Polymer: upload bucket indices through pinned host memory.

git-svn-id: https://svn.eduke32.com/eduke32@5298 1a8010ca-5511-0410-912e-c29ae57300e0

polymer/eduke32/build/include/baselayer.h

@@ -78,6 +78,7 @@ struct glinfo_t {
     char occlusionqueries;
     char glsl;
     char debugoutput;
+    char bufferstorage;
     char dumped;
 };
 

polymer/eduke32/build/include/glbuild.h

@@ -492,6 +492,14 @@ extern bglMapBufferARBProcPtr bglMapBufferARB;
 typedef GLboolean   (APIENTRY * bglUnmapBufferARBProcPtr)(GLenum target);
 extern bglUnmapBufferARBProcPtr bglUnmapBufferARB;
 
+// ARB_buffer_storage
+typedef void        (APIENTRY * bglBufferStorageProcPtr)(GLenum target, GLintptrARB size, const GLvoid * data, GLbitfield flags);
+extern bglBufferStorageProcPtr bglBufferStorage;
+
+// ARB_map_buffer_range
+typedef void*       (APIENTRY * bglMapBufferRangeProcPtr) (GLenum target, GLintptr offset, GLsizeiptr length, GLbitfield access);
+extern bglMapBufferRangeProcPtr bglMapBufferRange;
+
 // Occlusion queries
 typedef void (APIENTRY * bglGenQueriesARBProcPtr)(GLsizei n, GLuint *ids);
 extern bglGenQueriesARBProcPtr bglGenQueriesARB;

polymer/eduke32/build/include/polymer.h

@@ -202,6 +202,7 @@ typedef struct      s_prbucket {
     GLuint*          indices;
     uint32_t         count;
     uint32_t         buffersize;
+    GLuint*          indiceoffset;
 
     struct s_prbucket* next;
 }                   _prbucket;

polymer/eduke32/build/src/baselayer.c

@@ -435,6 +435,7 @@ int32_t osdcmd_glinfo(const osdfuncparm_t *parm)
                " Occlusion queries:       %s\n"
                " GLSL:                    %s\n"
                " Debug Output:            %s\n"
+               " Buffer Storage:          %s\n"
 #endif
                " Extensions:\n",
                glinfo.maxanisotropy, glinfo.maxanisotropy>1.0?"":" (no anisotropic filtering)",
@@ -456,7 +457,8 @@ int32_t osdcmd_glinfo(const osdfuncparm_t *parm)
                glinfo.sm4 ? "supported": "not supported",
                glinfo.occlusionqueries ? "supported": "not supported",
                glinfo.glsl ? "supported": "not supported",
-               glinfo.debugoutput ? "supported": "not supported"
+               glinfo.debugoutput ? "supported": "not supported",
+               glinfo.bufferstorage ? "supported" : "not supported"
 #endif
               );
 

polymer/eduke32/build/src/glbuild.c

@@ -178,6 +178,12 @@ bglBufferSubDataARBProcPtr bglBufferSubDataARB;
 bglMapBufferARBProcPtr bglMapBufferARB;
 bglUnmapBufferARBProcPtr bglUnmapBufferARB;
 
+// ARB_buffer_storage
+bglBufferStorageProcPtr bglBufferStorage;
+
+// ARB_map_buffer_range
+bglMapBufferRangeProcPtr bglMapBufferRange;
+
 // Occlusion queries
 bglGenQueriesARBProcPtr bglGenQueriesARB;
 bglDeleteQueriesARBProcPtr bglDeleteQueriesARB;
@@ -579,6 +585,12 @@ int32_t loadglextensions(void)
     bglMapBufferARB = (bglMapBufferARBProcPtr) GETPROCEXTSOFT("glMapBufferARB");
     bglUnmapBufferARB = (bglUnmapBufferARBProcPtr) GETPROCEXTSOFT("glUnmapBufferARB");
 
+    // ARB_buffer_storage
+    bglBufferStorage = (bglBufferStorageProcPtr)GETPROCEXTSOFT("glBufferStorage");
+
+    // ARB_map_buffer_range
+    bglMapBufferRange = (bglMapBufferRangeProcPtr)GETPROCEXTSOFT("glMapBufferRange");
+
     // Occlusion queries
     bglGenQueriesARB = (bglGenQueriesARBProcPtr) GETPROCEXTSOFT("glGenQueriesARB");
     bglDeleteQueriesARB = (bglDeleteQueriesARBProcPtr) GETPROCEXTSOFT("glDeleteQueriesARB");
@@ -886,6 +898,12 @@ int32_t unloadgldriver(void)
     bglMapBufferARB = (bglMapBufferARBProcPtr) NULL;
     bglUnmapBufferARB = (bglUnmapBufferARBProcPtr) NULL;
 
+    // ARB_buffer_storage
+    bglBufferStorage = (bglBufferStorageProcPtr) NULL;
+
+    // ARB_map_buffer_range
+    bglMapBufferRange = (bglMapBufferRangeProcPtr)NULL;
+
     // Occlusion queries
     bglGenQueriesARB = (bglGenQueriesARBProcPtr) NULL;
     bglDeleteQueriesARB = (bglDeleteQueriesARBProcPtr) NULL;

polymer/eduke32/build/src/polymer.c

@@ -74,6 +74,13 @@ const GLsizeiptrARB proneplanesize = sizeof(_prvert) * 4;
 const GLintptrARB prwalldatasize = sizeof(_prvert)* 4 * 3; // wall, over and mask planes for every wall
 GLintptrARB prwalldataoffset;
 
+GLuint          prindexringvbo;
+GLuint          *prindexring;
+const GLsizeiptrARB prindexringsize = 65535;
+GLintptrARB prindexringoffset;
+
+const GLbitfield prindexringmapflags = GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT;
+
 _prbucket       *prbuckethead;
 int32_t         prcanbucket;
 
@@ -846,7 +853,7 @@ int32_t             polymer_init(void)
     if (glinfo.debugoutput) {
         // Enable everything.
         bglDebugMessageControlARB(GL_DONT_CARE, GL_DONT_CARE, GL_DONT_CARE, 0, NULL, GL_TRUE);
-        bglDebugMessageCallbackARB(polymer_debugoutputcallback, NULL);
+        bglDebugMessageCallbackARB((GLDEBUGPROCARB)polymer_debugoutputcallback, NULL);
         bglEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
     }
 #endif
@@ -1023,6 +1030,14 @@ void                polymer_loadboard(void)
     bglBufferDataARB(GL_ARRAY_BUFFER_ARB, prwalldataoffset + (numwalls * prwalldatasize), NULL, mapvbousage);
     bglBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
 
+    bglGenBuffersARB(1, &prindexringvbo);
+    bglBindBufferARB(GL_ELEMENT_ARRAY_BUFFER, prindexringvbo);
+
+    bglBufferStorage(GL_ELEMENT_ARRAY_BUFFER, prindexringsize * sizeof(GLuint), NULL, prindexringmapflags | GL_DYNAMIC_STORAGE_BIT);
+    prindexring = (GLuint*)bglMapBufferRange(GL_ELEMENT_ARRAY_BUFFER, 0, prindexringsize * sizeof(GLuint), prindexringmapflags);
+
+    bglBindBufferARB(GL_ELEMENT_ARRAY_BUFFER, 0);
+
     i = 0;
     while (i < numsectors)
     {
@@ -1815,7 +1830,7 @@ static void         polymer_displayrooms(const int16_t dacursectnum)
         {
             // if we have a level boundary somewhere in the sector,
             // consider these walls as visportals
-            if (wall[sec->wallptr + i].nextsector < 0)
+            if (wall[sec->wallptr + i].nextsector < 0 && pr_buckets == 0)
                 doquery = 1;
         }
         while (--i >= 0);
@@ -2123,6 +2138,46 @@ static void         polymer_emptybuckets(void)
     bglVertexPointer(3, GL_FLOAT, sizeof(_prvert), NULL);
     bglTexCoordPointer(2, GL_FLOAT, sizeof(_prvert), (GLvoid *)(3 * sizeof(GLfloat)));
 
+    bglBindBufferARB(GL_ELEMENT_ARRAY_BUFFER, prindexringvbo);
+
+    uint32_t indexcount = 0;
+    while (bucket != NULL)
+    {
+        indexcount += bucket->count;
+
+        bucket = bucket->next;
+    }
+
+    // ensure space in index ring, wrap otherwise
+    if (indexcount + prindexringoffset >= prindexringsize)
+    {
+        bglUnmapBufferARB(GL_ELEMENT_ARRAY_BUFFER);
+        prindexring = (GLuint *)bglMapBufferRange(GL_ELEMENT_ARRAY_BUFFER, 0, prindexringsize * sizeof(GLuint), GL_MAP_INVALIDATE_BUFFER_BIT | prindexringmapflags);
+        prindexringoffset = 0;
+    }
+
+    // put indices in the ring, all at once
+    bucket = prbuckethead;
+
+    while (bucket != NULL)
+    {
+        if (bucket->count == 0)
+        {
+            bucket = bucket->next;
+            continue;
+        }
+
+        memcpy(&prindexring[prindexringoffset], bucket->indices, bucket->count * sizeof(GLuint));
+
+        bucket->indiceoffset = (GLuint*)(prindexringoffset * sizeof(GLuint));
+
+        prindexringoffset += bucket->count;
+
+        bucket = bucket->next;
+    }
+
+    bucket = prbuckethead;
+
     while (bucket != NULL)
     {
         if (bucket->count == 0)
@@ -2133,7 +2188,7 @@ static void         polymer_emptybuckets(void)
 
         int32_t materialbits = polymer_bindmaterial(&bucket->material, NULL, 0);
 
-        bglDrawElements(GL_TRIANGLES, bucket->count, GL_UNSIGNED_INT, bucket->indices);
+        bglDrawElements(GL_TRIANGLES, bucket->count, GL_UNSIGNED_INT, bucket->indiceoffset);
 
         polymer_unbindmaterial(materialbits);
 
@@ -2142,6 +2197,7 @@ static void         polymer_emptybuckets(void)
         bucket = bucket->next;
     }
 
+    bglBindBufferARB(GL_ELEMENT_ARRAY_BUFFER, 0);
     bglBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
 
     prcanbucket = 0;

polymer/eduke32/build/src/sdlayer.c

@@ -1318,6 +1318,7 @@ void sdlayer_setvideomode_opengl(void)
     glinfo.occlusionqueries = !!Bstrstr(glinfo.extensions, "GL_ARB_occlusion_query");
     glinfo.glsl = !!Bstrstr(glinfo.extensions, "GL_ARB_shader_objects");
     glinfo.debugoutput = !!Bstrstr(glinfo.extensions, "GL_ARB_debug_output");
+    glinfo.bufferstorage = !!Bstrstr(glinfo.extensions, "GL_ARB_buffer_storage");
 
     if (Bstrstr(glinfo.extensions, "WGL_3DFX_gamma_control"))
     {

polymer/eduke32/build/src/winlayer.c

@@ -2939,6 +2939,10 @@ static int32_t SetupOpenGL(int32_t width, int32_t height, int32_t bitspp)
             {
                 glinfo.debugoutput = 1;
             }
+            else if (!Bstrcmp((char *)p2, "GL_ARB_buffer_storage"))
+            {
+                glinfo.bufferstorage = 1;
+            }
         }
         Bfree(p);
     }