VideoBackends:D3D12: Dynamic vertex loader support

8 files changed, 129 insertions, 27 deletions

diff --git a/Source/Core/VideoBackends/D3D12/D3D12Renderer.cpp b/Source/Core/VideoBackends/D3D12/D3D12Renderer.cpp
index bfe44a3a58..80872086a9 100644
--- a/Source/Core/VideoBackends/D3D12/D3D12Renderer.cpp
+++ b/Source/Core/VideoBackends/D3D12/D3D12Renderer.cpp
@@ -180,7 +180,8 @@ void Renderer::SetPipeline(const AbstractPipeline* pipeline)
       m_state.root_signature = dx_pipeline->GetRootSignature();
       m_dirty_bits |= DirtyState_RootSignature | DirtyState_PS_CBV | DirtyState_VS_CBV |
                       DirtyState_GS_CBV | DirtyState_SRV_Descriptor |
-                      DirtyState_Sampler_Descriptor | DirtyState_UAV_Descriptor;
+                      DirtyState_Sampler_Descriptor | DirtyState_UAV_Descriptor |
+                      DirtyState_VS_SRV_Descriptor;
     }
     if (dx_pipeline->UseIntegerRTV() != m_state.using_integer_rtv)
     {
@@ -362,6 +363,11 @@ void Renderer::DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex)
   if (!ApplyState())
     return;
 
+  // DX12 is great and doesn't include the base vertex in SV_VertexID
+  if (static_cast<const DXPipeline*>(m_current_pipeline)->GetUsage() ==
+      AbstractPipelineUsage::GXUber)
+    g_dx_context->GetCommandList()->SetGraphicsRoot32BitConstant(
+        ROOT_PARAMETER_BASE_VERTEX_CONSTANT, base_vertex, 0);
   g_dx_context->GetCommandList()->DrawIndexedInstanced(num_indices, 1, base_index, base_vertex, 0);
 }
 
@@ -494,18 +500,22 @@ void Renderer::SetPixelShaderUAV(D3D12_CPU_DESCRIPTOR_HANDLE handle)
   m_dirty_bits |= DirtyState_PS_UAV;
 }
 
-void Renderer::SetVertexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 stride, u32 size)
+void Renderer::SetVertexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, D3D12_CPU_DESCRIPTOR_HANDLE srv,
+                               u32 stride, u32 size)
 {
-  if (m_state.vertex_buffer.BufferLocation == address &&
-      m_state.vertex_buffer.StrideInBytes == stride && m_state.vertex_buffer.SizeInBytes == size)
+  if (m_state.vertex_buffer.BufferLocation != address ||
+      m_state.vertex_buffer.StrideInBytes != stride || m_state.vertex_buffer.SizeInBytes != size)
   {
-    return;
+    m_state.vertex_buffer.BufferLocation = address;
+    m_state.vertex_buffer.StrideInBytes = stride;
+    m_state.vertex_buffer.SizeInBytes = size;
+    m_dirty_bits |= DirtyState_VertexBuffer;
+  }
+  if (m_state.vs_srv.ptr != srv.ptr)
+  {
+    m_state.vs_srv = srv;
+    m_dirty_bits |= DirtyState_VS_SRV;
   }
-
-  m_state.vertex_buffer.BufferLocation = address;
-  m_state.vertex_buffer.StrideInBytes = stride;
-  m_state.vertex_buffer.SizeInBytes = size;
-  m_dirty_bits |= DirtyState_VertexBuffer;
 }
 
 void Renderer::SetIndexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 size, DXGI_FORMAT format)
@@ -535,15 +545,17 @@ bool Renderer::ApplyState()
   // Clear bits before actually changing state. Some state (e.g. cbuffers) can't be set
   // if utility pipelines are bound.
   const u32 dirty_bits = m_dirty_bits;
-  m_dirty_bits &= ~(
-      DirtyState_Framebuffer | DirtyState_Pipeline | DirtyState_Viewport | DirtyState_ScissorRect |
-      DirtyState_PS_UAV | DirtyState_PS_CBV | DirtyState_VS_CBV | DirtyState_GS_CBV |
-      DirtyState_SRV_Descriptor | DirtyState_Sampler_Descriptor | DirtyState_UAV_Descriptor |
-      DirtyState_VertexBuffer | DirtyState_IndexBuffer | DirtyState_PrimitiveTopology);
+  m_dirty_bits &=
+      ~(DirtyState_Framebuffer | DirtyState_Pipeline | DirtyState_Viewport |
+        DirtyState_ScissorRect | DirtyState_PS_UAV | DirtyState_PS_CBV | DirtyState_VS_CBV |
+        DirtyState_GS_CBV | DirtyState_SRV_Descriptor | DirtyState_Sampler_Descriptor |
+        DirtyState_UAV_Descriptor | DirtyState_VertexBuffer | DirtyState_IndexBuffer |
+        DirtyState_PrimitiveTopology | DirtyState_VS_SRV_Descriptor);
 
   auto* const cmdlist = g_dx_context->GetCommandList();
+  auto* const pipeline = static_cast<const DXPipeline*>(m_current_pipeline);
   if (dirty_bits & DirtyState_Pipeline)
-    cmdlist->SetPipelineState(static_cast<const DXPipeline*>(m_current_pipeline)->GetPipeline());
+    cmdlist->SetPipelineState(pipeline->GetPipeline());
 
   if (dirty_bits & DirtyState_Framebuffer)
     BindFramebuffer(static_cast<DXFramebuffer*>(m_current_framebuffer));
@@ -572,8 +584,7 @@ bool Renderer::ApplyState()
                                             m_state.sampler_descriptor_base);
   }
 
-  if (static_cast<const DXPipeline*>(m_current_pipeline)->GetUsage() !=
-      AbstractPipelineUsage::Utility)
+  if (pipeline->GetUsage() != AbstractPipelineUsage::Utility)
   {
     if (dirty_bits & DirtyState_VS_CBV)
     {
@@ -590,6 +601,13 @@ bool Renderer::ApplyState()
       }
     }
 
+    if (dirty_bits & DirtyState_VS_SRV_Descriptor &&
+        pipeline->GetUsage() == AbstractPipelineUsage::GXUber)
+    {
+      cmdlist->SetGraphicsRootDescriptorTable(ROOT_PARAMETER_VS_SRV,
+                                              m_state.vertex_srv_descriptor_base);
+    }
+
     if (dirty_bits & DirtyState_GS_CBV)
     {
       cmdlist->SetGraphicsRootConstantBufferView(ROOT_PARAMETER_GS_CBV,
@@ -642,7 +660,9 @@ void Renderer::UpdateDescriptorTables()
   const bool sampler_update_failed =
       (m_dirty_bits & DirtyState_Samplers) && !UpdateSamplerDescriptorTable();
   const bool uav_update_failed = (m_dirty_bits & DirtyState_PS_UAV) && !UpdateUAVDescriptorTable();
-  if (texture_update_failed || sampler_update_failed || uav_update_failed)
+  const bool srv_update_failed =
+      (m_dirty_bits & DirtyState_VS_SRV) && !UpdateVSSRVDescriptorTable();
+  if (texture_update_failed || sampler_update_failed || uav_update_failed || srv_update_failed)
   {
     WARN_LOG_FMT(VIDEO, "Executing command list while waiting for temporary {}",
                  texture_update_failed ? "descriptors" : "samplers");
@@ -652,6 +672,7 @@ void Renderer::UpdateDescriptorTables()
     UpdateSRVDescriptorTable();
     UpdateSamplerDescriptorTable();
     UpdateUAVDescriptorTable();
+    UpdateVSSRVDescriptorTable();
   }
 }
 
@@ -701,6 +722,26 @@ bool Renderer::UpdateUAVDescriptorTable()
   return true;
 }
 
+bool Renderer::UpdateVSSRVDescriptorTable()
+{
+  if (!g_ActiveConfig.backend_info.bSupportsDynamicVertexLoader ||
+      static_cast<const DXPipeline*>(m_current_pipeline)->GetUsage() !=
+          AbstractPipelineUsage::GXUber)
+  {
+    return true;
+  }
+
+  DescriptorHandle handle;
+  if (!g_dx_context->GetDescriptorAllocator()->Allocate(1, &handle))
+    return false;
+
+  g_dx_context->GetDevice()->CopyDescriptorsSimple(1, handle.cpu_handle, m_state.vs_srv,
+                                                   D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV);
+  m_state.vertex_srv_descriptor_base = handle.gpu_handle;
+  m_dirty_bits = (m_dirty_bits & ~DirtyState_VS_SRV) | DirtyState_VS_SRV_Descriptor;
+  return true;
+}
+
 bool Renderer::UpdateComputeUAVDescriptorTable()
 {
   DescriptorHandle handle;
diff --git a/Source/Core/VideoBackends/D3D12/D3D12Renderer.h b/Source/Core/VideoBackends/D3D12/D3D12Renderer.h
index 9624199aa4..19188a5acf 100644
--- a/Source/Core/VideoBackends/D3D12/D3D12Renderer.h
+++ b/Source/Core/VideoBackends/D3D12/D3D12Renderer.h
@@ -88,7 +88,8 @@ public:
   void SetPixelShaderUAV(D3D12_CPU_DESCRIPTOR_HANDLE handle);
 
   // Graphics vertex/index buffer binding.
-  void SetVertexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 stride, u32 size);
+  void SetVertexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, D3D12_CPU_DESCRIPTOR_HANDLE srv,
+                       u32 stride, u32 size);
   void SetIndexBuffer(D3D12_GPU_VIRTUAL_ADDRESS address, u32 size, DXGI_FORMAT format);
 
   // Binds all dirty state
@@ -126,6 +127,8 @@ private:
     DirtyState_RootSignature = (1 << 17),
     DirtyState_ComputeRootSignature = (1 << 18),
     DirtyState_DescriptorHeaps = (1 << 19),
+    DirtyState_VS_SRV = (1 << 20),
+    DirtyState_VS_SRV_Descriptor = (1 << 21),
 
     DirtyState_All =
         DirtyState_Framebuffer | DirtyState_Pipeline | DirtyState_Textures | DirtyState_Samplers |
@@ -133,7 +136,8 @@ private:
         DirtyState_PS_UAV | DirtyState_PS_CBV | DirtyState_VS_CBV | DirtyState_GS_CBV |
         DirtyState_SRV_Descriptor | DirtyState_Sampler_Descriptor | DirtyState_UAV_Descriptor |
         DirtyState_VertexBuffer | DirtyState_IndexBuffer | DirtyState_PrimitiveTopology |
-        DirtyState_RootSignature | DirtyState_ComputeRootSignature | DirtyState_DescriptorHeaps
+        DirtyState_RootSignature | DirtyState_ComputeRootSignature | DirtyState_DescriptorHeaps |
+        DirtyState_VS_SRV | DirtyState_VS_SRV_Descriptor
   };
 
   void CheckForSwapChainChanges();
@@ -144,6 +148,7 @@ private:
   void UpdateDescriptorTables();
   bool UpdateSRVDescriptorTable();
   bool UpdateUAVDescriptorTable();
+  bool UpdateVSSRVDescriptorTable();
   bool UpdateComputeUAVDescriptorTable();
   bool UpdateSamplerDescriptorTable();
 
@@ -157,11 +162,13 @@ private:
     DXShader* compute_shader = nullptr;
     std::array<D3D12_GPU_VIRTUAL_ADDRESS, 3> constant_buffers = {};
     std::array<D3D12_CPU_DESCRIPTOR_HANDLE, MAX_TEXTURES> textures = {};
+    D3D12_CPU_DESCRIPTOR_HANDLE vs_srv = {};
     D3D12_CPU_DESCRIPTOR_HANDLE ps_uav = {};
     SamplerStateSet samplers = {};
     const DXTexture* compute_image_texture = nullptr;
     D3D12_VIEWPORT viewport = {};
     D3D12_RECT scissor = {};
+    D3D12_GPU_DESCRIPTOR_HANDLE vertex_srv_descriptor_base = {};
     D3D12_GPU_DESCRIPTOR_HANDLE srv_descriptor_base = {};
     D3D12_GPU_DESCRIPTOR_HANDLE sampler_descriptor_base = {};
     D3D12_GPU_DESCRIPTOR_HANDLE uav_descriptor_base = {};
diff --git a/Source/Core/VideoBackends/D3D12/D3D12VertexManager.cpp b/Source/Core/VideoBackends/D3D12/D3D12VertexManager.cpp
index b85e4e459c..5a9efc0d6c 100644
--- a/Source/Core/VideoBackends/D3D12/D3D12VertexManager.cpp
+++ b/Source/Core/VideoBackends/D3D12/D3D12VertexManager.cpp
@@ -64,6 +64,18 @@ bool VertexManager::Initialize()
                                                         &srv_desc, dh.cpu_handle);
   }
 
+  if (!g_dx_context->GetDescriptorHeapManager().Allocate(&m_vertex_srv))
+  {
+    PanicAlertFmt("Failed to allocate descriptor for vertex srv");
+    return false;
+  }
+
+  D3D12_SHADER_RESOURCE_VIEW_DESC srv_desc = {DXGI_FORMAT_R32_UINT, D3D12_SRV_DIMENSION_BUFFER,
+                                              D3D12_DEFAULT_SHADER_4_COMPONENT_MAPPING};
+  srv_desc.Buffer.NumElements = m_vertex_stream_buffer.GetSize() / sizeof(u32);
+  g_dx_context->GetDevice()->CreateShaderResourceView(m_vertex_stream_buffer.GetBuffer(), &srv_desc,
+                                                      m_vertex_srv.cpu_handle);
+
   UploadAllConstants();
   return true;
 }
@@ -115,7 +127,8 @@ void VertexManager::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_in
   ADDSTAT(g_stats.this_frame.bytes_vertex_streamed, static_cast<int>(vertex_data_size));
   ADDSTAT(g_stats.this_frame.bytes_index_streamed, static_cast<int>(index_data_size));
 
-  Renderer::GetInstance()->SetVertexBuffer(m_vertex_stream_buffer.GetGPUPointer(), vertex_stride,
+  Renderer::GetInstance()->SetVertexBuffer(m_vertex_stream_buffer.GetGPUPointer(),
+                                           m_vertex_srv.cpu_handle, vertex_stride,
                                            m_vertex_stream_buffer.GetSize());
   Renderer::GetInstance()->SetIndexBuffer(m_index_stream_buffer.GetGPUPointer(),
                                           m_index_stream_buffer.GetSize(), DXGI_FORMAT_R16_UINT);
diff --git a/Source/Core/VideoBackends/D3D12/D3D12VertexManager.h b/Source/Core/VideoBackends/D3D12/D3D12VertexManager.h
index 6bcff92bb7..2dd2cdfee9 100644
--- a/Source/Core/VideoBackends/D3D12/D3D12VertexManager.h
+++ b/Source/Core/VideoBackends/D3D12/D3D12VertexManager.h
@@ -46,6 +46,7 @@ protected:
   StreamBuffer m_uniform_stream_buffer;
   StreamBuffer m_texel_stream_buffer;
   std::array<DescriptorHandle, NUM_TEXEL_BUFFER_FORMATS> m_texel_buffer_views = {};
+  DescriptorHandle m_vertex_srv = {};
 };
 
 }  // namespace DX12
diff --git a/Source/Core/VideoBackends/D3D12/DX12Context.cpp b/Source/Core/VideoBackends/D3D12/DX12Context.cpp
index e3ff8a7417..5cf0e414c4 100644
--- a/Source/Core/VideoBackends/D3D12/DX12Context.cpp
+++ b/Source/Core/VideoBackends/D3D12/DX12Context.cpp
@@ -261,6 +261,16 @@ bool DXContext::CreateDescriptorHeaps()
   return true;
 }
 
+static void SetRootParamConstant(D3D12_ROOT_PARAMETER* rp, u32 shader_reg, u32 num_values,
+                                 D3D12_SHADER_VISIBILITY visibility)
+{
+  rp->ParameterType = D3D12_ROOT_PARAMETER_TYPE_32BIT_CONSTANTS;
+  rp->Constants.Num32BitValues = num_values;
+  rp->Constants.ShaderRegister = shader_reg;
+  rp->Constants.RegisterSpace = 0;
+  rp->ShaderVisibility = visibility;
+}
+
 static void SetRootParamCBV(D3D12_ROOT_PARAMETER* rp, u32 shader_reg,
                             D3D12_SHADER_VISIBILITY visibility)
 {
@@ -345,6 +355,11 @@ bool DXContext::CreateGXRootSignature()
   param_count++;
   SetRootParamCBV(&params[param_count], 0, D3D12_SHADER_VISIBILITY_GEOMETRY);
   param_count++;
+  SetRootParamTable(&params[param_count], &ranges[param_count], D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 3,
+                    1, D3D12_SHADER_VISIBILITY_VERTEX);
+  param_count++;
+  SetRootParamConstant(&params[param_count], 2, 1, D3D12_SHADER_VISIBILITY_VERTEX);
+  param_count++;
 
   // Since these must be contiguous, pixel lighting goes to bbox if not enabled.
   if (g_ActiveConfig.bBBoxEnable)
diff --git a/Source/Core/VideoBackends/D3D12/DX12Context.h b/Source/Core/VideoBackends/D3D12/DX12Context.h
index d1c5b38798..ad9ef9fe74 100644
--- a/Source/Core/VideoBackends/D3D12/DX12Context.h
+++ b/Source/Core/VideoBackends/D3D12/DX12Context.h
@@ -27,6 +27,8 @@ enum ROOT_PARAMETER
   ROOT_PARAMETER_VS_CBV,
   ROOT_PARAMETER_VS_CBV2,
   ROOT_PARAMETER_GS_CBV,
+  ROOT_PARAMETER_VS_SRV,
+  ROOT_PARAMETER_BASE_VERTEX_CONSTANT,
   ROOT_PARAMETER_PS_UAV_OR_CBV2,
   ROOT_PARAMETER_PS_CBV2,  // ROOT_PARAMETER_PS_UAV_OR_CBV2 if bbox is not enabled
   NUM_ROOT_PARAMETERS
diff --git a/Source/Core/VideoBackends/D3D12/VideoBackend.cpp b/Source/Core/VideoBackends/D3D12/VideoBackend.cpp
index 99cf3955c9..b901b994e0 100644
--- a/Source/Core/VideoBackends/D3D12/VideoBackend.cpp
+++ b/Source/Core/VideoBackends/D3D12/VideoBackend.cpp
@@ -87,7 +87,7 @@ void VideoBackend::FillBackendInfo()
   g_Config.backend_info.bSupportsLodBiasInSampler = true;
   g_Config.backend_info.bSupportsSettingObjectNames = true;
   g_Config.backend_info.bSupportsPartialMultisampleResolve = true;
-  g_Config.backend_info.bSupportsDynamicVertexLoader = false;
+  g_Config.backend_info.bSupportsDynamicVertexLoader = true;
 
   // We can only check texture support once we have a device.
   if (g_dx_context)
diff --git a/Source/Core/VideoCommon/UberShaderVertex.cpp b/Source/Core/VideoCommon/UberShaderVertex.cpp
index b49a5d8954..04f2b27a14 100644
--- a/Source/Core/VideoCommon/UberShaderVertex.cpp
+++ b/Source/Core/VideoCommon/UberShaderVertex.cpp
@@ -60,11 +60,34 @@ ShaderCode GenVertexShader(APIType api_type, const ShaderHostConfig& host_config
 SSBO_BINDING(1) readonly restrict buffer Vertices {{
   uint vertex_buffer[];
 }};
+)");
+    if (api_type == APIType::D3D)
+    {
+      // Write a function to get an offset into vertex_buffer corresponding to this vertex.
+      // This must be done differently for D3D compared to OpenGL/Vulkan/Metal, as on OpenGL, etc.,
+      // gl_VertexID starts counting at the base vertex specified in glDrawElementsBaseVertex,
+      // while on D3D, SV_VertexID (which spirv-cross translates gl_VertexID into) starts counting
+      // at 0 regardless of the BaseVertexLocation value passed to DrawIndexed.  In both cases,
+      // offset 0 of vertex_buffer corresponds to index 0 with basevertex set to 0, so we have to
+      // manually apply the basevertex offset for D3D
+      // D3D12 uses a root constant for this uniform, since it changes with every draw.
+      // D3D11 doesn't currently support dynamic vertex loader, and we'll have to figure something
+      // out for it if we want to support it in the future.
+      out.Write("UBO_BINDING(std140, 3) uniform DX_Constants {{\n"
+                "  uint base_vertex;\n"
+                "}};\n\n"
+                "uint GetVertexBaseOffset() {{\n"
+                "  return (gl_VertexID + base_vertex) * vertex_stride;\n"
+                "}}\n");
+    }
+    else
+    {
+      out.Write("uint GetVertexBaseOffset() {{\n"
+                "  return gl_VertexID * vertex_stride;\n"
+                "}}\n");
+    }
 
-uint GetVertexBaseOffset() {{
-  return gl_VertexID * vertex_stride;
-}}
-
+    out.Write(R"(
 uint4 load_input_uint4_ubyte4(uint vtx_offset, uint attr_offset) {{
   uint value = vertex_buffer[vtx_offset + attr_offset];
   return uint4(value & 0xff, (value >> 8) & 0xff, (value >> 16) & 0xff, value >> 24);