[D3D12] CP: Framework for multiple command lists per frame

12 files changed, 402 insertions, 239 deletions

diff --git a/src/xenia/gpu/d3d12/d3d12_command_processor.cc b/src/xenia/gpu/d3d12/d3d12_command_processor.cc
index c90a2b65c..a3a7d7bd1 100644
--- a/src/xenia/gpu/d3d12/d3d12_command_processor.cc
+++ b/src/xenia/gpu/d3d12/d3d12_command_processor.cc
@@ -55,7 +55,7 @@ namespace xe {
 namespace gpu {
 namespace d3d12 {
 
-constexpr uint32_t D3D12CommandProcessor::kQueuedFrames;
+constexpr uint32_t D3D12CommandProcessor::kQueueFrames;
 constexpr uint32_t
     D3D12CommandProcessor::RootExtraParameterIndices::kUnavailable;
 constexpr uint32_t D3D12CommandProcessor::kSwapTextureWidth;
@@ -441,7 +441,7 @@ uint64_t D3D12CommandProcessor::RequestViewDescriptors(
     D3D12_GPU_DESCRIPTOR_HANDLE& gpu_handle_out) {
   uint32_t descriptor_index;
   uint64_t current_heap_index = view_heap_pool_->Request(
-      fence_current_value_, previous_heap_index, count_for_partial_update,
+      frame_current_, previous_heap_index, count_for_partial_update,
       count_for_full_update, descriptor_index);
   if (current_heap_index == ui::d3d12::DescriptorHeapPool::kHeapIndexInvalid) {
     // There was an error.
@@ -467,7 +467,7 @@ uint64_t D3D12CommandProcessor::RequestSamplerDescriptors(
     D3D12_GPU_DESCRIPTOR_HANDLE& gpu_handle_out) {
   uint32_t descriptor_index;
   uint64_t current_heap_index = sampler_heap_pool_->Request(
-      fence_current_value_, previous_heap_index, count_for_partial_update,
+      frame_current_, previous_heap_index, count_for_partial_update,
       count_for_full_update, descriptor_index);
   if (current_heap_index == ui::d3d12::DescriptorHeapPool::kHeapIndexInvalid) {
     // There was an error.
@@ -520,7 +520,7 @@ ID3D12Resource* D3D12CommandProcessor::RequestScratchGPUBuffer(
   if (scratch_buffer_ != nullptr) {
     BufferForDeletion buffer_for_deletion;
     buffer_for_deletion.buffer = scratch_buffer_;
-    buffer_for_deletion.last_usage_fence_value = fence_current_value_;
+    buffer_for_deletion.last_usage_submission = submission_current_;
     buffers_for_deletion_.push_back(buffer_for_deletion);
   }
   scratch_buffer_ = buffer;
@@ -552,10 +552,7 @@ void D3D12CommandProcessor::SetSamplePositions(MsaaSamples sample_positions) {
   if (cvars::d3d12_ssaa_custom_sample_positions && !IsROVUsedForEDRAM()) {
     auto provider = GetD3D12Context()->GetD3D12Provider();
     auto tier = provider->GetProgrammableSamplePositionsTier();
-    uint32_t command_list_index =
-        uint32_t((fence_current_value_ + (kQueuedFrames - 1)) % kQueuedFrames);
-    if (tier >= 2 &&
-        command_lists_[command_list_index]->GetCommandList1() != nullptr) {
+    if (tier >= 2) {
       // Depth buffer transitions are affected by sample positions.
       SubmitBarriers();
       // Standard sample positions in Direct3D 10.1, but adjusted to take the
@@ -611,13 +608,14 @@ void D3D12CommandProcessor::SetComputePipeline(ID3D12PipelineState* pipeline) {
   }
 }
 
-void D3D12CommandProcessor::UnbindRenderTargets() {
-  render_target_cache_->UnbindRenderTargets();
+void D3D12CommandProcessor::FlushAndUnbindRenderTargets() {
+  render_target_cache_->FlushAndUnbindRenderTargets();
 }
 
 void D3D12CommandProcessor::SetExternalGraphicsPipeline(
-    ID3D12PipelineState* pipeline, bool reset_viewport, bool reset_blend_factor,
-    bool reset_stencil_ref) {
+    ID3D12PipelineState* pipeline, bool changing_rts_and_sample_positions,
+    bool changing_viewport, bool changing_blend_factor,
+    bool changing_stencil_ref) {
   if (current_external_pipeline_ != pipeline) {
     deferred_command_list_->D3DSetPipelineState(pipeline);
     current_external_pipeline_ = pipeline;
@@ -626,14 +624,17 @@ void D3D12CommandProcessor::SetExternalGraphicsPipeline(
   current_graphics_root_signature_ = nullptr;
   current_graphics_root_up_to_date_ = 0;
   primitive_topology_ = D3D_PRIMITIVE_TOPOLOGY_UNDEFINED;
-  if (reset_viewport) {
+  if (changing_rts_and_sample_positions) {
+    render_target_cache_->ForceApplyOnNextUpdate();
+  }
+  if (changing_viewport) {
     ff_viewport_update_needed_ = true;
     ff_scissor_update_needed_ = true;
   }
-  if (reset_blend_factor) {
+  if (changing_blend_factor) {
     ff_blend_factor_update_needed_ = true;
   }
-  if (reset_stencil_ref) {
+  if (changing_stencil_ref) {
     ff_stencil_ref_update_needed_ = true;
   }
 }
@@ -672,7 +673,9 @@ std::unique_ptr<xe::ui::RawImage> D3D12CommandProcessor::Capture() {
   deferred_command_list_->CopyTexture(location_dest, location_source);
   PushTransitionBarrier(swap_texture_, D3D12_RESOURCE_STATE_COPY_SOURCE,
                         D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE);
-  EndSubmission(false);
+  if (!EndSubmission(false)) {
+    return nullptr;
+  }
   AwaitAllSubmissionsCompletion();
   D3D12_RANGE readback_range;
   readback_range.Begin = swap_texture_copy_footprint_.Offset;
@@ -711,27 +714,30 @@ bool D3D12CommandProcessor::SetupContext() {
   auto device = provider->GetDevice();
   auto direct_queue = provider->GetDirectQueue();
 
+  submission_open_ = false;
+  submission_current_ = 1;
+  submission_completed_ = 0;
   if (FAILED(device->CreateFence(0, D3D12_FENCE_FLAG_NONE,
-                                 IID_PPV_ARGS(&fence_)))) {
-    XELOGE("Failed to create the fence");
+                                 IID_PPV_ARGS(&submission_fence_)))) {
+    XELOGE("Failed to create the submission fence");
     return false;
   }
-  fence_completion_event_ = CreateEvent(nullptr, false, false, nullptr);
-  if (fence_completion_event_ == nullptr) {
-    XELOGE("Failed to create the fence completion event");
+  submission_fence_completion_event_ =
+      CreateEvent(nullptr, false, false, nullptr);
+  if (submission_fence_completion_event_ == nullptr) {
+    XELOGE("Failed to create the submission fence completion event");
     return false;
   }
-  fence_current_value_ = 1;
-  fence_completed_value_ = 0;
 
-  for (uint32_t i = 0; i < kQueuedFrames; ++i) {
-    command_lists_[i] = ui::d3d12::CommandList::Create(
-        device, direct_queue, D3D12_COMMAND_LIST_TYPE_DIRECT);
-    if (command_lists_[i] == nullptr) {
-      XELOGE("Failed to create the command lists");
-      return false;
-    }
-  }
+  frame_open_ = false;
+  frame_current_ = 1;
+  frame_completed_ = 0;
+  std::memset(closed_frame_submissions_, 0, sizeof(closed_frame_submissions_));
+
+  command_list_writable_first_ = nullptr;
+  command_list_writable_last_ = nullptr;
+  command_list_submitted_first_ = nullptr;
+  command_list_submitted_last_ = nullptr;
   deferred_command_list_ = std::make_unique<DeferredCommandList>(this);
 
   constant_buffer_pool_ =
@@ -806,9 +812,9 @@ bool D3D12CommandProcessor::SetupContext() {
     return false;
   }
   // Get the layout for the upload buffer.
-  gamma_ramp_desc.DepthOrArraySize = kQueuedFrames;
+  gamma_ramp_desc.DepthOrArraySize = kQueueFrames;
   UINT64 gamma_ramp_upload_size;
-  device->GetCopyableFootprints(&gamma_ramp_desc, 0, kQueuedFrames * 2, 0,
+  device->GetCopyableFootprints(&gamma_ramp_desc, 0, kQueueFrames * 2, 0,
                                 gamma_ramp_footprints_, nullptr, nullptr,
                                 &gamma_ramp_upload_size);
   // Create the upload buffer for the gamma ramp.
@@ -892,9 +898,6 @@ bool D3D12CommandProcessor::SetupContext() {
       swap_texture_, &swap_srv_desc,
       swap_texture_srv_descriptor_heap_->GetCPUDescriptorHandleForHeapStart());
 
-  submission_open_ = false;
-  submission_frame_open_ = false;
-
   pix_capture_requested_.store(false, std::memory_order_relaxed);
   pix_capturing_ = false;
 
@@ -967,18 +970,22 @@ void D3D12CommandProcessor::ShutdownContext() {
   shared_memory_.reset();
 
   deferred_command_list_.reset();
-  for (uint32_t i = 0; i < kQueuedFrames; ++i) {
-    command_lists_[i].reset();
-  }
+  ClearCommandListCache();
+
+  frame_open_ = false;
+  frame_current_ = 1;
+  frame_completed_ = 0;
+  std::memset(closed_frame_submissions_, 0, sizeof(closed_frame_submissions_));
 
   // First release the fence since it may reference the event.
-  ui::d3d12::util::ReleaseAndNull(fence_);
-  if (fence_completion_event_) {
-    CloseHandle(fence_completion_event_);
-    fence_completion_event_ = nullptr;
+  ui::d3d12::util::ReleaseAndNull(submission_fence_);
+  if (submission_fence_completion_event_) {
+    CloseHandle(submission_fence_completion_event_);
+    submission_fence_completion_event_ = nullptr;
   }
-  fence_current_value_ = 1;
-  fence_completed_value_ = 0;
+  submission_open_ = false;
+  submission_current_ = 1;
+  submission_completed_ = 0;
 
   CommandProcessor::ShutdownContext();
 }
@@ -988,7 +995,7 @@ void D3D12CommandProcessor::WriteRegister(uint32_t index, uint32_t value) {
 
   if (index >= XE_GPU_REG_SHADER_CONSTANT_000_X &&
       index <= XE_GPU_REG_SHADER_CONSTANT_511_W) {
-    if (submission_open_) {
+    if (frame_open_) {
       uint32_t float_constant_index =
           (index - XE_GPU_REG_SHADER_CONSTANT_000_X) >> 2;
       if (float_constant_index >= 256) {
@@ -1038,12 +1045,13 @@ void D3D12CommandProcessor::PerformSwap(uint32_t frontbuffer_ptr,
   auto provider = GetD3D12Context()->GetD3D12Provider();
   auto device = provider->GetDevice();
 
-  // Upload the new gamma ramps.
-  uint32_t command_list_index =
-      uint32_t((fence_current_value_ + (kQueuedFrames - 1)) % kQueuedFrames);
+  // Upload the new gamma ramps, using the upload buffer for the current frame
+  // (will close the frame after this anyway, so can't write multiple times per
+  // frame).
+  uint32_t gamma_ramp_frame = uint32_t(frame_current_ % kQueueFrames);
   if (dirty_gamma_ramp_normal_) {
     const D3D12_PLACED_SUBRESOURCE_FOOTPRINT& gamma_ramp_footprint =
-        gamma_ramp_footprints_[command_list_index * 2];
+        gamma_ramp_footprints_[gamma_ramp_frame * 2];
     volatile uint32_t* mapping = reinterpret_cast<uint32_t*>(
         gamma_ramp_upload_mapping_ + gamma_ramp_footprint.Offset);
     for (uint32_t i = 0; i < 256; ++i) {
@@ -1069,7 +1077,7 @@ void D3D12CommandProcessor::PerformSwap(uint32_t frontbuffer_ptr,
   }
   if (dirty_gamma_ramp_pwl_) {
     const D3D12_PLACED_SUBRESOURCE_FOOTPRINT& gamma_ramp_footprint =
-        gamma_ramp_footprints_[command_list_index * 2 + 1];
+        gamma_ramp_footprints_[gamma_ramp_frame * 2 + 1];
     volatile uint32_t* mapping = reinterpret_cast<uint32_t*>(
         gamma_ramp_upload_mapping_ + gamma_ramp_footprint.Offset);
     for (uint32_t i = 0; i < 128; ++i) {
@@ -1102,7 +1110,7 @@ void D3D12CommandProcessor::PerformSwap(uint32_t frontbuffer_ptr,
     TextureFormat frontbuffer_format;
     if (texture_cache_->RequestSwapTexture(descriptor_cpu_start,
                                            frontbuffer_format)) {
-      render_target_cache_->UnbindRenderTargets();
+      render_target_cache_->FlushAndUnbindRenderTargets();
 
       // Create the gamma ramp texture descriptor.
       // This is according to D3D::InitializePresentationParameters from a game
@@ -1161,8 +1169,8 @@ void D3D12CommandProcessor::PerformSwap(uint32_t frontbuffer_ptr,
           descriptor_gpu_start, &gamma_ramp_gpu_handle,
           use_pwl_gamma_ramp ? (1.0f / 128.0f) : (1.0f / 256.0f),
           *deferred_command_list_);
-      // Ending the current frame's command list anyway, so no need to unbind
-      // the render targets when using ROV.
+      // Ending the current frame anyway, so no need to reset the current render
+      // targets when using ROV.
 
       PushTransitionBarrier(swap_texture_, D3D12_RESOURCE_STATE_RENDER_TARGET,
                             D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE);
@@ -1683,7 +1691,6 @@ bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
                 memexport_range.base_address_dwords << 2, memexport_range_size);
             readback_buffer_offset += memexport_range_size;
           }
-          EndSubmission(false);
           AwaitAllSubmissionsCompletion();
           D3D12_RANGE readback_range;
           readback_range.Begin = 0;
@@ -1713,13 +1720,13 @@ bool D3D12CommandProcessor::IssueDraw(PrimitiveType primitive_type,
 
 void D3D12CommandProcessor::InitializeTrace() {
   BeginSubmission(false);
-  bool any_submitted = false;
-  any_submitted |= shared_memory_->InitializeTraceSubmitDownloads();
-  if (any_submitted) {
-    EndSubmission(false);
-    AwaitAllSubmissionsCompletion();
-    shared_memory_->InitializeTraceCompleteDownloads();
+  bool any_downloads_submitted = false;
+  any_downloads_submitted |= shared_memory_->InitializeTraceSubmitDownloads();
+  if (!any_downloads_submitted || !EndSubmission(false)) {
+    return;
   }
+  AwaitAllSubmissionsCompletion();
+  shared_memory_->InitializeTraceCompleteDownloads();
 }
 
 void D3D12CommandProcessor::FinalizeTrace() {}
@@ -1746,18 +1753,19 @@ bool D3D12CommandProcessor::IssueCopy() {
       deferred_command_list_->D3DCopyBufferRegion(
           readback_buffer, 0, shared_memory_buffer, written_address,
           written_length);
-      EndSubmission(false);
-      AwaitAllSubmissionsCompletion();
-      D3D12_RANGE readback_range;
-      readback_range.Begin = 0;
-      readback_range.End = written_length;
-      void* readback_mapping;
-      if (SUCCEEDED(
-              readback_buffer->Map(0, &readback_range, &readback_mapping))) {
-        std::memcpy(memory_->TranslatePhysical(written_address),
-                    readback_mapping, written_length);
-        D3D12_RANGE readback_write_range = {};
-        readback_buffer->Unmap(0, &readback_write_range);
+      if (EndSubmission(false)) {
+        AwaitAllSubmissionsCompletion();
+        D3D12_RANGE readback_range;
+        readback_range.Begin = 0;
+        readback_range.End = written_length;
+        void* readback_mapping;
+        if (SUCCEEDED(
+                readback_buffer->Map(0, &readback_range, &readback_mapping))) {
+          std::memcpy(memory_->TranslatePhysical(written_address),
+                      readback_mapping, written_length);
+          D3D12_RANGE readback_write_range = {};
+          readback_buffer->Unmap(0, &readback_write_range);
+        }
       }
     }
   }
@@ -1769,20 +1777,75 @@ void D3D12CommandProcessor::BeginSubmission(bool is_guest_command) {
   SCOPE_profile_cpu_f("gpu");
 #endif  // FINE_GRAINED_DRAW_SCOPES
 
-  if (!submission_open_) {
-    submission_open_ = true;
+  bool is_opening_frame = is_guest_command && !frame_open_;
+  if (submission_open_ && !is_opening_frame) {
+    return;
+  }
+
+  // Check the fence - needed for all kinds of submissions (to reclaim transient
+  // resources early) and specifically for frames (not to queue too many).
+  submission_completed_ = submission_fence_->GetCompletedValue();
+  if (is_opening_frame) {
+    // Await the availability of the current frame.
+    uint64_t frame_current_last_submission =
+        closed_frame_submissions_[frame_current_ % kQueueFrames];
+    if (frame_current_last_submission > submission_completed_) {
+      submission_fence_->SetEventOnCompletion(
+          frame_current_last_submission, submission_fence_completion_event_);
+      WaitForSingleObject(submission_fence_completion_event_, INFINITE);
+      submission_completed_ = submission_fence_->GetCompletedValue();
+    }
+    // Update the completed frame index, also obtaining the actual completed
+    // frame number (since the CPU may be actually less than 3 frames behind)
+    // before reclaiming resources tracked with the frame number.
+    frame_completed_ =
+        std::max(frame_current_, uint64_t(kQueueFrames)) - kQueueFrames;
+    for (uint64_t frame = frame_completed_ + 1; frame < frame_current_;
+         ++frame) {
+      if (closed_frame_submissions_[frame % kQueueFrames] >
+          submission_completed_) {
+        break;
+      }
+      frame_completed_ = frame;
+    }
+  }
+
+  // Reclaim command lists.
+  while (command_list_submitted_first_) {
+    if (command_list_submitted_first_->last_usage_submission >
+        submission_completed_) {
+      break;
+    }
+    if (command_list_writable_last_) {
+      command_list_writable_last_->next = command_list_submitted_first_;
+    } else {
+      command_list_writable_first_ = command_list_submitted_first_;
+    }
+    command_list_writable_last_ = command_list_submitted_first_;
+    command_list_submitted_first_ = command_list_submitted_first_->next;
+    command_list_writable_last_->next = nullptr;
+  }
+  if (!command_list_submitted_first_) {
+    command_list_submitted_last_ = nullptr;
+  }
 
-    // Wait for a swap command list to become free.
-    // Command list 0 is used when fence_current_value_ is 1, 4, 7...
-    fence_completed_value_ = fence_->GetCompletedValue();
-    if (fence_completed_value_ + kQueuedFrames < fence_current_value_) {
-      fence_->SetEventOnCompletion(fence_current_value_ - kQueuedFrames,
-                                   fence_completion_event_);
-      WaitForSingleObject(fence_completion_event_, INFINITE);
-      fence_completed_value_ = fence_->GetCompletedValue();
+  // Delete transient buffers marked for deletion.
+  auto erase_buffers_end = buffers_for_deletion_.begin();
+  while (erase_buffers_end != buffers_for_deletion_.end()) {
+    if (erase_buffers_end->last_usage_submission > submission_completed_) {
+      ++erase_buffers_end;
+      break;
     }
+    erase_buffers_end->buffer->Release();
+    ++erase_buffers_end;
+  }
+  buffers_for_deletion_.erase(buffers_for_deletion_.begin(), erase_buffers_end);
 
-    // Start a new command list.
+  if (!submission_open_) {
+    submission_open_ = true;
+
+    // Start a new deferred command list - will create the real one in the end
+    // of submission.
     deferred_command_list_->Reset();
 
     // Reset cached state of the command list.
@@ -1799,32 +1862,17 @@ void D3D12CommandProcessor::BeginSubmission(bool is_guest_command) {
     current_sampler_heap_ = nullptr;
     primitive_topology_ = D3D_PRIMITIVE_TOPOLOGY_UNDEFINED;
 
+    shared_memory_->BeginSubmission();
+
     render_target_cache_->BeginSubmission();
 
     primitive_converter_->BeginSubmission();
   }
 
-  if (!submission_frame_open_) {
-    submission_frame_open_ = true;
+  if (is_opening_frame) {
+    frame_open_ = true;
 
-    // TODO(Triang3l): Move fence checking and command list releasing here.
-
-    // Cleanup resources after checking the fence.
-
-    auto erase_buffers_end = buffers_for_deletion_.begin();
-    while (erase_buffers_end != buffers_for_deletion_.end()) {
-      if (erase_buffers_end->last_usage_fence_value > fence_completed_value_) {
-        ++erase_buffers_end;
-        break;
-      }
-      erase_buffers_end->buffer->Release();
-      ++erase_buffers_end;
-    }
-    buffers_for_deletion_.erase(buffers_for_deletion_.begin(),
-                                erase_buffers_end);
-
-    // Reset bindings that depend on the resources with lifetime tracked with
-    // the fence.
+    // Reset bindings that depend on the data stored in the pools.
     std::memset(current_float_constant_map_vertex_, 0,
                 sizeof(current_float_constant_map_vertex_));
     std::memset(current_float_constant_map_pixel_, 0,
@@ -1841,9 +1889,11 @@ void D3D12CommandProcessor::BeginSubmission(bool is_guest_command) {
     samplers_written_vertex_ = false;
     samplers_written_pixel_ = false;
 
-    constant_buffer_pool_->Reclaim(fence_completed_value_);
-    view_heap_pool_->Reclaim(fence_completed_value_);
-    sampler_heap_pool_->Reclaim(fence_completed_value_);
+    // Reclaim pool pages - no need to do this every small submission since some
+    // may be reused.
+    constant_buffer_pool_->Reclaim(frame_completed_);
+    view_heap_pool_->Reclaim(frame_completed_);
+    sampler_heap_pool_->Reclaim(frame_completed_);
 
     pix_capturing_ =
         pix_capture_requested_.exchange(false, std::memory_order_relaxed);
@@ -1855,18 +1905,36 @@ void D3D12CommandProcessor::BeginSubmission(bool is_guest_command) {
       }
     }
 
-    shared_memory_->BeginFrame();
-
     texture_cache_->BeginFrame();
 
     primitive_converter_->BeginFrame();
   }
 }
 
-void D3D12CommandProcessor::EndSubmission(bool is_swap) {
+bool D3D12CommandProcessor::EndSubmission(bool is_swap) {
   auto provider = GetD3D12Context()->GetD3D12Provider();
 
-  if (is_swap && submission_frame_open_) {
+  // Make sure there is a command list to submit to.
+  if (submission_open_ && !command_list_writable_first_) {
+    std::unique_ptr<ui::d3d12::CommandList> new_command_list =
+        ui::d3d12::CommandList::Create(provider->GetDevice(),
+                                       provider->GetDirectQueue(),
+                                       D3D12_COMMAND_LIST_TYPE_DIRECT);
+    if (!new_command_list) {
+      // Try to submit later. Completely dropping the submission is not
+      // permitted because resources would be left in an undefined state.
+      return false;
+    }
+    command_list_writable_first_ = new CommandList;
+    command_list_writable_first_->command_list = std::move(new_command_list);
+    command_list_writable_first_->last_usage_submission = 0;
+    command_list_writable_first_->next = nullptr;
+    command_list_writable_last_ = command_list_writable_first_;
+  }
+
+  bool is_closing_frame = is_swap && frame_open_;
+
+  if (is_closing_frame) {
     texture_cache_->EndFrame();
   }
 
@@ -1875,27 +1943,37 @@ void D3D12CommandProcessor::EndSubmission(bool is_swap) {
 
     pipeline_cache_->EndSubmission();
 
-    render_target_cache_->EndSubmission();
-
     // Submit barriers now because resources with the queued barriers may be
     // destroyed between frames.
     SubmitBarriers();
 
     // Submit the command list.
-    uint32_t command_list_index =
-        uint32_t((fence_current_value_ + (kQueuedFrames - 1)) % kQueuedFrames);
-    auto current_command_list = command_lists_[command_list_index].get();
+    auto current_command_list =
+        command_list_writable_first_->command_list.get();
     current_command_list->BeginRecording();
     deferred_command_list_->Execute(current_command_list->GetCommandList(),
                                     current_command_list->GetCommandList1());
     current_command_list->Execute();
+    command_list_writable_first_->last_usage_submission = submission_current_;
+    if (command_list_submitted_last_) {
+      command_list_submitted_last_->next = command_list_writable_first_;
+    } else {
+      command_list_submitted_first_ = command_list_writable_first_;
+    }
+    command_list_submitted_last_ = command_list_writable_first_;
+    command_list_writable_first_ = command_list_writable_first_->next;
+    command_list_submitted_last_->next = nullptr;
+    if (!command_list_writable_first_) {
+      command_list_writable_last_ = nullptr;
+    }
 
-    provider->GetDirectQueue()->Signal(fence_, fence_current_value_++);
+    provider->GetDirectQueue()->Signal(submission_fence_,
+                                       submission_current_++);
 
     submission_open_ = false;
   }
 
-  if (is_swap && submission_frame_open_) {
+  if (is_closing_frame) {
     // Close the capture after submitting.
     if (pix_capturing_) {
       IDXGraphicsAnalysis* graphics_analysis = provider->GetGraphicsAnalysis();
@@ -1904,12 +1982,17 @@ void D3D12CommandProcessor::EndSubmission(bool is_swap) {
       }
       pix_capturing_ = false;
     }
-    submission_frame_open_ = false;
+    frame_open_ = false;
+    // Submission already closed now, so minus 1.
+    closed_frame_submissions_[(frame_current_++) % kQueueFrames] =
+        submission_current_ - 1;
 
     if (cache_clear_requested_) {
       cache_clear_requested_ = false;
       AwaitAllSubmissionsCompletion();
 
+      ClearCommandListCache();
+
       ui::d3d12::util::ReleaseAndNull(scratch_buffer_);
       scratch_buffer_size_ = 0;
 
@@ -1934,18 +2017,35 @@ void D3D12CommandProcessor::EndSubmission(bool is_swap) {
       // shared_memory_->ClearCache();
     }
   }
+
+  return true;
 }
 
 void D3D12CommandProcessor::AwaitAllSubmissionsCompletion() {
   // May be called if shutting down without everything set up.
-  if ((fence_completed_value_ + 1) >= fence_current_value_ || !fence_ ||
-      GetD3D12Context()->WasLost()) {
+  if ((submission_completed_ + 1) >= submission_current_ ||
+      !submission_fence_ || GetD3D12Context()->WasLost()) {
     return;
   }
-  fence_->SetEventOnCompletion(fence_current_value_ - 1,
-                               fence_completion_event_);
-  WaitForSingleObject(fence_completion_event_, INFINITE);
-  fence_completed_value_ = fence_current_value_ - 1;
+  submission_fence_->SetEventOnCompletion(submission_current_ - 1,
+                                          submission_fence_completion_event_);
+  WaitForSingleObject(submission_fence_completion_event_, INFINITE);
+  submission_completed_ = submission_current_ - 1;
+}
+
+void D3D12CommandProcessor::ClearCommandListCache() {
+  while (command_list_submitted_first_) {
+    auto next = command_list_submitted_first_->next;
+    delete command_list_submitted_first_;
+    command_list_submitted_first_ = next;
+  }
+  command_list_submitted_last_ = nullptr;
+  while (command_list_writable_first_) {
+    auto next = command_list_writable_first_->next;
+    delete command_list_writable_first_;
+    command_list_writable_first_ = next;
+  }
+  command_list_writable_last_ = nullptr;
 }
 
 void D3D12CommandProcessor::UpdateFixedFunctionState(bool primitive_two_faced) {
@@ -2819,9 +2919,8 @@ bool D3D12CommandProcessor::UpdateBindings(
   // Update constant buffers.
   if (!cbuffer_bindings_system_.up_to_date) {
     uint8_t* system_constants = constant_buffer_pool_->Request(
-        fence_current_value_,
-        xe::align(uint32_t(sizeof(system_constants_)), 256u), nullptr, nullptr,
-        &cbuffer_bindings_system_.buffer_address);
+        frame_current_, xe::align(uint32_t(sizeof(system_constants_)), 256u),
+        nullptr, nullptr, &cbuffer_bindings_system_.buffer_address);
     if (system_constants == nullptr) {
       return false;
     }
@@ -2832,7 +2931,7 @@ bool D3D12CommandProcessor::UpdateBindings(
   }
   if (!cbuffer_bindings_float_vertex_.up_to_date) {
     uint8_t* float_constants = constant_buffer_pool_->Request(
-        fence_current_value_, float_constant_size_vertex, nullptr, nullptr,
+        frame_current_, float_constant_size_vertex, nullptr, nullptr,
         &cbuffer_bindings_float_vertex_.buffer_address);
     if (float_constants == nullptr) {
       return false;
@@ -2857,7 +2956,7 @@ bool D3D12CommandProcessor::UpdateBindings(
   }
   if (!cbuffer_bindings_float_pixel_.up_to_date) {
     uint8_t* float_constants = constant_buffer_pool_->Request(
-        fence_current_value_, float_constant_size_pixel, nullptr, nullptr,
+        frame_current_, float_constant_size_pixel, nullptr, nullptr,
         &cbuffer_bindings_float_pixel_.buffer_address);
     if (float_constants == nullptr) {
       return false;
@@ -2887,7 +2986,7 @@ bool D3D12CommandProcessor::UpdateBindings(
   if (!cbuffer_bindings_bool_loop_.up_to_date) {
     uint32_t* bool_loop_constants =
         reinterpret_cast<uint32_t*>(constant_buffer_pool_->Request(
-            fence_current_value_, 768, nullptr, nullptr,
+            frame_current_, 768, nullptr, nullptr,
             &cbuffer_bindings_bool_loop_.buffer_address));
     if (bool_loop_constants == nullptr) {
       return false;
@@ -2906,9 +3005,9 @@ bool D3D12CommandProcessor::UpdateBindings(
     write_bool_loop_constant_view = true;
   }
   if (!cbuffer_bindings_fetch_.up_to_date) {
-    uint8_t* fetch_constants = constant_buffer_pool_->Request(
-        fence_current_value_, 768, nullptr, nullptr,
-        &cbuffer_bindings_fetch_.buffer_address);
+    uint8_t* fetch_constants =
+        constant_buffer_pool_->Request(frame_current_, 768, nullptr, nullptr,
+                                       &cbuffer_bindings_fetch_.buffer_address);
     if (fetch_constants == nullptr) {
       return false;
     }
diff --git a/src/xenia/gpu/d3d12/d3d12_command_processor.h b/src/xenia/gpu/d3d12/d3d12_command_processor.h
index 1852d928a..8558f6f26 100644
--- a/src/xenia/gpu/d3d12/d3d12_command_processor.h
+++ b/src/xenia/gpu/d3d12/d3d12_command_processor.h
@@ -53,7 +53,8 @@ class D3D12CommandProcessor : public CommandProcessor {
     return static_cast<xe::ui::d3d12::D3D12Context*>(context_.get());
   }
 
-  // Returns the deferred drawing command list for the currently open frame.
+  // Returns the deferred drawing command list for the currently open
+  // submission.
   DeferredCommandList* GetDeferredCommandList() {
     return deferred_command_list_.get();
   }
@@ -63,8 +64,11 @@ class D3D12CommandProcessor : public CommandProcessor {
   // targets.
   bool IsROVUsedForEDRAM() const;
 
-  uint64_t GetCurrentFenceValue() const { return fence_current_value_; }
-  uint64_t GetCompletedFenceValue() const { return fence_completed_value_; }
+  uint64_t GetCurrentSubmission() const { return submission_current_; }
+  uint64_t GetCompletedSubmission() const { return submission_completed_; }
+
+  uint64_t GetCurrentFrame() const { return frame_current_; }
+  uint64_t GetCompletedFrame() const { return frame_completed_; }
 
   // 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
@@ -106,8 +110,8 @@ class D3D12CommandProcessor : public CommandProcessor {
       D3D12_CPU_DESCRIPTOR_HANDLE& cpu_handle_out,
       D3D12_GPU_DESCRIPTOR_HANDLE& gpu_handle_out);
 
-  // Returns a single temporary GPU-side buffer within a frame for tasks like
-  // texture untiling and resolving.
+  // Returns a single temporary GPU-side buffer within a submission for tasks
+  // like texture untiling and resolving.
   ID3D12Resource* RequestScratchGPUBuffer(uint32_t size,
                                           D3D12_RESOURCE_STATES state);
   // This must be called when done with the scratch buffer, to notify the
@@ -127,22 +131,23 @@ class D3D12CommandProcessor : public CommandProcessor {
   }
 
   // Sets the current pipeline state to a compute pipeline. This is for cache
-  // invalidation primarily. A frame must be open.
+  // invalidation primarily. A submission must be open.
   void SetComputePipeline(ID3D12PipelineState* pipeline);
 
   // Stores and unbinds render targets before binding changing render targets
   // externally. This is separate from SetExternalGraphicsPipeline because it
   // causes computations to be dispatched, and the scratch buffer may also be
   // used.
-  void UnbindRenderTargets();
+  void FlushAndUnbindRenderTargets();
 
   // Sets the current pipeline state to a special drawing pipeline, invalidating
-  // various cached state variables. UnbindRenderTargets may be needed before
-  // calling this. A frame must be open.
-  void SetExternalGraphicsPipeline(ID3D12PipelineState* pipeline,
-                                   bool reset_viewport = true,
-                                   bool reset_blend_factor = false,
-                                   bool reset_stencil_ref = false);
+  // various cached state variables. FlushAndUnbindRenderTargets may be needed
+  // before calling this. A submission must be open.
+  void SetExternalGraphicsPipeline(
+      ID3D12PipelineState* pipeline,
+      bool changing_rts_and_sample_positions = true,
+      bool changing_viewport = true, bool changing_blend_factor = false,
+      bool changing_stencil_ref = false);
 
   // Returns the text to display in the GPU backend name in the window title.
   std::wstring GetWindowTitleText() const;
@@ -170,7 +175,7 @@ class D3D12CommandProcessor : public CommandProcessor {
   void FinalizeTrace() override;
 
  private:
-  static constexpr uint32_t kQueuedFrames = 3;
+  static constexpr uint32_t kQueueFrames = 3;
 
   enum RootParameter : UINT {
     // These are always present.
@@ -226,9 +231,12 @@ class D3D12CommandProcessor : public CommandProcessor {
   // opposed to simply resuming after mid-frame synchronization).
   void BeginSubmission(bool is_guest_command);
   // If is_swap is true, a full frame is closed - with, if needed, cache
-  // clearing and stopping capturing.
-  void EndSubmission(bool is_swap);
+  // clearing and stopping capturing. Returns whether the submission was done
+  // successfully, if it has failed, leaves it open.
+  bool EndSubmission(bool is_swap);
   void AwaitAllSubmissionsCompletion();
+  // Need to await submission completion before calling.
+  void ClearCommandListCache();
 
   void UpdateFixedFunctionState(bool primitive_two_faced);
   void UpdateSystemConstantValues(
@@ -253,12 +261,30 @@ class D3D12CommandProcessor : public CommandProcessor {
 
   bool cache_clear_requested_ = false;
 
-  uint64_t fence_current_value_ = 1;
-  uint64_t fence_completed_value_ = 0;
-  HANDLE fence_completion_event_ = nullptr;
-  ID3D12Fence* fence_ = nullptr;
-
-  std::unique_ptr<ui::d3d12::CommandList> command_lists_[kQueuedFrames] = {};
+  bool submission_open_ = false;
+  // Values of submission_fence_.
+  uint64_t submission_current_ = 1;
+  uint64_t submission_completed_ = 0;
+  HANDLE submission_fence_completion_event_ = nullptr;
+  ID3D12Fence* submission_fence_ = nullptr;
+
+  bool frame_open_ = false;
+  // Guest frame index, since some transient resources can be reused across
+  // submissions. Values updated in the beginning of a frame.
+  uint64_t frame_current_ = 1;
+  uint64_t frame_completed_ = 0;
+  // Submission indices of frames that have already been submitted.
+  uint64_t closed_frame_submissions_[kQueueFrames] = {};
+
+  struct CommandList {
+    std::unique_ptr<ui::d3d12::CommandList> command_list;
+    uint64_t last_usage_submission;
+    CommandList* next;
+  };
+  CommandList* command_list_writable_first_ = nullptr;
+  CommandList* command_list_writable_last_ = nullptr;
+  CommandList* command_list_submitted_first_ = nullptr;
+  CommandList* command_list_submitted_last_ = nullptr;
   std::unique_ptr<DeferredCommandList> deferred_command_list_ = nullptr;
 
   std::unique_ptr<SharedMemory> shared_memory_ = nullptr;
@@ -283,10 +309,10 @@ class D3D12CommandProcessor : public CommandProcessor {
   ID3D12Resource* gamma_ramp_texture_ = nullptr;
   D3D12_RESOURCE_STATES gamma_ramp_texture_state_;
   // Upload buffer for an image that is the same as gamma_ramp_, but with
-  // kQueuedFrames array layers.
+  // kQueueFrames array layers.
   ID3D12Resource* gamma_ramp_upload_ = nullptr;
   uint8_t* gamma_ramp_upload_mapping_ = nullptr;
-  D3D12_PLACED_SUBRESOURCE_FOOTPRINT gamma_ramp_footprints_[kQueuedFrames * 2];
+  D3D12_PLACED_SUBRESOURCE_FOOTPRINT gamma_ramp_footprints_[kQueueFrames * 2];
 
   static constexpr uint32_t kSwapTextureWidth = 1280;
   static constexpr uint32_t kSwapTextureHeight = 720;
@@ -308,7 +334,7 @@ class D3D12CommandProcessor : public CommandProcessor {
 
   struct BufferForDeletion {
     ID3D12Resource* buffer;
-    uint64_t last_usage_fence_value;
+    uint64_t last_usage_submission;
   };
   std::deque<BufferForDeletion> buffers_for_deletion_;
 
@@ -322,9 +348,6 @@ class D3D12CommandProcessor : public CommandProcessor {
   ID3D12Resource* readback_buffer_ = nullptr;
   uint32_t readback_buffer_size_ = 0;
 
-  bool submission_open_ = false;
-  bool submission_frame_open_ = false;
-
   std::atomic<bool> pix_capture_requested_ = false;
   bool pix_capturing_;
 
diff --git a/src/xenia/gpu/d3d12/primitive_converter.cc b/src/xenia/gpu/d3d12/primitive_converter.cc
index 0732d6f23..a3dfec579 100644
--- a/src/xenia/gpu/d3d12/primitive_converter.cc
+++ b/src/xenia/gpu/d3d12/primitive_converter.cc
@@ -112,7 +112,7 @@ bool PrimitiveConverter::Initialize() {
   }
   static_ib_upload_->Unmap(0, nullptr);
   // Not uploaded yet.
-  static_ib_upload_fence_value_ = UINT64_MAX;
+  static_ib_upload_submission_ = UINT64_MAX;
   if (FAILED(device->CreateCommittedResource(
           &ui::d3d12::util::kHeapPropertiesDefault, D3D12_HEAP_FLAG_NONE,
           &static_ib_desc, D3D12_RESOURCE_STATE_COPY_DEST, nullptr,
@@ -145,29 +145,26 @@ void PrimitiveConverter::ClearCache() { buffer_pool_->ClearCache(); }
 void PrimitiveConverter::BeginSubmission() {
   // Got a command list now - upload and transition the static index buffer if
   // needed.
-  if (static_ib_upload_ && static_ib_upload_fence_value_ == UINT64_MAX) {
-    // Not uploaded yet - upload.
-    command_processor_->GetDeferredCommandList()->D3DCopyResource(
-        static_ib_, static_ib_upload_);
-    command_processor_->PushTransitionBarrier(
-        static_ib_, D3D12_RESOURCE_STATE_COPY_DEST,
-        D3D12_RESOURCE_STATE_INDEX_BUFFER);
-    static_ib_upload_fence_value_ = command_processor_->GetCurrentFenceValue();
+  if (static_ib_upload_) {
+    if (static_ib_upload_submission_ == UINT64_MAX) {
+      // Not uploaded yet - upload.
+      command_processor_->GetDeferredCommandList()->D3DCopyResource(
+          static_ib_, static_ib_upload_);
+      command_processor_->PushTransitionBarrier(
+          static_ib_, D3D12_RESOURCE_STATE_COPY_DEST,
+          D3D12_RESOURCE_STATE_INDEX_BUFFER);
+      static_ib_upload_submission_ = command_processor_->GetCurrentSubmission();
+    } else if (command_processor_->GetCompletedSubmission() >=
+               static_ib_upload_submission_) {
+      // Completely uploaded - release the upload buffer.
+      static_ib_upload_->Release();
+      static_ib_upload_ = nullptr;
+    }
   }
 }
 
 void PrimitiveConverter::BeginFrame() {
-  uint64_t completed_fence_value = command_processor_->GetCompletedFenceValue();
-
-  if (static_ib_upload_ && static_ib_upload_fence_value_ != UINT64_MAX &&
-      completed_fence_value >= static_ib_upload_fence_value_) {
-    // Completely uploaded - release the upload buffer.
-    static_ib_upload_->Release();
-    static_ib_upload_ = nullptr;
-  }
-
-  buffer_pool_->Reclaim(command_processor_->GetCompletedFenceValue());
-
+  buffer_pool_->Reclaim(command_processor_->GetCompletedFrame());
   converted_indices_cache_.clear();
   memory_regions_used_ = 0;
 }
@@ -698,7 +695,7 @@ void* PrimitiveConverter::AllocateIndices(
   }
   D3D12_GPU_VIRTUAL_ADDRESS gpu_address;
   uint8_t* mapping =
-      buffer_pool_->Request(command_processor_->GetCurrentFenceValue(), size,
+      buffer_pool_->Request(command_processor_->GetCurrentFrame(), size,
                             nullptr, nullptr, &gpu_address);
   if (mapping == nullptr) {
     XELOGE("Failed to allocate space for %u converted %u-bit vertex indices",
diff --git a/src/xenia/gpu/d3d12/primitive_converter.h b/src/xenia/gpu/d3d12/primitive_converter.h
index c9f6b192e..f45c36146 100644
--- a/src/xenia/gpu/d3d12/primitive_converter.h
+++ b/src/xenia/gpu/d3d12/primitive_converter.h
@@ -112,7 +112,7 @@ class PrimitiveConverter {
   // CPU-side, used only for uploading - destroyed once the copy commands have
   // been completed.
   ID3D12Resource* static_ib_upload_ = nullptr;
-  uint64_t static_ib_upload_fence_value_;
+  uint64_t static_ib_upload_submission_;
   // GPU-side - used for drawing.
   ID3D12Resource* static_ib_ = nullptr;
   D3D12_GPU_VIRTUAL_ADDRESS static_ib_gpu_address_;
diff --git a/src/xenia/gpu/d3d12/render_target_cache.cc b/src/xenia/gpu/d3d12/render_target_cache.cc
index a68a79e0c..494cdba54 100644
--- a/src/xenia/gpu/d3d12/render_target_cache.cc
+++ b/src/xenia/gpu/d3d12/render_target_cache.cc
@@ -452,15 +452,22 @@ void RenderTargetCache::ClearCache() {
 }
 
 void RenderTargetCache::BeginSubmission() {
-  // A submission does not always end in a resolve (for example, when memexport
-  // readback happens) or something else that would surely submit the UAV
-  // barrier, so we need to preserve the `current_` variables.
-  if (!command_processor_->IsROVUsedForEDRAM()) {
-    ClearBindings();
-  }
+  // With the ROV, a submission does not always end in a resolve (for example,
+  // when memexport readback happens) or something else that would surely submit
+  // the UAV barrier, so we need to preserve the `current_` variables.
+  //
+  // With RTVs, simply going to a different command list doesn't have to cause
+  // storing the render targets to the EDRAM buffer, however, the new command
+  // list doesn't have the needed RTVs/DSV bound yet.
+  //
+  // Just make sure they are bound to the new command list.
+  ForceApplyOnNextUpdate();
 }
 
-void RenderTargetCache::EndSubmission() { UnbindRenderTargets(); }
+void RenderTargetCache::EndFrame() {
+  // May be clearing the cache after this.
+  FlushAndUnbindRenderTargets();
+}
 
 bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
   // There are two kinds of render target binding updates in this implementation
@@ -754,6 +761,8 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
     }
   }
 
+  bool sample_positions_set = false;
+
   // Need to change the bindings.
   if (full_update || render_targets_to_attach) {
 #if 0
@@ -895,7 +904,8 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
     if (!rov_used) {
       // Sample positions when loading depth must match sample positions when
       // drawing.
-      command_processor_->SetSamplePositions(rb_surface_info.msaa_samples);
+      command_processor_->SetSamplePositions(current_msaa_samples_);
+      sample_positions_set = true;
 
       // Load the contents of the new render targets from the EDRAM buffer (will
       // change the state of the render targets to copy destination).
@@ -921,13 +931,16 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
 
       // Transition the render targets to the appropriate state if needed,
       // compress the list of the render target because null RTV descriptors are
-      // broken in Direct3D 12 and bind the render targets to the command list.
-      D3D12_CPU_DESCRIPTOR_HANDLE rtv_handles[4];
+      // broken in Direct3D 12 and update the list of the render targets to bind
+      // to the command list.
       uint32_t rtv_count = 0;
       for (uint32_t i = 0; i < 4; ++i) {
         const RenderTargetBinding& binding = current_bindings_[i];
+        if (!binding.is_bound) {
+          continue;
+        }
         RenderTarget* render_target = binding.render_target;
-        if (!binding.is_bound || render_target == nullptr) {
+        if (render_target == nullptr) {
           continue;
         }
         XELOGGPU("RT Color %u: base %u, format %u", i, edram_bases[i],
@@ -936,7 +949,6 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
             render_target->resource, render_target->state,
             D3D12_RESOURCE_STATE_RENDER_TARGET);
         render_target->state = D3D12_RESOURCE_STATE_RENDER_TARGET;
-        rtv_handles[rtv_count] = render_target->handle;
         current_pipeline_render_targets_[rtv_count].guest_render_target = i;
         current_pipeline_render_targets_[rtv_count].format =
             GetColorDXGIFormat(ColorRenderTargetFormat(formats[i]));
@@ -946,7 +958,6 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
         current_pipeline_render_targets_[i].guest_render_target = i;
         current_pipeline_render_targets_[i].format = DXGI_FORMAT_UNKNOWN;
       }
-      const D3D12_CPU_DESCRIPTOR_HANDLE* dsv_handle;
       const RenderTargetBinding& depth_binding = current_bindings_[4];
       RenderTarget* depth_render_target = depth_binding.render_target;
       current_pipeline_render_targets_[4].guest_render_target = 4;
@@ -956,17 +967,42 @@ bool RenderTargetCache::UpdateRenderTargets(const D3D12Shader* pixel_shader) {
             depth_render_target->resource, depth_render_target->state,
             D3D12_RESOURCE_STATE_DEPTH_WRITE);
         depth_render_target->state = D3D12_RESOURCE_STATE_DEPTH_WRITE;
-        dsv_handle = &depth_binding.render_target->handle;
         current_pipeline_render_targets_[4].format =
             GetDepthDXGIFormat(DepthRenderTargetFormat(formats[4]));
       } else {
-        dsv_handle = nullptr;
         current_pipeline_render_targets_[4].format = DXGI_FORMAT_UNKNOWN;
       }
       command_processor_->SubmitBarriers();
-      command_processor_->GetDeferredCommandList()->D3DOMSetRenderTargets(
-          rtv_count, rtv_handles, FALSE, dsv_handle);
+      apply_to_command_list_ = true;
+    }
+  }
+
+  // Bind the render targets to the command list, either in case of an update or
+  // if asked to externally.
+  if (!rov_used && apply_to_command_list_) {
+    apply_to_command_list_ = false;
+    if (!sample_positions_set) {
+      command_processor_->SetSamplePositions(current_msaa_samples_);
+    }
+    D3D12_CPU_DESCRIPTOR_HANDLE rtv_handles[4];
+    uint32_t rtv_count;
+    for (rtv_count = 0; rtv_count < 4; ++rtv_count) {
+      const PipelineRenderTarget& pipeline_render_target =
+          current_pipeline_render_targets_[rtv_count];
+      if (pipeline_render_target.format == DXGI_FORMAT_UNKNOWN) {
+        break;
+      }
+      const RenderTargetBinding& binding =
+          current_bindings_[pipeline_render_target.guest_render_target];
+      rtv_handles[rtv_count] = binding.render_target->handle;
     }
+    const RenderTargetBinding& depth_binding = current_bindings_[4];
+    const D3D12_CPU_DESCRIPTOR_HANDLE* dsv_handle =
+        current_pipeline_render_targets_[4].format != DXGI_FORMAT_UNKNOWN
+            ? &depth_binding.render_target->handle
+            : nullptr;
+    command_processor_->GetDeferredCommandList()->D3DOMSetRenderTargets(
+        rtv_count, rtv_handles, FALSE, dsv_handle);
   }
 
   // Update the dirty regions.
@@ -2098,7 +2134,7 @@ RenderTargetCache::ResolveTarget* RenderTargetCache::FindOrCreateResolveTarget(
   return resolve_target;
 }
 
-void RenderTargetCache::UnbindRenderTargets() {
+void RenderTargetCache::FlushAndUnbindRenderTargets() {
   if (command_processor_->IsROVUsedForEDRAM()) {
     return;
   }
@@ -2217,6 +2253,7 @@ void RenderTargetCache::ClearBindings() {
   current_msaa_samples_ = MsaaSamples::k1X;
   current_edram_max_rows_ = 0;
   std::memset(current_bindings_, 0, sizeof(current_bindings_));
+  apply_to_command_list_ = true;
 }
 
 #if 0
diff --git a/src/xenia/gpu/d3d12/render_target_cache.h b/src/xenia/gpu/d3d12/render_target_cache.h
index e5e7cb128..f7580d30e 100644
--- a/src/xenia/gpu/d3d12/render_target_cache.h
+++ b/src/xenia/gpu/d3d12/render_target_cache.h
@@ -256,7 +256,7 @@ class RenderTargetCache {
   void ClearCache();
 
   void BeginSubmission();
-  void EndSubmission();
+  void EndFrame();
   // Called in the beginning of a draw call - may bind pipelines.
   bool UpdateRenderTargets(const D3D12Shader* pixel_shader);
   // Returns the host-to-guest mappings and host formats of currently bound
@@ -272,10 +272,13 @@ class RenderTargetCache {
   bool Resolve(SharedMemory* shared_memory, TextureCache* texture_cache,
                Memory* memory, uint32_t& written_address_out,
                uint32_t& written_length_out);
+  // Makes sure the render targets are re-attached to the command list for which
+  // the next update will take place.
+  void ForceApplyOnNextUpdate() { apply_to_command_list_ = true; }
   // Flushes the render targets to EDRAM and unbinds them, for instance, when
   // the command processor takes over framebuffer bindings to draw something
   // special.
-  void UnbindRenderTargets();
+  void FlushAndUnbindRenderTargets();
   void WriteEDRAMUint32UAVDescriptor(D3D12_CPU_DESCRIPTOR_HANDLE handle);
 
   // Totally necessary to rely on the base format - Too Human switches between
@@ -636,6 +639,7 @@ class RenderTargetCache {
   // current_edram_max_rows_ is for RTV/DSV only (render target texture size).
   uint32_t current_edram_max_rows_ = 0;
   RenderTargetBinding current_bindings_[5] = {};
+  bool apply_to_command_list_ = true;
 
   PipelineRenderTarget current_pipeline_render_targets_[5];
 
diff --git a/src/xenia/gpu/d3d12/shared_memory.cc b/src/xenia/gpu/d3d12/shared_memory.cc
index 6fc00bce5..1b80cb271 100644
--- a/src/xenia/gpu/d3d12/shared_memory.cc
+++ b/src/xenia/gpu/d3d12/shared_memory.cc
@@ -165,8 +165,8 @@ void SharedMemory::Shutdown() {
   }
 }
 
-void SharedMemory::BeginFrame() {
-  upload_buffer_pool_->Reclaim(command_processor_->GetCompletedFenceValue());
+void SharedMemory::BeginSubmission() {
+  upload_buffer_pool_->Reclaim(command_processor_->GetCompletedSubmission());
 }
 
 SharedMemory::GlobalWatchHandle SharedMemory::RegisterGlobalWatch(
@@ -367,7 +367,7 @@ bool SharedMemory::RequestRange(uint32_t start, uint32_t length) {
       ID3D12Resource* upload_buffer;
       uint32_t upload_buffer_offset, upload_buffer_size;
       uint8_t* upload_buffer_mapping = upload_buffer_pool_->RequestPartial(
-          command_processor_->GetCurrentFenceValue(),
+          command_processor_->GetCurrentSubmission(),
           upload_range_length << page_size_log2_, &upload_buffer,
           &upload_buffer_offset, &upload_buffer_size, nullptr);
       if (upload_buffer_mapping == nullptr) {
diff --git a/src/xenia/gpu/d3d12/shared_memory.h b/src/xenia/gpu/d3d12/shared_memory.h
index c2189a10f..9b24b01b1 100644
--- a/src/xenia/gpu/d3d12/shared_memory.h
+++ b/src/xenia/gpu/d3d12/shared_memory.h
@@ -44,7 +44,7 @@ class SharedMemory {
     return buffer_gpu_address_;
   }
 
-  void BeginFrame();
+  void BeginSubmission();
 
   typedef void (*GlobalWatchCallback)(void* context, uint32_t address_first,
                                       uint32_t address_last,
diff --git a/src/xenia/gpu/d3d12/texture_cache.cc b/src/xenia/gpu/d3d12/texture_cache.cc
index 74fe4211d..e9124d89e 100644
--- a/src/xenia/gpu/d3d12/texture_cache.cc
+++ b/src/xenia/gpu/d3d12/texture_cache.cc
@@ -1169,7 +1169,7 @@ void TextureCache::BeginFrame() {
   texture_current_usage_time_ = xe::Clock::QueryHostUptimeMillis();
 
   // If memory usage is too high, destroy unused textures.
-  uint64_t completed_fence_value = command_processor_->GetCompletedFenceValue();
+  uint64_t completed_frame = command_processor_->GetCompletedFrame();
   uint32_t limit_soft_mb = cvars::d3d12_texture_cache_limit_soft;
   uint32_t limit_hard_mb = cvars::d3d12_texture_cache_limit_hard;
   if (IsResolutionScale2X()) {
@@ -1186,7 +1186,7 @@ void TextureCache::BeginFrame() {
       break;
     }
     Texture* texture = texture_used_first_;
-    if (texture->last_usage_fence_value > completed_fence_value) {
+    if (texture->last_usage_frame > completed_frame) {
       break;
     }
     if (!limit_hard_exceeded &&
@@ -2311,7 +2311,7 @@ TextureCache::Texture* TextureCache::FindOrCreateTexture(TextureKey key) {
   texture->resource_size =
       device->GetResourceAllocationInfo(0, 1, &desc).SizeInBytes;
   texture->state = state;
-  texture->last_usage_fence_value = command_processor_->GetCurrentFenceValue();
+  texture->last_usage_frame = command_processor_->GetCurrentFrame();
   texture->last_usage_time = texture_current_usage_time_;
   texture->used_previous = texture_used_last_;
   texture->used_next = nullptr;
@@ -2606,7 +2606,7 @@ bool TextureCache::LoadTextureData(Texture* texture) {
       }
       D3D12_GPU_VIRTUAL_ADDRESS cbuffer_gpu_address;
       uint8_t* cbuffer_mapping = cbuffer_pool->Request(
-          command_processor_->GetCurrentFenceValue(),
+          command_processor_->GetCurrentFrame(),
           xe::align(uint32_t(sizeof(load_constants)), 256u), nullptr, nullptr,
           &cbuffer_gpu_address);
       if (cbuffer_mapping == nullptr) {
@@ -2684,10 +2684,10 @@ bool TextureCache::LoadTextureData(Texture* texture) {
 }
 
 void TextureCache::MarkTextureUsed(Texture* texture) {
-  uint64_t current_fence_value = command_processor_->GetCurrentFenceValue();
+  uint64_t current_frame = command_processor_->GetCurrentFrame();
   // This is called very frequently, don't relink unless needed for caching.
-  if (texture->last_usage_fence_value != current_fence_value) {
-    texture->last_usage_fence_value = current_fence_value;
+  if (texture->last_usage_frame != current_frame) {
+    texture->last_usage_frame = current_frame;
     texture->last_usage_time = texture_current_usage_time_;
     if (texture->used_next == nullptr) {
       // Simplify the code a bit - already in the end of the list.
diff --git a/src/xenia/gpu/d3d12/texture_cache.h b/src/xenia/gpu/d3d12/texture_cache.h
index be6ccf4cf..9690a50fb 100644
--- a/src/xenia/gpu/d3d12/texture_cache.h
+++ b/src/xenia/gpu/d3d12/texture_cache.h
@@ -343,7 +343,7 @@ class TextureCache {
     uint64_t resource_size;
     D3D12_RESOURCE_STATES state;
 
-    uint64_t last_usage_fence_value;
+    uint64_t last_usage_frame;
     uint64_t last_usage_time;
     Texture* used_previous;
     Texture* used_next;
diff --git a/src/xenia/ui/d3d12/pools.cc b/src/xenia/ui/d3d12/pools.cc
index 2b5275fbd..2229d42be 100644
--- a/src/xenia/ui/d3d12/pools.cc
+++ b/src/xenia/ui/d3d12/pools.cc
@@ -24,9 +24,9 @@ UploadBufferPool::UploadBufferPool(ID3D12Device* device, uint32_t page_size)
 
 UploadBufferPool::~UploadBufferPool() { ClearCache(); }
 
-void UploadBufferPool::Reclaim(uint64_t completed_fence_value) {
+void UploadBufferPool::Reclaim(uint64_t completed_submission_index) {
   while (submitted_first_) {
-    if (submitted_first_->last_usage_fence_value > completed_fence_value) {
+    if (submitted_first_->last_submission_index > completed_submission_index) {
       break;
     }
     if (writable_last_) {
@@ -67,7 +67,7 @@ void UploadBufferPool::ClearCache() {
   writable_last_ = nullptr;
 }
 
-uint8_t* UploadBufferPool::Request(uint64_t usage_fence_value, uint32_t size,
+uint8_t* UploadBufferPool::Request(uint64_t submission_index, uint32_t size,
                                    ID3D12Resource** buffer_out,
                                    uint32_t* offset_out,
                                    D3D12_GPU_VIRTUAL_ADDRESS* gpu_address_out) {
@@ -76,9 +76,9 @@ uint8_t* UploadBufferPool::Request(uint64_t usage_fence_value, uint32_t size,
     return nullptr;
   }
   assert_true(!current_page_used_ ||
-              usage_fence_value >= writable_first_->last_usage_fence_value);
+              submission_index >= writable_first_->last_submission_index);
   assert_true(!submitted_last_ ||
-              usage_fence_value >= submitted_last_->last_usage_fence_value);
+              submission_index >= submitted_last_->last_submission_index);
   if (page_size_ - current_page_used_ < size || !writable_first_) {
     // Start a new page if can't fit all the bytes or don't have an open page.
     if (writable_first_) {
@@ -122,13 +122,13 @@ uint8_t* UploadBufferPool::Request(uint64_t usage_fence_value, uint32_t size,
       writable_first_->buffer = new_buffer;
       writable_first_->gpu_address = new_buffer->GetGPUVirtualAddress();
       writable_first_->mapping = new_buffer_mapping;
-      writable_first_->last_usage_fence_value = usage_fence_value;
+      writable_first_->last_submission_index = submission_index;
       writable_first_->next = nullptr;
       writable_last_ = writable_first_;
     }
     current_page_used_ = 0;
   }
-  writable_first_->last_usage_fence_value = usage_fence_value;
+  writable_first_->last_submission_index = submission_index;
   if (buffer_out) {
     *buffer_out = writable_first_->buffer;
   }
@@ -145,7 +145,7 @@ uint8_t* UploadBufferPool::Request(uint64_t usage_fence_value, uint32_t size,
 }
 
 uint8_t* UploadBufferPool::RequestPartial(
-    uint64_t usage_fence_value, uint32_t size, ID3D12Resource** buffer_out,
+    uint64_t submission_index, uint32_t size, ID3D12Resource** buffer_out,
     uint32_t* offset_out, uint32_t* size_out,
     D3D12_GPU_VIRTUAL_ADDRESS* gpu_address_out) {
   size = std::min(size, page_size_);
@@ -153,7 +153,7 @@ uint8_t* UploadBufferPool::RequestPartial(
     size = std::min(size, page_size_ - current_page_used_);
   }
   uint8_t* mapping =
-      Request(usage_fence_value, size, buffer_out, offset_out, gpu_address_out);
+      Request(submission_index, size, buffer_out, offset_out, gpu_address_out);
   if (!mapping) {
     return nullptr;
   }
@@ -172,9 +172,9 @@ DescriptorHeapPool::DescriptorHeapPool(ID3D12Device* device,
 
 DescriptorHeapPool::~DescriptorHeapPool() { ClearCache(); }
 
-void DescriptorHeapPool::Reclaim(uint64_t completed_fence_value) {
+void DescriptorHeapPool::Reclaim(uint64_t completed_submission_index) {
   while (submitted_first_) {
-    if (submitted_first_->last_usage_fence_value > completed_fence_value) {
+    if (submitted_first_->last_submission_index > completed_submission_index) {
       break;
     }
     if (writable_last_) {
@@ -213,7 +213,7 @@ void DescriptorHeapPool::ClearCache() {
   writable_last_ = nullptr;
 }
 
-uint64_t DescriptorHeapPool::Request(uint64_t usage_fence_value,
+uint64_t DescriptorHeapPool::Request(uint64_t submission_index,
                                      uint64_t previous_heap_index,
                                      uint32_t count_for_partial_update,
                                      uint32_t count_for_full_update,
@@ -225,9 +225,9 @@ uint64_t DescriptorHeapPool::Request(uint64_t usage_fence_value,
     return kHeapIndexInvalid;
   }
   assert_true(!current_page_used_ ||
-              usage_fence_value >= writable_first_->last_usage_fence_value);
+              submission_index >= writable_first_->last_submission_index);
   assert_true(!submitted_last_ ||
-              usage_fence_value >= submitted_last_->last_usage_fence_value);
+              submission_index >= submitted_last_->last_submission_index);
   // If the last full update happened on the current page, a partial update is
   // possible.
   uint32_t count = previous_heap_index == current_heap_index_
@@ -271,11 +271,11 @@ uint64_t DescriptorHeapPool::Request(uint64_t usage_fence_value,
     writable_first_->heap = new_heap;
     writable_first_->cpu_start = new_heap->GetCPUDescriptorHandleForHeapStart();
     writable_first_->gpu_start = new_heap->GetGPUDescriptorHandleForHeapStart();
-    writable_first_->last_usage_fence_value = usage_fence_value;
+    writable_first_->last_submission_index = submission_index;
     writable_first_->next = nullptr;
     writable_last_ = writable_first_;
   }
-  writable_first_->last_usage_fence_value = usage_fence_value;
+  writable_first_->last_submission_index = submission_index;
   index_out = current_page_used_;
   current_page_used_ += count;
   return current_heap_index_;
diff --git a/src/xenia/ui/d3d12/pools.h b/src/xenia/ui/d3d12/pools.h
index 6e6f65512..4499bd08f 100644
--- a/src/xenia/ui/d3d12/pools.h
+++ b/src/xenia/ui/d3d12/pools.h
@@ -18,21 +18,24 @@ namespace xe {
 namespace ui {
 namespace d3d12 {
 
+// Submission index is the fence value or a value derived from it (if reclaiming
+// less often than once per fence value, for instance).
+
 class UploadBufferPool {
  public:
   UploadBufferPool(ID3D12Device* device, uint32_t page_size);
   ~UploadBufferPool();
 
-  void Reclaim(uint64_t completed_fence_value);
+  void Reclaim(uint64_t completed_submission_index);
   void ClearCache();
 
   // Request to write data in a single piece, creating a new page if the current
   // one doesn't have enough free space.
-  uint8_t* Request(uint64_t usage_fence_value, uint32_t size,
+  uint8_t* Request(uint64_t submission_index, uint32_t size,
                    ID3D12Resource** buffer_out, uint32_t* offset_out,
                    D3D12_GPU_VIRTUAL_ADDRESS* gpu_address_out);
   // Request to write data in multiple parts, filling the buffer entirely.
-  uint8_t* RequestPartial(uint64_t usage_fence_value, uint32_t size,
+  uint8_t* RequestPartial(uint64_t submission_index, uint32_t size,
                           ID3D12Resource** buffer_out, uint32_t* offset_out,
                           uint32_t* size_out,
                           D3D12_GPU_VIRTUAL_ADDRESS* gpu_address_out);
@@ -45,7 +48,7 @@ class UploadBufferPool {
     ID3D12Resource* buffer;
     D3D12_GPU_VIRTUAL_ADDRESS gpu_address;
     void* mapping;
-    uint64_t last_usage_fence_value;
+    uint64_t last_submission_index;
     Page* next;
   };
 
@@ -68,7 +71,7 @@ class DescriptorHeapPool {
                      uint32_t page_size);
   ~DescriptorHeapPool();
 
-  void Reclaim(uint64_t completed_fence_value);
+  void Reclaim(uint64_t completed_submission_index);
   void ClearCache();
 
   // Because all descriptors for a single draw call must be in the same heap,
@@ -96,7 +99,7 @@ class DescriptorHeapPool {
   // This MUST be called even if there's nothing to write in a partial update
   // (with count_for_partial_update being 0), because a full update may still be
   // required.
-  uint64_t Request(uint64_t usage_fence_value, uint64_t previous_heap_index,
+  uint64_t Request(uint64_t submission_index, uint64_t previous_heap_index,
                    uint32_t count_for_partial_update,
                    uint32_t count_for_full_update, uint32_t& index_out);
 
@@ -122,7 +125,7 @@ class DescriptorHeapPool {
     ID3D12DescriptorHeap* heap;
     D3D12_CPU_DESCRIPTOR_HANDLE cpu_start;
     D3D12_GPU_DESCRIPTOR_HANDLE gpu_start;
-    uint64_t last_usage_fence_value;
+    uint64_t last_submission_index;
     Page* next;
   };