rendering fixes

sschoen · Nov 25, 2023 · aaf6aae · aaf6aae
1 parent ec73604
commit aaf6aae
Show file tree

Hide file tree

Showing 5 changed files with 57 additions and 32 deletions.
diff --git a/WickedEngine/shaders/impostorVS.hlsl b/WickedEngine/shaders/impostorVS.hlsl
@@ -24,6 +24,6 @@ VSOut main(uint vertexID : SV_VertexID)
  Out.slice = data.x & 0xFFFFFF;
  Out.dither = float((data.x >> 24u) & 0xFF) / 255.0;
  Out.instanceColor = data.y;
- Out.primitiveID = impostorID * 2;
+ Out.primitiveID = vertexID / 2u;
  return Out;
 }
diff --git a/WickedEngine/shaders/impostor_prepareCS.hlsl b/WickedEngine/shaders/impostor_prepareCS.hlsl
@@ -60,12 +60,12 @@ void main(uint3 DTid : SV_DispatchThreadID)
  const uint vertexOffset = impostorOffset * 4u;
 
  // Write out indices:
- output_indices[indexOffset + 0] = vertexOffset + 0;
+ output_indices[indexOffset + 0] = vertexOffset + 0; // provoking vertex prim0
  output_indices[indexOffset + 1] = vertexOffset + 1;
  output_indices[indexOffset + 2] = vertexOffset + 2;
- output_indices[indexOffset + 3] = vertexOffset + 2;
- output_indices[indexOffset + 4] = vertexOffset + 1;
- output_indices[indexOffset + 5] = vertexOffset + 3;
+ output_indices[indexOffset + 3] = vertexOffset + 3; // provoking vertex prim1
+ output_indices[indexOffset + 4] = vertexOffset + 2;
+ output_indices[indexOffset + 5] = vertexOffset + 1;
 
  // We rotate the billboard to face camera, but unlike emitted particles, 
  // they don't rotate according to camera rotation, but the camera position relative

diff --git a/WickedEngine/wiRenderPath3D.cpp b/WickedEngine/wiRenderPath3D.cpp
@@ -117,7 +117,11 @@ namespace wi
  {
  TextureDesc desc;
  desc.format = wi::renderer::format_idbuffer;
- desc.bind_flags = BindFlag::RENDER_TARGET | BindFlag::SHADER_RESOURCE | BindFlag::UNORDERED_ACCESS;
+ desc.bind_flags = BindFlag::RENDER_TARGET | BindFlag::SHADER_RESOURCE;
+ if (getMSAASampleCount() > 1)
+ {
+ desc.bind_flags |= BindFlag::UNORDERED_ACCESS;
+ }
  desc.width = internalResolution.x;
  desc.height = internalResolution.y;
  desc.sample_count = 1;
@@ -817,6 +821,8 @@ namespace wi
 
  wi::renderer::OcclusionCulling_Reset(visibility_main, cmd); // must be outside renderpass!
 
+ wi::renderer::RefreshImpostors(*scene, cmd);
+
  RenderPassImage rp[] = {
  RenderPassImage::DepthStencil(
  &depthBuffer_Main,
@@ -1036,7 +1042,6 @@ namespace wi
  );
  wi::renderer::RefreshLightmaps(*scene, cmd);
  wi::renderer::RefreshEnvProbes(visibility_main, cmd);
- wi::renderer::RefreshImpostors(*scene, cmd);
  });
 
  if (visibility_main.IsRequestedPlanarReflections())
@@ -2042,7 +2047,7 @@ namespace wi
  }
  }
 
- if (getDepthOfFieldEnabled() && camera->aperture_size > 0 && getDepthOfFieldStrength() > 0 && depthoffieldResources.IsValid())
+ if (getDepthOfFieldEnabled() && camera->aperture_size > 0.001f && getDepthOfFieldStrength() > 0.001f && depthoffieldResources.IsValid())
  {
  wi::renderer::Postprocess_DepthOfField(
  depthoffieldResources,

diff --git a/WickedEngine/wiRenderer.cpp b/WickedEngine/wiRenderer.cpp
@@ -9944,6 +9944,25 @@ void ComputeBloom(
  device->EventBegin("Bloom", cmd);
  auto range = wi::profiler::BeginRangeGPU("Bloom", cmd);
 
+ {
+ GPUBarrier barriers[] = {
+ GPUBarrier::Image(&res.texture_bloom, res.texture_bloom.desc.layout, ResourceState::UNORDERED_ACCESS),
+ GPUBarrier::Image(&res.texture_temp, res.texture_temp.desc.layout, ResourceState::UNORDERED_ACCESS),
+ };
+ device->Barrier(barriers, arraysize(barriers), cmd);
+ }
+
+ device->ClearUAV(&res.texture_bloom, 0, cmd);
+ device->ClearUAV(&res.texture_temp, 0, cmd);
+
+ {
+ GPUBarrier barriers[] = {
+ GPUBarrier::Memory(&res.texture_bloom),
+ GPUBarrier::Memory(&res.texture_temp),
+ };
+ device->Barrier(barriers, arraysize(barriers), cmd);
+ }
+
  // Separate bright parts of image to bloom texture:
  {
  device->EventBegin("Bloom Separate", cmd);
@@ -9962,31 +9981,24 @@ void ComputeBloom(
  bloom.buffer_input_luminance = device->GetDescriptorIndex((buffer_luminance == nullptr) ? &luminance_dummy : buffer_luminance, SubresourceType::SRV);
  device->PushConstants(&bloom, sizeof(bloom), cmd);
 
- {
- GPUBarrier barriers[] = {
- GPUBarrier::Image(&res.texture_bloom, res.texture_bloom.desc.layout, ResourceState::UNORDERED_ACCESS),
- };
- device->Barrier(barriers, arraysize(barriers), cmd);
- }
-
  device->Dispatch(
  (desc.width + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
  (desc.height + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
  1,
  cmd
  );
 
- {
- GPUBarrier barriers[] = {
- GPUBarrier::Memory(),
- GPUBarrier::Image(&res.texture_bloom, ResourceState::UNORDERED_ACCESS, res.texture_bloom.desc.layout),
- };
- device->Barrier(barriers, arraysize(barriers), cmd);
- }
-
  device->EventEnd(cmd);
  }
 
+ {
+ GPUBarrier barriers[] = {
+ GPUBarrier::Image(&res.texture_bloom, ResourceState::UNORDERED_ACCESS, res.texture_bloom.desc.layout),
+ GPUBarrier::Image(&res.texture_temp, ResourceState::UNORDERED_ACCESS, res.texture_temp.desc.layout),
+ };
+ device->Barrier(barriers, arraysize(barriers), cmd);
+ }
+
  device->EventBegin("Bloom Mipchain", cmd);
  MIPGEN_OPTIONS mipopt;
  mipopt.gaussian_temp = &res.texture_temp;
@@ -14977,6 +14989,22 @@ void Postprocess_Tonemap(
 
  device->EventBegin("Postprocess_Tonemap", cmd);
 
+ {
+ GPUBarrier barriers[] = {
+ GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
+ };
+ device->Barrier(barriers, arraysize(barriers), cmd);
+ }
+
+ device->ClearUAV(&output, 0, cmd);
+
+ {
+ GPUBarrier barriers[] = {
+ GPUBarrier::Memory(&output),
+ };
+ device->Barrier(barriers, arraysize(barriers), cmd);
+ }
+
  device->BindComputeShader(&shaders[CSTYPE_POSTPROCESS_TONEMAP], cmd);
 
  const TextureDesc& desc = output.GetDesc();
@@ -15008,13 +15036,6 @@ void Postprocess_Tonemap(
  tonemap_push.display_colorspace = static_cast<uint32_t>(display_colorspace);
  device->PushConstants(&tonemap_push, sizeof(tonemap_push), cmd);
 
- {
- GPUBarrier barriers[] = {
- GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
- };
- device->Barrier(barriers, arraysize(barriers), cmd);
- }
-
  device->Dispatch(
  (desc.width + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
  (desc.height + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
@@ -15024,7 +15045,6 @@ void Postprocess_Tonemap(
 
  {
  GPUBarrier barriers[] = {
- GPUBarrier::Memory(),
  GPUBarrier::Image(&output, ResourceState::UNORDERED_ACCESS, output.desc.layout),
  };
  device->Barrier(barriers, arraysize(barriers), cmd);

diff --git a/WickedEngine/wiVersion.cpp b/WickedEngine/wiVersion.cpp
@@ -9,7 +9,7 @@ namespace wi::version
  // minor features, major updates, breaking compatibility changes
  const int minor = 71;
  // minor bug fixes, alterations, refactors, updates
- const int revision = 352;
+ const int revision = 353;
 
  const std::string version_string = std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(revision);