vsm: 2x2 pages support

#681

game/camera.cpp

@@ -279,7 +279,7 @@ Matrix4x4 Camera::viewShadowVsm(const Tempest::Vec3& lightDir) const {
 
 Matrix4x4 Camera::viewShadowVsm(const Vec3& cameraPos, float rotation, const Tempest::Matrix4x4& viewProj, const Vec3& ldir) const {
   float smWidth    = 1024; // ~4 pixels per santimeter
-  float smDepth    = 10*5120;
+  float smDepth    = 5*5120;
   (void)rotation;
 
   float smWidthInv = 1.f/smWidth;

game/graphics/drawcommands.cpp

@@ -65,6 +65,10 @@ DrawCommands::DrawCommands(VisualObjects& owner, DrawBuckets& buckets, DrawClust
     tasks.emplace_back(std::move(cmd));
     }
 
+  if(virtualShadowSys) {
+    Tempest::DispatchIndirectCommand cmd = {2000,1,1};
+    vsmIndirectCmd = Resources::device().ssbo(&cmd, sizeof(cmd));
+    }
   }
 
 DrawCommands::~DrawCommands() {
@@ -184,16 +188,13 @@ bool DrawCommands::commit() {
       v.visClusters = device.ssbo(nullptr, size);
       }
     if(visChg && v.viewport==SceneGlobals::V_Vsm) {
+      Resources::recycle(std::move(v.vsmClusters));
       v.vsmClusters = device.ssbo(nullptr, v.visClusters.byteSize());
       }
     if(cmdChg) {
       Resources::recycle(std::move(v.indirectCmd));
       v.indirectCmd = device.ssbo(cx.data(), sizeof(IndirectCmd)*cx.size());
-      }
-
-    if(cmdChg) {
-      Resources::recycle(std::move(v.pkgOffsets));
-      v.pkgOffsets  = device.ssbo(nullptr, sizeof(uint32_t)*cx.size());
+      visChg |= (v.viewport==SceneGlobals::V_Vsm);
       }
 
     Resources::recycle(std::move(v.descInit));
@@ -239,13 +240,20 @@ void DrawCommands::updateTasksUniforms() {
   for(auto& v:views) {
     if(v.viewport!=SceneGlobals::V_Vsm)
       continue;
-    Resources::recycle(std::move(v.descPackDraw));
-    v.descPackDraw = device.descriptors(Shaders::inst().vsmPackDraw0);
-    v.descPackDraw.set(1, v.vsmClusters);
-    v.descPackDraw.set(2, v.visClusters);
-    v.descPackDraw.set(3, v.indirectCmd);
-    v.descPackDraw.set(4, *scene.vsmPageList);
-    v.descPackDraw.set(5, v.pkgOffsets);
+    Resources::recycle(std::move(v.descPackDraw0));
+    v.descPackDraw0 = device.descriptors(Shaders::inst().vsmPackDraw0);
+    v.descPackDraw0.set(1, v.vsmClusters);
+    v.descPackDraw0.set(2, v.visClusters);
+    v.descPackDraw0.set(3, v.indirectCmd);
+    v.descPackDraw0.set(4, *scene.vsmPageList);
+    v.descPackDraw0.set(5, vsmIndirectCmd);
+
+    Resources::recycle(std::move(v.descPackDraw1));
+    v.descPackDraw1 = device.descriptors(Shaders::inst().vsmPackDraw1);
+    v.descPackDraw1.set(1, v.vsmClusters);
+    v.descPackDraw1.set(2, v.visClusters);
+    v.descPackDraw1.set(3, v.indirectCmd);
+    v.descPackDraw1.set(4, *scene.vsmPageList);
     }
 
   updateVsmUniforms();
@@ -327,7 +335,7 @@ void DrawCommands::updateCommandUniforms() {
           }
 
         if(v==SceneGlobals::V_Vsm) {
-          desc[v].set(L_CmdOffsets, views[v].pkgOffsets);
+          desc[v].set(L_CmdOffsets, views[v].indirectCmd);
           desc[v].set(L_VsmPages,   *scene.vsmPageList);
           }
         }
@@ -477,11 +485,12 @@ void DrawCommands::visibilityVsm(Encoder<CommandBuffer>& cmd, uint8_t fId) {
     cmd.dispatch(push.meshletCount);
     }
 
-  cmd.setUniforms(Shaders::inst().vsmPackDraw0, views[SceneGlobals::V_Vsm].descPackDraw);
+  cmd.setUniforms(Shaders::inst().vsmPackDraw0, views[SceneGlobals::V_Vsm].descPackDraw0);
   cmd.dispatch(1);
 
-  cmd.setUniforms(Shaders::inst().vsmPackDraw1, views[SceneGlobals::V_Vsm].descPackDraw);
-  cmd.dispatch(2000); // TODO: indirect
+  cmd.setUniforms(Shaders::inst().vsmPackDraw1, views[SceneGlobals::V_Vsm].descPackDraw1);
+  cmd.dispatch(8096); // TODO: indirect
+  // cmd.dispatchIndirect(vsmIndirectCmd, 0);
   }
 
 void DrawCommands::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fId) {

game/graphics/drawcommands.h

@@ -123,8 +123,9 @@ class DrawCommands {
       Tempest::DescriptorSet  descInit;
       Tempest::StorageBuffer  visClusters, indirectCmd;
 
-      Tempest::DescriptorSet  descPackDraw;
-      Tempest::StorageBuffer  pkgOffsets, vsmClusters;
+      Tempest::DescriptorSet  descPackDraw0;
+      Tempest::DescriptorSet  descPackDraw1;
+      Tempest::StorageBuffer  vsmClusters;
 
       bool                    isEnabled() const;
       };
@@ -144,4 +145,6 @@ class DrawCommands {
     bool                     cmdDurtyBit = false;
 
     View                     views[SceneGlobals::V_Count];
+
+    Tempest::StorageBuffer   vsmIndirectCmd;
   };

game/graphics/renderer.cpp

@@ -203,13 +203,12 @@ void Renderer::resetSwapchain() {
     vsm.pagesDbgPso     = &Shaders::inst().vsmDbg;
     vsm.uboDbg          = device.descriptors(*vsm.pagesDbgPso);
 
-    vsm.pageTbl         = device.image3d(TextureFormat::R32U, 64, 64, 32);
-    vsm.pageData        = device.image2d(TextureFormat::RGBA8, 4096, 4096); // NOTE: not yet sure on what to use: depth or atomics
-    vsm.pageDataZ       = device.zbuffer(shadowFormat, 4096, 4096);
+    vsm.pageTbl         = device.image3d(TextureFormat::R32U, 32, 32, 16);
+    vsm.pageData        = device.zbuffer(shadowFormat, 4096, 4096);
     vsm.shadowMask      = device.image2d(Tempest::RGBA8, w, h);
 
-    auto pageCount      = uint32_t((vsm.pageDataZ.w()+128-1)/128) * uint32_t((vsm.pageDataZ.h()+128-1)/128);
-    vsm.pageList        = device.ssbo(nullptr, (pageCount + 4)*sizeof(uint32_t));
+    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));
     }
 
   if(settings.swrEnabled) {
@@ -466,15 +465,15 @@ void Renderer::prepareUniforms() {
     vsm.uboLight.set(2, gbufNormal,  Sampler::nearest());
     vsm.uboLight.set(3, zbuffer,     Sampler::nearest());
     vsm.uboLight.set(4, vsm.pageTbl);
-    vsm.uboLight.set(5, vsm.pageDataZ);
+    vsm.uboLight.set(5, vsm.pageData);
     vsm.uboLight.set(6, vsm.shadowMask);
 
     vsm.uboDbg.set(0, wview->sceneGlobals().uboGlobal[SceneGlobals::V_Main]);
     vsm.uboDbg.set(1, gbufDiffuse, Sampler::nearest());
     vsm.uboDbg.set(2, gbufNormal,  Sampler::nearest());
     vsm.uboDbg.set(3, zbuffer,     Sampler::nearest());
     vsm.uboDbg.set(4, vsm.pageTbl);
-    vsm.uboDbg.set(5, vsm.pageDataZ);
+    vsm.uboDbg.set(5, vsm.pageData);
     vsm.uboDbg.set(6, vsm.shadowMask);
     }
 
@@ -492,7 +491,7 @@ void Renderer::prepareUniforms() {
       sh[i] = &textureCast<const Texture2d&>(shadowMap[i]);
       }
   wview->setShadowMaps(sh);
-  wview->setVirtualShadowMap(vsm.pageData, vsm.pageList);
+  wview->setVirtualShadowMap(StorageImage(), vsm.pageList);
   wview->setSwRenderingImage(swr.outputImage);
 
   wview->setHiZ(textureCast<const Texture2d&>(hiz.hiZ));
@@ -619,7 +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.pageDataZ.w(), vsm.pageDataZ.h()); //TODO: mip0 resolution
+    frustrum[SceneGlobals::V_Vsm].make(shadowMatrixVsm, vsm.pageData.w(), vsm.pageData.h()); //TODO: remove
     wview->updateFrustrum(frustrum);
     }
 
@@ -859,8 +858,12 @@ 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.dispatchThreads(size_t(vsm.pageTbl.w()), size_t(vsm.pageTbl.h()), size_t(vsm.pageTbl.d()));
+
   cmd.setUniforms(*vsm.pagesListPso, vsm.uboList);
   cmd.dispatchThreads(size_t(vsm.pageTbl.w()), size_t(vsm.pageTbl.h()), size_t(vsm.pageTbl.d()));
+
   // sort for debug purpose only
   cmd.setDebugMarker("VSM-sort(debug)");
   cmd.setUniforms(Shaders::inst().vsmSortPages, vsm.uboList);
@@ -870,13 +873,8 @@ void Renderer::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fI
   view.visibilityVsm(cmd,fId);
 
   cmd.setDebugMarker("VSM-rendering");
-  cmd.setFramebuffer({}, {vsm.pageDataZ, 0.f, Tempest::Preserve});
+  cmd.setFramebuffer({}, {vsm.pageData, 0.f, Tempest::Preserve});
   view.drawVsm(cmd,fId);
-
-  //cmd.setDebugMarker("VSM-compose");
-  //cmd.setFramebuffer({});
-  //cmd.setUniforms(*vsm.pagesComposePso, vsm.uboCompose);
-  //cmd.dispatchThreads(zbuffer.size());
   }
 
 void Renderer::drawSwr(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fId, WorldView& view) {

game/graphics/renderer.h

@@ -227,8 +227,7 @@ class Renderer final {
       Tempest::DescriptorSet    uboDbg;
 
       Tempest::StorageImage     pageTbl;
-      Tempest::StorageImage     pageData;
-      Tempest::ZBuffer          pageDataZ;
+      Tempest::ZBuffer          pageData;
       Tempest::StorageBuffer    pageList;
       Tempest::StorageImage     shadowMask;
       } vsm;

game/graphics/shaders.cpp

@@ -198,6 +198,7 @@ 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");
     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

@@ -76,10 +76,10 @@ class Shaders {
 
     // Virtual shadow
     Tempest::ComputePipeline vsmClusterTask;
-    Tempest::ComputePipeline vsmClear, vsmMarkPages, vsmListPages, vsmSortPages;
+    Tempest::ComputePipeline vsmClear, vsmMarkPages, vsmClumpPages, vsmListPages;
     Tempest::ComputePipeline vsmPackDraw0, vsmPackDraw1;
-    // Tempest::ComputePipeline vsmRendering;
     Tempest::RenderPipeline  vsmDirectLight;
+    Tempest::ComputePipeline vsmSortPages;
     Tempest::RenderPipeline  vsmDbg;
 
     // Software rendering

shader/CMakeLists.txt

@@ -299,9 +299,10 @@ add_shader(cmaa2_deferred_color_apply_2x2.frag     antialiasing/cmaa2/deferred_c
 
 # 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(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_dbg.frag          virtual_shadow/vsm_mark_pages.comp -DDEBUG -S frag)
+add_shader(vsm_clump_pages.comp  virtual_shadow/vsm_clump_pages.comp)
 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/materials/main.frag

@@ -17,10 +17,6 @@ layout(location = 0) in flat uint bucketId;
 layout(location = 1) in Varyings  shInp;
 #endif
 
-#if defined(VIRTUAL_SHADOW)
-layout(location = 3) in flat uint vsmPageId;
-#endif
-
 #if DEBUG_DRAW
 layout(location = DEBUG_DRAW_LOC) in flat uint debugId;
 #endif

shader/materials/main.vert

@@ -52,29 +52,25 @@ layout(location = 0) out flat uint bucketIdOut[]; //TODO: per-primitive
 layout(location = 1) out Varyings  shOut[];
 #endif
 
-#if defined(GL_VERTEX_SHADER) && defined(VIRTUAL_SHADOW)
-layout(location = 3) out flat uint vsmPageIdOut;
-#elif defined(VIRTUAL_SHADOW)
-layout(location = 3) out flat uint vsmPageIdOut[];
-#endif
-
 #if defined(VIRTUAL_SHADOW)
 uint shadowPageId = 0;
 #endif
 
 #if defined(VIRTUAL_SHADOW)
 vec4 mapViewport(vec4 pos, out float clipDistance[4]) {
-  const ivec3 page = unpackVsmPageInfo(vsm.pageList[shadowPageId]);
+  const uint  data = vsm.pageList[shadowPageId];
+  const ivec3 page = unpackVsmPageInfo(data);
+  const ivec2 sz   = unpackVsmPageSize(data);
   pos.xy /= float(1u << page.z);
 
   pos.xy = (pos.xy*0.5+0.5); // [0..1]
   pos.xy = (pos.xy*VSM_PAGE_TBL_SIZE - page.xy);
 
   {
-    clipDistance[0] = 0+pos.x;
-    clipDistance[1] = 1-pos.x;
-    clipDistance[2] = 0+pos.y;
-    clipDistance[3] = 1-pos.y;
+    clipDistance[0] = 0   +pos.x;
+    clipDistance[1] = sz.x-pos.x;
+    clipDistance[2] = 0   +pos.y;
+    clipDistance[3] = sz.y-pos.y;
   }
 
   const vec2 pageId = vec2(unpackVsmPageId(shadowPageId));
@@ -87,12 +83,7 @@ vec4 mapViewport(vec4 pos, out float clipDistance[4]) {
 
 #if defined(VIRTUAL_SHADOW)
 void initVsm(uvec4 task) {
-  shadowPageId = task.w;
-#if defined(GL_VERTEX_SHADER)
-  vsmPageIdOut = task.w;
-#else
-  vsmPageIdOut[gl_LocalInvocationIndex] = task.w;
-#endif
+  shadowPageId = task.w & 0xFFFF;
   }
 #endif
 
@@ -243,12 +234,12 @@ void main() {
 #endif
 
 #if defined(VIRTUAL_SHADOW)
-  const uint firstMeshlet = cmdOffsets[push.commandId];
+  const uint firstMeshlet = cmd[push.commandId].writeOffset;
 #else
   const uint firstMeshlet = push.firstMeshlet;
 #endif
 
-  const uvec4 task     = payload[workIndex + firstMeshlet];
+  const uvec4 task = payload[workIndex + firstMeshlet];
 
 #if defined(VIRTUAL_SHADOW)
   initVsm(task);

shader/materials/materials_common.glsl

@@ -1,6 +1,10 @@
 #ifndef MATERIALS_COMMON_GLSL
 #define MATERIALS_COMMON_GLSL
 
+#if defined(VIRTUAL_SHADOW)
+#include "virtual_shadow/vsm_common.glsl"
+#endif
+
 #include "common.glsl"
 #include "scene.glsl"
 
@@ -201,8 +205,8 @@ layout(binding = L_GDepth  )         uniform sampler2D gbufferDepth;
 #endif
 
 #if defined(VIRTUAL_SHADOW) && !defined(CLUSTER)
-layout(binding = L_CmdOffsets, std430) readonly buffer Offsets { uint cmdOffsets[]; };
-layout(binding = L_VsmPages,   std430) readonly buffer Pages   { uvec4 header; uint  pageList[]; } vsm;
+layout(binding = L_CmdOffsets, std430) readonly buffer IndirectBuf { IndirectCmd cmd[]; };
+layout(binding = L_VsmPages,   std430) readonly buffer Pages       { VsmHeader header; uint pageList[]; } vsm;
 #endif
 
 #if !defined(CLUSTER) && (MESH_TYPE!=T_PFX)

shader/virtual_shadow/vsm_clear.comp

@@ -4,13 +4,14 @@
 #extension GL_ARB_separate_shader_objects : enable
 #extension GL_EXT_samplerless_texture_functions : enable
 
+#include "virtual_shadow/vsm_common.glsl"
 #include "scene.glsl"
 
 layout(local_size_x = 8, local_size_y = 8) in;
 
 layout(binding = 0, std430) buffer CB0 {
-  uvec4 header;
-  uint  pageList[];
+  VsmHeader header;
+  uint      pageList[];
   };
 layout(binding = 1, r32ui) uniform uimage3D pageTbl;
 
@@ -22,6 +23,7 @@ void main() {
 
   imageStore(pageTbl, at, uvec4(0));
   if(at==ivec3(0)) {
-    header = uvec4(0,1,1,0);
+    header.pageCount    = 0;
+    header.meshletCount = 0;
     }
   }