introducing NVidia accelerated OpenGL videosinks :)

Original commit message from CVS:
introducing NVidia accelerated OpenGL videosinks :)

1 files changed, 134 insertions, 122 deletions

diff --git a/sys/glsink/gstgl_nvimage.c b/sys/glsink/gstgl_nvimage.c
index faa5224c..61d18749 100644
--- a/sys/glsink/gstgl_nvimage.c
+++ b/sys/glsink/gstgl_nvimage.c
@@ -13,19 +13,18 @@
 #include <stdlib.h>
 #include <sys/types.h>
 
-#define GL_GLEXT_PROTOTYPES
-
-// VERY dangerous:
-#define GL_WRITE_PIXEL_DATA_RANGE_NV 0x8878
 
 #include <X11/Xlib.h>
 #include <X11/Xutil.h>
 #include <GL/glx.h>
 #include <GL/gl.h>
-#include <GL/glext.h>
+//#include <GL/glext.h>
 #include <GL/glu.h>
 #include <string.h>
 
+// too lazy to write an API for this ;) 
+#include "regcomb_yuvrgb.c"
+
 #include "gstglsink.h"
 
 typedef struct _GstGLImageConnection GstGLImageConnection;
@@ -102,8 +101,12 @@ gst_gl_nvimage_connection (GstImageConnection *conn)
 gboolean
 gst_gl_nvimage_check_xvideo ()
 {
-  int ver, rel, req, ev, err;
+  //int ver, rel, req, ev, err;
   
+  if (!GL_ARB_multitexture_Init()) return FALSE;
+  if (!GL_EXT_paletted_texture_Init()) return FALSE;
+  if (!GL_NV_register_combiners_Init()) return FALSE;
+
 #if 0 
   if (display == NULL)
     return FALSE;
@@ -111,15 +114,15 @@ gst_gl_nvimage_check_xvideo ()
     return TRUE;
 #endif
 
-  return FALSE;
+  return TRUE;
 }
 
 static GstCaps * 	
 gst_gl_nvimage_get_caps (GstImageInfo *info)
 {
-  gint i;
-  int adaptors;
-  int formats;
+  //gint i;
+  //int adaptors;
+  //int formats;
   GstCaps *caps = NULL;
   GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);
   
@@ -128,14 +131,14 @@ gst_gl_nvimage_get_caps (GstImageInfo *info)
 
   if (gst_gl_nvimage_check_xvideo () == FALSE)
   {
-    g_warning("GL_NVImage: Server has no NVidia extension support\n");
+    g_warning("GL_NVImage: Server has no NVidia extension support\n"); 
     return NULL;
   }
 
   caps = gst_caps_append (caps, GST_CAPS_NEW (
 					      "xvimage_caps",
 					      "video/raw",
-					      "format",  GST_PROPS_FOURCC (GST_MAKE_FOURCC ('Y', 'C', '1', '2')),
+					      "format",  GST_PROPS_FOURCC (GST_MAKE_FOURCC ('Y', 'V', '1', '2')),
 					      "width",   GST_PROPS_INT_RANGE (0, 1024),
 					      "height",  GST_PROPS_INT_RANGE (0, 1024))
 			  );
@@ -145,9 +148,9 @@ gst_gl_nvimage_get_caps (GstImageInfo *info)
 static GstImageConnection *
 gst_gl_nvimage_set_caps (GstImageInfo *info, GstCaps *caps)
 {
-  gint i, j = 0;
-  int adaptors;
-  int formats;
+  //gint i, j = 0;
+  //int adaptors;
+  //int formats;
   GstGLImageConnection *conn;
   GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);
   guint32 format;
@@ -166,6 +169,12 @@ gst_gl_nvimage_set_caps (GstImageInfo *info, GstCaps *caps)
 		  "format", &format,
                   NULL);
 
+  // maybe I should a bit more checking here, e.g. maximum size smaller than maximum texture extents
+  if (format != GST_MAKE_FOURCC ('Y', 'V', '1', '2')) 
+    { 
+      GST_DEBUG (GST_CAT_PLUGIN_INFO, "GL_NVImage: Format is invalid !\n");
+      return NULL;
+    }
   if (0) //conn->port == (XvPortID) -1)
   {
     /* this happens if the plugin can't handle the caps, so no warning */
@@ -173,6 +182,7 @@ gst_gl_nvimage_set_caps (GstImageInfo *info, GstCaps *caps)
     return NULL;
   }
   
+  GST_DEBUG (GST_CAT_PLUGIN_INFO, "GL_NVImage: caps %p are ok, creating image", caps);
   return (GstImageConnection *) conn;
 }
 
@@ -182,35 +192,22 @@ gst_gl_nvimage_get_image (GstImageInfo *info, GstImageConnection *conn)
   GstNvImage *image;
   GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);  
   GstGLImageConnection *nvconn = gst_gl_nvimage_connection (conn);  
-  int slot = 0;
   
   /* checks */
   if (xinfo == NULL) return NULL;
   if (nvconn == NULL) return NULL;
 
-  // I should also check the current GLX context ! Don't have to, I am guarantueed to always be in the same thread
+  // I should also check the current GLX context ! 
+  // Ah, Don't have to, I am guarantueed to always be in the same thread
 
   image = g_new0(GstNvImage, 1);
 
-  for (slot = 0; slot < AGP_BUFSLOTS; slot++)
-    {
-      if (!nvconn->m_bufslots[slot]) break;
-    }
-
   image->data.size = nvconn->w * nvconn->h * 3/2;
 
-  if (slot < AGP_BUFSLOTS) // found an AGP buffer slot
-    {
-      image->data.data = nvconn->m_memory + slot * YUVTEX_SIZE; 
-      image->slot = slot;  // store for freeing
-      nvconn->m_bufslots[slot] = 1; // it is now taken
-    }
-  else
-    {
-      g_warning("Allocating from main memory !");
-      image->data.data = g_malloc(image->data.size);
-      image->slot = AGP_BUFSLOTS; // no AGP slot
-    }
+  //g_warning("Allocating %d bytes from main memory !", image->data.size);
+  image->data.data = g_malloc(image->data.size);
+  //image->slot = AGP_BUFSLOTS; // no AGP slot
+
   image->conn = nvconn;
 
   if (image->data.data == NULL)
@@ -229,52 +226,125 @@ gst_gl_nvimage_put_image (GstImageInfo *info, GstImageData *image)
   GstNvImage *im = (GstNvImage *) image;
   GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);  
   
+  int img_width = im->conn->w; 
+  int img_height = im->conn->h; 
+  int uv_width = img_width >> 1;
+  int uv_height = img_height >> 1;
+  
+  unsigned char *buf_y = im->data.data;
+  unsigned char *buf_v = (buf_y + img_width * img_height);
+  unsigned char *buf_u = buf_v + ((img_width/2) * (img_height/2));      
+
   /* checks omitted for speed (and lazyness), do we need them? */
   g_assert (xinfo != NULL);
   
-  /* Upload the texture here */
-  g_warning("PUTTING IMAGE - BROOOKEN");
-
   // both upload the video, and redraw the screen
   glClear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT);
 
+  glMatrixMode(GL_MODELVIEW);
+  glLoadIdentity();
+  glTranslatef(0.0, 0.0, -50.0);
   glEnable(GL_TEXTURE_2D);
 
   glPushMatrix();
   //glTranslatef(0,1,0);
-  glRotatef(xinfo->rotX-250,1,0,0);
+  glRotatef(xinfo->rotX,1,0,0);
   glRotatef(xinfo->rotY,0,1,0);
-  int zoom = xinfo->zoom;
-  glScaled(zoom,zoom,zoom);
-  //Draws the surface rectangle
 
-  glBindTexture(GL_TEXTURE_2D, im->conn->ytex_id);
-  glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, im->conn->w, im->conn->h, GL_RGBA, 
-  		  GL_UNSIGNED_BYTE, im->data.data);  
-  float xmax = (float)im->conn->w/TEX_XSIZE;
-  float ymax = (float)im->conn->h/TEX_YSIZE;
+  if (xinfo->info.demo)
+    {
+      //xinfo->rotX += 0.2;
+      //xinfo->rotY -= 0.2;
+      float zoom = xinfo->zoom;
+      glScalef(zoom,zoom,zoom); 
+      
+      if (xinfo->zoom > 2.0)
+	xinfo->zoomdir = -0.05;
+
+      if (xinfo->zoom < 1.0)
+	xinfo->zoomdir = 0.05;
 
-  glColor4f(1,1,1,1);
-  glBegin(GL_QUADS);
+      xinfo->zoom += xinfo->zoomdir;
+    }  
 
-  glNormal3f(0, -1, 0);
-  
-  glTexCoord2f(xmax, 0);
-  glVertex3f(4,0,-4);
+  //Draws the surface rectangle
+
+  if (Ywidth != im->conn->w || Yheight != im->conn->h)
+  {
+    Ywidth = im->conn->w; Yheight = im->conn->h;  UVwidth = im->conn->w/2; UVheight = im->conn->h/2;
+    Initialize_Backend(Ywidth,Yheight,UVwidth,UVheight,GL_LINEAR);
+  }
 
-  glTexCoord2f(0, 0);
-  glVertex3f(-4,0,-4);
+  LoadYUVPlanes(Yhandle,Uhandle,Vhandle,img_width,img_height,uv_width,uv_height,buf_y,buf_u,buf_v);
+  float xmax = (float)(im->conn->w-1)/tex_xsize;
+  float ymax = (float)(im->conn->h-1)/tex_ysize;
 
-  glTexCoord2f(0, ymax);
-  glVertex3f(-4,0,4);
+  /* Upload the texture here */
+  //g_warning("PUTTING IMAGE %f %f %d %d\n", xmax, ymax, tex_xsize, tex_ysize);
 
-  glTexCoord2f(xmax, ymax);
-  glVertex3f(4,0,4);
+  //glColor4f(1,1,1,1); // do NOT set a color here ! Done by Initialize_Backend, or actually SetConst !
+  glBegin(GL_QUADS);
+
+  glNormal3f(0, -1, 0);
+  glMultiTexCoord2fARB(GL_TEXTURE0_ARB,0,0); 
+  glMultiTexCoord2fARB(GL_TEXTURE1_ARB,0,0); 
+  glMultiTexCoord2fARB(GL_TEXTURE2_ARB,0,0); 
+  glVertex3f(-4,4,0);
+
+  glMultiTexCoord2fARB(GL_TEXTURE0_ARB,0,ymax); 
+  glMultiTexCoord2fARB(GL_TEXTURE1_ARB,0,ymax); 
+  glMultiTexCoord2fARB(GL_TEXTURE2_ARB,0,ymax); 
+  glVertex3f(-4,-4,0);
+
+  glMultiTexCoord2fARB(GL_TEXTURE0_ARB,xmax,ymax); 
+  glMultiTexCoord2fARB(GL_TEXTURE1_ARB,xmax,ymax); 
+  glMultiTexCoord2fARB(GL_TEXTURE2_ARB,xmax,ymax); 
+  glVertex3f(4,-4,0);
+
+  glMultiTexCoord2fARB(GL_TEXTURE0_ARB,xmax,0); 
+  glMultiTexCoord2fARB(GL_TEXTURE1_ARB,xmax,0); 
+  glMultiTexCoord2fARB(GL_TEXTURE2_ARB,xmax,0); 
+  glVertex3f(4,4,0);
 
   glEnd();
 
   glPopMatrix();
 
+  if (xinfo->info.dumpvideo)
+    {
+      static int framenr = 0; 
+      char capfilename[255];
+      static guint8 *cap_image_data = NULL, *cap_image_data2 = NULL;
+      int i;
+      
+      // hmmmm, is this reentrant ?!
+      if (cap_image_data == NULL)
+	cap_image_data = (guint8 *)malloc(img_width * img_height * 3);
+
+      if (cap_image_data2 == NULL)
+	cap_image_data2 = (guint8 *)malloc(img_width * img_height * 3);
+      
+      printf("Recording frame #%d\n", framenr);
+      glReadPixels(0,0,img_width,img_height,GL_RGB,GL_UNSIGNED_BYTE,cap_image_data);
+      // invert the pixels
+      for (i = 0; i < img_height; i++)
+	memcpy(cap_image_data2 + i * img_width * 3, cap_image_data + (img_height-1-i) * img_width * 3, img_width*3);
+      
+      sprintf(capfilename, "cap%04d.ppm", framenr);
+      FILE *outfile = fopen(capfilename, "wb");
+      if (outfile != NULL)
+	{
+	  fprintf(outfile, "P6\n"); 
+	  fprintf(outfile,"# created by raw_zb\n"); 
+	  fprintf(outfile,"%d %d\n",img_width,img_height); 
+	  fprintf(outfile,"255\n"); 
+	  fwrite(cap_image_data2, sizeof(char), img_width*img_height*3, outfile);
+	  fclose(outfile);
+	}
+      framenr++;
+    }
+
+
   glXSwapBuffers(xinfo->dpy, xinfo->win);
 }
 
@@ -301,76 +371,18 @@ gst_gl_nvimage_open_conn (GstImageConnection *conn, GstImageInfo *info)
   GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);  
   GstGLImageConnection *xconn = gst_gl_nvimage_connection (conn);
 
-  unsigned char data_sep[2][2] = {{0, 255}, {0, 255}}; 
-  int slot;
-
-  g_warning("Opening NVidia Connection");
-  xconn->m_memory = (unsigned char*)glXAllocateMemoryNV(AGP_BUFSLOTS*YUVTEX_SIZE, 0, 1.0, 1.0);
-  
-  if (!xconn->m_memory)
-    {
-      printf("Unable to acquire graphics card mem... will acquire in normal memory.\n");
-      for (slot = 0; slot < AGP_BUFSLOTS; slot++)
-	xconn->m_bufslots[slot] = 1;
-    }
-  else
-    {
-      // maybe this fast writable memory, awfully slow to read from, though
-      glPixelDataRangeNV(GL_WRITE_PIXEL_DATA_RANGE_NV, AGP_BUFSLOTS*YUVTEX_SIZE, xconn->m_memory);
-      glEnableClientState(GL_WRITE_PIXEL_DATA_RANGE_NV);
-
-      for (slot = 0; slot < AGP_BUFSLOTS; slot++)
-	xconn->m_bufslots[slot] = 0;
-    }
-
-  glGenTextures(1, &xconn->ytex_id);
-  glBindTexture(GL_TEXTURE_2D, xconn->ytex_id);
-  glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
-  glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_REPLACE);
-  glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8_ALPHA8, TEX_XSIZE, TEX_YSIZE, 0, GL_LUMINANCE_ALPHA, GL_UNSIGNED_BYTE, NULL);
-
-  glActiveTextureARB(GL_TEXTURE1_ARB);
-  glGenTextures(1, &xconn->uvtex_id);
-  glBindTexture(GL_TEXTURE_2D, xconn->uvtex_id);
-  glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
-  glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8_ALPHA8, TEX_XSIZE/2, TEX_YSIZE/2, 0, GL_LUMINANCE_ALPHA, GL_UNSIGNED_BYTE, NULL);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
-  glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_ADD);
-
-  glActiveTextureARB(GL_TEXTURE2_ARB);
-  glGenTextures(1, &xconn->septex_id);
-  glBindTexture(GL_TEXTURE_2D, xconn->septex_id);
-  glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
-  glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8, 2, 2, 0, GL_LUMINANCE, GL_UNSIGNED_BYTE, data_sep);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
-  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
-  glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_ADD);
-
-  glFlushPixelDataRangeNV(GL_WRITE_PIXEL_DATA_RANGE_NV);
-  //glEnable(GL_TEXTURE_2D);
-  glActiveTextureARB(GL_TEXTURE0_ARB);
-  glEnable(GL_TEXTURE_2D);
-  glActiveTextureARB(GL_TEXTURE1_ARB);
-  glEnable(GL_TEXTURE_2D);
-  glActiveTextureARB(GL_TEXTURE2_ARB);
-  glEnable(GL_TEXTURE_2D);
-  glActiveTextureARB(GL_TEXTURE0_ARB);
+  g_warning("Opening NVidia GL context connection, using register combiners.\n");
+  {
+    Ywidth = TEX_XSIZE; Yheight = TEX_YSIZE;  UVwidth = TEX_XSIZE/2; UVheight = TEX_YSIZE/2;
+    Initialize_Backend(Ywidth,Yheight,UVwidth,UVheight,GL_LINEAR);
+  }
 }
 
 static void
 gst_gl_nvimage_close_conn (GstImageConnection *conn, GstImageInfo *info)
 {
   GstGLImageConnection *xconn = gst_gl_nvimage_connection (conn);
-  GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);  
+  //GstGLImageInfo *xinfo = gst_gl_nvimage_info (info);  
 
   // anything needed in here ? Oh, maybe drawing de-init, or something
   glDeleteTextures(1, &xconn->ytex_id);