[D3D12] Banjo-Kazooie/Tooie top tile clear fix (rendering still broken in Tooie)

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -59,6 +59,21 @@ ID3D12GraphicsCommandList1* D3D12CommandProcessor::GetCurrentCommandList1()
   return command_lists_[current_queue_frame_]->GetCommandList1();
 }
 
+uint32_t D3D12CommandProcessor::GetCurrentColorMask(
+    const D3D12Shader* pixel_shader) const {
+  if (pixel_shader == nullptr) {
+    return 0;
+  }
+  auto& regs = *register_file_;
+  uint32_t color_mask = regs[XE_GPU_REG_RB_COLOR_MASK].u32 & 0xFFFF;
+  for (uint32_t i = 0; i < 4; ++i) {
+    if (!pixel_shader->writes_color_target(i)) {
+      color_mask &= ~(0xF << (i * 4));
+    }
+  }
+  return color_mask;
+}
+
 void D3D12CommandProcessor::PushTransitionBarrier(
     ID3D12Resource* resource, D3D12_RESOURCE_STATES old_state,
     D3D12_RESOURCE_STATES new_state, UINT subresource) {
@@ -848,12 +863,6 @@ bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
     // Doesn't actually draw.
     return true;
   }
-  uint32_t color_mask = enable_mode == xenos::ModeControl::kColorDepth ?
-                        regs[XE_GPU_REG_RB_COLOR_MASK].u32 & 0xFFFF : 0;
-  if (!color_mask && !(regs[XE_GPU_REG_RB_DEPTHCONTROL].u32 & (0x1 | 0x4))) {
-    // Not writing to color, depth or doing stencil test, so doesn't draw.
-    return true;
-  }
   if ((regs[XE_GPU_REG_PA_SU_SC_MODE_CNTL].u32 & 0x3) == 0x3 &&
       primitive_type != PrimitiveType::kPointList &&
       primitive_type != PrimitiveType::kRectangleList) {
@@ -901,11 +910,17 @@ bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
     return false;
   }
 
+  uint32_t color_mask = GetCurrentColorMask(pixel_shader);
+  if (!color_mask && !(regs[XE_GPU_REG_RB_DEPTHCONTROL].u32 & (0x1 | 0x4))) {
+    // Not writing to color, depth or doing stencil test, so doesn't draw.
+    return true;
+  }
+
   bool new_frame = BeginFrame();
   auto command_list = GetCurrentCommandList();
 
   // Set up the render targets - this may bind pipelines.
-  if (!render_target_cache_->UpdateRenderTargets()) {
+  if (!render_target_cache_->UpdateRenderTargets(pixel_shader)) {
     // Doesn't actually draw.
     return true;
   }

src/xenia/gpu/d3d12/d3d12_command_processor.h

@@ -48,6 +48,16 @@ class D3D12CommandProcessor : public CommandProcessor {
   ID3D12GraphicsCommandList* GetCurrentCommandList() const;
   ID3D12GraphicsCommandList1* GetCurrentCommandList1() const;
 
+  // Gets the current color write mask, taking the pixel shader's write mask
+  // into account. If a shader doesn't write to a render target, it shouldn't be
+  // written to and it shouldn't be even bound - otherwise, in Halo 3, one
+  // render target is being destroyed by a shader not writing anything, and in
+  // Banjo-Tooie, the result of clearing the top tile is being ignored because
+  // there are 4 render targets bound with the same EDRAM base (clearly not
+  // correct usage), but the shader only clears 1, and then EDRAM buffer stores
+  // conflict with each other.
+  uint32_t GetCurrentColorMask(const D3D12Shader* pixel_shader) const;
+
   void PushTransitionBarrier(
       ID3D12Resource* resource, D3D12_RESOURCE_STATES old_state,
       D3D12_RESOURCE_STATES new_state,

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -335,20 +335,7 @@ PipelineCache::UpdateStatus PipelineCache::UpdateBlendStateAndRenderTargets(
     dirty |= regs.render_targets[i].format != render_targets[i].format;
     regs.render_targets[i].format = render_targets[i].format;
   }
-  uint32_t color_mask;
+  uint32_t color_mask = command_processor_->GetCurrentColorMask(pixel_shader);
-  if (pixel_shader != nullptr) {
-    color_mask = register_file_->values[XE_GPU_REG_RB_COLOR_MASK].u32 & 0xFFFF;
-    for (uint32_t i = 0; i < 4; ++i) {
-      // If the pixel shader doesn't write to a render target, writing to it is
-      // disabled in the blend state. Otherwise, in Halo 3, one important render
-      // target is destroyed by a shader not writing to one of the outputs.
-      if (!pixel_shader->writes_color_target(i)) {
-        color_mask &= ~(0xF << (i * 4));
-      }
-    }
-  } else {
-    color_mask = 0;
-  }
   dirty |= regs.color_mask != color_mask;
   regs.color_mask = color_mask;
   bool blend_enable =

src/xenia/gpu/d3d12/render_target_cache.cc

@@ -323,7 +323,7 @@ void RenderTargetCache::BeginFrame() {
   // stable D24F==D32F comparison.
 }
 
-bool RenderTargetCache::UpdateRenderTargets() {
+bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
   // There are two kinds of render target binding updates in this implementation
   // in case something has been changed - full and partial.
   //
@@ -416,16 +416,12 @@ bool RenderTargetCache::UpdateRenderTargets() {
   uint32_t edram_bases[5];
   uint32_t formats[5];
   bool formats_are_64bpp[5];
-  uint32_t rb_color_mask = regs[XE_GPU_REG_RB_COLOR_MASK].u32;
+  uint32_t color_mask = command_processor_->GetCurrentColorMask(pixel_shader);
-  if (xenos::ModeControl(regs[XE_GPU_REG_RB_MODECONTROL].u32 & 0x7) !=
-      xenos::ModeControl::kColorDepth) {
-    rb_color_mask = 0;
-  }
   uint32_t rb_color_info[4] = {
       regs[XE_GPU_REG_RB_COLOR_INFO].u32, regs[XE_GPU_REG_RB_COLOR1_INFO].u32,
       regs[XE_GPU_REG_RB_COLOR2_INFO].u32, regs[XE_GPU_REG_RB_COLOR3_INFO].u32};
   for (uint32_t i = 0; i < 4; ++i) {
-    enabled[i] = (rb_color_mask & (0xF << (i * 4))) != 0;
+    enabled[i] = (color_mask & (0xF << (i * 4))) != 0;
     edram_bases[i] = std::min(rb_color_info[i] & 0xFFF, 2048u);
     formats[i] = (rb_color_info[i] >> 16) & 0xF;
     formats_are_64bpp[i] =
@@ -2027,7 +2023,13 @@ void RenderTargetCache::StoreRenderTargetsToEDRAM() {
         if (current_bindings_[a].edram_base < current_bindings_[b].edram_base) {
           return true;
         }
-        return a < b;
+        // If EDRAM bases are the same (not really a valid usage, but happens in
+        // Banjo-Tooie - in case color writing was enabled for invalid render
+        // targets in some draw call), treat the render targets with the lowest
+        // index as more important (it's the primary one after all, while the
+        // rest are additional). Also treat the depth buffer as highest-priority
+        // (in the comparison, treat depth as 0 and color as 1-4).
+        return ((a + 1) % 5) > ((b + 1) % 5);
       });
 
   // Calculate the dispatch width.

src/xenia/gpu/d3d12/render_target_cache.h

@@ -12,6 +12,7 @@
 
 #include <unordered_map>
 
+#include "xenia/gpu/d3d12/d3d12_shader.h"
 #include "xenia/gpu/d3d12/shared_memory.h"
 #include "xenia/gpu/d3d12/texture_cache.h"
 #include "xenia/gpu/register_file.h"
@@ -209,7 +210,7 @@ class RenderTargetCache {
 
   void BeginFrame();
   // Called in the beginning of a draw call - may bind pipelines.
-  bool UpdateRenderTargets();
+  bool UpdateRenderTargets(const D3D12Shader* pixel_shader);
   // Returns the host-to-guest mappings and host formats of currently bound
   // render targets for pipeline creation and remapping in shaders. They are
   // consecutive, and format DXGI_FORMAT_UNKNOWN terminates the list. Depth