Proper(ish) VdSwap - fixes a bunch of things.

Caching is working a bit better, now.

src/xenia/gpu/buffer_resource.cc

@@ -30,10 +30,9 @@ int BufferResource::Prepare() {
     }
   }
 
-  // DISABLED
-  //if (!dirtied_) {
-  //  return 0;
-  //}
+  if (!dirtied_) {
+    return 0;
+  }
   dirtied_ = false;
 
   // pass dirty regions?

src/xenia/gpu/command_processor.cc

@@ -300,6 +300,16 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
         }
         break;
 
+      case PM4_XE_SWAP:
+        // Xenia-specific VdSwap hook.
+        // VdSwap will post this to tell us we need to swap the screen/fire an interrupt.
+        XETRACECP("[%.8X] Packet(%.8X): PM4_XE_SWAP",
+                  packet_ptr, packet);
+        LOG_DATA(count);
+        ADVANCE_PTR(count);
+        graphics_system_->Swap();
+        break;
+
       case PM4_INDIRECT_BUFFER:
         // indirect buffer dispatch
         {
@@ -334,14 +344,11 @@ uint32_t CommandProcessor::ExecutePacket(PacketArgs& args) {
             } else {
               // Register.
               XEASSERT(poll_reg_addr < RegisterFile::kRegisterCount);
-
-              if (poll_reg_addr == XE_GPU_REG_COHER_STATUS_HOST) {
-                // Waiting for coherency. We should have all the info we need
-                // now (base+size+mode), so kick it off.
-                MakeCoherent();
-              }
-
               value = regs->values[poll_reg_addr].u32;
+              if (poll_reg_addr == XE_GPU_REG_COHER_STATUS_HOST) {
+                MakeCoherent();
+                value = regs->values[poll_reg_addr].u32;
+              }
             }
             switch (wait_info & 0x7) {
             case 0x0: // Never.
@@ -768,16 +775,23 @@ void CommandProcessor::WriteRegister(
 }
 
 void CommandProcessor::MakeCoherent() {
-  RegisterFile* regs = driver_->register_file();
-  auto status_host = regs->values[XE_GPU_REG_COHER_STATUS_HOST].u32;
-  auto base_host = regs->values[XE_GPU_REG_COHER_BASE_HOST].u32;
-  auto size_host = regs->values[XE_GPU_REG_COHER_SIZE_HOST].u32;
-
   // Status host often has 0x01000000 or 0x03000000.
   // This is likely toggling VC (vertex cache) or TC (texture cache).
   // Or, it also has a direction in here maybe - there is probably
   // some way to check for dest coherency (what all the COHER_DEST_BASE_*
   // registers are for).
+  // Best docs I've found on this are here:
+  // http://amd-dev.wpengine.netdna-cdn.com/wordpress/media/2013/10/R6xx_R7xx_3D.pdf
+  // http://cgit.freedesktop.org/xorg/driver/xf86-video-radeonhd/tree/src/r6xx_accel.c?id=3f8b6eccd9dba116cc4801e7f80ce21a879c67d2#n454
+
+  RegisterFile* regs = driver_->register_file();
+  auto status_host = regs->values[XE_GPU_REG_COHER_STATUS_HOST].u32;
+  auto base_host = regs->values[XE_GPU_REG_COHER_BASE_HOST].u32;
+  auto size_host = regs->values[XE_GPU_REG_COHER_SIZE_HOST].u32;
+
+  if (!(status_host & 0x80000000ul)) {
+    return;
+  }
 
   // TODO(benvanik): notify resource cache of base->size and type.
   XETRACECP("Make %.8X -> %.8X (%db) coherent",

src/xenia/gpu/d3d11/d3d11_graphics_system.cc

@@ -24,7 +24,7 @@ D3D11GraphicsSystem::D3D11GraphicsSystem(Emulator* emulator)
     : GraphicsSystem(emulator),
       window_(nullptr), dxgi_factory_(nullptr), device_(nullptr),
       timer_queue_(nullptr), vsync_timer_(nullptr),
-      interrupt_pending_(true) {
+      last_swap_time_(0.0) {
 }
 
 D3D11GraphicsSystem::~D3D11GraphicsSystem() {
@@ -141,36 +141,26 @@ void D3D11GraphicsSystem::Initialize() {
 void D3D11GraphicsSystem::Pump() {
   SCOPE_profile_cpu_f("gpu");
 
-  if (swap_pending_) {
-    swap_pending_ = false;
-
-    // TODO(benvanik): remove this when commands are understood.
-    driver_->Resolve();
-
-    // Swap window.
-    // If we are set to vsync this will block.
-    window_->Swap();
-
-    DispatchInterruptCallback(0);
-    interrupt_pending_ = false;
-  } else if (interrupt_pending_) {
-    DispatchInterruptCallback(0);
-    interrupt_pending_ = false;
-  } else {
-    double time_since_last_interrupt = xe_pal_now() - last_interrupt_time_;
-    if (time_since_last_interrupt > 0.5) {
-      // If we have gone too long without an interrupt, fire one.
-      DispatchInterruptCallback(0);
-    }
-    if (time_since_last_interrupt > 0.3) {
-      // Force a swap when profiling.
-      if (Profiler::is_enabled()) {
-        window_->Swap();
-      }
+  double time_since_last_swap = xe_pal_now() - last_swap_time_;
+  if (time_since_last_swap > 1.0) {
+    // Force a swap when profiling.
+    if (Profiler::is_enabled()) {
+      window_->Swap();
     }
   }
 }
 
+void D3D11GraphicsSystem::Swap() {
+  // TODO(benvanik): remove this when commands are understood.
+  driver_->Resolve();
+
+  // Swap window.
+  // If we are set to vsync this will block.
+  window_->Swap();
+
+  last_swap_time_ = xe_pal_now();
+}
+
 void __stdcall D3D11GraphicsSystem::VsyncCallback(D3D11GraphicsSystem* gs,
                                                   BOOLEAN) {
   static bool thread_name_set = false;
@@ -185,7 +175,6 @@ void __stdcall D3D11GraphicsSystem::VsyncCallback(D3D11GraphicsSystem* gs,
   // TODO(benvanik): we shouldn't need to do the dispatch here, but there's
   //     something wrong and the CP will block waiting for code that
   //     needs to be run in the interrupt.
-  // gs->interrupt_pending_ = true;
   gs->DispatchInterruptCallback(0);
 }
 

src/xenia/gpu/d3d11/d3d11_graphics_system.h

@@ -35,6 +35,8 @@ public:
 
   virtual void Shutdown();
 
+  void Swap() override;
+
 protected:
   virtual void Initialize();
   virtual void Pump();
@@ -49,7 +51,7 @@ private:
   HANDLE          timer_queue_;
   HANDLE          vsync_timer_;
 
-  bool            interrupt_pending_;
+  double          last_swap_time_;
 };
 
 

src/xenia/gpu/graphics_system.cc

@@ -28,7 +28,7 @@ GraphicsSystem::GraphicsSystem(Emulator* emulator) :
     thread_(nullptr), running_(false), driver_(nullptr),
     command_processor_(nullptr),
     interrupt_callback_(0), interrupt_callback_data_(0),
-    last_interrupt_time_(0), swap_pending_(false), thread_wait_(nullptr) {
+    last_interrupt_time_(0), thread_wait_(nullptr) {
   // Create the run loop used for any windows/etc.
   // This must be done on the thread we create the driver.
   run_loop_ = xe_run_loop_create();

src/xenia/gpu/graphics_system.h

@@ -45,8 +45,7 @@ public:
 
   void MarkVblank();
   void DispatchInterruptCallback(uint32_t source, uint32_t cpu = 0xFFFFFFFF);
-  bool swap_pending() const { return swap_pending_; }
-  void set_swap_pending(bool value) { swap_pending_ = value; }
+  virtual void Swap() = 0;
 
 protected:
   virtual void Initialize();
@@ -83,7 +82,6 @@ protected:
   uint32_t          interrupt_callback_;
   uint32_t          interrupt_callback_data_;
   double            last_interrupt_time_;
-  bool              swap_pending_;
   HANDLE            thread_wait_;
 };
 

src/xenia/gpu/nop/nop_graphics_system.h

@@ -28,6 +28,8 @@ public:
 
   virtual void Shutdown();
 
+  void Swap() override {}
+
 protected:
   virtual void Initialize();
   virtual void Pump();

src/xenia/gpu/resource_cache.cc

@@ -9,6 +9,8 @@
 
 #include <xenia/gpu/resource_cache.h>
 
+#include <algorithm>
+
 
 using namespace std;
 using namespace xe;
@@ -110,6 +112,8 @@ uint64_t ResourceCache::HashRange(const MemoryRange& memory_range) {
 }
 
 void ResourceCache::SyncRange(uint32_t address, int length) {
+  SCOPE_profile_cpu_f("gpu");
+
   // Scan the page table in sync with our resource list. This means
   // we have O(n) complexity for updates, though we could definitely
   // make this faster/cleaner.
@@ -118,15 +122,12 @@ void ResourceCache::SyncRange(uint32_t address, int length) {
   // will not be changing, which allows us to do a foreach(res) and reload
   // and then clear the table.
 
-  // DISABLED
-  return;
-
   // total bytes = (512 * 1024 * 1024) / (16 * 1024) = 32768
   // each byte = 1 page
   // Walk as qwords so we can clear things up faster.
   uint64_t* page_table = reinterpret_cast<uint64_t*>(
       memory_->Translate(memory_->page_table()));
-  int page_size = 16 * 1024;  // 16KB pages
+  uint32_t page_size = 16 * 1024;  // 16KB pages
 
   uint32_t lo_address = address % 0x20000000;
   uint32_t hi_address = lo_address + length;
@@ -134,24 +135,38 @@ void ResourceCache::SyncRange(uint32_t address, int length) {
   int start_page = lo_address / page_size;
   int end_page = hi_address / page_size;
 
-  auto it = paged_resources_.upper_bound(lo_address);
-  auto end_it = paged_resources_.lower_bound(hi_address);
-  while (it != end_it) {
-    const auto& memory_range = it->second->memory_range();
-    int lo_page = (memory_range.guest_base % 0x20000000) / page_size;
-    int hi_page = lo_page + (memory_range.length / page_size);
-    for (int i = lo_page / 8; i <= hi_page / 8; ++i) {
-      uint64_t page_flags = page_table[i];
-      if (page_flags) {
-        // Dirty!
-        it->second->MarkDirty(i * 8 * page_size, (i * 8 + 7) * page_size);
+  {
+    SCOPE_profile_cpu_i("gpu", "SyncRange:mark");
+    auto it = lo_address > page_size ?
+        paged_resources_.upper_bound(lo_address - page_size) :
+        paged_resources_.begin();
+    auto end_it = paged_resources_.lower_bound(hi_address + page_size);
+    while (it != end_it) {
+      const auto& memory_range = it->second->memory_range();
+      int lo_page = (memory_range.guest_base % 0x20000000) / page_size;
+      int hi_page = lo_page + (memory_range.length / page_size);
+      lo_page = std::max(lo_page, start_page);
+      hi_page = std::min(hi_page, end_page);
+      if (lo_page > hi_page) {
+        ++it;
+        continue;
       }
+      for (int i = lo_page / 8; i <= hi_page / 8; ++i) {
+        uint64_t page_flags = page_table[i];
+        if (page_flags) {
+          // Dirty!
+          it->second->MarkDirty(i * 8 * page_size, (i * 8 + 7) * page_size);
+        }
+      }
+      ++it;
     }
-    ++it;
   }
 
   // Reset page table.
-  for (auto i = start_page / 8; i <= end_page / 8; ++i) {
-    page_table[i] = 0;
+  {
+    SCOPE_profile_cpu_i("gpu", "SyncRange:reset");
+    for (auto i = start_page / 8; i <= end_page / 8; ++i) {
+      page_table[i] = 0;
+    }
   }
 }

src/xenia/gpu/texture_resource.cc

@@ -292,10 +292,9 @@ int TextureResource::Prepare() {
     }
   }
 
-  // DISABLED
-  //if (!dirtied_) {
-  //  return 0;
-  //}
+  if (!dirtied_) {
+    return 0;
+  }
   dirtied_ = false;
 
   // pass dirty regions?

src/xenia/gpu/xenos/packets.h

@@ -70,6 +70,8 @@ enum Type3Opcode {
   PM4_CONTEXT_UPDATE        = 0x5e,   // updates the current context, if needed
   PM4_INTERRUPT             = 0x54,   // generate interrupt from the command stream
 
+  PM4_XE_SWAP               = 0x55,   // Xenia only: VdSwap uses this to trigger a swap.
+
   PM4_IM_STORE              = 0x2c,   // copy sequencer instruction memory to system memory
 
   // Tiled rendering:

src/xenia/kernel/xboxkrnl_video.cc

@@ -12,6 +12,7 @@
 #include <xenia/emulator.h>
 #include <xenia/cpu/cpu.h>
 #include <xenia/gpu/gpu.h>
+#include <xenia/gpu/xenos/packets.h>
 #include <xenia/kernel/kernel_state.h>
 #include <xenia/kernel/xboxkrnl_private.h>
 #include <xenia/kernel/xboxkrnl_rtl.h>
@@ -422,19 +423,16 @@ SHIM_CALL VdSwap_shim(
       unk6,
       unk7);
 
-  KernelState* kernel_state = shared_kernel_state_;
-  XEASSERTNOTNULL(kernel_state);
-  GraphicsSystem* gs = kernel_state->emulator()->graphics_system();
-  if (!gs) {
-    return;
-  }
-
-  gs->set_swap_pending(true);
-
   // The caller seems to reserve 64 words (256b) in the primary ringbuffer
-  // for this method to do what it needs. We just zero them out. We could
-  // encode the parameters in the stream for the ringbuffer, if needed.
+  // for this method to do what it needs. We just zero them out and send a
+  // token value. It'd be nice to figure out what this is really doing so
+  // that we could simulate it, though due to TCR I bet all games need to
+  // use this method.
   xe_zero_struct(SHIM_MEM_ADDR(unk0), 64 * 4);
+  auto dwords = reinterpret_cast<uint32_t*>(SHIM_MEM_ADDR(unk0));
+  dwords[0] = XESWAP32((0x03 << 30) |
+                       ((1 - 1) << 16) |
+                       (xenos::PM4_XE_SWAP << 8));
 
   SHIM_SET_RETURN_64(0);
 }