Merge pull request #4471 from Armada651/hybrid-depth

VideoCommon: Don't process the depth range in the vertex shader if it's not oversized.

Source/Core/VideoBackends/D3D/Render.cpp

@@ -559,6 +559,10 @@ void Renderer::SetViewport()
   float Y = Renderer::EFBToScaledYf(xfmem.viewport.yOrig + xfmem.viewport.ht - scissorYOff);
   float Wd = Renderer::EFBToScaledXf(2.0f * xfmem.viewport.wd);
   float Ht = Renderer::EFBToScaledYf(-2.0f * xfmem.viewport.ht);
+  float range = MathUtil::Clamp<float>(xfmem.viewport.zRange, 0.0f, 16777215.0f);
+  float min_depth =
+      MathUtil::Clamp<float>(xfmem.viewport.farZ - range, 0.0f, 16777215.0f) / 16777216.0f;
+  float max_depth = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
   if (Wd < 0.0f)
   {
     X += Wd;
@@ -570,18 +574,24 @@ void Renderer::SetViewport()
     Ht = -Ht;
   }
 
+  // If an inverted depth range is used, which D3D doesn't support,
+  // we need to calculate the depth range in the vertex shader.
+  if (xfmem.viewport.zRange < 0.0f)
+  {
+    min_depth = 0.0f;
+    max_depth = GX_MAX_DEPTH;
+  }
+
   // In D3D, the viewport rectangle must fit within the render target.
   X = (X >= 0.f) ? X : 0.f;
   Y = (Y >= 0.f) ? Y : 0.f;
   Wd = (X + Wd <= GetTargetWidth()) ? Wd : (GetTargetWidth() - X);
   Ht = (Y + Ht <= GetTargetHeight()) ? Ht : (GetTargetHeight() - Y);
 
-  // We do depth clipping and depth range in the vertex shader instead of relying
-  // on the graphics API. However we still need to ensure depth values don't exceed
-  // the maximum value supported by the console GPU. We also need to account for the
-  // fact that the entire depth buffer is inverted on D3D, so we set GX_MAX_DEPTH as
-  // an inverted near value.
-  D3D11_VIEWPORT vp = CD3D11_VIEWPORT(X, Y, Wd, Ht, 1.0f - GX_MAX_DEPTH, D3D11_MAX_DEPTH);
+  // We use an inverted depth range here to apply the Reverse Z trick.
+  // This trick makes sure we match the precision provided by the 1:0
+  // clipping depth range on the hardware.
+  D3D11_VIEWPORT vp = CD3D11_VIEWPORT(X, Y, Wd, Ht, 1.0f - max_depth, 1.0f - min_depth);
   D3D::context->RSSetViewports(1, &vp);
 }
 

Source/Core/VideoBackends/D3D12/Render.cpp

@@ -464,6 +464,10 @@ void Renderer::SetViewport()
   float y = Renderer::EFBToScaledYf(xfmem.viewport.yOrig + xfmem.viewport.ht - scissor_y_offset);
   float width = Renderer::EFBToScaledXf(2.0f * xfmem.viewport.wd);
   float height = Renderer::EFBToScaledYf(-2.0f * xfmem.viewport.ht);
+  float range = MathUtil::Clamp<float>(xfmem.viewport.zRange, 0.0f, 16777215.0f);
+  float min_depth =
+      MathUtil::Clamp<float>(xfmem.viewport.farZ - range, 0.0f, 16777215.0f) / 16777216.0f;
+  float max_depth = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
   if (width < 0.0f)
   {
     x += width;
@@ -475,18 +479,24 @@ void Renderer::SetViewport()
     height = -height;
   }
 
+  // If an inverted depth range is used, which D3D doesn't support,
+  // we need to calculate the depth range in the vertex shader.
+  if (xfmem.viewport.zRange < 0.0f)
+  {
+    min_depth = 0.0f;
+    max_depth = GX_MAX_DEPTH;
+  }
+
   // In D3D, the viewport rectangle must fit within the render target.
   x = (x >= 0.f) ? x : 0.f;
   y = (y >= 0.f) ? y : 0.f;
   width = (x + width <= GetTargetWidth()) ? width : (GetTargetWidth() - x);
   height = (y + height <= GetTargetHeight()) ? height : (GetTargetHeight() - y);
 
-  // We do depth clipping and depth range in the vertex shader instead of relying
-  // on the graphics API. However we still need to ensure depth values don't exceed
-  // the maximum value supported by the console GPU. We also need to account for the
-  // fact that the entire depth buffer is inverted on D3D, so we set GX_MAX_DEPTH as
-  // an inverted near value.
-  D3D12_VIEWPORT vp = {x, y, width, height, 1.0f - GX_MAX_DEPTH, D3D12_MAX_DEPTH};
+  // We use an inverted depth range here to apply the Reverse Z trick.
+  // This trick makes sure we match the precision provided by the 1:0
+  // clipping depth range on the hardware.
+  D3D12_VIEWPORT vp = {x, y, width, height, 1.0f - max_depth, 1.0f - min_depth};
   D3D::current_command_list->RSSetViewports(1, &vp);
 }
 

Source/Core/VideoBackends/OGL/Render.cpp

@@ -1122,12 +1122,10 @@ void Renderer::SetViewport()
                           (float)scissorYOff);
   float Width = EFBToScaledXf(2.0f * xfmem.viewport.wd);
   float Height = EFBToScaledYf(-2.0f * xfmem.viewport.ht);
-  float GLNear = MathUtil::Clamp<float>(
-                     xfmem.viewport.farZ -
-                         MathUtil::Clamp<float>(xfmem.viewport.zRange, -16777216.0f, 16777216.0f),
-                     0.0f, 16777215.0f) /
-                 16777216.0f;
-  float GLFar = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
+  float range = MathUtil::Clamp<float>(xfmem.viewport.zRange, -16777215.0f, 16777215.0f);
+  float min_depth =
+      MathUtil::Clamp<float>(xfmem.viewport.farZ - range, 0.0f, 16777215.0f) / 16777216.0f;
+  float max_depth = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
   if (Width < 0)
   {
     X += Width;
@@ -1154,17 +1152,7 @@ void Renderer::SetViewport()
   // vertex shader we only need to ensure depth values don't exceed the maximum
   // value supported by the console GPU. If not, we simply clamp the near/far values
   // themselves to the maximum value as done above.
-  if (g_ActiveConfig.backend_info.bSupportsDepthClamp)
-  {
-    if (xfmem.viewport.zRange < 0.0f)
-      glDepthRangef(0.0f, GX_MAX_DEPTH);
-    else
-      glDepthRangef(GX_MAX_DEPTH, 0.0f);
-  }
-  else
-  {
-    glDepthRangef(GLFar, GLNear);
-  }
+  glDepthRangef(max_depth, min_depth);
 }
 
 void Renderer::ClearScreen(const EFBRectangle& rc, bool colorEnable, bool alphaEnable, bool zEnable,

Source/Core/VideoBackends/Vulkan/Renderer.cpp

@@ -1635,6 +1635,10 @@ void Renderer::SetViewport()
   float y = Renderer::EFBToScaledYf(xfmem.viewport.yOrig + xfmem.viewport.ht - scissor_y_offset);
   float width = Renderer::EFBToScaledXf(2.0f * xfmem.viewport.wd);
   float height = Renderer::EFBToScaledYf(-2.0f * xfmem.viewport.ht);
+  float range = MathUtil::Clamp<float>(xfmem.viewport.zRange, -16777215.0f, 16777215.0f);
+  float min_depth =
+      MathUtil::Clamp<float>(xfmem.viewport.farZ - range, 0.0f, 16777215.0f) / 16777216.0f;
+  float max_depth = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
   if (width < 0.0f)
   {
     x += width;
@@ -1646,28 +1650,19 @@ void Renderer::SetViewport()
     height = -height;
   }
 
-  // If we do depth clipping and depth range in the vertex shader we only need to ensure
-  // depth values don't exceed the maximum value supported by the console GPU. If not,
-  // we simply clamp the near/far values themselves to the maximum value as done above.
-  float min_depth, max_depth;
-  if (g_ActiveConfig.backend_info.bSupportsDepthClamp)
+  // If an inverted depth range is used, which the Vulkan drivers don't
+  // support, we need to calculate the depth range in the vertex shader.
+  // TODO: Make this into a DriverDetails bug and write a test for CTS.
+  if (xfmem.viewport.zRange < 0.0f)
   {
-    min_depth = 1.0f - GX_MAX_DEPTH;
-    max_depth = 1.0f;
-  }
-  else
-  {
-    float near_val = MathUtil::Clamp<float>(xfmem.viewport.farZ -
-                                                MathUtil::Clamp<float>(xfmem.viewport.zRange,
-                                                                       -16777216.0f, 16777216.0f),
-                                            0.0f, 16777215.0f) /
-                     16777216.0f;
-    float far_val = MathUtil::Clamp<float>(xfmem.viewport.farZ, 0.0f, 16777215.0f) / 16777216.0f;
-    min_depth = near_val;
-    max_depth = far_val;
+    min_depth = 0.0f;
+    max_depth = GX_MAX_DEPTH;
   }
 
-  VkViewport viewport = {x, y, width, height, min_depth, max_depth};
+  // We use an inverted depth range here to apply the Reverse Z trick.
+  // This trick makes sure we match the precision provided by the 1:0
+  // clipping depth range on the hardware.
+  VkViewport viewport = {x, y, width, height, 1.0f - max_depth, 1.0f - min_depth};
   StateTracker::GetInstance()->SetViewport(viewport);
 }
 

Source/Core/VideoCommon/VertexShaderGen.cpp

@@ -30,6 +30,10 @@ VertexShaderUid GetVertexShaderUid()
   uid_data->msaa = g_ActiveConfig.iMultisamples > 1;
   uid_data->ssaa = g_ActiveConfig.iMultisamples > 1 && g_ActiveConfig.bSSAA;
   uid_data->numColorChans = xfmem.numChan.numColorChans;
+  uid_data->vertex_depth =
+      g_ActiveConfig.backend_info.bSupportsDepthClamp &&
+      ((fabs(xfmem.viewport.zRange) > 16777215.0f || fabs(xfmem.viewport.farZ) > 16777215.0f) ||
+       (xfmem.viewport.zRange < 0.0f && !g_ActiveConfig.backend_info.bSupportsReversedDepthRange));
 
   GetLightingShaderUid(uid_data->lighting);
 
@@ -416,13 +420,11 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const vertex_shader_uid_da
       out.Write("o.colors_1 = color1;\n");
   }
 
-  // Write the true depth value. If the game uses depth textures, then the pixel shader will
-  // override it with the correct values if not then early z culling will improve speed.
+  // If we can disable the incorrect depth clipping planes using depth clamping, then we can do
+  // our own depth clipping and calculate the depth range before the perspective divide if
+  // necessary.
   if (g_ActiveConfig.backend_info.bSupportsDepthClamp)
   {
-    // If we can disable the incorrect depth clipping planes using depth clamping, then we can do
-    // our own depth clipping and calculate the depth range before the perspective divide.
-
     // Since we're adjusting z for the depth range before the perspective divide, we have to do our
     // own clipping. We want to clip so that -w <= z <= 0, which matches the console -1..0 range.
     // We adjust our depth value for clipping purposes to match the perspective projection in the
@@ -430,7 +432,14 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const vertex_shader_uid_da
     out.Write("float clipDepth = o.pos.z * (1.0 - 1e-7);\n");
     out.Write("o.clipDist0 = clipDepth + o.pos.w;\n");  // Near: z < -w
     out.Write("o.clipDist1 = -clipDepth;\n");           // Far: z > 0
+  }
 
+  // Write the true depth value. If the game uses depth textures, then the pixel shader will
+  // override it with the correct values if not then early z culling will improve speed.
+  // There are two different ways to do this, when the depth range is oversized, we process
+  // the depth range in the vertex shader, if not we let the host driver handle it.
+  if (uid_data->vertex_depth)
+  {
     // Adjust z for the depth range. We're using an equation which incorperates a depth inversion,
     // so we can map the console -1..0 range to the 0..1 range used in the depth buffer.
     // We have to handle the depth range in the vertex shader instead of after the perspective
@@ -442,11 +451,8 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const vertex_shader_uid_da
   }
   else
   {
-    // If we can't disable the incorrect depth clipping planes, then we need to rely on the
-    // graphics API to handle the depth range after the perspective divide. This can result in
-    // inaccurate depth values due to the missing depth bias, but that can be least corrected by
-    // overriding depth values in the pixel shader. We still need to take care of the reversed depth
-    // though, so we do that here.
+    // Here we rely on the host driver to process the depth range, however we still need to invert
+    // the console -1..0 range to the 0..1 range used in the depth buffer.
     out.Write("o.pos.z = -o.pos.z;\n");
   }
 

Source/Core/VideoCommon/VertexShaderGen.h

@@ -43,7 +43,8 @@ struct vertex_shader_uid_data
   u32 texMtxInfo_n_projection : 16;  // Stored separately to guarantee that the texMtxInfo struct is
                                      // 8 bits wide
   u32 ssaa : 1;
-  u32 pad : 15;
+  u32 vertex_depth : 1;
+  u32 pad : 14;
 
   struct
   {