task emulation cleanup

#38

Engine/gapi/spirv/meshconverter.cpp

@@ -1131,11 +1131,11 @@ void MeshConverter::emitPayoadLoad(libspirv::MutableBytecode& comp, uint32_t eng
     });
 
   // Constants
-  const uint32_t const0    = constants[0];
-  const uint32_t const1    = constants[1];
-  const uint32_t const7    = constants[7];
-  const uint32_t const16   = comp.OpConstant(fn, uint_t, 16);
-  const uint32_t constFFFF = comp.OpConstant(fn, uint_t, 0xFFFF);
+  const uint32_t const0     = constants[0];
+  const uint32_t const1     = constants[1];
+  const uint32_t const7     = constants[7];
+  const uint32_t const20    = comp.OpConstant(fn, uint_t, 20);
+  const uint32_t constFFFFF = comp.OpConstant(fn, uint_t, 0xFFFFF);
 
   // Function
   fn = comp.end();
@@ -1187,13 +1187,13 @@ void MeshConverter::emitPayoadLoad(libspirv::MutableBytecode& comp, uint32_t eng
   uint32_t rAt = comp.fetchAddBound();
   {
     // firstTask_lutPtr
-    fn.insert(spv::OpBitwiseAnd, {uint_t, rAt, firstTask_lutPtr, constFFFF});
+    fn.insert(spv::OpBitwiseAnd, {uint_t, rAt, firstTask_lutPtr, constFFFFF});
   }
 
   if(usr_WorkGroupID!=0) {
     // firstTask_lutPtr
     uint32_t rWg0 = comp.fetchAddBound();
-    fn.insert(spv::OpShiftRightLogical, {uint_t, rWg0, firstTask_lutPtr, const16});
+    fn.insert(spv::OpShiftRightLogical, {uint_t, rWg0, firstTask_lutPtr, const20});
 
     const uint32_t ptrWgId = comp.fetchAddBound();
     fn.insert(spv::OpAccessChain, {_ptr_Input_uint, ptrWgId, gl_WorkGroupID, const0});

Engine/gapi/vulkan/vcommandbuffer.cpp

@@ -1074,6 +1074,7 @@ void VMeshCommandBuffer::pushChunk() {
     ch.impl = cbTask;
     chunks.push(ch);
     cbTask = nullptr;
+    taskIndirectId = 0;
     }
   if(cbMesh!=nullptr) {
     auto& ms = *device.meshHelper;
@@ -1084,6 +1085,7 @@ void VMeshCommandBuffer::pushChunk() {
     ch.impl = cbMesh;
     chunks.push(ch);
     cbMesh = nullptr;
+    meshIndirectId = 0;
     }
   VCommandBuffer::pushChunk();
   }
@@ -1096,11 +1098,6 @@ void VMeshCommandBuffer::setPipeline(AbstractGraphicsApi::Pipeline& p) {
 
   auto& ms = *device.meshHelper;
 
-  if(cbTask==VK_NULL_HANDLE && cbMesh==VK_NULL_HANDLE) {
-    taskIndirectId = 0;
-    meshIndirectId = 0;
-    }
-
   if(cbTask==VK_NULL_HANDLE && px.taskPipeline()!=VK_NULL_HANDLE) {
     VkCommandBufferAllocateInfo allocInfo = {};
     allocInfo.sType              = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;

Engine/gapi/vulkan/vmeshlethelper.cpp

@@ -56,7 +56,7 @@ VMeshletHelper::VMeshletHelper(VDevice& dev) : dev(dev) {
       uint32_t indirectRate;
       uint32_t indirectCmdCount;
       } push;
-    push.indirectRate     = 2;
+    push.indirectRate     = indirectRate;
     push.indirectCmdCount = 0;
 
     auto c = dev.dataMgr().get();
@@ -126,7 +126,7 @@ void VMeshletHelper::bindVS(VkCommandBuffer impl, VkPipelineLayout lay) {
   }
 
 void VMeshletHelper::drawIndirect(VkCommandBuffer impl, uint32_t drawId) {
-  assert(drawId<IndirectCmdCount);
+  assert(drawId*indirectRate < IndirectCmdCount);
   uint32_t off = drawId*indirectOffset + 2*sizeof(uint32_t);
   vkCmdDrawIndexedIndirect(impl, indirect.impl, off, 1, 0);
   }
@@ -171,10 +171,10 @@ void VMeshletHelper::taskEpiloguePass(VkCommandBuffer impl, uint32_t meshCallsCo
     uint32_t indirectRate;
     uint32_t indirectCmdCount;
     } push;
-  push.indirectRate     = 2;
+  push.indirectRate     = indirectRate;
   push.indirectCmdCount = meshCallsCount;
 
-  // prefix summ pass
+  // post-process pass
   barrier(impl, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
           VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
           VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT);
@@ -211,7 +211,7 @@ void VMeshletHelper::sortPass(VkCommandBuffer impl, uint32_t meshCallsCount) {
     uint32_t indirectRate;
     uint32_t indirectCmdCount;
     } push;
-  push.indirectRate     = 2;
+  push.indirectRate     = indirectRate;
   push.indirectCmdCount = meshCallsCount;
 
   // prefix summ pass

Engine/gapi/vulkan/vmeshlethelper.h

@@ -14,10 +14,6 @@ class VCompPipeline;
 
 class VMeshletHelper {
   private:
-    struct IndirectCommand {
-      uint32_t indexCount;
-      };
-
     struct DrawIndexedIndirectCommand {
       uint32_t drawId;
       uint32_t indexCountSrc;
@@ -26,7 +22,7 @@ class VMeshletHelper {
       uint32_t firstIndex;    // prefix sum
       int32_t  vertexOffset;  // can be abused to offset into var_buffer
       uint32_t firstInstance; // caps: should be zero
-      uint32_t vboOffset;     // prefix sum
+      uint32_t lutPtr;        // pointer to task-payload
       };
 
   public:
@@ -36,9 +32,7 @@ class VMeshletHelper {
       PipelinewordsCount = 32*1024*1024,
       PipelineMemorySize = PipelinewordsCount*4,
       MeshletsMemorySize = MeshletsMaxCount*3*4,
-
-      IndirectScratchSize = IndirectCmdCount*sizeof(IndirectCommand),
-      IndirectMemorySize  = IndirectCmdCount*sizeof(VkDrawIndexedIndirectCommand),
+      IndirectMemorySize = IndirectCmdCount*sizeof(DrawIndexedIndirectCommand),
       };
     explicit VMeshletHelper(VDevice& dev);
     ~VMeshletHelper();

Engine/shaders/mesh_compactage.comp

@@ -20,7 +20,7 @@ struct DrawIndexedIndirectCommand
   uint firstIndex;    // prefix sum
   int  vertexOffset;  // can be abused to offset into var_buffer
   uint firstInstance; // caps: should be zero
-  uint padding0;
+  uint lutPtr;
 };
 
 layout(binding = 0, std430) restrict buffer EngineInternal0

Engine/shaders/mesh_prefix_pass.comp

@@ -78,7 +78,7 @@ void main() {
   for(uint i=b; i<e; ++i) {
     uint idx        = i*indirectRate;
     uint indexCount = indirect[idx].indexCountSrc;
-    uint firstIndex = grow + prefixIbo;
+    uint firstIndex = indexCount>0 ? grow + prefixIbo : 0;
     uint inst       = indexCount>0 ? 1 : 0;
 
     prefixIbo += indexCount;

Engine/shaders/task_lut_pass.comp

@@ -65,18 +65,22 @@ void main() {
     if(at>=total)
       break;
 
-    const Descriptor d = mesh.desc[at+first];
+    const Descriptor d   = mesh.desc[at+first];
+    const uint       idx = d.drawId*indirectRate;
+    if(indirect[idx].lutPtr==0) {
+      continue;
+      }
+
     if(gl_LocalInvocationIndex==0) {
-      uint idx      = d.drawId*indirectRate;
       uint meshlets = atomicAdd(indirect[idx].lutIterator, d.dispatchSz);
       lutOffset     = meshlets + indirect[idx].lutPtr;
       }
     barrier();
 
     [[loop]]
     for(uint i=gl_LocalInvocationIndex; i<d.dispatchSz; i+=gl_WorkGroupSize.x) {
-      const uint usrWorkGroup = d.ptr!=0 ? i : 0;
-      var.heap[lutOffset+i] = (usrWorkGroup << 16u) | (d.ptr & 0xFFFF);
+      const uint usrWorkGroup = i;
+      var.heap[lutOffset+i] = (usrWorkGroup << 20u) | (d.ptr & 0xFFFFF);
       }
     }
 

Engine/shaders/task_post_pass.comp

@@ -48,6 +48,11 @@ layout(push_constant, std430) uniform UboPush {
   };
 
 void main() {
+  if(var.grow==0) {
+    var.grow = 1; // zero is reserved as null-value
+    }
+  barrier();
+
   uint index = gl_LocalInvocationID.x;
   uint len   = indirectCmdCount;