Add VideoMemoryAllocator

This commit is contained in:
Grayson Riffe 2025-02-19 03:03:08 -06:00
parent a31ef34262
commit ae707f1f3d
11 changed files with 219 additions and 57 deletions

View File

@ -1,5 +1,5 @@
# NF library CMakeLists.txt # NF library CMakeLists.txt
add_library(NothinFancy STATIC "src/Engine.cpp" "src/include/nf.h" "src/pch.h" "src/util.h" "src/util/log.h" "src/util/log.cpp" "src/include/nf/config.h" "src/util/util.cpp" "src/util/file.h" "src/util/file.cpp" "src/client/Client.h" "src/client/Client.cpp" "src/client/Window.h" "src/client/Window.cpp" "src/client/render/RenderEngine.h" "src/client/render/RenderEngine.cpp" "src/client/render/ShaderModule.h" "src/client/render/ShaderModule.cpp" "src/client/render/VulkanResource.h" "src/client/render/Buffer.h" "src/client/render/Buffer.cpp") add_library(NothinFancy STATIC "src/Engine.cpp" "src/include/nf.h" "src/pch.h" "src/util.h" "src/util/log.h" "src/util/log.cpp" "src/include/nf/config.h" "src/util/util.cpp" "src/util/file.h" "src/util/file.cpp" "src/client/Client.h" "src/client/Client.cpp" "src/client/Window.h" "src/client/Window.cpp" "src/client/render/RenderEngine.h" "src/client/render/RenderEngine.cpp" "src/client/render/ShaderModule.h" "src/client/render/ShaderModule.cpp" "src/client/render/Resource.h" "src/client/render/Buffer.h" "src/client/render/Buffer.cpp" "src/client/render/VideoMemoryAllocator.h" "src/client/render/VideoMemoryAllocator.cpp")
# Use C++20 # Use C++20
set_property(TARGET NothinFancy PROPERTY CXX_STANDARD 20) set_property(TARGET NothinFancy PROPERTY CXX_STANDARD 20)

View File

@ -18,8 +18,18 @@ namespace nf::client {
std::thread inputThread(&Client::runInputThread, this, std::move(windowPromise)); std::thread inputThread(&Client::runInputThread, this, std::move(windowPromise));
m_renderEngine = std::make_unique<render::RenderEngine>(std::move(windowFuture.get()), m_config.display); m_renderEngine = std::make_unique<render::RenderEngine>(std::move(windowFuture.get()), m_config.display);
auto fpsClock1 = std::chrono::high_resolution_clock::now(), fpsClock2 = fpsClock1;
unsigned int frame = 0;
while (m_running) { while (m_running) {
m_renderEngine->doFrame(); m_renderEngine->doFrame();
frame++;
fpsClock2 = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> duration = fpsClock2 - fpsClock1;
if (duration.count() >= 1.0) {
NFLog(std::format("FPS: {}", frame));
frame = 0;
fpsClock1 = std::chrono::high_resolution_clock::now();
}
} }
inputThread.join(); inputThread.join();

View File

@ -5,18 +5,24 @@
#include "util.h" #include "util.h"
namespace nf::client::render { namespace nf::client::render {
Buffer::Buffer(BufferType type, const VkDevice& device, const VkPhysicalDevice& physicalDevice, const VkCommandPool& commandPool, const VkQueue& queue, void* bufferData, size_t bufferSize) Buffer::Buffer(BufferType type, const VkDevice& device, VideoMemoryAllocator& allocator, const VkCommandPool& commandPool, const VkQueue& queue, void* bufferData, size_t bufferSize)
: VulkanResource(device) : Resource(device)
, m_handle() , m_handle()
, m_memory() , m_memory()
, m_offset()
, m_allocator(allocator)
{ {
VkBuffer stagingBuffer = nullptr; VkBuffer stagingBuffer = nullptr;
VkDeviceMemory stagingBufferMemory = nullptr; createBuffer(bufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, stagingBuffer);
createBuffer(bufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, physicalDevice, stagingBuffer, stagingBufferMemory);
void* bufferPointer = nullptr; VkDeviceMemory stagingBufferMemory = nullptr;
vkMapMemory(m_device, stagingBufferMemory, 0, bufferSize, 0, &bufferPointer); VkDeviceSize stagingBufferMemoryOffset = 0;
std::memcpy(bufferPointer, bufferData, bufferSize); m_allocator.allocateForBuffer(stagingBuffer, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, stagingBufferMemory, stagingBufferMemoryOffset);
vkBindBufferMemory(m_device, stagingBuffer, stagingBufferMemory, stagingBufferMemoryOffset);
void* stagingBufferPointer = nullptr;
vkMapMemory(m_device, stagingBufferMemory, stagingBufferMemoryOffset, bufferSize, 0, &stagingBufferPointer);
std::memcpy(stagingBufferPointer, bufferData, bufferSize);
vkUnmapMemory(m_device, stagingBufferMemory); vkUnmapMemory(m_device, stagingBufferMemory);
VkBufferUsageFlags mainUsage = NULL; VkBufferUsageFlags mainUsage = NULL;
@ -31,50 +37,26 @@ namespace nf::client::render {
break; break;
} }
createBuffer(bufferSize, VK_BUFFER_USAGE_TRANSFER_DST_BIT | mainUsage, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, physicalDevice, m_handle, m_memory); createBuffer(bufferSize, VK_BUFFER_USAGE_TRANSFER_DST_BIT | mainUsage, m_handle);
m_allocator.allocateForBuffer(m_handle, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, m_memory, m_offset);
vkBindBufferMemory(m_device, m_handle, m_memory, m_offset);
copyBuffer(stagingBuffer, m_handle, bufferSize, commandPool, queue); copyBuffer(stagingBuffer, m_handle, bufferSize, commandPool, queue);
destroyBuffer(stagingBuffer, stagingBufferMemory); destroyBuffer(stagingBuffer, stagingBufferMemory, stagingBufferMemoryOffset);
} }
const VkBuffer& Buffer::getHandle() const { const VkBuffer& Buffer::getHandle() const {
return m_handle; return m_handle;
} }
void Buffer::createBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags memoryProperties, VkPhysicalDevice physicalDevice, VkBuffer& buffer, VkDeviceMemory& bufferMemory) { void Buffer::createBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkBuffer& buffer) {
VkBufferCreateInfo bufferCI = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO }; VkBufferCreateInfo bufferCI = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
bufferCI.size = size; bufferCI.size = size;
bufferCI.usage = usage; bufferCI.usage = usage;
if (vkCreateBuffer(m_device, &bufferCI, nullptr, &buffer) != VK_SUCCESS) if (vkCreateBuffer(m_device, &bufferCI, nullptr, &buffer) != VK_SUCCESS)
NFError("Could not create vertex buffer."); NFError("Could not create vertex buffer.");
VkMemoryRequirements memoryRequirements = {};
vkGetBufferMemoryRequirements(m_device, buffer, &memoryRequirements);
VkMemoryAllocateInfo allocateInfo = { VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
allocateInfo.allocationSize = memoryRequirements.size;
allocateInfo.memoryTypeIndex = findMemoryType(physicalDevice, memoryRequirements.memoryTypeBits, memoryProperties);
if (allocateInfo.memoryTypeIndex == -1)
NFError("Could not find suitable memory type.");
if (vkAllocateMemory(m_device, &allocateInfo, nullptr, &bufferMemory) != VK_SUCCESS)
NFError("Could not allocate buffer memory.");
vkBindBufferMemory(m_device, buffer, bufferMemory, 0);
}
uint32_t Buffer::findMemoryType(VkPhysicalDevice physicalDevice, uint32_t typeFilter, VkMemoryPropertyFlags properties) {
VkPhysicalDeviceMemoryProperties memoryProperties = {};
vkGetPhysicalDeviceMemoryProperties(physicalDevice, &memoryProperties);
for (int i = 0; i < memoryProperties.memoryTypeCount; i++)
if (typeFilter & (1 << i) && (memoryProperties.memoryTypes[i].propertyFlags & properties) == properties)
return i;
return -1;
} }
void Buffer::copyBuffer(VkBuffer bufferSource, VkBuffer bufferDestination, VkDeviceSize size, VkCommandPool commandPool, VkQueue queue) { void Buffer::copyBuffer(VkBuffer bufferSource, VkBuffer bufferDestination, VkDeviceSize size, VkCommandPool commandPool, VkQueue queue) {
@ -111,12 +93,12 @@ namespace nf::client::render {
vkFreeCommandBuffers(m_device, commandPool, 1, &commandBuffer); vkFreeCommandBuffers(m_device, commandPool, 1, &commandBuffer);
} }
void Buffer::destroyBuffer(VkBuffer buffer, VkDeviceMemory bufferMemory) { void Buffer::destroyBuffer(VkBuffer buffer, VkDeviceMemory bufferMemory, VkDeviceSize offset) {
vkFreeMemory(m_device, bufferMemory, nullptr); m_allocator.deallocate(bufferMemory, offset);
vkDestroyBuffer(m_device, buffer, nullptr); vkDestroyBuffer(m_device, buffer, nullptr);
} }
Buffer::~Buffer() { Buffer::~Buffer() {
destroyBuffer(m_handle, m_memory); destroyBuffer(m_handle, m_memory, m_offset);
} }
} }

View File

@ -1,7 +1,8 @@
// Buffer class header // Buffer class header
#pragma once #pragma once
#include "VulkanResource.h" #include "Resource.h"
#include "VideoMemoryAllocator.h"
namespace nf::client::render { namespace nf::client::render {
enum class BufferType { enum class BufferType {
@ -9,21 +10,24 @@ namespace nf::client::render {
Index Index
}; };
class Buffer : VulkanResource { class Buffer : Resource {
public: public:
Buffer(BufferType type, const VkDevice& device, const VkPhysicalDevice& physicalDevice, const VkCommandPool& commandPool, const VkQueue& queue, void* bufferData, size_t bufferSize); Buffer(BufferType type, const VkDevice& device, VideoMemoryAllocator& allocator, const VkCommandPool& commandPool, const VkQueue& queue, void* bufferData, size_t bufferSize);
const VkBuffer& getHandle() const; const VkBuffer& getHandle() const;
~Buffer(); ~Buffer();
private: private:
void createBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags memoryProperties, VkPhysicalDevice physicalDevice, VkBuffer& buffer, VkDeviceMemory& bufferMemory); void createBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkBuffer& buffer);
uint32_t findMemoryType(VkPhysicalDevice physicalDevice, uint32_t typeFilter, VkMemoryPropertyFlags properties);
void copyBuffer(VkBuffer sourceBuffer, VkBuffer destinationBuffer, VkDeviceSize size, VkCommandPool commandPool, VkQueue queue); void copyBuffer(VkBuffer sourceBuffer, VkBuffer destinationBuffer, VkDeviceSize size, VkCommandPool commandPool, VkQueue queue);
void destroyBuffer(VkBuffer buffer, VkDeviceMemory bufferMemory); void destroyBuffer(VkBuffer buffer, VkDeviceMemory bufferMemory, VkDeviceSize offset);
VkBuffer m_handle; VkBuffer m_handle;
VkDeviceMemory m_memory; VkDeviceMemory m_memory;
VkDeviceSize m_offset;
VideoMemoryAllocator& m_allocator;
}; };
} }

View File

@ -36,6 +36,7 @@ namespace nf::client::render {
, m_numFramesInFlight(2) , m_numFramesInFlight(2)
, m_frameExecutors() , m_frameExecutors()
, m_currentFrameExecutor() , m_currentFrameExecutor()
, m_allocator()
, m_bufferVertex() , m_bufferVertex()
, m_bufferIndex() , m_bufferIndex()
{ {
@ -52,6 +53,7 @@ namespace nf::client::render {
createOutputPipeline(); createOutputPipeline();
createFrameExecutors(); createFrameExecutors();
m_allocator = std::make_unique<VideoMemoryAllocator>(m_device, m_physicalDevice);
createBuffers(); createBuffers();
m_window->show(); m_window->show();
@ -452,14 +454,14 @@ namespace nf::client::render {
}; };
size_t triangleVerticesSize = sizeof(triangleVertices[0]) * triangleVertices.size(); size_t triangleVerticesSize = sizeof(triangleVertices[0]) * triangleVertices.size();
m_bufferVertex = new Buffer(BufferType::Vertex, m_device, m_physicalDevice, m_commandPool, m_queueGraphics, triangleVertices.data(), triangleVerticesSize); m_bufferVertex = std::make_unique<Buffer>(BufferType::Vertex, m_device, *m_allocator, m_commandPool, m_queueGraphics, triangleVertices.data(), triangleVerticesSize);
std::vector<uint32_t> triangleIndices = { std::vector<uint32_t> triangleIndices = {
0, 1, 2, 2, 3, 0 0, 1, 2, 2, 3, 0
}; };
size_t triangleIndicesSize = sizeof(triangleIndices[0]) * triangleIndices.size(); size_t triangleIndicesSize = sizeof(triangleIndices[0]) * triangleIndices.size();
m_bufferIndex = new Buffer(BufferType::Index, m_device, m_physicalDevice, m_commandPool, m_queueGraphics, triangleIndices.data(), triangleIndicesSize); m_bufferIndex = std::make_unique<Buffer>(BufferType::Index, m_device, *m_allocator, m_commandPool, m_queueGraphics, triangleIndices.data(), triangleIndicesSize);
} }
void RenderEngine::doFrame() { void RenderEngine::doFrame() {
@ -591,8 +593,8 @@ namespace nf::client::render {
RenderEngine::~RenderEngine() { RenderEngine::~RenderEngine() {
waitIdle(); waitIdle();
delete m_bufferIndex; m_bufferIndex.reset();
delete m_bufferVertex; m_bufferVertex.reset();
for (int i = 0; i < m_frameExecutors.size(); i++) { for (int i = 0; i < m_frameExecutors.size(); i++) {
vkDestroyFence(m_device, m_frameExecutors[i].fenceFrameInFlight, nullptr); vkDestroyFence(m_device, m_frameExecutors[i].fenceFrameInFlight, nullptr);

View File

@ -3,6 +3,7 @@
#include "client/Window.h" #include "client/Window.h"
#include "nf/config.h" #include "nf/config.h"
#include "VideoMemoryAllocator.h"
#include "Buffer.h" #include "Buffer.h"
namespace nf::client::render { namespace nf::client::render {
@ -68,7 +69,8 @@ namespace nf::client::render {
std::vector<FrameExecutor> m_frameExecutors; std::vector<FrameExecutor> m_frameExecutors;
uint32_t m_currentFrameExecutor; uint32_t m_currentFrameExecutor;
Buffer* m_bufferVertex; std::unique_ptr<VideoMemoryAllocator> m_allocator;
Buffer* m_bufferIndex; std::unique_ptr<Buffer> m_bufferVertex;
std::unique_ptr<Buffer> m_bufferIndex;
}; };
} }

View File

@ -2,9 +2,9 @@
#pragma once #pragma once
namespace nf::client::render { namespace nf::client::render {
class VulkanResource { class Resource {
public: public:
VulkanResource(const VkDevice& device) Resource(const VkDevice& device)
: m_device(device) : m_device(device)
{} {}

View File

@ -6,7 +6,7 @@
namespace nf::client::render { namespace nf::client::render {
ShaderModule::ShaderModule(const VkDevice& device, const std::string& shaderBinary) ShaderModule::ShaderModule(const VkDevice& device, const std::string& shaderBinary)
: VulkanResource(device) : Resource(device)
, m_shaderModule() , m_shaderModule()
{ {
VkShaderModuleCreateInfo shaderModuleCI = { VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO }; VkShaderModuleCreateInfo shaderModuleCI = { VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO };

View File

@ -1,10 +1,10 @@
// ShaderModule class header // ShaderModule class header
#pragma once #pragma once
#include "VulkanResource.h" #include "Resource.h"
namespace nf::client::render { namespace nf::client::render {
class ShaderModule : VulkanResource { class ShaderModule : Resource {
public: public:
ShaderModule(const VkDevice& device, const std::string& shaderBinary); ShaderModule(const VkDevice& device, const std::string& shaderBinary);

View File

@ -0,0 +1,130 @@
// VideoMemoryAllocator class implementation
#include "pch.h"
#include "VideoMemoryAllocator.h"
#include "util.h"
// 1 GiB
#define HEAP_SIZE_LARGE 1024ull * 1024 * 1024
// 50 MiB
#define DEFAULT_ALLOCATION_AMOUNT 50ull * 1024 * 1024
namespace nf::client::render {
VideoMemoryAllocator::VideoMemoryAllocator(const VkDevice& device, const VkPhysicalDevice& physicalDevice)
: m_device(device)
, m_physicalDevice(physicalDevice)
, m_blocks()
{}
void VideoMemoryAllocator::allocateForBuffer(VkBuffer buffer, VkMemoryPropertyFlags memoryPropertyFlags, VkDeviceMemory& outDeviceMemory, VkDeviceSize& outOffset) {
VkMemoryRequirements memoryRequirements = {};
vkGetBufferMemoryRequirements(m_device, buffer, &memoryRequirements);
VkMemoryAllocateInfo allocateInfo = { VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
allocateInfo.allocationSize = memoryRequirements.size;
allocateInfo.memoryTypeIndex = findMemoryType(memoryRequirements.memoryTypeBits, memoryPropertyFlags);
if (allocateInfo.memoryTypeIndex == -1)
NFError("Could not find suitable memory type.");
allocateFromBlock(allocateInfo, memoryRequirements.alignment, outDeviceMemory, outOffset);
}
void VideoMemoryAllocator::deallocate(VkDeviceMemory memoryBlock, VkDeviceSize offset) {
const auto blockIterator = std::find_if(m_blocks.begin(), m_blocks.end(), [&](const auto& currentBlock) {
if (currentBlock.memory == memoryBlock)
return true;
else return false;
});
if (blockIterator == m_blocks.end())
NFError("Could not find video memory block to deallocate from.");
MemoryBlock& block = *blockIterator;
const auto allocationIterator = block.allocations.find(offset);
if (allocationIterator == block.allocations.end())
NFError("Could not find video memory allocation to deallocate.");
block.allocations.erase(allocationIterator);
// If block is now empty, free it
if (block.allocations.empty()) {
vkFreeMemory(m_device, block.memory, nullptr);
m_blocks.erase(blockIterator);
}
}
uint32_t VideoMemoryAllocator::findMemoryType(uint32_t typeFilter, VkMemoryPropertyFlags desiredPropertyFlags) {
VkPhysicalDeviceMemoryProperties memoryProperties = {};
vkGetPhysicalDeviceMemoryProperties(m_physicalDevice, &memoryProperties);
for (int i = 0; i < memoryProperties.memoryTypeCount; i++)
if (typeFilter & (1 << i) && (memoryProperties.memoryTypes[i].propertyFlags & desiredPropertyFlags) == desiredPropertyFlags)
return i;
return -1;
}
void VideoMemoryAllocator::allocateFromBlock(VkMemoryAllocateInfo& allocateInfo, uint32_t alignment, VkDeviceMemory& outDeviceMemory, VkDeviceSize& outOffset) {
while (true) {
for (auto& block : m_blocks) {
if (block.memoryTypeIndex == allocateInfo.memoryTypeIndex) {
VkDeviceSize lastOffset = 0, lastSize = 0;
// Check between existing allocations
for (const auto& [currentOffset, currentSize] : block.allocations) {
// Check for alignment
VkDeviceSize checkOffset = lastOffset + lastSize, remainder = checkOffset % alignment;
checkOffset = remainder ? checkOffset + (alignment - remainder) : checkOffset;
// Check for enough space
if (currentOffset - lastOffset - lastSize >= allocateInfo.allocationSize) {
outDeviceMemory = block.memory;
outOffset = checkOffset;
block.allocations[checkOffset] = allocateInfo.allocationSize;
return;
}
lastOffset = currentOffset, lastSize = currentSize;
}
// Check end
VkDeviceSize checkOffset = lastOffset + lastSize, remainder = checkOffset % alignment;
checkOffset = remainder ? checkOffset + (alignment - remainder) : checkOffset;
if ((block.size - 1) - lastOffset - lastSize >= allocateInfo.allocationSize) {
outDeviceMemory = block.memory;
outOffset = checkOffset;
block.allocations[checkOffset] = allocateInfo.allocationSize;
return;
}
// If not enough space, continue
continue;
}
}
// No suitable block found, allocate one
VkMemoryAllocateInfo newBlockAllocateInfo = { VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO };
newBlockAllocateInfo.memoryTypeIndex = allocateInfo.memoryTypeIndex;
newBlockAllocateInfo.allocationSize = calculateBlockSize(allocateInfo.memoryTypeIndex);
allocateBlock(newBlockAllocateInfo);
}
}
size_t VideoMemoryAllocator::calculateBlockSize(uint32_t memoryTypeIndex) const {
VkPhysicalDeviceMemoryProperties memoryProperties = {};
vkGetPhysicalDeviceMemoryProperties(m_physicalDevice, &memoryProperties);
size_t heapSize = memoryProperties.memoryHeaps[memoryProperties.memoryTypes[memoryTypeIndex].heapIndex].size;
return heapSize > HEAP_SIZE_LARGE ? DEFAULT_ALLOCATION_AMOUNT : (heapSize / 8);
}
void VideoMemoryAllocator::allocateBlock(VkMemoryAllocateInfo& allocateInfo) {
VkDeviceMemory memoryBlock = nullptr;
if (vkAllocateMemory(m_device, &allocateInfo, nullptr, &memoryBlock) != VK_SUCCESS)
NFError("Could not allocate video memory.");
m_blocks.emplace_back(memoryBlock, allocateInfo.memoryTypeIndex, allocateInfo.allocationSize);
}
}

View File

@ -0,0 +1,32 @@
// VideoMemoryAllocator class header
#pragma once
namespace nf::client::render {
struct MemoryBlock {
VkDeviceMemory memory;
uint32_t memoryTypeIndex;
size_t size;
std::map<VkDeviceSize, VkDeviceSize> allocations;
// Offset Size
};
class VideoMemoryAllocator {
public:
VideoMemoryAllocator(const VkDevice& device, const VkPhysicalDevice& physicalDevice);
void allocateForBuffer(VkBuffer buffer, VkMemoryPropertyFlags memoryPropertyFlags, VkDeviceMemory& outDeviceMemory, VkDeviceSize& outOffset);
void deallocate(VkDeviceMemory memoryBlock, VkDeviceSize offset);
private:
uint32_t findMemoryType(uint32_t typeFilter, VkMemoryPropertyFlags desiredPropertyFlags);
void allocateFromBlock(VkMemoryAllocateInfo& allocateInfo, uint32_t alignment, VkDeviceMemory& outDeviceMemory, VkDeviceSize& outOffset);
size_t calculateBlockSize(uint32_t memoryTypeIndex) const;
void allocateBlock(VkMemoryAllocateInfo& allocateInfo);
const VkDevice& m_device;
const VkPhysicalDevice& m_physicalDevice;
std::vector<MemoryBlock> m_blocks;
};
}