[Memory] Pass exact_range to watch callbacks

src/xenia/gpu/d3d12/primitive_converter.cc

@@ -700,7 +700,7 @@ void* PrimitiveConverter::AllocateIndices(
 }
 
 std::pair<uint32_t, uint32_t> PrimitiveConverter::MemoryWriteCallback(
-    uint32_t physical_address_start, uint32_t length) {
+    uint32_t physical_address_start, uint32_t length, bool exact_range) {
   // 1 bit = (512 / 64) MB = 8 MB. Invalidate a region of this size.
   uint32_t bit_index_first = physical_address_start >> 23;
   uint32_t bit_index_last = (physical_address_start + length - 1) >> 23;
@@ -713,9 +713,10 @@ std::pair<uint32_t, uint32_t> PrimitiveConverter::MemoryWriteCallback(
 }
 
 std::pair<uint32_t, uint32_t> PrimitiveConverter::MemoryWriteCallbackThunk(
-    void* context_ptr, uint32_t physical_address_start, uint32_t length) {
+    void* context_ptr, uint32_t physical_address_start, uint32_t length,
+    bool exact_range) {
   return reinterpret_cast<PrimitiveConverter*>(context_ptr)
-      ->MemoryWriteCallback(physical_address_start, length);
+      ->MemoryWriteCallback(physical_address_start, length, exact_range);
 }
 
 D3D12_GPU_VIRTUAL_ADDRESS PrimitiveConverter::GetStaticIndexBuffer(

src/xenia/gpu/d3d12/primitive_converter.h

@@ -90,9 +90,10 @@ class PrimitiveConverter {
 
   // Callback for invalidating buffers mid-frame.
   std::pair<uint32_t, uint32_t> MemoryWriteCallback(
-      uint32_t physical_address_start, uint32_t length);
+      uint32_t physical_address_start, uint32_t length, bool exact_range);
   static std::pair<uint32_t, uint32_t> MemoryWriteCallbackThunk(
-      void* context_ptr, uint32_t physical_address_start, uint32_t length);
+      void* context_ptr, uint32_t physical_address_start, uint32_t length,
+      bool exact_range);
 
   D3D12CommandProcessor* command_processor_;
   RegisterFile* register_file_;

src/xenia/gpu/d3d12/shared_memory.cc

@@ -578,13 +578,14 @@ void SharedMemory::GetRangesToUpload(uint32_t request_page_first,
 }
 
 std::pair<uint32_t, uint32_t> SharedMemory::MemoryWriteCallbackThunk(
-    void* context_ptr, uint32_t physical_address_start, uint32_t length) {
+    void* context_ptr, uint32_t physical_address_start, uint32_t length,
+    bool exact_range) {
   return reinterpret_cast<SharedMemory*>(context_ptr)
-      ->MemoryWriteCallback(physical_address_start, length);
+      ->MemoryWriteCallback(physical_address_start, length, exact_range);
 }
 
 std::pair<uint32_t, uint32_t> SharedMemory::MemoryWriteCallback(
-    uint32_t physical_address_start, uint32_t length) {
+    uint32_t physical_address_start, uint32_t length, bool exact_range) {
   uint32_t page_first = physical_address_start >> page_size_log2_;
   uint32_t page_last = (physical_address_start + length - 1) >> page_size_log2_;
   assert_true(page_first < page_count_ && page_last < page_count_);
@@ -593,12 +594,13 @@ std::pair<uint32_t, uint32_t> SharedMemory::MemoryWriteCallback(
 
   auto global_lock = global_critical_region_.Acquire();
 
+  if (!exact_range) {
     // Check if a somewhat wider range (up to 256 KB with 4 KB pages) can be
     // invalidated - if no GPU-written data nearby that was not intended to be
-  // invalidated since it's not in sync with CPU memory and can't be reuploaded.
+    // invalidated since it's not in sync with CPU memory and can't be
-  // It's a lot cheaper to upload some excess data than to catch access
+    // reuploaded. It's a lot cheaper to upload some excess data than to catch
-  // violations - with 4 KB callbacks, the original Doom runs at 4 FPS on
+    // access violations - with 4 KB callbacks, the original Doom runs at 4 FPS
-  // Intel Core i7-3770, with 64 KB the CPU game code takes 3 ms to run per
+    // on Intel Core i7-3770, with 64 KB the CPU game code takes 3 ms to run per
     // frame, but with 256 KB it's 0.7 ms.
     if (page_first & 63) {
       uint64_t gpu_written_start =
@@ -614,6 +616,7 @@ std::pair<uint32_t, uint32_t> SharedMemory::MemoryWriteCallback(
       page_last = (page_last & ~uint32_t(63)) +
                   (std::max(xe::tzcnt(gpu_written_end), uint8_t(1)) - 1);
     }
+  }
 
   for (uint32_t i = block_first; i <= block_last; ++i) {
     uint64_t invalidate_bits = UINT64_MAX;

src/xenia/gpu/d3d12/shared_memory.h

@@ -191,9 +191,10 @@ class SharedMemory {
 
   // Memory access callback.
   static std::pair<uint32_t, uint32_t> MemoryWriteCallbackThunk(
-      void* context_ptr, uint32_t physical_address_start, uint32_t length);
+      void* context_ptr, uint32_t physical_address_start, uint32_t length,
+      bool exact_range);
   std::pair<uint32_t, uint32_t> MemoryWriteCallback(
-      uint32_t physical_address_start, uint32_t length);
+      uint32_t physical_address_start, uint32_t length, bool exact_range);
 
   struct GlobalWatch {
     GlobalWatchCallback callback;

src/xenia/memory.cc

@@ -1643,7 +1643,8 @@ bool PhysicalHeap::TriggerWatches(uint32_t virtual_address, uint32_t length,
   for (auto physical_write_watch : memory_->physical_write_watches_) {
     std::pair<uint32_t, uint32_t> callback_unwatch_range =
         physical_write_watch->callback(physical_write_watch->callback_context,
-                                       physical_address_start, physical_length);
+                                       physical_address_start, physical_length,
+                                       unwatch_exact_range);
     if (!unwatch_exact_range) {
       unwatch_first = std::max(unwatch_first, callback_unwatch_range.first);
       unwatch_last = std::min(

src/xenia/memory.h

@@ -338,7 +338,8 @@ class Memory {
   // the watched region that can be safely unwatched, if it doesn't matter,
   // return (0, UINT32_MAX).
   typedef std::pair<uint32_t, uint32_t> (*PhysicalWriteWatchCallback)(
-      void* context_ptr, uint32_t physical_address_start, uint32_t length);
+      void* context_ptr, uint32_t physical_address_start, uint32_t length,
+      bool exact_range);
 
   // Physical memory write watching, allowing subsystems to invalidate cached
   // data that depends on memory contents.