vsm: configurable page size; sw-shadows initial

#681

game/camera.cpp

@@ -281,7 +281,7 @@ Matrix4x4 Camera::viewShadowVsmLwc(const Tempest::Vec3& ldir) const {
 
 Matrix4x4 Camera::mkViewShadowVsm(const Vec3& cameraPos, const Vec3& ldir) const {
   float smWidth    = 1024; // ~4 pixels per santimeter
-  float smDepth    = 5*5120;
+  float smDepth    = 10*5120;
 
   float smWidthInv = 1.f/smWidth;
   float zScale     = 1.f/smDepth;

game/graphics/drawcommands.cpp

@@ -69,6 +69,8 @@ DrawCommands::DrawCommands(VisualObjects& owner, DrawBuckets& buckets, DrawClust
     Tempest::DispatchIndirectCommand cmd = {2000,1,1};
     vsmIndirectCmd = Resources::device().ssbo(&cmd, sizeof(cmd));
     }
+  // vsmSwrImage = Resources::device().image2d(TextureFormat::R16,  4096, 4096);
+  // vsmSwrImage = Resources::device().image2d(TextureFormat::R32U, 4096, 4096);
   }
 
 DrawCommands::~DrawCommands() {
@@ -349,10 +351,7 @@ void DrawCommands::updateCommandUniforms() {
   }
 
 void DrawCommands::updateVsmUniforms() {
-  if(Gothic::options().swRenderingPreset==0)
-    return;
-
-  if(scene.swMainImage==nullptr)
+  if(Gothic::options().swRenderingPreset==0 && Gothic::options().doVirtualShadow==false)
     return;
 
   auto& device = Resources::device();
@@ -377,7 +376,7 @@ void DrawCommands::updateVsmUniforms() {
     }
 
   const uint32_t preset = Gothic::options().swRenderingPreset;
-  if(preset>0 && !Shaders::inst().swRendering.isEmpty()) {
+  if(preset>0 && !Shaders::inst().swRendering.isEmpty() && scene.swMainImage!=nullptr) {
     Resources::recycle(std::move(swrDesc));
     swrDesc = device.descriptors(Shaders::inst().swRendering);
     swrDesc.set(0, *scene.swMainImage);
@@ -394,6 +393,21 @@ void DrawCommands::updateVsmUniforms() {
       swrDesc.set(10, owner.instanceSsbo());
       }
     }
+
+  if(false && Gothic::options().doVirtualShadow && scene.vsmPageList!=nullptr) {
+    Resources::recycle(std::move(vsmDesc));
+    vsmDesc = device.descriptors(Shaders::inst().vsmRendering);
+
+    vsmDesc.set(0, vsmSwrImage);
+    vsmDesc.set(1, scene.uboGlobal[SceneGlobals::V_Vsm]);
+    vsmDesc.set(2, *scene.vsmPageList);
+    vsmDesc.set(3, clusters.ssbo());
+    vsmDesc.set(4, owner.instanceSsbo());
+    vsmDesc.set(5, ibo);
+    vsmDesc.set(6, vbo);
+    vsmDesc.set(7, tex);
+    vsmDesc.set(8, Sampler::bilinear());
+    }
   }
 
 void DrawCommands::prepareUniforms() {
@@ -517,6 +531,15 @@ void DrawCommands::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_
       cmd.dispatchMeshIndirect(view.indirectCmd, sizeof(IndirectCmd)*id + sizeof(uint32_t)); else
       cmd.drawIndirect(view.indirectCmd, sizeof(IndirectCmd)*id);
     }
+
+  if(false) {
+    cmd.setFramebuffer({});
+    struct Push { uint32_t meshletCount; } push = {};
+    push.meshletCount = uint32_t(clusters.size());
+    cmd.setUniforms(Shaders::inst().vsmRendering, vsmDesc, &push, sizeof(push));
+    // const auto sz = Shaders::inst().vsmRendering.workGroupSize();
+    cmd.dispatch(1024u);
+    }
   }
 
 void DrawCommands::drawHiZ(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fId) {

game/graphics/drawcommands.h

@@ -147,4 +147,6 @@ class DrawCommands {
     View                     views[SceneGlobals::V_Count];
 
     Tempest::StorageBuffer   vsmIndirectCmd;
+    Tempest::StorageImage    vsmSwrImage;
+    Tempest::DescriptorSet   vsmDesc;
   };

game/graphics/renderer.cpp

@@ -207,7 +207,8 @@ void Renderer::resetSwapchain() {
     vsm.pageData        = device.zbuffer(shadowFormat, 4096, 4096);
     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);
+    const int32_t VSM_PAGE_SIZE = 128;
+    auto pageCount      = uint32_t((vsm.pageData.w()+VSM_PAGE_SIZE-1)/VSM_PAGE_SIZE) * uint32_t((vsm.pageData.h()+VSM_PAGE_SIZE-1)/VSM_PAGE_SIZE);
     vsm.pageList        = device.ssbo(nullptr, (pageCount + 4)*sizeof(uint32_t));
     }
 
@@ -858,6 +859,7 @@ void Renderer::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fI
   cmd.setUniforms(*vsm.pagesMarkPso, vsm.uboPages);
   cmd.dispatchThreads(zbuffer.size());
 
+  //TODO: trimming
   //cmd.setUniforms(Shaders::inst().vsmClumpPages0, vsm.uboList);
   //cmd.dispatch(1);
 

game/graphics/shaders.cpp

@@ -206,7 +206,7 @@ Shaders::Shaders() {
     vsmPackDraw1   = computeShader("vsm_pack_draws1.comp.sprv");
     vsmDirectLight = postEffect("copy", "direct_light_vsm", RenderState::ZTestMode::NoEqual);
     vsmDbg         = postEffect("copy", "vsm_dbg", RenderState::ZTestMode::Always);
-    // vsmRendering  = computeShader("vsm_rendering.comp.sprv");
+    vsmRendering   = computeShader("vsm_rendering.comp.sprv");
     }
 
   if(Gothic::options().swRenderingPreset>0) {

game/graphics/shaders.h

@@ -83,6 +83,8 @@ class Shaders {
     Tempest::ComputePipeline vsmSortPages;
     Tempest::RenderPipeline  vsmDbg;
 
+    Tempest::ComputePipeline vsmRendering;
+
     // Software rendering
     Tempest::ComputePipeline swRendering;
     Tempest::RenderPipeline  swRenderingDbg;

shader/materials/materials_common.glsl

@@ -38,17 +38,11 @@ const vec3 debugColors[MAX_DEBUG_COLORS] = {
   };
 #endif
 
-#define MAX_NUM_SKELETAL_NODES 96
-#define MAX_MORPH_LAYERS       4
-#define MaxVert                64
-#define MaxPrim                64
-#define MaxInd                 (MaxPrim*3)
-
-#define T_LANDSCAPE 0
-#define T_OBJ       1
-#define T_SKINING   2
-#define T_MORPH     3
-#define T_PFX       4
+const uint MAX_NUM_SKELETAL_NODES = 96;
+const uint MAX_MORPH_LAYERS       = 4;
+const uint MaxVert                = 64;
+const uint MaxPrim                = 64;
+const uint MaxInd                 = MaxPrim*3;
 
 const uint L_Scene      = 0;
 const uint L_Payload    = 1;
@@ -68,6 +62,12 @@ const uint L_GDepth     = 13;
 const uint L_CmdOffsets = 14;
 const uint L_VsmPages   = 15;
 
+#define T_LANDSCAPE 0
+#define T_OBJ       1
+#define T_SKINING   2
+#define T_MORPH     3
+#define T_PFX       4
+
 #ifndef MESH_TYPE
 #define MESH_TYPE 255
 #endif

shader/software_rendering/sw_light.comp

@@ -68,7 +68,7 @@ layout(push_constant, std430) uniform UboPush {
   } push;
 
 //layout(binding = 0, rgba8) uniform image2D   outputImage;
-layout(binding = 0, r32ui) uniform uimage2D outputImage;
+layout(binding = 0, r32ui)  uniform uimage2D outputImage;
 layout(binding = 1, std140) uniform UboScene {
   SceneDesc scene;
   };

shader/virtual_shadow/vsm_clump_pages.comp

@@ -36,7 +36,7 @@ void trimMip(int mip) {
       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) {
+      if(a+b+c+d == 4) {
         imageStore(pageTbl, ax, uvec4(0));
         }
       }
@@ -62,7 +62,7 @@ void mainGroups() {
     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) {
+    if(a+b+c+d >= 2) {
       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));
@@ -77,7 +77,7 @@ void mainGroups() {
     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) {
+    if(a+b+c+d >= 3) {
       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));
@@ -95,62 +95,3 @@ void main() {
 #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);
-
-#if 0
-  if(at.z==1) {
-    uint pageS = 0;
-    if(all(lessThan(at, size)))
-      pageS = imageLoad(pageTbl, at).r;
-
-    const int step = VSM_PAGE_TBL_SIZE/4;
-    if(all(lessThan(ivec2(step), at.xy)) && all(lessThan(at.xy, ivec2(VSM_PAGE_TBL_SIZE-step)))) {
-      ivec3 x = ivec3(ivec2(at.xy - step)*2, 0);
-
-      bool t = (imageLoad(pageTbl, x+ivec3(0,0,0)).r>0) &&
-               (imageLoad(pageTbl, x+ivec3(1,0,0)).r>0) &&
-               (imageLoad(pageTbl, x+ivec3(0,1,0)).r>0) &&
-               (imageLoad(pageTbl, x+ivec3(1,1,0)).r>0);
-
-      if(t)
-        ;//imageStore(pageTbl, at, uvec4(0));
-      }
-    }
-#endif
-
-  uint pageS = 0;
-  if(all(lessThan(at, size)))
-    pageS = imageLoad(pageTbl, at).r;
-  pageVal[id.x][id.y] = pageS;
-  barrier();
-
-  // only 2x2 for now
-  if(id.x%2==0 && id.y%2==0) {
-    uint pageL = pageVal[id.x+1][id.y+0];
-    uint pageR = pageVal[id.x+0][id.y+1];
-    uint pageD = pageVal[id.x+1][id.y+1];
-
-    if(pageS==1 && pageL==1 && pageR==1 && pageD==1) {
-      pageS                   = 2;
-      pageVal[id.x+0][id.y+0] = 2;
-      pageVal[id.x+1][id.y+0] = 0;
-      pageVal[id.x+0][id.y+1] = 0;
-      pageVal[id.x+1][id.y+1] = 0;
-      }
-    }
-  barrier();
-
-  if(pageS>1) {
-    imageStore(pageTbl, at+ivec3(0,0,0), uvec4(pageS,0,0,0));
-    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));
-    }
-  }
-*/

shader/virtual_shadow/vsm_cluster_task.comp

@@ -56,7 +56,7 @@ bool projectSphere(const vec4 sphere, out vec4 aabb, out float depthMin) {
   const float R = sphere.w;
 
   const float smWidthInv = 1.0/(1024.0); // NOTE: from camera.cpp
-  const float zScale     = 1.0/(5*5120.0);
+  const float zScale     = 1.0/(10*5120.0);
 
   float Rz = R * zScale;
   if(c.z - Rz > 1 || c.z + Rz < 0)

shader/virtual_shadow/vsm_common.glsl

@@ -1,10 +1,10 @@
 #ifndef VSM_COMMON_GLSL
 #define VSM_COMMON_GLSL
 
-const int VSM_MAX_PAGES     = 1024;
 const int VSM_PAGE_SIZE     = 128;
-const int VSM_PAGE_TBL_SIZE = 32;   // small for testing, 64 can be better
-const int VSM_PAGE_PER_ROW  = 32;
+const int VSM_PAGE_TBL_SIZE = 32;  // small for testing, 64 can be better
+const int VSM_PAGE_PER_ROW  = 4096/VSM_PAGE_SIZE;
+const int VSM_MAX_PAGES     = VSM_PAGE_PER_ROW * VSM_PAGE_PER_ROW; // 1024;
 const int VSM_CLIPMAP_SIZE  = VSM_PAGE_SIZE * VSM_PAGE_TBL_SIZE;
 
 struct VsmHeader {