vsm: pages 4x4

#681

game/camera.cpp

@@ -271,16 +271,17 @@ Matrix4x4 Camera::viewShadowLwc(const Tempest::Vec3& lightDir, size_t layer) con
   return mkViewShadow(cameraPos-origin,rotation,vp,lightDir,layer);
   }
 
-Matrix4x4 Camera::viewShadowVsm(const Tempest::Vec3& lightDir) const {
-  auto  vp       = viewProj();
-  float rotation = (180+src.spin.y-rotOffset.y);
-  return viewShadowVsm(cameraPos,rotation,vp,lightDir);
+Matrix4x4 Camera::viewShadowVsm(const Tempest::Vec3& ldir) const {
+  return mkViewShadowVsm(cameraPos,ldir);
+  }
+
+Matrix4x4 Camera::viewShadowVsmLwc(const Tempest::Vec3& ldir) const {
+  return mkViewShadowVsm(cameraPos-origin,ldir);
   }
 
-Matrix4x4 Camera::viewShadowVsm(const Vec3& cameraPos, float rotation, const Tempest::Matrix4x4& viewProj, const Vec3& ldir) const {
+Matrix4x4 Camera::mkViewShadowVsm(const Vec3& cameraPos, const Vec3& ldir) const {
   float smWidth    = 1024; // ~4 pixels per santimeter
   float smDepth    = 5*5120;
-  (void)rotation;
 
   float smWidthInv = 1.f/smWidth;
   float zScale     = 1.f/smDepth;

game/camera.h

@@ -110,7 +110,7 @@ class Camera final {
     Tempest::Matrix4x4 viewShadowLwc(const Tempest::Vec3& ldir, size_t layer) const;
 
     Tempest::Matrix4x4 viewShadowVsm(const Tempest::Vec3& ldir) const;
-    Tempest::Matrix4x4 viewShadowVsm(const Tempest::Vec3& cameraPos, float rotation, const Tempest::Matrix4x4& viewProj, const Tempest::Vec3& ldir) const;
+    Tempest::Matrix4x4 viewShadowVsmLwc(const Tempest::Vec3& ldir) const;
 
     ListenerPos        listenerPosition() const;
     Tempest::Vec3      originLwc() const { return origin; };
@@ -175,6 +175,7 @@ class Camera final {
     Tempest::Matrix4x4    mkRotation(const Tempest::Vec3& spin) const;
     Tempest::Matrix4x4    mkViewShadow(const Tempest::Vec3& cameraPos, float rotation,
                                        const Tempest::Matrix4x4& viewProj, const Tempest::Vec3& lightDir, size_t layer) const;
+    Tempest::Matrix4x4    mkViewShadowVsm(const Tempest::Vec3& cameraPos, const Tempest::Vec3& ldir) const;
     void                  resetDst();
 
     void                  clampRotation(Tempest::Vec3& spin);

game/commandline.cpp

@@ -127,6 +127,12 @@ CommandLine::CommandLine(int argc, const char** argv) {
       if(i<argc)
         isBindlessSh = (std::string_view(argv[i])!="0" && std::string_view(argv[i])!="false");
       }
+    else if(arg=="-vsm") {
+      // not to document - debug only
+      ++i;
+      if(i<argc)
+        isVsm = (std::string_view(argv[i])!="0" && std::string_view(argv[i])!="false");
+      }
     else {
       Log::i("unreacognized commandline option: \"", arg, "\"");
       }

game/commandline.h

@@ -39,6 +39,7 @@ class CommandLine {
     bool                isRtGi()           const { return isGi;         }
     bool                isMeshShading()    const { return isMeshSh;     }
     bool                isBindless()       const { return isBindlessSh; }
+    bool                isVirtualShadow()  const { return isVsm; }
     bool                doStartMenu()      const { return !noMenu;      }
     bool                doForceG1()        const { return forceG1;      }
     bool                doForceG2()        const { return forceG2;      }
@@ -68,6 +69,7 @@ class CommandLine {
     bool                isMeshSh     = true;
 #endif
     bool                isBindlessSh = true;
+    bool                isVsm        = false;
     bool                isGi         = false;
     bool                forceG1      = false;
     bool                forceG2      = false;

game/gothic.cpp

@@ -89,7 +89,8 @@ Gothic::Gothic() {
     }
 
   if(hasBindless() && gpu.compute.maxInvocations>=1024) {
-    opts.doVirtualShadow = false; //TODO: productize
+    //TODO: productize
+    opts.doVirtualShadow = CommandLine::inst().isVirtualShadow();
     }
 
   opts.aaPreset = CommandLine::inst().aaPreset();

game/graphics/drawcommands.cpp

@@ -194,7 +194,7 @@ bool DrawCommands::commit() {
     if(cmdChg) {
       Resources::recycle(std::move(v.indirectCmd));
       v.indirectCmd = device.ssbo(cx.data(), sizeof(IndirectCmd)*cx.size());
-      visChg |= (v.viewport==SceneGlobals::V_Vsm);
+      visChg |= (v.viewport==SceneGlobals::V_Vsm); //FIXME
       }
 
     Resources::recycle(std::move(v.descInit));

game/graphics/renderer.cpp

@@ -208,7 +208,7 @@ void Renderer::resetSwapchain() {
     vsm.shadowMask      = device.image2d(Tempest::RGBA8, w, h);
 
     auto pageCount      = uint32_t((vsm.pageData.w()+128-1)/128) * uint32_t((vsm.pageData.h()+128-1)/128);
-    vsm.pageList        = device.ssbo(nullptr, (pageCount + 2)*sizeof(uint32_t));
+    vsm.pageList        = device.ssbo(nullptr, (pageCount + 4)*sizeof(uint32_t));
     }
 
   if(settings.swrEnabled) {
@@ -618,7 +618,7 @@ void Renderer::draw(Tempest::Attachment& result, Encoder<CommandBuffer>& cmd, ui
       }
     frustrum[SceneGlobals::V_Main].make(viewProj,zbuffer.w(),zbuffer.h());
     frustrum[SceneGlobals::V_HiZ] = frustrum[SceneGlobals::V_Main];
-    frustrum[SceneGlobals::V_Vsm].make(shadowMatrixVsm, vsm.pageData.w(), vsm.pageData.h()); //TODO: remove
+    frustrum[SceneGlobals::V_Vsm] = frustrum[SceneGlobals::V_Shadow1]; //TODO: remove
     wview->updateFrustrum(frustrum);
     }
 
@@ -858,7 +858,10 @@ void Renderer::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fI
   cmd.setUniforms(*vsm.pagesMarkPso, vsm.uboPages);
   cmd.dispatchThreads(zbuffer.size());
 
-  cmd.setUniforms(Shaders::inst().vsmClumpPages, vsm.uboList);
+  //cmd.setUniforms(Shaders::inst().vsmClumpPages0, vsm.uboList);
+  //cmd.dispatch(1);
+
+  cmd.setUniforms(Shaders::inst().vsmClumpPages1, vsm.uboList);
   cmd.dispatchThreads(size_t(vsm.pageTbl.w()), size_t(vsm.pageTbl.h()), size_t(vsm.pageTbl.d()));
 
   cmd.setUniforms(*vsm.pagesListPso, vsm.uboList);

game/graphics/sceneglobals.cpp

@@ -127,8 +127,9 @@ void SceneGlobals::setViewLwc(const Tempest::Matrix4x4& view, const Tempest::Mat
     uboGlobalCpu.viewShadowLwc[i] = sh[i];
   }
 
-void SceneGlobals::setViewVsm(const Tempest::Matrix4x4& view) {
-  uboGlobalCpu.viewVirtualShadow = view;
+void SceneGlobals::setViewVsm(const Tempest::Matrix4x4& view, const Tempest::Matrix4x4& viewLwc) {
+  uboGlobalCpu.viewVirtualShadow    = view;
+  uboGlobalCpu.viewVirtualShadowLwc = viewLwc;
   }
 
 void SceneGlobals::setSky(const Sky& s) {

game/graphics/sceneglobals.h

@@ -30,7 +30,7 @@ class SceneGlobals final {
                         float zNear, float zFar,
                         const Tempest::Matrix4x4 *sh);
     void setViewLwc(const Tempest::Matrix4x4& view, const Tempest::Matrix4x4& proj, const Tempest::Matrix4x4 *sh);
-    void setViewVsm(const Tempest::Matrix4x4& view);
+    void setViewVsm(const Tempest::Matrix4x4& view, const Tempest::Matrix4x4& viewLwc);
     void setSky(const Sky& s);
     void setUnderWater(bool w);
 
@@ -86,6 +86,7 @@ class SceneGlobals final {
       Tempest::Matrix4x4              viewProjectLwcInv;
       Tempest::Matrix4x4              viewShadowLwc[Resources::ShadowLayers];
       Tempest::Matrix4x4              viewVirtualShadow;
+      Tempest::Matrix4x4              viewVirtualShadowLwc;
       Tempest::Matrix4x4              view, project, projectInv;
       Tempest::Vec3                   sunDir        = {0,0,1};
       float                           waveAnim      = 0;

game/graphics/shaders.cpp

@@ -198,7 +198,8 @@ Shaders::Shaders() {
     vsmClusterTask = computeShader("vsm_cluster_task.comp.sprv");
     vsmClear       = computeShader("vsm_clear.comp.sprv");
     vsmMarkPages   = computeShader("vsm_mark_pages.comp.sprv");
-    vsmClumpPages  = computeShader("vsm_clump_pages.comp.sprv");
+    vsmClumpPages0 = computeShader("vsm_clump_pages0.comp.sprv");
+    vsmClumpPages1 = computeShader("vsm_clump_pages1.comp.sprv");
     vsmListPages   = computeShader("vsm_list_pages.comp.sprv");
     vsmSortPages   = computeShader("vsm_sort_pages.comp.sprv");
     vsmPackDraw0   = computeShader("vsm_pack_draws0.comp.sprv");

game/graphics/shaders.h

@@ -76,7 +76,8 @@ class Shaders {
 
     // Virtual shadow
     Tempest::ComputePipeline vsmClusterTask;
-    Tempest::ComputePipeline vsmClear, vsmMarkPages, vsmClumpPages, vsmListPages;
+    Tempest::ComputePipeline vsmClear, vsmMarkPages, vsmListPages;
+    Tempest::ComputePipeline vsmClumpPages0, vsmClumpPages1;
     Tempest::ComputePipeline vsmPackDraw0, vsmPackDraw1;
     Tempest::RenderPipeline  vsmDirectLight;
     Tempest::ComputePipeline vsmSortPages;

game/graphics/worldview.cpp

@@ -51,7 +51,7 @@ void WorldView::preFrameUpdate(const Camera& camera, uint64_t tickCount, uint8_t
   sGlobal.setSky(gSky);
   sGlobal.setViewProject(camera.view(),camera.projective(),camera.zNear(),camera.zFar(),shadow);
   sGlobal.setViewLwc(camera.viewLwc(),camera.projective(),shadowLwc);
-  sGlobal.setViewVsm(camera.viewShadowVsm(ldir));
+  sGlobal.setViewVsm(camera.viewShadowVsm(ldir), camera.viewShadowVsmLwc(ldir));
   sGlobal.originLwc = camera.originLwc();
   sGlobal.setUnderWater(camera.isInWater());
 

shader/CMakeLists.txt

@@ -298,11 +298,12 @@ add_shader(cmaa2_deferred_color_apply_2x2.vert     antialiasing/cmaa2/deferred_c
 add_shader(cmaa2_deferred_color_apply_2x2.frag     antialiasing/cmaa2/deferred_color_apply_2x2.frag)
 
 # virtual shadows
-add_shader(direct_light_vsm.frag virtual_shadow/vsm_mark_pages.comp -DCOMPOSE -S frag)
-add_shader(vsm_dbg.frag          virtual_shadow/vsm_mark_pages.comp -DDEBUG -S frag)
+add_shader(direct_light_vsm.frag virtual_shadow/vsm_mark_pages.comp -DCOMPOSE    -DLWC -S frag)
+add_shader(vsm_dbg.frag          virtual_shadow/vsm_mark_pages.comp -DDEBUG      -DLWC -S frag)
+add_shader(vsm_mark_pages.comp   virtual_shadow/vsm_mark_pages.comp -DMARK_PAGES -DLWC)
 add_shader(vsm_clear.comp        virtual_shadow/vsm_clear.comp)
-add_shader(vsm_mark_pages.comp   virtual_shadow/vsm_mark_pages.comp -DMARK_PAGES)
-add_shader(vsm_clump_pages.comp  virtual_shadow/vsm_clump_pages.comp)
+add_shader(vsm_clump_pages0.comp virtual_shadow/vsm_clump_pages.comp -DPASS0)
+add_shader(vsm_clump_pages1.comp virtual_shadow/vsm_clump_pages.comp -DPASS1)
 add_shader(vsm_list_pages.comp   virtual_shadow/vsm_list_pages.comp)
 add_shader(vsm_sort_pages.comp   virtual_shadow/vsm_sort_pages.comp)
 add_shader(vsm_pack_draws0.comp  virtual_shadow/vsm_pack_draws.comp   -DPASS0)

shader/scene.glsl

@@ -9,6 +9,7 @@ struct SceneDesc {
   mat4  viewProjectLwcInv;
   mat4  viewShadowLwc[2];
   mat4  viewVirtualShadow;
+  mat4  viewVirtualShadowLwc;
   mat4  view;
   mat4  project;
   mat4  projectInv;

shader/virtual_shadow/vsm_clear.comp

@@ -25,5 +25,7 @@ void main() {
   if(at==ivec3(0)) {
     header.pageCount    = 0;
     header.meshletCount = 0;
+    header.counterM     = 0;
+    header.counterV     = 0;
     }
   }

shader/virtual_shadow/vsm_clump_pages.comp

@@ -8,14 +8,97 @@
 #include "scene.glsl"
 #include "common.glsl"
 
+#if defined(PASS0)
+layout(local_size_x = 16, local_size_y = 8) in;
+#else
 layout(local_size_x = 8, local_size_y = 8) in;
+#endif
 
 layout(binding = 0, std430) buffer Pages { VsmHeader header; uint  pageList[]; } vsm;
 layout(binding = 1, r32ui)  uniform uimage3D pageTbl;
 
-shared uint pageVal[gl_WorkGroupSize.x][gl_WorkGroupSize.y];
+void trimMip(int mip) {
+  const ivec3 size = imageSize(pageTbl);
+  const ivec3 id   = ivec3(gl_LocalInvocationID);
+
+  const ivec2 b = ivec2(VSM_PAGE_TBL_SIZE/4);
+  const ivec2 h = ivec2(VSM_PAGE_TBL_SIZE/2);
+
+  for(int i=id.x; i<h.x; i+=int(gl_WorkGroupSize.x)) {
+    for(int r=id.y; r<h.y; r+=int(gl_WorkGroupSize.y)) {
+      ivec3 ax = ivec3(i+b.x, r+b.y, mip);
+      uint  st = imageLoad(pageTbl, ax).r;
+      if(st==0)
+        continue;
+
+      ivec3 at = ivec3(i*2, r*2, mip-1);
+      uint  a  = imageLoad(pageTbl, at+ivec3(0,0,0)).r;
+      uint  b  = imageLoad(pageTbl, at+ivec3(1,0,0)).r;
+      uint  c  = imageLoad(pageTbl, at+ivec3(0,1,0)).r;
+      uint  d  = imageLoad(pageTbl, at+ivec3(1,1,0)).r;
+      if(a>0 && b>0 && c>0 && d>0) {
+        imageStore(pageTbl, ax, uvec4(0));
+        }
+      }
+    }
+  }
+
+void mainTrim() {
+  const ivec3 size = imageSize(pageTbl);
+  // NOTE: need only about 5 mips
+  for(int i=1; i<size.z; ++i) {
+    trimMip(i);
+    barrier();
+    }
+  }
+
+void mainGroups() {
+  const ivec3 size = imageSize(pageTbl);
+  const ivec3 at   = ivec3(gl_GlobalInvocationID);
+  const ivec3 id   = ivec3(gl_LocalInvocationID);
+
+  if(true && at.x%2==0 && at.y%2==0) {
+    uint a = imageLoad(pageTbl, at+ivec3(0,0,0)).r;
+    uint b = imageLoad(pageTbl, at+ivec3(1,0,0)).r;
+    uint c = imageLoad(pageTbl, at+ivec3(1,1,0)).r;
+    uint d = imageLoad(pageTbl, at+ivec3(0,1,0)).r;
+    if(a==1 && b==1 && c==1 && d==1) {
+      imageStore(pageTbl, at+ivec3(0,0,0), uvec4(2));
+      imageStore(pageTbl, at+ivec3(1,0,0), uvec4(0));
+      imageStore(pageTbl, at+ivec3(0,1,0), uvec4(0));
+      imageStore(pageTbl, at+ivec3(1,1,0), uvec4(0));
+      }
+    }
+  groupMemoryBarrier();
+  barrier();
+
+  if(true && at.x%4==0 && at.y%4==0) {
+    uint a = imageLoad(pageTbl, at+ivec3(0,0,0)).r;
+    uint b = imageLoad(pageTbl, at+ivec3(2,0,0)).r;
+    uint c = imageLoad(pageTbl, at+ivec3(2,2,0)).r;
+    uint d = imageLoad(pageTbl, at+ivec3(0,2,0)).r;
+    if(a==2 && b==2 && c==2 && d==2) {
+      imageStore(pageTbl, at+ivec3(0,0,0), uvec4(4));
+      imageStore(pageTbl, at+ivec3(2,0,0), uvec4(0));
+      imageStore(pageTbl, at+ivec3(0,2,0), uvec4(0));
+      imageStore(pageTbl, at+ivec3(2,2,0), uvec4(0));
+      }
+    }
+  }
 
 void main() {
+#if defined(PASS0)
+  mainTrim();
+#elif defined(PASS1)
+  mainGroups();
+#else
+#error "invalid pass-id"
+#endif
+  }
+
+/*
+shared uint pageVal[gl_WorkGroupSize.x][gl_WorkGroupSize.y];
+void main2() {
   const ivec3 size = imageSize(pageTbl);
   const ivec3 at   = ivec3(gl_GlobalInvocationID);
   const ivec3 id   = ivec3(gl_LocalInvocationID);
@@ -70,3 +153,4 @@ void main() {
     imageStore(pageTbl, at+ivec3(1,1,0), uvec4(0));
     }
   }
+*/

shader/virtual_shadow/vsm_cluster_task.comp

@@ -40,6 +40,17 @@ layout(push_constant, std430) uniform UboPush {
   int meshletCount;
   } push;
 
+shared bool workgroupBool;
+
+bool workgroupAny(bool a) {
+  workgroupBool = false;
+  barrier();
+  if(a)
+    workgroupBool = true;
+  barrier();
+  return workgroupBool;
+  }
+
 bool projectSphere(const vec4 sphere, out vec4 aabb, out float depthMin) {
   const vec3  c = (scene.viewProject * vec4(sphere.xyz, 1)).xyz;
   const float R = sphere.w;
@@ -109,8 +120,12 @@ void runCluster(const uint clusterId) {
   if(cluster.sphere.w<=0.f)
     return; // disabled or deleted
 
-  //if(!frustrumTest(cluster.sphere))
-  //  return;
+  if(frustrumTest(cluster.sphere)) {
+    if(gl_LocalInvocationIndex==0)
+      atomicAdd(vsm.header.counterV, cluster.meshletCount);
+    } else {
+    // return;
+    }
 
   vec4  aabb     = vec4(0);
   float depthMin = 1;
@@ -119,6 +134,7 @@ void runCluster(const uint clusterId) {
 
   const uint commandId = cluster.bucketId_commandId & 0xFFFF;
   const uint pageCount = vsm.header.pageCount;
+  bool anyPage = false;
 
   for(uint i=gl_LocalInvocationIndex; i<pageCount; i+=gl_WorkGroupSize.y) {
     // if(i!=13)
@@ -135,7 +151,11 @@ void runCluster(const uint clusterId) {
     if(!emitCluster(cluster, i))
       break;
     atomicAdd(cmd[commandId].instanceCount, cluster.meshletCount);
+    anyPage = true;
     }
+
+  if(workgroupAny(anyPage) && gl_LocalInvocationIndex==0)
+    atomicAdd(vsm.header.counterM, cluster.meshletCount);
   }
 
 void main() {

shader/virtual_shadow/vsm_common.glsl

@@ -10,6 +10,8 @@ const int VSM_CLIPMAP_SIZE  = VSM_PAGE_SIZE * VSM_PAGE_TBL_SIZE;
 struct VsmHeader {
   uint pageCount;
   uint meshletCount;
+  uint counterM;
+  uint counterV;
   };
 
 uint packVsmPageInfo(ivec3 at, ivec2 size) {