Text archives Help

---

Author: knolla
Date: Sun Jun 11 02:48:54 2006
New Revision: 1116

Added:
   trunk/Interface/SSERayPacket.h
Modified:
   trunk/Core/Math/SSEDefs.h
   trunk/Interface/RayPacket.h
   trunk/Model/Intersections/IsosurfaceImplicit.cc
   trunk/Model/Intersections/IsosurfaceImplicit.h
   trunk/Model/Primitives/IsosurfaceOctreeVolume.cc
   trunk/Model/Primitives/IsosurfaceOctreeVolume.h
   trunk/Model/Primitives/OctreeVolume.h
Log:
Early ray termination working in IsosurfaceOctreeVolume; slight speedup but 
still only marginally faster than single-ray. Far too much traversal time 
spent.

Modified: trunk/Core/Math/SSEDefs.h
==============================================================================
--- trunk/Core/Math/SSEDefs.h   (original)
+++ trunk/Core/Math/SSEDefs.h   Sun Jun 11 02:48:54 2006
@@ -253,7 +253,7 @@
         int nonzeros = 0;
         #pragma unroll(4)
         for(int i=0; i<4; i++)
-            nonzeros += (mask & (1 << (i<<2))) ? 1 : 0;
+            nonzeros += (mask & (1 << i)) ? 1 : 0;
         return nonzeros;
     }
     
@@ -263,7 +263,7 @@
         int zeros = 0;
         #pragma unroll(4)
         for(int i=0; i<4; i++)
-            zeros += (mask & (1 << (i<<2))) ? 0 : 1;
+            zeros += (mask & (1 << i)) ? 0 : 1;
         return zeros;
     }
 

Modified: trunk/Interface/RayPacket.h
==============================================================================
--- trunk/Interface/RayPacket.h (original)
+++ trunk/Interface/RayPacket.h Sun Jun 11 02:48:54 2006
@@ -78,22 +78,6 @@
     // Char-based arrays
     char scratchpad_data[MaxSize][MaxScratchpadSize];
   };
-  
-#ifdef MANTA_SSE
-  //A wrapper structure for a typical SSE/packet traversal, such as a BVH
-  struct MANTA_ALIGN(16) SSERayPacket
-  {
-    sse_t activeMask[RayPacketData::SSE_MaxSize];
-    int activeRays;
-    sse_t* orig[3];
-    sse_t* dir[3];
-    sse_t* inv_dir[3];
-    sse_t* signs[3];
-    sse_t* normal[3];
-    sse_t* minT;
-    RayPacket* rp;
-  };
-#endif  
 
   class RayPacket {
   public:

Added: trunk/Interface/SSERayPacket.h
==============================================================================
--- (empty file)
+++ trunk/Interface/SSERayPacket.h      Sun Jun 11 02:48:54 2006
@@ -0,0 +1,29 @@
+
+//A wrapper structure for a typical SSE/packet traversal, such as a BVH
+
+#ifndef Manta_SSERayPacket_h
+#define Manta_SSERayPacket_h
+
+#ifdef MANTA_SSE
+
+#include <Interface/RayPacket.h>
+
+namespace Manta {
+
+  struct MANTA_ALIGN(16) SSERayPacket
+  {
+    sse_t activeMask[RayPacketData::SSE_MaxSize];
+    int activeRays;
+    sse_t* orig[3];
+    sse_t* dir[3];
+    sse_t* inv_dir[3];
+    sse_t* signs[3];
+    sse_t* normal[3];
+    sse_t* minT;
+    RayPacket* rp;
+  };
+}
+
+#endif // MANTA_SSE
+
+#endif

Modified: trunk/Model/Intersections/IsosurfaceImplicit.cc
==============================================================================
--- trunk/Model/Intersections/IsosurfaceImplicit.cc     (original)
+++ trunk/Model/Intersections/IsosurfaceImplicit.cc     Sun Jun 11 02:48:54 
2006
@@ -220,7 +220,8 @@
             //int nonzeros = count_nonzeros(sse_thisvoxelmask);
             //cerr << "nonzeros=" << nonzeros << endl;
             
-            //srp.activeRays -= count_nonzeros(sse_thisvoxelmask);
+            //this line of code should implement early termination, but this 
is broken.
+            srp.activeRays -= count_nonzeros(sse_thisvoxelmask);
             
             //active rays in this smd are ones that were active before, and 
did NOT intersect.
             srp.activeMask[smd] = andnot4(sse_thisvoxelmask, 
srp.activeMask[smd]);

Modified: trunk/Model/Intersections/IsosurfaceImplicit.h
==============================================================================
--- trunk/Model/Intersections/IsosurfaceImplicit.h      (original)
+++ trunk/Model/Intersections/IsosurfaceImplicit.h      Sun Jun 11 02:48:54 
2006
@@ -10,6 +10,7 @@
 
 #ifdef MANTA_SSE
 #include <Core/Math/SSEDefs.h>
+#include <Interface/SSERayPacket.h>
 #endif
 
 namespace Manta

Modified: trunk/Model/Primitives/IsosurfaceOctreeVolume.cc
==============================================================================
--- trunk/Model/Primitives/IsosurfaceOctreeVolume.cc    (original)
+++ trunk/Model/Primitives/IsosurfaceOctreeVolume.cc    Sun Jun 11 02:48:54 
2006
@@ -11,7 +11,7 @@
 #include <Core/Math/SSEDefs.h>
 #endif
 
-#define USE_OCTREE_DATA
+//#define USE_OCTREE_DATA
 
 #define MIN4(a,b,c,d) min(min(a,b), min(c,d));
 #define MAX4(a,b,c,d) max(max(a,b), max(c,d));
@@ -114,6 +114,7 @@
 void IsosurfaceOctreeVolume::intersect(RenderContext const &context, 
RayPacket &packet) const
 {
 #ifdef MANTA_SSE
+//#if 1
        packet_intersect_implicit_bvh(packet);
 #else
     for ( int i = packet.rayBegin; i < packet.rayEnd; i++ )
@@ -912,7 +913,7 @@
                                 bvh_octnode(srp, newfirst, last, child_cell, 
stop_depth, depth+1, child_idx, index_trace);
                             }
                         }
-                        if (srp.activeRays==0)
+                        if (srp.activeRays<=0)
                             return;
                     }
 
@@ -1116,7 +1117,7 @@
                         bvh_octleaf(srp, newfirst, last, child_cell, 
stop_depth, depth+1, 
                             leaf_depth, leaf_value, leaf_base_cell, 
index_trace);
                     }
-                    if (srp.activeRays==0)
+                    if (srp.activeRays<=0)
                         return;
                 }
             }
@@ -1186,7 +1187,7 @@
                 Vector cmax(child_cell.data[0]+1, child_cell.data[1]+1, 
child_cell.data[2]+1);
                 char newfirst, newlast;
                 intersect_cap_octant(srp, first, last, newfirst, newlast, 
cmin, cmax, child_tenter, child_texit, hitmask);
-#if 1
+#ifndef IOV_BOX_CELLS
                 if (newfirst > newlast)
                     continue;
                     
@@ -1221,7 +1222,7 @@
                     //cerr << "in cap " << (unsigned long)(&cap) << ", 
octant " << target_child << endl;
                     IsosurfaceImplicit::sse_intersect(srp, newfirst, 
newlast, cmin, cmax, rho, 
                         octdata->get_isovalue(), child_tenter, child_texit, 
hitmask, this, PrimitiveCommon::getMaterial());
-                    if (srp.activeRays==0)
+                    if (srp.activeRays<=0)
                         return;
                 }
 #endif                

Modified: trunk/Model/Primitives/IsosurfaceOctreeVolume.h
==============================================================================
--- trunk/Model/Primitives/IsosurfaceOctreeVolume.h     (original)
+++ trunk/Model/Primitives/IsosurfaceOctreeVolume.h     Sun Jun 11 02:48:54 
2006
@@ -14,8 +14,12 @@
 
 #ifdef MANTA_SSE
 #include <Core/Math/SSEDefs.h>
+#include <Interface/SSERayPacket.h>
 #endif
 
+//define this to show box cells, not the cap-value cells (and the implicit)
+//#define IOV_BOX_CELLS
+
 
 namespace Manta
 {
@@ -173,6 +177,9 @@
             {
                 sse_t boxmin[3];
                 sse_t boxmax[3];
+                sse_t tnear[3];
+                sse_t tfar[3];
+                sse_t dgt0[3];
                 #pragma unroll(3)
                 for(int axis=0; axis<3; axis++)
                 {
@@ -187,31 +194,31 @@
                     tenter[smd] = zero4();
                     texit[smd] = srp.minT[smd];
                     
-                    sse_t signs = cmp4_ge(srp.dir[0][smd],zero4());
-                    const sse_t b0_x = mask4(signs, boxmin[0], boxmax[0]);
-                    const sse_t b1_x = mask4(signs, boxmax[0], boxmin[0]);
-                    signs = cmp4_ge(srp.dir[1][smd],zero4());
-                    const sse_t b0_y = mask4(signs, boxmin[1], boxmax[1]);
-                    const sse_t b1_y = mask4(signs, boxmax[1], boxmin[1]);
-                    signs = cmp4_ge(srp.dir[2][smd],zero4());
-                    const sse_t b0_z = mask4(signs, boxmin[2], boxmax[2]);
-                    const sse_t b1_z = mask4(signs, boxmax[2], boxmin[2]);
-                    
-                    const sse_t tBoxNearX = mul4(sub4(b0_x, 
srp.orig[0][smd]), srp.inv_dir[0][smd]);
-                    const sse_t tBoxNearY = mul4(sub4(b0_y, 
srp.orig[1][smd]), srp.inv_dir[1][smd]);
-                    const sse_t tBoxNearZ = mul4(sub4(b0_z, 
srp.orig[2][smd]), srp.inv_dir[2][smd]);
-
-                    tenter[smd] = max4(tenter[smd],tBoxNearX);
-                    tenter[smd] = max4(tenter[smd],tBoxNearY);
-                    tenter[smd] = max4(tenter[smd],tBoxNearZ);
-
-                    const sse_t tBoxFarX = mul4(sub4(b1_x, 
srp.orig[0][smd]), srp.inv_dir[0][smd]);
-                    const sse_t tBoxFarY = mul4(sub4(b1_y, 
srp.orig[1][smd]), srp.inv_dir[1][smd]);
-                    const sse_t tBoxFarZ = mul4(sub4(b1_z, 
srp.orig[2][smd]), srp.inv_dir[2][smd]);
-
-                    texit[smd] = min4(texit[smd],tBoxFarX);
-                    texit[smd] = min4(texit[smd],tBoxFarY);
-                    texit[smd] = min4(texit[smd],tBoxFarZ);
+                    dgt0[0] = cmp4_ge(srp.dir[0][smd],zero4());
+                    const sse_t b0_x = mask4(dgt0[0], boxmin[0], boxmax[0]);
+                    const sse_t b1_x = mask4(dgt0[0], boxmax[0], boxmin[0]);
+                    dgt0[1] = cmp4_ge(srp.dir[1][smd],zero4());
+                    const sse_t b0_y = mask4(dgt0[1], boxmin[1], boxmax[1]);
+                    const sse_t b1_y = mask4(dgt0[1], boxmax[1], boxmin[1]);
+                    dgt0[2] = cmp4_ge(srp.dir[2][smd],zero4());
+                    const sse_t b0_z = mask4(dgt0[2], boxmin[2], boxmax[2]);
+                    const sse_t b1_z = mask4(dgt0[2], boxmax[2], boxmin[2]);
+                    
+                    tnear[0] = mul4(sub4(b0_x, srp.orig[0][smd]), 
srp.inv_dir[0][smd]);
+                    tnear[1] = mul4(sub4(b0_y, srp.orig[1][smd]), 
srp.inv_dir[1][smd]);
+                    tnear[2] = mul4(sub4(b0_z, srp.orig[2][smd]), 
srp.inv_dir[2][smd]);
+
+                    tenter[smd] = max4(tenter[smd],tnear[0]);
+                    tenter[smd] = max4(tenter[smd],tnear[1]);
+                    tenter[smd] = max4(tenter[smd],tnear[2]);
+
+                    tfar[0] = mul4(sub4(b1_x, srp.orig[0][smd]), 
srp.inv_dir[0][smd]);
+                    tfar[1] = mul4(sub4(b1_y, srp.orig[1][smd]), 
srp.inv_dir[1][smd]);
+                    tfar[2] = mul4(sub4(b1_z, srp.orig[2][smd]), 
srp.inv_dir[2][smd]);
+
+                    texit[smd] = min4(texit[smd],tfar[0]);
+                    texit[smd] = min4(texit[smd],tfar[1]);
+                    texit[smd] = min4(texit[smd],tfar[2]);
                     
                     hitmask[smd] = cmp4_lt(tenter[smd], texit[smd]);
                     if (_mm_movemask_ps(hitmask[smd]))
@@ -220,7 +227,7 @@
                         newlast = smd;
                     } 
                     
-#if 0
+#ifdef IOV_BOX_CELLS
                     sse_t hitmask2 = and4(hitmask[smd], cmp4_lt(tenter[smd], 
srp.minT[smd]));
                     srp.minT[smd] = mask4(hitmask2, tenter[smd], 
srp.minT[smd]);
                     if (_mm_movemask_ps(hitmask2))

Modified: trunk/Model/Primitives/OctreeVolume.h
==============================================================================
--- trunk/Model/Primitives/OctreeVolume.h       (original)
+++ trunk/Model/Primitives/OctreeVolume.h       Sun Jun 11 02:48:54 2006
@@ -33,7 +33,7 @@
 #endif
 
 //enable this only when testing octree data and original data side by side
-//#define TEST_INDATA
+#define TEST_INDATA
 
 //enable dynamic multires via the "stop_depth"
 //#define OCTVOL_DYNAMIC_MULTIRES

---

• [MANTA] r1116 - in trunk: Core/Math Interface Model/Intersections Model/Primitives, knolla, 06/11/2006

---

Archive powered by MHonArc 2.6.16.