         "paddw %%mm4, %%mm2 \n\t"
  55                         "psrlw $1, %%mm1 \n\t"
  56                         "psrlw $1, %%mm2 \n\t"
  57
  58                         "movq (%2), %%mm3 \n\t"
  59                         "movq (%3), %%mm5 \n\t"
  60                         "punpcklbw %%mm0, %%mm3 \n\t"
  61                         "punpcklbw %%mm0, %%mm5 \n\t"
  62                         "movq %%mm3, %%mm4 \n\t"
  63                         "movq %%mm5, %%mm6 \n\t"
  64                         "punpcklwd %%mm0, %%mm3 \n\t"
  65                         "punpckhwd %%mm0, %%mm4 \n\t"
  66                         "punpcklwd %%mm0, %%mm5 \n\t"
  67                         "punpckhwd %%mm0, %%mm6 \n\t"
  68                         "pslld $8, %%mm3 \n\t"
  69                         "pslld $8, %%mm4 \n\t"
  70                         "pslld $24, %%mm5 \n\t"
  71                         "pslld $24, %%mm6 \n\t"
  72
  73                         "por %%mm3, %%mm1 \n\t"
  74                         "por %%mm4, %%mm2 \n\t"
  75                         "por %%mm5, %%mm1 \n\t"
  76                         "por %%mm6, %%mm2 \n\t"
  77
  78                         "addl $8, %0 \n\t"
  79                         "addl $8, %1 \n\t"
  80                         "addl $4, %2 \n\t"
  81                         "addl $4, %3 \n\t"
  82                         "movq %%mm1, (%8) \n\t"
  83                         "movq %%mm2, 8(%8) \n\t"
  84                         "addl $16, %8 \n\t"
  85                         "decl %9 \n\t"
  86                         "jnz 1b \n\t"
  87                         : "=r" (y1), "=r" (y2), "=r" (u), "=r" (v)
  88                         : "0" (y1), "1" (y2), "2" (u), "3" (v), "r" (dst), "r" (w/8)
  89                         : "memory"
  90                 );
  91                 for (j = (w&7)/2; j; j--) {
  92                         *dst++ = (*y1++ + *y2++)/2;
  93                         *dst++ = *u++;
  94                         *dst++ = (*y1++ + *y2++)/2;
  95                         *dst++ = *v++;
  96                 }
  97                 y1 += yinc;
  98                 y2 += yinc;
  99                 u += uinc;
 100                 v += vinc;
 101                 dst += dstinc;
 102         }
 103         asm volatile ( "emms \n\t" ::: "memory" );
 104 }
 105 #endif
 106
 107
 108
 109 static void halfpack_C(unsigned char *dst, unsigned char *src[3],
 110                      unsigned int dststride, unsigned int srcstride[3],
 111                      int w, int h)
 112 {
 113         int i, j;
 114         unsigned char *y1, *y2, *u, *v;
 115         unsigned int dstinc, yinc, uinc, vinc;
 116
 117         y1 = src[0];
 118         y2 = src[0] + srcstride[0];
 119         u = src[1];
 120         v = src[2];
 121
 122         dstinc = dststride - 2*w;
 123         yinc = 2*srcstride[0] - w;
 124         uinc = srcstride[1] - w/2;
 125         vinc = srcstride[2] - w/2;
 126
 127         for (i = h/2; i; i--) {
 128                 for (j = w/2; j; j--) {
 129                         *dst++ = (*y1++ + *y2++)>>1;
 130                         *dst++ = *u++;
 131                         *dst++ = (*y1++ + *y2++)>>1;
 132                         *dst++ = *v++;
 133                 }
 134                 y1 += yinc;
 135                 y2 += yinc;
 136                 u += uinc;
 137                 v += vinc;
 138                 dst += dstinc;
 139         }
 140 }
 141
 142 static void (*halfpack)(unsigned char *dst, unsigned char *src[3],
 143         unsigned int dststride, unsigned int srcstride[3], int w, int h);
 144
 145
 146 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi)
 147 {
 148         mp_image_t *dmpi;
 149
 150         // hope we'll get DR buffer:
 151         dmpi=vf_get_image(vf->next, IMGFMT_YUY2,
 152                           MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 153                           mpi->w, mpi->h/2);
 154
 155         switch(vf->priv->field) {
 156         case 0:
 157         case 1:
 158                 yuv422ptoyuy2(mpi->planes[0] + mpi->stride[0]*vf->priv->field,
 159                         mpi->planes[1], mpi->planes[2], dmpi->planes[0],
 160                         mpi->w, mpi->h/2, mpi->stride[0]*2, mpi->stride[1], dmpi->stride[0]);
 161                 break;
 162         default:
 163                 halfpack(dmpi->planes[0], mpi->planes, dmpi->stride[0],
 164                         mpi->stride, mpi->w, mpi->h);
 165         }
 166
 167         return vf_next_put_image(vf,dmpi);
 168 }
 169
 170 static int config(struct vf_instance_s* vf,
 171                   int width, int height, int d_width, int d_height,
 172                   unsigned int flags, unsigned int outfmt)
 173 {
 174         /* FIXME - also support UYVY output? */
 175         return vf_next_config(vf, width, height/2, d_width, d_height, flags, IMGFMT_YUY2);
 176 }
 177
 178
 179 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
 180 {
 181         /* FIXME - really any YUV 4:2:0 input format should work */
 182         switch (fmt) {
 183         case IMGFMT_YV12:
 184         case IMGFMT_IYUV:
 185         case IMGFMT_I420:
 186                 return vf_next_query_format(vf,IMGFMT_YUY2);
 187         }
 188         return 0;
 189 }
 190
 191 static void uninit(struct vf_instance_s* vf)
 192 {
 193         free(vf->priv);
 194 }
 195
 196 static int open(vf_instance_t *vf, char* args)
 197 {
 198         vf->config=config;
 199         vf->query_format=query_format;
 200         vf->put_image=put_image;
 201         vf->uninit=uninit;
 202
 203         vf->priv = calloc(1, sizeof (struct vf_priv_s));
 204         vf->priv->field = 2;
 205         if (args) sscanf(args, "%d", &vf->priv->field);
 206
 207         halfpack = halfpack_C;
 208 #ifdef HAVE_MMX
 209         if(gCpuCaps.hasMMX) halfpack = halfpack_MMX;
 210 #endif
 211         return 1;
 212 }
 213
 214 vf_info_t vf_info_halfpack = {
 215         "yuv planar 4:2:0 -> packed 4:2:2, half height",
 216         "halfpack",
 217         "Richard Felker",
 218         "",
 219         open,
 220         NULL
 221 };
 222