path: root/libvo/vo_dga.c

#define DISP

/*
 * $Id$
 * 
 * video_out_dga.c, X11 interface
 *
 *
 * Copyright ( C ) 2001, Andreas Ackermann. All Rights Reserved.
 *
 * <acki@acki-netz.de>
 *
 * Sourceforge username: acki2
 * 
 * note well: 
 *   
 * - covers only common video card formats i.e. 
 *      BGR_16_15_555
 *      BGR_16_16_565
 *      BGR_24_24_888
 *      BGR_32_24_888
 *
 * - works only on x86 architectures
 *
 * $Log$
 * Revision 1.44  2002/02/17 08:24:43  nick
 * I don't like such reports: '-vo dga:vidix or -vo x11:vidix works fine for me'
 *
 * Revision 1.42  2002/02/12 23:19:37  michael
 * use mem2agpcpy() instead of fast_memcpy()
 *
 * Revision 1.41  2002/02/09 01:21:48  arpi
 * 10000hl to Holm... control MUST BE static...
 *
 * Revision 1.40  2002/02/09 00:47:26  arpi
 * query_ stuff replaced by new control() - patch by David Holm
 *
 * Revision 1.39  2002/01/31 11:45:25  alex
 * removed obsoleted Terminate_Display_Process
 *
 * Revision 1.38  2002/01/31 09:52:45  nick
 * new info for tuning
 *
 * Revision 1.35  2001/12/28 20:52:54  alex
 * use XF86VidMode later in init (at line 1031) only if we've got support (if have_vm==1)
 *
 * Revision 1.34  2001/12/04 17:24:25  alex
 * do not crash if can't get modelines (dga2.0)
 *
 * Revision 1.33  2001/11/06 11:21:08  nick
 * Move yuv2rgb to postprocess
 *
 *
 * Revision 1.31  2001/10/30 17:04:31  nick
 * Using new stuff of rgb15to16
 *
 * Revision 1.30  2001/08/13 11:08:18  atlka
 * changes according to -utf8 option, draw_osd() function added
 *
 * Revision 1.29  2001/07/16 18:41:52  jkeil
 * vo_dga doesn't compile on non-x86 architecture due to x86 asm usage.
 *
 * Revision 1.28  2001/07/03 23:45:49  arpi
 * extern vo_doublebuffering cleanup
 *
 * Revision 1.27  2001/06/22 19:51:25  atmosfear
 * Fixed pointer->integer cast warning.
 *
 * Revision 1.26  2001/06/18 16:38:06  acki2
 * - just modified an error message
 *
 * Revision 1.25  2001/06/17 22:21:47  acki2
 * - if DGA fails to report some valid modes, vo_dga now exits gracefully
 *   instead of crashing ... (100000x100000 bug ...)
 *
 * Revision 1.24  2001/06/17 20:59:39  acki2
 * - doublebuffering now can be switched on and off with the -(no)double switch.
 *   Default in libvo is disabled.
 *
 * Revision 1.23  2001/05/24 20:48:45  arpi_esp
 * removed redundant osd.h includes
 *
 * Revision 1.22  2001/05/07 19:16:04  acki2
 * - now chooses mode with highest ymax (enables doublebuffering in some cases
 *   it didn't work before)
 * - use my own memcopy() on non MMX machines again
 * - do memcpy() in one single block if stride==0
 *
 * Revision 1.21  2001/05/03 22:39:38  acki2
 * - finally: 15to16 conversion included!!!
 *
 * Revision 1.20  2001/05/02 23:21:27  acki2
 * - now we use fastmemcpy() for copying. Saves about 25% of copying time on K6-2+
 *
 * Revision 1.19  2001/05/01 22:37:37  acki2
 * - now features 24->32 conversion (this is actually faster than letting the
 *   codec produce depth 32 in the first place for avis :-))) )
 *
 * Revision 1.18  2001/05/01 20:24:31  acki2
 * - now mpeg is fast again (no more offscreen buffer rubbish) But is it really ok?
 *
 * Revision 1.17  2001/04/24 11:42:04  pontscho
 * clean up
 *
 * Revision 1.16  2001/04/24 10:21:12  szabii
 * some warnings killed
 *
 * Revision 1.15  2001/04/19 21:39:10  arpi_esp
 * driver info now depends on detected DGA version
 *
 * Revision 1.14  2001/04/17 22:28:09  acki2
 * - now also supports OSD for YV12 (big speed penalty by having to build image
 *   in offscreen memory and then copying;
 * - OSD still works just with doublebuffering enabled :-(
 *
 * Revision 1.13  2001/04/17 20:51:58  acki2
 * - query_format() now uses new return value concept
 * - now support for OSD :-))) for RGB modes
 *   YV12 is flickering in quite an ugly fashion; have to fix this, but
 *   will cost an extra copying of image data ... :-(((
 *
 * Revision 1.12  2001/04/13 22:11:08  acki2
 * - fixed bug with depth and mpg when current bpp of XServer was != 32
 * - when -bpp is selected, I accept only query_modes() for THIS particular depth
 *   (if it's supported by hardware)
 *
 * Revision 1.10  2001/04/01 22:01:28  acki2
 * - still more debug output to be able to fix 15/16 bpp problem
 *
 * Revision 1.9  2001/04/01 08:07:14  acki2
 * - added detection of memsize of graphics card to check if double buffering is possible
 * - fixed resolution switching a little and added more debug output
 * - resolution switching is still according to d_width and d_height which
 *   is not always a good idea ...
 *
 * 
 * 30/02/2001
 *
 * o query_format(): with DGA 2.0 it returns all depths it supports
 *   (even 16 when running 32 and vice versa)
 *   Checks for (hopefully!) compatible RGBmasks in 15/16 bit modes
 * o added some more criterions for resolution switching
 * o cleanup
 * o with DGA2.0 present, ONLY DGA2.0 functions are used
 * o for 15/16 modes ONLY RGB 555 is supported, since the divx-codec
 *   happens to map the data this way. If your graphics card supports
 *   this, you're well off and may use these modes; for mpeg 
 *   movies things could be different, but I was too lazy to implement 
 *   it ...
 * o you may define VO_DGA_FORCE_DEPTH to the depth you desire 
 *   if you don't like the choice the driver makes
 *   Beware: unless you can use DGA2.0 this has to be your X Servers
 *           depth!!!
 * o Added double buffering :-))
 * o included VidMode switching support for DGA1.0, written by  Michael Graffam
 *    mgraffam@idsi.net
 * 
 */

//#define VO_DGA_DBG 1
//#undef HAVE_DGA2
//#undef HAVE_XF86VM

#include <stdio.h>
#include <stdlib.h>
#include <stdarg.h>
#include <string.h>
#include <errno.h>

#include "config.h"
#include "video_out.h"
#include "video_out_internal.h"
#include "../postproc/swscale.h"
#include "../postproc/rgb2rgb.h"
#include "aspect.h"

LIBVO_EXTERN( dga )

#include <X11/Xlib.h>
#include <X11/extensions/xf86dga.h>

#ifdef HAVE_XF86VM
#include <X11/extensions/xf86vmode.h>
#endif


#include "x11_common.h"
#include "../postproc/rgb2rgb.h"
#include "fastmemcpy.h"

static vo_info_t vo_info =
{
#ifdef HAVE_DGA2
        "DGA ( Direct Graphic Access V2.0 )",
#else
        "DGA ( Direct Graphic Access V1.0+XF86VidModeExtension )",
#endif
        "dga",
        "Andreas Ackermann <acki@acki-netz.de>",
        ""
};


//------------------------------------------------------------------


//#define BITSPP (vo_dga_modes[vo_dga_active_mode].vdm_bitspp)
//#define BYTESPP (vo_dga_modes[vo_dga_active_mode].vdm_bytespp)

#define VO_DGA_INVALID_RES 100000

#define HW_MODE (vo_dga_modes[vo_dga_hw_mode])
#define SRC_MODE (vo_dga_modes[vo_dga_src_mode]) 

struct vd_modes {
  int    vdm_mplayer_depth;
  int    vdm_supported;
  int    vdm_depth;
  int    vdm_bitspp;
  int    vdm_bytespp;
  int    vdm_rmask;
  int    vdm_gmask;
  int    vdm_bmask;
  int    vdm_hw_mode;
  int    vdm_conversion_func;
};

//------------------------------------------------------------------

#define VDM_CONV_NATIVE 0
#define VDM_CONV_15TO16 1
#define VDM_CONV_24TO32 2

static struct vd_modes vo_dga_modes[] = {
  // these entries describe HW modes
  // however, we use the same entries to tell mplayer what we support
  // so the last two values describe, which HW mode to use and which conversion 
  // function to use for a mode that is not supported by HW

  {  0,  0,  0,  0, 0,          0,          0, 0,      0, 0},
  { 15,  0, 15, 16, 2,     0x7c00,     0x03e0, 0x001f, 2, VDM_CONV_15TO16 },
  { 16,  0, 16, 16, 2,     0xf800,     0x07e0, 0x001f, 2, VDM_CONV_NATIVE },
  { 24,  0, 24, 24, 3,   0xff0000,   0x00ff00, 0x0000ff, 4, VDM_CONV_24TO32},
  { 32,  0, 24, 32, 4, 0x00ff0000, 0x0000ff00, 0x000000ff, 4, VDM_CONV_NATIVE}
};

static int vo_dga_mode_num = sizeof(vo_dga_modes)/sizeof(struct vd_modes);

// enable a HW mode (by description)
int vd_EnableMode( int depth, int bitspp, 
                    int rmask, int gmask, int bmask){
  int i;
  for(i=1; i<vo_dga_mode_num; i++){
    if(vo_dga_modes[i].vdm_depth == depth &&
       vo_dga_modes[i].vdm_bitspp == bitspp &&
       vo_dga_modes[i].vdm_rmask == rmask &&
       vo_dga_modes[i].vdm_gmask == gmask &&
       vo_dga_modes[i].vdm_bmask == bmask){
       vo_dga_modes[i].vdm_supported = 1;
       vo_dga_modes[i].vdm_hw_mode = i;
       vo_dga_modes[i].vdm_conversion_func = VDM_CONV_NATIVE;
       return i;
    }
  }
  return 0;
}

int vd_ModeEqual(int depth, int bitspp, 
		 int rmask, int gmask, int bmask, int index){
  return (
     (vo_dga_modes[index].vdm_depth == depth &&
     vo_dga_modes[index].vdm_bitspp == bitspp &&
     vo_dga_modes[index].vdm_rmask == rmask &&
     vo_dga_modes[index].vdm_gmask == gmask &&
     vo_dga_modes[index].vdm_bmask == bmask)
     ? 1 : 0); 
}


// enable a HW mode (mplayer_depth decides which)
int vd_ValidateMode( int mplayer_depth){
  int i;
  if(mplayer_depth == 0)return 0;
  for(i=1; i<vo_dga_mode_num; i++){
    if(vo_dga_modes[i].vdm_mplayer_depth == mplayer_depth ){ 
      vo_dga_modes[i].vdm_supported = 1;
      vo_dga_modes[i].vdm_hw_mode = i;
      vo_dga_modes[i].vdm_conversion_func = VDM_CONV_NATIVE;
      return i;
    }
  }
  return 0;
}

// do we support this mode? (not important whether native or conversion)
int vd_ModeValid( int mplayer_depth){
  int i;
  if(mplayer_depth == 0)return 0;
  for(i=1; i<vo_dga_mode_num; i++){
    if(vo_dga_modes[i].vdm_mplayer_depth == mplayer_depth && 
       vo_dga_modes[i].vdm_supported != 0){
      return i;
    }
  }
  return 0;
}

char *vd_GetModeString(int index){

#define VO_DGA_MAX_STRING_LEN 100
  static char stringbuf[VO_DGA_MAX_STRING_LEN]; 
  stringbuf[VO_DGA_MAX_STRING_LEN-1]=0;
  snprintf(stringbuf, VO_DGA_MAX_STRING_LEN-2, 
    "depth=%d, bpp=%d, r=%06x, g=%06x, b=%06x, %s (-bpp %d)",
    vo_dga_modes[index].vdm_depth,
    vo_dga_modes[index].vdm_bitspp,
    vo_dga_modes[index].vdm_rmask,
    vo_dga_modes[index].vdm_gmask,
    vo_dga_modes[index].vdm_bmask,
    vo_dga_modes[index].vdm_supported ? 
    (vo_dga_modes[index].vdm_conversion_func == VDM_CONV_NATIVE ? 
        "native (fast),    " : "conversion (slow),") :
        "not supported :-( ",
    vo_dga_modes[index].vdm_mplayer_depth);
  return stringbuf;
}

//-----------------------------------------------------------------

#ifdef HAVE_XF86VM
static XF86VidModeModeInfo **vo_dga_vidmodes=NULL;
#endif


extern int       verbose;          

static int       vo_dga_src_format;
static int       vo_dga_width;           // bytes per line in framebuffer
static int       vo_dga_vp_width;        // visible pixels per line in 
                                         // framebuffer
static int       vo_dga_vp_height;       // visible lines in framebuffer
static int       vo_dga_is_running = 0; 
static int       vo_dga_src_width;       // width of video in pixels
static int       vo_dga_src_height;      // height of video in pixels
static int       vo_dga_src_offset=0;    // offset in src
static int       vo_dga_vp_offset=0;     // offset in dest
static int       vo_dga_bytes_per_line;  // bytes per line to copy
static int       vo_dga_src_skip;        // bytes to skip after copying one 
                                         // line 
                                         // (not supported yet) in src
static int       vo_dga_vp_skip;         // dto. for dest 
static int       vo_dga_lines;           // num of lines to copy                                
static int       vo_dga_hw_mode = 0;     // index in mode list that is actually
                                         // used by framebuffer
static int       vo_dga_src_mode = 0;    // index in mode list that is used by 
                                         // codec
static int       vo_dga_XServer_mode = 0;// index in mode list for resolution
                                         // XServer is running

static int       vo_dga_dbf_mem_offset;  // offset in bytes for alternative 
                                         // framebuffer (0 if dbf is not 
					 // possible)
static int       vo_dga_dbf_y_offset;    // y offset (in scanlines)
static int       
                 vo_dga_dbf_current;     // current buffer (0 or 1)

static unsigned char     *vo_dga_base;
static Display  *vo_dga_dpy;

/* saved src and dst dimensions for SwScaler */
static unsigned int scale_srcW = 0,
                    scale_dstW = 0,
                    scale_srcH = 0,
                    scale_dstH = 0;


//---------------------------------------------------------

#define VD_INFO  0
#define VD_ERR   0
#define VD_DBG   2
#define VD_RES   1

void vd_printf( int level, const char *str, ...){
  va_list ap;

#ifndef VO_DGA_DBG
  // show resolution and DBG-messages only in verbose mode ...
  if( !verbose && level)return;         
#endif
  
  va_start(ap, str);
  vprintf(str, ap);
  va_end(ap);
}

//---------------------------------------------------------

static void draw_alpha( int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride ){

  char *d;
  unsigned int offset;
  unsigned int buffer_stride;

  offset = vo_dga_width * y0 +x0;
  buffer_stride = vo_dga_width;
  d = (&((char *)vo_dga_base)[vo_dga_vp_offset + vo_dga_dbf_current * vo_dga_dbf_mem_offset]);
     
  switch( HW_MODE.vdm_mplayer_depth ){

  case 32: 
    vo_draw_alpha_rgb32(w,h,src,srca,stride, d+4*offset , 4*buffer_stride); 
    break;
  case 24: 
    vo_draw_alpha_rgb24(w,h,src,srca,stride, d+3*offset , 3*buffer_stride);
    break;
  case 15:
    vo_draw_alpha_rgb15(w,h,src,srca,stride, d+2*offset , 2*buffer_stride);
    break;
  case 16:        
    vo_draw_alpha_rgb16(w,h,src,srca,stride, d+2*offset , 2*buffer_stride);
    break;
  }
}


//---------------------------------------------------------



// I had tried to work with mmx/3dnow copy code but
// there was not much speed gain and I didn't know
// how to save the FPU/mmx registers - so the copy
// code interferred with sound output ...
// removed the leftovers
// acki2 on 30/3/2001


#define rep_movsl(dest, src, numwords, d_add, count) \
__asm__ __volatile__( \
" \
1:                     \n\t\
                  movl %%edx, %%ecx \n\t \
                  cld\n\t \
                  rep\n\t \
                  movsl \n\t\
                  add %%eax, %%edi \n\t\
                  dec %%ebx \n\t\
                  jnz 1b \n\t\
" \
                  : \
                  : "a" (d_add), "b" (count), "S" (src), "D" (dest), \
		    "d" (numwords) \
                  : "memory" )

// quick & dirty - for debugging only 

void fillblock(char *strt, int yoff, int lines, int val){
  char *i;
  for(i = strt + yoff * vo_dga_width *HW_MODE.vdm_bytespp; 
      i< strt + (lines+yoff) * vo_dga_width *HW_MODE.vdm_bytespp;  ){
    *i++ = val;
  }
}


//---------------------------------------------------------

static uint32_t draw_frame( uint8_t *src[] ){

  int vp_skip = vo_dga_vp_skip;
  int lpl = vo_dga_bytes_per_line >> 2; 
  int numlines = vo_dga_lines;     

  char *s, *d;

  s = *src;
  d = (&((char *)vo_dga_base)[vo_dga_vp_offset + vo_dga_dbf_current * vo_dga_dbf_mem_offset]);
  
  switch(SRC_MODE.vdm_conversion_func){
  case VDM_CONV_NATIVE:

    mem2agpcpy_pic(
    	d, s, 
	vo_dga_bytes_per_line, 
	numlines,
	vo_dga_bytes_per_line+vo_dga_vp_skip, 
	vo_dga_bytes_per_line);
	  
  // DBG-COde

#if 0
  d = (&((char *)vo_dga_base)[vo_dga_vp_offset + vo_dga_dbf_current * vo_dga_dbf_mem_offset]);
  fillblock(d, 0, 10, 0x800000ff);
  fillblock(d, 10, 10, 0x8000ff00);
  fillblock(d, 20, 10, 0x80ff0000);
  fillblock(d, 30, 10, 0xff0000ff);
  fillblock(d, 40, 10, 0x800000ff);
  fillblock(d, 50, 10, 0x0f0000ff);
#endif	  
    break;
  case VDM_CONV_15TO16:
        {
	  int i;
	  char *e;
	  for(i=0; i< vo_dga_lines; i++){
            rgb15to16( s, d, vo_dga_bytes_per_line);
	    d+=vo_dga_bytes_per_line;
	    s+=vo_dga_bytes_per_line;
            d+= vo_dga_vp_skip;
	  }
	}
	break;
  case VDM_CONV_24TO32:

    {
      int i,k,l,m;
      for(i = 0; i< vo_dga_lines; i++ ){
	for(k = 0; k< vo_dga_src_width; k+=2 ){
          l = *(((uint32_t *)s)++);
          m = (l & 0xff000000)>> 24 ;
          *(((uint32_t *)d)++) = (l & 0x00ffffff); // | 0x80000000;
          m |= *(((uint16_t *)s)++) << 8;           
          *(((uint32_t *)d)++) = m; // | 0x80000000 ;
	}
        d+= vp_skip;
      }
    }
    //printf("vo_dga: 24 to 32 not implemented yet!!!\n");
    break;
  }
  return 0;
}

//---------------------------------------------------------

static void check_events(void)
{
  int e=vo_x11_check_events(vo_dga_dpy);
}

//---------------------------------------------------------

#include "sub.h"

static void draw_osd(void)
{ vo_draw_text(vo_dga_src_width,vo_dga_src_height,draw_alpha); }

static void flip_page( void ){

  if(vo_dga_dbf_mem_offset != 0){

#ifdef HAVE_DGA2
    XDGASetViewport (vo_dga_dpy, XDefaultScreen(vo_dga_dpy), 
		    0, vo_dga_dbf_current * vo_dga_dbf_y_offset, 
		    XDGAFlipRetrace);
#else
    XF86DGASetViewPort (vo_dga_dpy, XDefaultScreen(vo_dga_dpy),
		        0, vo_dga_dbf_current * vo_dga_dbf_y_offset);
#endif
    vo_dga_dbf_current = 1 - vo_dga_dbf_current;
  }
}

//---------------------------------------------------------

static uint32_t draw_slice( uint8_t *src[],int stride[],
                            int w,int h,int x,int y )
{
  if (scale_srcW) {
    uint8_t *dst[3] = {vo_dga_base + vo_dga_dbf_current * vo_dga_dbf_mem_offset + vo_dga_vp_offset, NULL, NULL};
    SwScale_YV12slice(src,stride,y,h,
          dst,
          /*scale_dstW*/ vo_dga_width * HW_MODE.vdm_bytespp, HW_MODE.vdm_bitspp,
		      scale_srcW, scale_srcH, scale_dstW, scale_dstH);
  } else {
    yuv2rgb( vo_dga_base + vo_dga_dbf_current * vo_dga_dbf_mem_offset + vo_dga_vp_offset + 
          (vo_dga_width * y +x) * HW_MODE.vdm_bytespp,
           src[0], src[1], src[2],
           w,h, vo_dga_width * HW_MODE.vdm_bytespp,
           stride[0],stride[1] );
  }
  return 0;
};

//---------------------------------------------------------

static const vo_info_t* get_info( void )
{ return &vo_info; }

//---------------------------------------------------------

static uint32_t query_format( uint32_t format )
{

#ifdef HAVE_DGA2	
 XDGAMode *modelines;
 int       modecount;
#endif
 Display  *qdisp;

 int i;
 static int dga_depths_init = 0;

 if(dga_depths_init == 0){

   if((qdisp = XOpenDisplay(0))==NULL){
     vd_printf(VD_ERR, "vo_dga: Can't open display!\n");
     return 0;
   }
   if( !vo_init() ){
    vd_printf(VD_ERR, "vo_dga: vo_init() failed!\n");
    return 1; 
   }
   vo_dga_XServer_mode = vd_ValidateMode(vo_depthonscreen);
 
   if(vo_dga_XServer_mode ==0){
#ifndef HAVE_DGA2
     vd_printf(VD_ERR, "vo_dga: Your X-Server is not running in a ");
     vd_printf(VD_ERR, "resolution supported by DGA driver!\n");
#endif     
   }//else{
   //  vd_printf(VD_INFO, "vo_dga: X running at: %s\n", 
   //            vd_GetModeString(vo_dga_XServer_mode));
   //}                                
 
#ifdef HAVE_DGA2
   modelines=XDGAQueryModes(qdisp, XDefaultScreen(qdisp),&modecount);
   if(modelines){
     for(i=0; i< mode