vsm: page hiz culling

#681

game/gothic.cpp

@@ -88,7 +88,7 @@ Gothic::Gothic() {
     opts.doBindless = CommandLine::inst().isBindless();
     }
 
-  if(hasBindless() && gpu.compute.maxInvocations>=1024 && gpu.memoryModel) {
+  if(hasBindless() && gpu.compute.maxInvocations>=1024) {
     //TODO: productize
     opts.doVirtualShadow = CommandLine::inst().isVirtualShadow();
     }

shader/virtual_shadow/vsm_clear.comp

@@ -23,6 +23,7 @@ void main() {
     return;
 
   imageStore(pageTbl, at, uvec4(0));
+  // imageStore(pageTblDepth, at, uvec4(floatBitsToUint(1)));
   imageStore(pageTblDepth, at, uvec4(0xFFFFFFFF));
 
   if(at==ivec3(0)) {

shader/virtual_shadow/vsm_clump_pages.comp

@@ -76,6 +76,10 @@ void main() {
   const uint  lane = gl_LocalInvocationIndex;
 
   poolPageTable();
+  // if(true) {
+  //   storePageTable();
+  //   return;
+  //   }
 
   if(lane < minX.length()) {
     minX  [lane] = 0xFFFF;

shader/virtual_shadow/vsm_cluster_task.comp

@@ -51,7 +51,7 @@ shared uint vsmMipMask;
 shared uint anyPageGlob;
 #endif
 
-bool projectSphere(const vec4 sphere, out vec4 aabb, out float depthMin) {
+bool projectSphere(const vec4 sphere, out vec4 aabb, out float depthMax) {
   const vec3  c = (scene.viewProject * vec4(sphere.xyz, 1)).xyz;
   const float R = sphere.w;
 
@@ -65,7 +65,7 @@ bool projectSphere(const vec4 sphere, out vec4 aabb, out float depthMin) {
   float Rx = R * smWidthInv;
   float Ry = R * smWidthInv;
   aabb     = vec4(c.xy-vec2(Rx,Ry), c.xy+vec2(Rx,Ry));
-  depthMin = c.z + Rz;
+  depthMax = c.z + Rz;
   // aabb = c.xyxy;
   return true;
   }
@@ -84,11 +84,11 @@ bool frustrumTest(const vec4 sphere) {
   }
 
 
-bool projectCluster(const Cluster cluster, out vec4 aabb, out float depthMin) {
+bool projectCluster(const Cluster cluster, out vec4 aabb, out float depthMax) {
   if(cluster.instanceId==0xFFFFFFFF)
-    return projectSphere(cluster.sphere, aabb, depthMin);
-  return projectSphere(cluster.sphere, aabb, depthMin);
-  // return projectInstance(cluster, aabb, depthMin);
+    return projectSphere(cluster.sphere, aabb, depthMax);
+  return projectSphere(cluster.sphere, aabb, depthMax);
+  // return projectInstance(cluster, aabb, depthMax);
   }
 
 bool emitCluster(const Cluster cluster, uint page) {
@@ -236,6 +236,30 @@ void pullPages(uint mip) {
   barrier();
   }
 
+bool pageHiZTest(vec4 aabb, float depthMax, const ivec3 at, const ivec2 sz) {
+  aabb /= (1u << at.z);
+  aabb = aabb*0.5+0.5; // [0..1]
+  aabb = aabb*VSM_PAGE_TBL_SIZE;
+
+  ivec4 iaabb = ivec4(ivec2(floor(aabb.xy)), ivec2(ceil(aabb.zw))+1);
+  iaabb.xy = max(iaabb.xy, ivec2(at.xy));
+  iaabb.zw = min(iaabb.zw, ivec2(at.xy+sz));
+  iaabb.zw = min(iaabb.zw, ivec2(VSM_PAGE_TBL_SIZE));
+
+  for(int x=iaabb.x; x<iaabb.z; ++x)
+    for(int y=iaabb.y; y<iaabb.w; ++y) {
+      const uint  p = imageLoad(pageTbl, ivec3(x,y,at.z)).x;
+      const float f = uintBitsToFloat(p);
+
+      // if(p!=0xFFFFFFFF)
+      //   return true;
+      if(f<depthMax)
+        return true;
+      }
+
+  return false;
+  }
+
 void runCluster(const uint clusterId) {
   const Cluster cluster = clusters[clusterId];
   if(cluster.sphere.w<=0.f)
@@ -249,8 +273,8 @@ void runCluster(const uint clusterId) {
     }
 
   vec4  aabb     = vec4(0);
-  float depthMin = 1;
-  if(!projectCluster(cluster, aabb, depthMin))
+  float depthMax = 1;
+  if(!projectCluster(cluster, aabb, depthMax))
     return;
 
   for(uint ix=0; ix<pageListSize; ++ix) {
@@ -261,10 +285,8 @@ void runCluster(const uint clusterId) {
     const ivec2 sz   = unpackVsmPageSize(data);
     if(!pageBoundsTest(aabb, page, sz))
       continue;
-    const uint  p = imageLoad(pageTbl, ivec3(page)).x;
-    const float f = uintBitsToFloat(p);
-    if(depthMin < f)
-      ;//continue;
+    if(!pageHiZTest(aabb, depthMax, page, sz))
+      continue;
     if(!emitCluster(cluster, i))
       break;
     }

shader/virtual_shadow/vsm_list_pages.comp

@@ -0,0 +1,34 @@
+#version 450
+
+#extension GL_GOOGLE_include_directive    : enable
+#extension GL_ARB_separate_shader_objects : enable
+#extension GL_EXT_samplerless_texture_functions : enable
+
+#include "virtual_shadow/vsm_common.glsl"
+#include "scene.glsl"
+#include "common.glsl"
+
+//layout(local_size_x = 64) in;
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(binding = 0, std430) buffer Pages { VsmHeader header; uint  pageList[]; } vsm;
+layout(binding = 1, r32ui) uniform uimage3D pageTbl;
+layout(binding = 2, r32ui) uniform readonly uimage3D pageTblDepth;
+//layout(binding = 1, std430) writeonly buffer Dst   { VsmHeader header; uint  pageList[]; } dst;
+
+const uint NumThreads = gl_WorkGroupSize.x*gl_WorkGroupSize.y*gl_WorkGroupSize.z;
+
+void main() {
+  const ivec3 at   = ivec3(gl_GlobalInvocationID.xyz);
+  const ivec3 size = imageSize(pageTbl);
+  if(any(greaterThanEqual(at, size)))
+    return;
+
+  uint v = imageLoad(pageTbl, at).x;
+  if(v==0)
+    return;
+
+  uint i = atomicAdd(vsm.header.pageCount, 1);
+  vsm.pageList[i] = 0xFFFFFFFF;
+  imageStore(pageTbl, at, uvec4(i));
+  }

shader/virtual_shadow/vsm_mark_pages.comp

@@ -66,10 +66,10 @@ uvec4 unpack565_16(uint p) {
   return ret;
   }
 
-void storeVal(uint v) {
+void storeHiZValue(uint v) {
   uvec4 dx = unpack565_16(v);
   ivec3 at = ivec3(dx.xyz);
-  uint  iz = dx.w;
+  uint  iz = floatBitsToUint(dx.w/float(0xFFFF));
 
   //NOTE: require vulkan memory-model extension or vulkan 1.2
   //imageAtomicStore(pageTbl, at, 1u, gl_ScopeDevice, gl_StorageSemanticsNone, gl_SemanticsRelaxed);
@@ -86,7 +86,7 @@ void setupHiZ() {
     pageHiZ[lane] = 0xFFFFFFFF;
   }
 
-void insertHiZ(ivec3 at, float z) {
+void markPage(ivec3 at, float z) {
   if(z<0 || z>=1)
     return;
   //imageAtomicStore(pageTbl, at, 1u, gl_ScopeDevice, gl_StorageSemanticsNone, gl_SemanticsRelaxed);
@@ -101,7 +101,7 @@ void insertHiZ(ivec3 at, float z) {
     cur = (v > cur) ? v : cur;
     }
   // imageAtomicAdd(pageTbl, ivec3(0), 1u); //counter
-  storeVal(cur);
+  storeHiZValue(cur);
   }
 
 void flushHiZ() {
@@ -111,7 +111,7 @@ void flushHiZ() {
   const uint v = pageHiZ[lane];
   if(v==0xFFFFFFFF)
     return;
-  storeVal(v);
+  storeHiZValue(v);
   }
 #else
 shared uint pageHiZ[NumThreads];
@@ -120,7 +120,7 @@ void setupHiZ() {
   pageHiZ[lane] = 0xFFFFFFFF;
   }
 
-void insertHiZ(ivec3 at, float z) {
+void markPage(ivec3 at, float z) {
   if(z<0 || z>=1)
     return;
 
@@ -136,15 +136,15 @@ void insertHiZ(ivec3 at, float z) {
     return; // update same entry
 
   // imageAtomicAdd(pageTbl, ivec3(0), 1u); //counter
-  storeVal(v);
+  storeHiZValue(v);
   }
 
 void flushHiZ() {
   const uint lane = gl_LocalInvocationIndex;
   const uint v    = pageHiZ[lane];
   if(v==0xFFFFFFFF)
     return;
-  storeVal(v);
+  storeHiZValue(v);
   }
 #endif
 
@@ -180,10 +180,6 @@ int shadowLod(vec2 dx, vec2 dy) {
   return max(0, int(minLod + bias + 0.5));
   }
 
-void markPage(ivec2 pageI, int mip, float depth) {
-  insertHiZ(ivec3(pageI, mip), depth);
-  }
-
 void realMain() {
   const ivec2 size = textureSize(depth, 0);
   if(any(greaterThanEqual(pixelPos, size)))
@@ -213,7 +209,7 @@ void realMain() {
     return;
 
   ivec2 pageI = ivec2((page*0.5+0.5)*VSM_PAGE_TBL_SIZE);
-  markPage(pageI, mip, shPos0.z);
+  markPage(ivec3(pageI, mip), shPos0.z);
   }
 
 void main() {