split page listing and page alocation

#681

game/graphics/renderer.cpp

@@ -211,7 +211,7 @@ void Renderer::resetSwapchain() {
 
     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 + 16*4)*sizeof(uint32_t));
+    vsm.pageList        = device.ssbo(nullptr, (pageCount + 4 + 16*4 + 16*4)*sizeof(uint32_t));
     }
 
   if(settings.swrEnabled) {
@@ -890,6 +890,10 @@ void Renderer::drawVsm(Tempest::Encoder<Tempest::CommandBuffer>& cmd, uint8_t fI
     cmd.dispatchThreads(size_t(vsm.pageDataCs.w()), size_t(vsm.pageDataCs.h()));
     }
 
+  // list?
+  cmd.setUniforms(shaders.vsmListPages, vsm.uboAlloc);
+  cmd.dispatch(size_t(vsm.pageTbl.d()));
+
   // alloc
   cmd.setUniforms(shaders.vsmAllocPages, vsm.uboAlloc);
   cmd.dispatch(1);

game/graphics/shaders.cpp

@@ -203,6 +203,7 @@ Shaders::Shaders() {
     vsmMarkSky     = computeShader("fog3d_vsm_mark_pages.comp.sprv");
     vsmTrimPages   = computeShader("vsm_trim_pages.comp.sprv");
     vsmClumpPages  = computeShader("vsm_clump_pages.comp.sprv");
+    vsmListPages   = computeShader("vsm_list_pages.comp.sprv");
     vsmAllocPages  = computeShader("vsm_alloc_pages.comp.sprv");
     vsmMergePages  = computeShader("vsm_merge_pages.comp.sprv");
     vsmPackDraw0   = computeShader("vsm_pack_draws0.comp.sprv");

game/graphics/shaders.h

@@ -77,7 +77,7 @@ class Shaders {
     // Virtual shadow
     Tempest::ComputePipeline vsmClusterTask;
     Tempest::ComputePipeline vsmClear, vsmClearPages, vsmMarkPages, vsmMarkSky;
-    Tempest::ComputePipeline vsmTrimPages, vsmClumpPages, vsmAllocPages, vsmMergePages;
+    Tempest::ComputePipeline vsmTrimPages, vsmListPages, vsmClumpPages, vsmAllocPages, vsmMergePages;
     Tempest::ComputePipeline vsmPackDraw0, vsmPackDraw1;
     Tempest::RenderPipeline  vsmDirectLight;
     Tempest::RenderPipeline  vsmReprojectSm;

shader/CMakeLists.txt

@@ -306,12 +306,12 @@ add_shader(vsm_clear.comp          virtual_shadow/vsm_clear.comp)
 add_shader(vsm_clear_pages.comp    virtual_shadow/vsm_clear_pages.comp)
 add_shader(vsm_trim_pages.comp     virtual_shadow/vsm_trim_pages.comp)
 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_alloc_pages.comp    virtual_shadow/vsm_alloc_pages.comp)
 add_shader(vsm_merge_pages.comp    virtual_shadow/vsm_merge_pages.comp)
 add_shader(vsm_pack_draws0.comp    virtual_shadow/vsm_pack_draws.comp   -DPASS0)
 add_shader(vsm_pack_draws1.comp    virtual_shadow/vsm_pack_draws.comp   -DPASS1)
 add_shader(vsm_cluster_task.comp   virtual_shadow/vsm_cluster_task.comp -DVIRTUAL_SHADOW)
-#add_shader(vsm_list_pages.comp     virtual_shadow/vsm_list_pages.comp)
 
 add_shader(vsm_rendering.comp      virtual_shadow/vsm_rendering.comp)
 

shader/virtual_shadow/vsm_alloc_pages.comp

@@ -24,12 +24,6 @@ shared uint bits[(NumThreads+31)/32];
 shared uint ladder  [VSM_PAGE_TBL_SIZE];
 shared uint ladderWr[VSM_PAGE_TBL_SIZE];
 
-ivec2 loadPageSize(ivec3 at) {
-  uint a = imageLoad(pageTbl, at).r;
-  a = a >> 1;
-  return ivec2(a, a >> 4) & 0xF;
-  }
-
 void storePage(ivec2 pId, ivec3 at, ivec2 tile, ivec2 size) {
   const uint pageId = pId.x + pId.y*VSM_PAGE_PER_ROW;
 
@@ -55,31 +49,12 @@ uint bitCountShared(uint till) {
   }
 
 void listPages() {
-  const ivec3 size = imageSize(pageTbl);
-  const uint  lane = gl_LocalInvocationIndex;
-  const ivec3 at   = ivec3(gl_GlobalInvocationID);
-  pageListSize = 0;
+  const uint lane = gl_LocalInvocationIndex;
 
-  for(int i=0; i<1111 && i<size.z; ++i) {
-    const ivec2 sz = loadPageSize(ivec3(at.xy, i));
-    if(lane < bits.length())
-      bits[lane] = 0;
-    barrier();
-
-    if(sz.x>0)
-      atomicOr(bits[lane/32], 1u<<(lane%32));
-    barrier();
-
-    if(sz.x>0) {
-      uint  id = pageListSize + bitCountShared(lane);
-      pageList[id] = packVsmPageInfo(ivec3(at.xy, i), sz);
-      }
-    barrier();
-
-    if(sz.x>0)
-      atomicAdd(pageListSize, 1);
-    barrier();
-    }
+  pageListSize = vsm.header.pageCount;
+  for(uint i=lane; i<pageListSize; i+=NumThreads)
+    pageList[i] = vsm.pageList[i];
+  barrier();
   }
 
 ivec2 findRowCol(in uint page, const uvec2 size) {
@@ -141,20 +116,21 @@ void main() {
   const ivec3 at   = ivec3(gl_GlobalInvocationID);
   const ivec3 id   = ivec3(gl_LocalInvocationID);
   const uint  lane = gl_LocalInvocationIndex;
+  if(lane < ladderWr.length())
+    ladderWr[lane] = 0;
+  barrier();
 
   imageStore(dbg, at.xy, uvec4(0));
   memoryBarrierImage();
   barrier();
 
-  if(lane < ladderWr.length())
-    ladderWr[lane] = 0;
+  listPages();
+
   if(lane < vsm.pageList.length())
     vsm.pageList[lane] = 0;
   memoryBarrierBuffer();
   barrier();
 
-  listPages();
-
   const uint  frag = lane<pageListSize ? pageList[lane] : 0;
   const uvec2 size = unpackVsmPageSize(frag);
   barrier();

shader/virtual_shadow/vsm_clear.comp

@@ -33,6 +33,8 @@ void main() {
     header.counterV     = 0;
     }
 
-  if(gl_LocalInvocationIndex<16)
-    header.pageBbox[gl_LocalInvocationIndex] = ivec4(0xFF, 0xFF, 0, 0);
+  if(gl_LocalInvocationIndex<16) {
+    header.pagePerMip[gl_LocalInvocationIndex] = 0;
+    header.pageBbox  [gl_LocalInvocationIndex] = ivec4(0xFF, 0xFF, 0, 0);
+    }
   }

shader/virtual_shadow/vsm_clump_pages.comp

@@ -70,6 +70,8 @@ void storePageTable() {
     uint  p = unpackPrimeBit(d);
     uvec2 s = uvec2(unpackSize(d));
     imageStore(pageTbl, at, uvec4(p | (s.x << 1) | (s.y << 5)));
+    if(s.x>0)
+      atomicAdd(vsm.header.pagePerMip[at.z], 1);
     }
   }
 

shader/virtual_shadow/vsm_common.glsl

@@ -18,6 +18,7 @@ struct VsmHeader {
   uint  meshletCount;
   uint  counterM;
   uint  counterV;
+  uint  pagePerMip[VSM_PAGE_MIPS];
   ivec4 pageBbox[VSM_PAGE_MIPS];
   };
 

shader/virtual_shadow/vsm_list_pages.comp

@@ -8,27 +8,71 @@
 #include "scene.glsl"
 #include "common.glsl"
 
-//layout(local_size_x = 64) in;
-layout(local_size_x = 8, local_size_y = 8) in;
+layout(local_size_x = 32, local_size_y = 32) 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;
+layout(binding = 1, r32ui)  uniform uimage3D pageTbl;
+layout(binding = 2, r32ui)  uniform uimage2D dbg;
 
 const uint NumThreads = gl_WorkGroupSize.x*gl_WorkGroupSize.y*gl_WorkGroupSize.z;
 
-void main() {
-  const ivec3 at   = ivec3(gl_GlobalInvocationID.xyz);
+shared uint pageListPrefix;
+shared uint bits[(NumThreads+31)/32];
+
+ivec2 loadPageSize(ivec3 at) {
+  uint a = imageLoad(pageTbl, at).r;
+  a = a >> 1;
+  return ivec2(a, a >> 4) & 0xF;
+  }
+
+uint bitCountShared(uint till) {
+  uint n = till/32;
+  uint f = till%32;
+  uint r = 0;
+  for(uint i=0; i<n; ++i)
+    r += bitCount(bits[i]);
+  r += bitCount(bits[n] & ((1u << f)-1u));
+  return r;
+  }
+
+void listPages(uint mip) {
   const ivec3 size = imageSize(pageTbl);
-  if(any(greaterThanEqual(at, size)))
-    return;
+  const uint  lane = gl_LocalInvocationIndex;
+  const ivec3 at   = ivec3(gl_LocalInvocationID.xy, mip);
+
+  if(lane < bits.length())
+    bits[lane] = 0;
+  barrier();
+
+  const ivec2 sz = loadPageSize(at);
+  if(sz.x>0)
+    atomicOr(bits[lane/32], 1u<<(lane%32));
+  barrier();
+
+  if(sz.x>0) {
+    const uint id = pageListPrefix + bitCountShared(lane);
+    vsm.pageList[id] = packVsmPageInfo(at, sz);
+    }
+  barrier();
+  }
+
+void main() {
+  const uint lane = gl_LocalInvocationIndex;
+  const uint mip  = gl_WorkGroupID.x;
+
+  pageListPrefix = 0;
+  barrier();
+
+  if(lane<mip)
+    atomicAdd(pageListPrefix, vsm.header.pagePerMip[lane]);
+  barrier();
+
+  if(mip+1==VSM_PAGE_MIPS && lane==mip) {
+    vsm.header.pageCount = pageListPrefix + vsm.header.pagePerMip[lane];
+    }
 
-  uint v = imageLoad(pageTbl, at).x;
-  if(v==0)
+  if(vsm.header.pagePerMip[mip]==0)
     return;
 
-  uint i = atomicAdd(vsm.header.pageCount, 1);
-  vsm.pageList[i] = 0xFFFFFFFF;
-  imageStore(pageTbl, at, uvec4(i));
+  listPages(mip);
   }