CircularBuffer: Allow users to bind their own memory to our buffer.

src/xenia/gpu/vulkan/buffer_cache.cc

@@ -28,11 +28,17 @@ constexpr VkDeviceSize kConstantRegisterUniformRange =
 BufferCache::BufferCache(RegisterFile* register_file,
                          ui::vulkan::VulkanDevice* device, size_t capacity)
     : register_file_(register_file), device_(*device) {
-  transient_buffer_ = std::make_unique<ui::vulkan::CircularBuffer>(device);
-  if (!transient_buffer_->Initialize(capacity,
-                                     VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
-                                         VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
-                                         VK_BUFFER_USAGE_VERTEX_BUFFER_BIT)) {
+  transient_buffer_ = std::make_unique<ui::vulkan::CircularBuffer>(
+      device,
+      VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
+          VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
+      capacity);
+
+  VkMemoryRequirements pool_reqs;
+  transient_buffer_->GetBufferMemoryRequirements(&pool_reqs);
+  gpu_memory_pool_ = device->AllocateMemory(pool_reqs);
+
+  if (!transient_buffer_->Initialize(gpu_memory_pool_, 0)) {
     assert_always();
   }
 
@@ -134,6 +140,10 @@ BufferCache::~BufferCache() {
   vkDestroyDescriptorSetLayout(device_, descriptor_set_layout_, nullptr);
   vkDestroyDescriptorPool(device_, descriptor_pool_, nullptr);
   transient_buffer_->Shutdown();
+
+  if (gpu_memory_pool_) {
+    vkFreeMemory(device_, gpu_memory_pool_, nullptr);
+  }
 }
 
 std::pair<VkDeviceSize, VkDeviceSize> BufferCache::UploadConstantRegisters(
@@ -221,8 +231,6 @@ std::pair<VkDeviceSize, VkDeviceSize> BufferCache::UploadConstantRegisters(
 std::pair<VkBuffer, VkDeviceSize> BufferCache::UploadIndexBuffer(
     const void* source_ptr, size_t source_length, IndexFormat format,
     std::shared_ptr<ui::vulkan::Fence> fence) {
-  // TODO(benvanik): check cache.
-
   // Allocate space in the buffer for our data.
   auto offset = AllocateTransientData(source_length, fence);
   if (offset == VK_WHOLE_SIZE) {
@@ -249,8 +257,6 @@ std::pair<VkBuffer, VkDeviceSize> BufferCache::UploadIndexBuffer(
 std::pair<VkBuffer, VkDeviceSize> BufferCache::UploadVertexBuffer(
     const void* source_ptr, size_t source_length, Endian endian,
     std::shared_ptr<ui::vulkan::Fence> fence) {
-  // TODO(benvanik): check cache.
-
   // Allocate space in the buffer for our data.
   auto offset = AllocateTransientData(source_length, fence);
   if (offset == VK_WHOLE_SIZE) {
@@ -324,7 +330,7 @@ void BufferCache::InvalidateCache() {
 }
 
 void BufferCache::ClearCache() {
-  // TODO(benvanik): caching.
+  transient_cache_.clear();
 }
 
 void BufferCache::Scavenge() { transient_buffer_->Scavenge(); }

src/xenia/gpu/vulkan/buffer_cache.h

@@ -17,6 +17,8 @@
 #include "xenia/ui/vulkan/vulkan.h"
 #include "xenia/ui/vulkan/vulkan_device.h"
 
+#include <unordered_map>
+
 namespace xe {
 namespace gpu {
 namespace vulkan {
@@ -101,9 +103,12 @@ class BufferCache {
   RegisterFile* register_file_ = nullptr;
   VkDevice device_ = nullptr;
 
+  VkDeviceMemory gpu_memory_pool_ = nullptr;
+
   // Staging ringbuffer we cycle through fast. Used for data we don't
   // plan on keeping past the current frame.
   std::unique_ptr<ui::vulkan::CircularBuffer> transient_buffer_ = nullptr;
+  std::unordered_map<uint64_t, VkDeviceSize> transient_cache_;
 
   VkDescriptorPool descriptor_pool_ = nullptr;
   VkDescriptorSetLayout descriptor_set_layout_ = nullptr;

src/xenia/gpu/vulkan/texture_cache.cc

@@ -111,7 +111,8 @@ TextureCache::TextureCache(Memory* memory, RegisterFile* register_file,
       register_file_(register_file),
       trace_writer_(trace_writer),
       device_(device),
-      staging_buffer_(device) {
+      staging_buffer_(device, VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
+                      kStagingBufferSize) {
   // Descriptor pool used for all of our cached descriptors.
   VkDescriptorPoolCreateInfo descriptor_pool_info;
   descriptor_pool_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
@@ -152,8 +153,7 @@ TextureCache::TextureCache(Memory* memory, RegisterFile* register_file,
                                     nullptr, &texture_descriptor_set_layout_);
   CheckResult(err, "vkCreateDescriptorSetLayout");
 
-  if (!staging_buffer_.Initialize(kStagingBufferSize,
-                                  VK_BUFFER_USAGE_TRANSFER_SRC_BIT)) {
+  if (!staging_buffer_.Initialize()) {
     assert_always();
   }
 

src/xenia/ui/vulkan/circular_buffer.cc

@@ -19,13 +19,10 @@ namespace xe {
 namespace ui {
 namespace vulkan {
 
-CircularBuffer::CircularBuffer(VulkanDevice* device) : device_(device) {}
-CircularBuffer::~CircularBuffer() { Shutdown(); }
-
-bool CircularBuffer::Initialize(VkDeviceSize capacity, VkBufferUsageFlags usage,
-                                VkDeviceSize alignment) {
+CircularBuffer::CircularBuffer(VulkanDevice* device, VkBufferUsageFlags usage,
+                               VkDeviceSize capacity, VkDeviceSize alignment)
+    : device_(device), capacity_(capacity) {
   VkResult status = VK_SUCCESS;
-  capacity = xe::round_up(capacity, alignment);
 
   // Create our internal buffer.
   VkBufferCreateInfo buffer_info;
@@ -40,15 +37,52 @@ bool CircularBuffer::Initialize(VkDeviceSize capacity, VkBufferUsageFlags usage,
   status = vkCreateBuffer(*device_, &buffer_info, nullptr, &gpu_buffer_);
   CheckResult(status, "vkCreateBuffer");
   if (status != VK_SUCCESS) {
+    assert_always();
+  }
+
+  VkMemoryRequirements reqs;
+  vkGetBufferMemoryRequirements(*device_, gpu_buffer_, &reqs);
+  alignment_ = reqs.alignment;
+}
+CircularBuffer::~CircularBuffer() { Shutdown(); }
+
+bool CircularBuffer::Initialize(VkDeviceMemory memory, VkDeviceSize offset) {
+  assert_true(offset % alignment_ == 0);
+  gpu_memory_ = memory;
+  gpu_base_ = offset;
+
+  VkResult status = VK_SUCCESS;
+
+  // Bind the buffer to its backing memory.
+  status = vkBindBufferMemory(*device_, gpu_buffer_, gpu_memory_, gpu_base_);
+  CheckResult(status, "vkBindBufferMemory");
+  if (status != VK_SUCCESS) {
+    XELOGE("CircularBuffer::Initialize - Failed to bind memory!");
+    Shutdown();
     return false;
   }
 
+  // Map the memory so we can access it.
+  status = vkMapMemory(*device_, gpu_memory_, gpu_base_, capacity_, 0,
+                       reinterpret_cast<void**>(&host_base_));
+  CheckResult(status, "vkMapMemory");
+  if (status != VK_SUCCESS) {
+    XELOGE("CircularBuffer::Initialize - Failed to map memory!");
+    Shutdown();
+    return false;
+  }
+
+  return true;
+}
+
+bool CircularBuffer::Initialize() {
+  VkResult status = VK_SUCCESS;
+
   VkMemoryRequirements reqs;
   vkGetBufferMemoryRequirements(*device_, gpu_buffer_, &reqs);
 
   // Allocate memory from the device to back the buffer.
-  assert_true(reqs.size == capacity);
-  reqs.alignment = std::max(alignment, reqs.alignment);
+  owns_gpu_memory_ = true;
   gpu_memory_ = device_->AllocateMemory(reqs);
   if (!gpu_memory_) {
     XELOGE("CircularBuffer::Initialize - Failed to allocate memory!");
@@ -56,7 +90,6 @@ bool CircularBuffer::Initialize(VkDeviceSize capacity, VkBufferUsageFlags usage,
     return false;
   }
 
-  alignment_ = reqs.alignment;
   capacity_ = reqs.size;
   gpu_base_ = 0;
 
@@ -92,12 +125,16 @@ void CircularBuffer::Shutdown() {
     vkDestroyBuffer(*device_, gpu_buffer_, nullptr);
     gpu_buffer_ = nullptr;
   }
-  if (gpu_memory_) {
+  if (gpu_memory_ && owns_gpu_memory_) {
     vkFreeMemory(*device_, gpu_memory_, nullptr);
     gpu_memory_ = nullptr;
   }
 }
 
+void CircularBuffer::GetBufferMemoryRequirements(VkMemoryRequirements* reqs) {
+  vkGetBufferMemoryRequirements(*device_, gpu_buffer_, reqs);
+}
+
 bool CircularBuffer::CanAcquire(VkDeviceSize length) {
   // Make sure the length is aligned.
   length = xe::round_up(length, alignment_);
@@ -166,7 +203,8 @@ CircularBuffer::Allocation* CircularBuffer::Acquire(
 
       return alloc;
     } else if ((read_head_ - 0) >= aligned_length) {
-      // Free space from begin -> read
+      // Not enough space from write -> capacity, but there is enough free space
+      // from begin -> read
       auto alloc = new Allocation();
       alloc->host_ptr = host_base_ + 0;
       alloc->gpu_memory = gpu_memory_;
@@ -220,6 +258,11 @@ void CircularBuffer::Scavenge() {
     delete *it;
     it = allocations_.erase(it);
   }
+
+  if (allocations_.empty()) {
+    // Reset R/W heads to work around fragmentation issues.
+    read_head_ = write_head_ = 0;
+  }
 }
 
 }  // namespace vulkan

src/xenia/ui/vulkan/circular_buffer.h

@@ -27,7 +27,8 @@ namespace vulkan {
 // ends of the buffer), where trailing older allocations are freed after use.
 class CircularBuffer {
  public:
-  CircularBuffer(VulkanDevice* device);
+  CircularBuffer(VulkanDevice* device, VkBufferUsageFlags usage,
+                 VkDeviceSize capacity, VkDeviceSize alignment = 256);
   ~CircularBuffer();
 
   struct Allocation {
@@ -42,10 +43,12 @@ class CircularBuffer {
     std::shared_ptr<Fence> fence;
   };
 
-  bool Initialize(VkDeviceSize capacity, VkBufferUsageFlags usage,
-                  VkDeviceSize alignment = 256);
+  bool Initialize(VkDeviceMemory memory, VkDeviceSize offset);
+  bool Initialize();
   void Shutdown();
 
+  void GetBufferMemoryRequirements(VkMemoryRequirements* reqs);
+
   VkDeviceSize alignment() const { return alignment_; }
   VkDeviceSize capacity() const { return capacity_; }
   VkBuffer gpu_buffer() const { return gpu_buffer_; }
@@ -66,12 +69,14 @@ class CircularBuffer {
   void Scavenge();
 
  private:
+  // All of these variables are relative to gpu_base
   VkDeviceSize capacity_ = 0;
   VkDeviceSize alignment_ = 0;
   VkDeviceSize write_head_ = 0;
   VkDeviceSize read_head_ = 0;
 
   VulkanDevice* device_;
+  bool owns_gpu_memory_ = false;
   VkBuffer gpu_buffer_ = nullptr;
   VkDeviceMemory gpu_memory_ = nullptr;
   VkDeviceSize gpu_base_ = 0;