Merge pull request #5297 from ReinUsesLisp/vulkan-allocator-common

vulkan_memory_allocator: Improvements to the memory allocator
This commit is contained in:
Rodrigo Locatti 2021-01-15 21:50:05 -03:00 committed by GitHub
commit c17ee0da5d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
19 changed files with 608 additions and 553 deletions

View file

@ -135,8 +135,6 @@ add_library(video_core STATIC
renderer_vulkan/vk_graphics_pipeline.h renderer_vulkan/vk_graphics_pipeline.h
renderer_vulkan/vk_master_semaphore.cpp renderer_vulkan/vk_master_semaphore.cpp
renderer_vulkan/vk_master_semaphore.h renderer_vulkan/vk_master_semaphore.h
renderer_vulkan/vk_memory_manager.cpp
renderer_vulkan/vk_memory_manager.h
renderer_vulkan/vk_pipeline_cache.cpp renderer_vulkan/vk_pipeline_cache.cpp
renderer_vulkan/vk_pipeline_cache.h renderer_vulkan/vk_pipeline_cache.h
renderer_vulkan/vk_query_cache.cpp renderer_vulkan/vk_query_cache.cpp
@ -259,6 +257,8 @@ add_library(video_core STATIC
vulkan_common/vulkan_instance.h vulkan_common/vulkan_instance.h
vulkan_common/vulkan_library.cpp vulkan_common/vulkan_library.cpp
vulkan_common/vulkan_library.h vulkan_common/vulkan_library.h
vulkan_common/vulkan_memory_allocator.cpp
vulkan_common/vulkan_memory_allocator.h
vulkan_common/vulkan_surface.cpp vulkan_common/vulkan_surface.cpp
vulkan_common/vulkan_surface.h vulkan_common/vulkan_surface.h
vulkan_common/vulkan_wrapper.cpp vulkan_common/vulkan_wrapper.cpp

View file

@ -23,7 +23,6 @@
#include "video_core/renderer_vulkan/renderer_vulkan.h" #include "video_core/renderer_vulkan/renderer_vulkan.h"
#include "video_core/renderer_vulkan/vk_blit_screen.h" #include "video_core/renderer_vulkan/vk_blit_screen.h"
#include "video_core/renderer_vulkan/vk_master_semaphore.h" #include "video_core/renderer_vulkan/vk_master_semaphore.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_rasterizer.h" #include "video_core/renderer_vulkan/vk_rasterizer.h"
#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_state_tracker.h" #include "video_core/renderer_vulkan/vk_state_tracker.h"
@ -32,6 +31,7 @@
#include "video_core/vulkan_common/vulkan_device.h" #include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_instance.h" #include "video_core/vulkan_common/vulkan_instance.h"
#include "video_core/vulkan_common/vulkan_library.h" #include "video_core/vulkan_common/vulkan_library.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_surface.h" #include "video_core/vulkan_common/vulkan_surface.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
@ -137,7 +137,7 @@ bool RendererVulkan::Init() try {
InitializeDevice(); InitializeDevice();
Report(); Report();
memory_manager = std::make_unique<VKMemoryManager>(*device); memory_allocator = std::make_unique<MemoryAllocator>(*device);
state_tracker = std::make_unique<StateTracker>(gpu); state_tracker = std::make_unique<StateTracker>(gpu);
@ -149,11 +149,11 @@ bool RendererVulkan::Init() try {
rasterizer = std::make_unique<RasterizerVulkan>(render_window, gpu, gpu.MemoryManager(), rasterizer = std::make_unique<RasterizerVulkan>(render_window, gpu, gpu.MemoryManager(),
cpu_memory, screen_info, *device, cpu_memory, screen_info, *device,
*memory_manager, *state_tracker, *scheduler); *memory_allocator, *state_tracker, *scheduler);
blit_screen = blit_screen =
std::make_unique<VKBlitScreen>(cpu_memory, render_window, *rasterizer, *device, std::make_unique<VKBlitScreen>(cpu_memory, render_window, *rasterizer, *device,
*memory_manager, *swapchain, *scheduler, screen_info); *memory_allocator, *swapchain, *scheduler, screen_info);
return true; return true;
} catch (const vk::Exception& exception) { } catch (const vk::Exception& exception) {
@ -172,7 +172,7 @@ void RendererVulkan::ShutDown() {
blit_screen.reset(); blit_screen.reset();
scheduler.reset(); scheduler.reset();
swapchain.reset(); swapchain.reset();
memory_manager.reset(); memory_allocator.reset();
device.reset(); device.reset();
} }

View file

@ -29,8 +29,8 @@ namespace Vulkan {
class Device; class Device;
class StateTracker; class StateTracker;
class MemoryAllocator;
class VKBlitScreen; class VKBlitScreen;
class VKMemoryManager;
class VKSwapchain; class VKSwapchain;
class VKScheduler; class VKScheduler;
@ -75,7 +75,7 @@ private:
vk::DebugUtilsMessenger debug_callback; vk::DebugUtilsMessenger debug_callback;
std::unique_ptr<Device> device; std::unique_ptr<Device> device;
std::unique_ptr<VKMemoryManager> memory_manager; std::unique_ptr<MemoryAllocator> memory_allocator;
std::unique_ptr<StateTracker> state_tracker; std::unique_ptr<StateTracker> state_tracker;
std::unique_ptr<VKScheduler> scheduler; std::unique_ptr<VKScheduler> scheduler;
std::unique_ptr<VKSwapchain> swapchain; std::unique_ptr<VKSwapchain> swapchain;

View file

@ -22,13 +22,13 @@
#include "video_core/renderer_vulkan/renderer_vulkan.h" #include "video_core/renderer_vulkan/renderer_vulkan.h"
#include "video_core/renderer_vulkan/vk_blit_screen.h" #include "video_core/renderer_vulkan/vk_blit_screen.h"
#include "video_core/renderer_vulkan/vk_master_semaphore.h" #include "video_core/renderer_vulkan/vk_master_semaphore.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_shader_util.h" #include "video_core/renderer_vulkan/vk_shader_util.h"
#include "video_core/renderer_vulkan/vk_swapchain.h" #include "video_core/renderer_vulkan/vk_swapchain.h"
#include "video_core/surface.h" #include "video_core/surface.h"
#include "video_core/textures/decoders.h" #include "video_core/textures/decoders.h"
#include "video_core/vulkan_common/vulkan_device.h" #include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan { namespace Vulkan {
@ -115,10 +115,10 @@ struct VKBlitScreen::BufferData {
VKBlitScreen::VKBlitScreen(Core::Memory::Memory& cpu_memory_, VKBlitScreen::VKBlitScreen(Core::Memory::Memory& cpu_memory_,
Core::Frontend::EmuWindow& render_window_, Core::Frontend::EmuWindow& render_window_,
VideoCore::RasterizerInterface& rasterizer_, const Device& device_, VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
VKMemoryManager& memory_manager_, VKSwapchain& swapchain_, MemoryAllocator& memory_allocator_, VKSwapchain& swapchain_,
VKScheduler& scheduler_, const VKScreenInfo& screen_info_) VKScheduler& scheduler_, const VKScreenInfo& screen_info_)
: cpu_memory{cpu_memory_}, render_window{render_window_}, rasterizer{rasterizer_}, : cpu_memory{cpu_memory_}, render_window{render_window_}, rasterizer{rasterizer_},
device{device_}, memory_manager{memory_manager_}, swapchain{swapchain_}, device{device_}, memory_allocator{memory_allocator_}, swapchain{swapchain_},
scheduler{scheduler_}, image_count{swapchain.GetImageCount()}, screen_info{screen_info_} { scheduler{scheduler_}, image_count{swapchain.GetImageCount()}, screen_info{screen_info_} {
resource_ticks.resize(image_count); resource_ticks.resize(image_count);
@ -150,8 +150,8 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
SetUniformData(data, framebuffer); SetUniformData(data, framebuffer);
SetVertexData(data, framebuffer); SetVertexData(data, framebuffer);
auto map = buffer_commit->Map(); const std::span<u8> map = buffer_commit.Map();
std::memcpy(map.Address(), &data, sizeof(data)); std::memcpy(map.data(), &data, sizeof(data));
if (!use_accelerated) { if (!use_accelerated) {
const u64 image_offset = GetRawImageOffset(framebuffer, image_index); const u64 image_offset = GetRawImageOffset(framebuffer, image_index);
@ -165,8 +165,8 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
constexpr u32 block_height_log2 = 4; constexpr u32 block_height_log2 = 4;
const u32 bytes_per_pixel = GetBytesPerPixel(framebuffer); const u32 bytes_per_pixel = GetBytesPerPixel(framebuffer);
Tegra::Texture::UnswizzleTexture( Tegra::Texture::UnswizzleTexture(
std::span(map.Address() + image_offset, size_bytes), std::span(host_ptr, size_bytes), map.subspan(image_offset, size_bytes), std::span(host_ptr, size_bytes), bytes_per_pixel,
bytes_per_pixel, framebuffer.width, framebuffer.height, 1, block_height_log2, 0); framebuffer.width, framebuffer.height, 1, block_height_log2, 0);
const VkBufferImageCopy copy{ const VkBufferImageCopy copy{
.bufferOffset = image_offset, .bufferOffset = image_offset,
@ -224,8 +224,6 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, write_barrier); VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, write_barrier);
}); });
} }
map.Release();
scheduler.Record([renderpass = *renderpass, framebuffer = *framebuffers[image_index], scheduler.Record([renderpass = *renderpass, framebuffer = *framebuffers[image_index],
descriptor_set = descriptor_sets[image_index], buffer = *buffer, descriptor_set = descriptor_sets[image_index], buffer = *buffer,
size = swapchain.GetSize(), pipeline = *pipeline, size = swapchain.GetSize(), pipeline = *pipeline,
@ -642,7 +640,7 @@ void VKBlitScreen::ReleaseRawImages() {
raw_images.clear(); raw_images.clear();
raw_buffer_commits.clear(); raw_buffer_commits.clear();
buffer.reset(); buffer.reset();
buffer_commit.reset(); buffer_commit = MemoryCommit{};
} }
void VKBlitScreen::CreateStagingBuffer(const Tegra::FramebufferConfig& framebuffer) { void VKBlitScreen::CreateStagingBuffer(const Tegra::FramebufferConfig& framebuffer) {
@ -659,7 +657,7 @@ void VKBlitScreen::CreateStagingBuffer(const Tegra::FramebufferConfig& framebuff
}; };
buffer = device.GetLogical().CreateBuffer(ci); buffer = device.GetLogical().CreateBuffer(ci);
buffer_commit = memory_manager.Commit(buffer, true); buffer_commit = memory_allocator.Commit(buffer, MemoryUsage::Upload);
} }
void VKBlitScreen::CreateRawImages(const Tegra::FramebufferConfig& framebuffer) { void VKBlitScreen::CreateRawImages(const Tegra::FramebufferConfig& framebuffer) {
@ -690,7 +688,7 @@ void VKBlitScreen::CreateRawImages(const Tegra::FramebufferConfig& framebuffer)
.pQueueFamilyIndices = nullptr, .pQueueFamilyIndices = nullptr,
.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
}); });
raw_buffer_commits[i] = memory_manager.Commit(raw_images[i], false); raw_buffer_commits[i] = memory_allocator.Commit(raw_images[i], MemoryUsage::DeviceLocal);
raw_image_views[i] = device.GetLogical().CreateImageView(VkImageViewCreateInfo{ raw_image_views[i] = device.GetLogical().CreateImageView(VkImageViewCreateInfo{
.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO, .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
.pNext = nullptr, .pNext = nullptr,

View file

@ -6,7 +6,7 @@
#include <memory> #include <memory>
#include "video_core/renderer_vulkan/vk_memory_manager.h" #include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Core { namespace Core {
@ -43,7 +43,7 @@ public:
explicit VKBlitScreen(Core::Memory::Memory& cpu_memory, explicit VKBlitScreen(Core::Memory::Memory& cpu_memory,
Core::Frontend::EmuWindow& render_window, Core::Frontend::EmuWindow& render_window,
VideoCore::RasterizerInterface& rasterizer, const Device& device, VideoCore::RasterizerInterface& rasterizer, const Device& device,
VKMemoryManager& memory_manager, VKSwapchain& swapchain, MemoryAllocator& memory_allocator, VKSwapchain& swapchain,
VKScheduler& scheduler, const VKScreenInfo& screen_info); VKScheduler& scheduler, const VKScreenInfo& screen_info);
~VKBlitScreen(); ~VKBlitScreen();
@ -86,7 +86,7 @@ private:
Core::Frontend::EmuWindow& render_window; Core::Frontend::EmuWindow& render_window;
VideoCore::RasterizerInterface& rasterizer; VideoCore::RasterizerInterface& rasterizer;
const Device& device; const Device& device;
VKMemoryManager& memory_manager; MemoryAllocator& memory_allocator;
VKSwapchain& swapchain; VKSwapchain& swapchain;
VKScheduler& scheduler; VKScheduler& scheduler;
const std::size_t image_count; const std::size_t image_count;
@ -104,14 +104,14 @@ private:
vk::Sampler sampler; vk::Sampler sampler;
vk::Buffer buffer; vk::Buffer buffer;
VKMemoryCommit buffer_commit; MemoryCommit buffer_commit;
std::vector<u64> resource_ticks; std::vector<u64> resource_ticks;
std::vector<vk::Semaphore> semaphores; std::vector<vk::Semaphore> semaphores;
std::vector<vk::Image> raw_images; std::vector<vk::Image> raw_images;
std::vector<vk::ImageView> raw_image_views; std::vector<vk::ImageView> raw_image_views;
std::vector<VKMemoryCommit> raw_buffer_commits; std::vector<MemoryCommit> raw_buffer_commits;
u32 raw_width = 0; u32 raw_width = 0;
u32 raw_height = 0; u32 raw_height = 0;
}; };

View file

@ -36,11 +36,11 @@ constexpr VkAccessFlags TRANSFORM_FEEDBACK_WRITE_ACCESS =
} // Anonymous namespace } // Anonymous namespace
Buffer::Buffer(const Device& device_, VKMemoryManager& memory_manager, VKScheduler& scheduler_, Buffer::Buffer(const Device& device_, MemoryAllocator& memory_allocator, VKScheduler& scheduler_,
VKStagingBufferPool& staging_pool_, VAddr cpu_addr_, std::size_t size_) StagingBufferPool& staging_pool_, VAddr cpu_addr_, std::size_t size_)
: BufferBlock{cpu_addr_, size_}, device{device_}, scheduler{scheduler_}, staging_pool{ : BufferBlock{cpu_addr_, size_}, device{device_}, scheduler{scheduler_}, staging_pool{
staging_pool_} { staging_pool_} {
const VkBufferCreateInfo ci{ buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO, .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
.pNext = nullptr, .pNext = nullptr,
.flags = 0, .flags = 0,
@ -49,22 +49,20 @@ Buffer::Buffer(const Device& device_, VKMemoryManager& memory_manager, VKSchedul
.sharingMode = VK_SHARING_MODE_EXCLUSIVE, .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
.queueFamilyIndexCount = 0, .queueFamilyIndexCount = 0,
.pQueueFamilyIndices = nullptr, .pQueueFamilyIndices = nullptr,
}; });
commit = memory_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
buffer.handle = device.GetLogical().CreateBuffer(ci);
buffer.commit = memory_manager.Commit(buffer.handle, false);
} }
Buffer::~Buffer() = default; Buffer::~Buffer() = default;
void Buffer::Upload(std::size_t offset, std::size_t data_size, const u8* data) { void Buffer::Upload(std::size_t offset, std::size_t data_size, const u8* data) {
const auto& staging = staging_pool.GetUnusedBuffer(data_size, true); const auto& staging = staging_pool.Request(data_size, MemoryUsage::Upload);
std::memcpy(staging.commit->Map(data_size), data, data_size); std::memcpy(staging.mapped_span.data(), data, data_size);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
const VkBuffer handle = Handle(); const VkBuffer handle = Handle();
scheduler.Record([staging = *staging.handle, handle, offset, data_size, scheduler.Record([staging = staging.buffer, handle, offset, data_size,
&device = device](vk::CommandBuffer cmdbuf) { &device = device](vk::CommandBuffer cmdbuf) {
const VkBufferMemoryBarrier read_barrier{ const VkBufferMemoryBarrier read_barrier{
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER, .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
@ -100,12 +98,12 @@ void Buffer::Upload(std::size_t offset, std::size_t data_size, const u8* data) {
} }
void Buffer::Download(std::size_t offset, std::size_t data_size, u8* data) { void Buffer::Download(std::size_t offset, std::size_t data_size, u8* data) {
const auto& staging = staging_pool.GetUnusedBuffer(data_size, true); auto staging = staging_pool.Request(data_size, MemoryUsage::Download);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
const VkBuffer handle = Handle(); const VkBuffer handle = Handle();
scheduler.Record( scheduler.Record(
[staging = *staging.handle, handle, offset, data_size](vk::CommandBuffer cmdbuf) { [staging = staging.buffer, handle, offset, data_size](vk::CommandBuffer cmdbuf) {
const VkBufferMemoryBarrier barrier{ const VkBufferMemoryBarrier barrier{
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER, .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
.pNext = nullptr, .pNext = nullptr,
@ -126,7 +124,7 @@ void Buffer::Download(std::size_t offset, std::size_t data_size, u8* data) {
}); });
scheduler.Finish(); scheduler.Finish();
std::memcpy(data, staging.commit->Map(data_size), data_size); std::memcpy(data, staging.mapped_span.data(), data_size);
} }
void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset, void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
@ -164,29 +162,29 @@ void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst
VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer_, VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer_,
Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_, Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
const Device& device_, VKMemoryManager& memory_manager_, const Device& device_, MemoryAllocator& memory_allocator_,
VKScheduler& scheduler_, VKStreamBuffer& stream_buffer_, VKScheduler& scheduler_, VKStreamBuffer& stream_buffer_,
VKStagingBufferPool& staging_pool_) StagingBufferPool& staging_pool_)
: VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer>{rasterizer_, gpu_memory_, : VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer>{rasterizer_, gpu_memory_,
cpu_memory_, stream_buffer_}, cpu_memory_, stream_buffer_},
device{device_}, memory_manager{memory_manager_}, scheduler{scheduler_}, staging_pool{ device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
staging_pool_} {} staging_pool{staging_pool_} {}
VKBufferCache::~VKBufferCache() = default; VKBufferCache::~VKBufferCache() = default;
std::shared_ptr<Buffer> VKBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) { std::shared_ptr<Buffer> VKBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
return std::make_shared<Buffer>(device, memory_manager, scheduler, staging_pool, cpu_addr, return std::make_shared<Buffer>(device, memory_allocator, scheduler, staging_pool, cpu_addr,
size); size);
} }
VKBufferCache::BufferInfo VKBufferCache::GetEmptyBuffer(std::size_t size) { VKBufferCache::BufferInfo VKBufferCache::GetEmptyBuffer(std::size_t size) {
size = std::max(size, std::size_t(4)); size = std::max(size, std::size_t(4));
const auto& empty = staging_pool.GetUnusedBuffer(size, false); const auto& empty = staging_pool.Request(size, MemoryUsage::DeviceLocal);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([size, buffer = *empty.handle](vk::CommandBuffer cmdbuf) { scheduler.Record([size, buffer = empty.buffer](vk::CommandBuffer cmdbuf) {
cmdbuf.FillBuffer(buffer, 0, size, 0); cmdbuf.FillBuffer(buffer, 0, size, 0);
}); });
return {*empty.handle, 0, 0}; return {empty.buffer, 0, 0};
} }
} // namespace Vulkan } // namespace Vulkan

View file

@ -8,21 +8,20 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/buffer_cache/buffer_cache.h" #include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h" #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
#include "video_core/renderer_vulkan/vk_stream_buffer.h" #include "video_core/renderer_vulkan/vk_stream_buffer.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan { namespace Vulkan {
class Device; class Device;
class VKMemoryManager;
class VKScheduler; class VKScheduler;
class Buffer final : public VideoCommon::BufferBlock { class Buffer final : public VideoCommon::BufferBlock {
public: public:
explicit Buffer(const Device& device, VKMemoryManager& memory_manager, VKScheduler& scheduler, explicit Buffer(const Device& device, MemoryAllocator& memory_allocator, VKScheduler& scheduler,
VKStagingBufferPool& staging_pool, VAddr cpu_addr_, std::size_t size_); StagingBufferPool& staging_pool, VAddr cpu_addr_, std::size_t size_);
~Buffer(); ~Buffer();
void Upload(std::size_t offset, std::size_t data_size, const u8* data); void Upload(std::size_t offset, std::size_t data_size, const u8* data);
@ -33,7 +32,7 @@ public:
std::size_t copy_size); std::size_t copy_size);
VkBuffer Handle() const { VkBuffer Handle() const {
return *buffer.handle; return *buffer;
} }
u64 Address() const { u64 Address() const {
@ -43,18 +42,19 @@ public:
private: private:
const Device& device; const Device& device;
VKScheduler& scheduler; VKScheduler& scheduler;
VKStagingBufferPool& staging_pool; StagingBufferPool& staging_pool;
VKBuffer buffer; vk::Buffer buffer;
MemoryCommit commit;
}; };
class VKBufferCache final : public VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer> { class VKBufferCache final : public VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer> {
public: public:
explicit VKBufferCache(VideoCore::RasterizerInterface& rasterizer, explicit VKBufferCache(VideoCore::RasterizerInterface& rasterizer,
Tegra::MemoryManager& gpu_memory, Core::Memory::Memory& cpu_memory, Tegra::MemoryManager& gpu_memory, Core::Memory::Memory& cpu_memory,
const Device& device, VKMemoryManager& memory_manager, const Device& device, MemoryAllocator& memory_allocator,
VKScheduler& scheduler, VKStreamBuffer& stream_buffer, VKScheduler& scheduler, VKStreamBuffer& stream_buffer,
VKStagingBufferPool& staging_pool); StagingBufferPool& staging_pool);
~VKBufferCache(); ~VKBufferCache();
BufferInfo GetEmptyBuffer(std::size_t size) override; BufferInfo GetEmptyBuffer(std::size_t size) override;
@ -64,9 +64,9 @@ protected:
private: private:
const Device& device; const Device& device;
VKMemoryManager& memory_manager; MemoryAllocator& memory_allocator;
VKScheduler& scheduler; VKScheduler& scheduler;
VKStagingBufferPool& staging_pool; StagingBufferPool& staging_pool;
}; };
} // namespace Vulkan } // namespace Vulkan

View file

@ -164,7 +164,7 @@ VkDescriptorSet VKComputePass::CommitDescriptorSet(
QuadArrayPass::QuadArrayPass(const Device& device_, VKScheduler& scheduler_, QuadArrayPass::QuadArrayPass(const Device& device_, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool_, VKDescriptorPool& descriptor_pool_,
VKStagingBufferPool& staging_buffer_pool_, StagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_) VKUpdateDescriptorQueue& update_descriptor_queue_)
: VKComputePass(device_, descriptor_pool_, BuildQuadArrayPassDescriptorSetLayoutBinding(), : VKComputePass(device_, descriptor_pool_, BuildQuadArrayPassDescriptorSetLayoutBinding(),
BuildQuadArrayPassDescriptorUpdateTemplateEntry(), BuildQuadArrayPassDescriptorUpdateTemplateEntry(),
@ -177,18 +177,18 @@ QuadArrayPass::~QuadArrayPass() = default;
std::pair<VkBuffer, VkDeviceSize> QuadArrayPass::Assemble(u32 num_vertices, u32 first) { std::pair<VkBuffer, VkDeviceSize> QuadArrayPass::Assemble(u32 num_vertices, u32 first) {
const u32 num_triangle_vertices = (num_vertices / 4) * 6; const u32 num_triangle_vertices = (num_vertices / 4) * 6;
const std::size_t staging_size = num_triangle_vertices * sizeof(u32); const std::size_t staging_size = num_triangle_vertices * sizeof(u32);
auto& buffer = staging_buffer_pool.GetUnusedBuffer(staging_size, false); const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
update_descriptor_queue.Acquire(); update_descriptor_queue.Acquire();
update_descriptor_queue.AddBuffer(*buffer.handle, 0, staging_size); update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue); const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
ASSERT(num_vertices % 4 == 0); ASSERT(num_vertices % 4 == 0);
const u32 num_quads = num_vertices / 4; const u32 num_quads = num_vertices / 4;
scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = *buffer.handle, num_quads, scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer,
first, set](vk::CommandBuffer cmdbuf) { num_quads, first, set](vk::CommandBuffer cmdbuf) {
constexpr u32 dispatch_size = 1024; constexpr u32 dispatch_size = 1024;
cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline); cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {}); cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {});
@ -208,11 +208,11 @@ std::pair<VkBuffer, VkDeviceSize> QuadArrayPass::Assemble(u32 num_vertices, u32
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, {barrier}, {}); VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, {barrier}, {});
}); });
return {*buffer.handle, 0}; return {staging_ref.buffer, 0};
} }
Uint8Pass::Uint8Pass(const Device& device, VKScheduler& scheduler_, Uint8Pass::Uint8Pass(const Device& device, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool, VKStagingBufferPool& staging_buffer_pool_, VKDescriptorPool& descriptor_pool, StagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_) VKUpdateDescriptorQueue& update_descriptor_queue_)
: VKComputePass(device, descriptor_pool, BuildInputOutputDescriptorSetBindings(), : VKComputePass(device, descriptor_pool, BuildInputOutputDescriptorSetBindings(),
BuildInputOutputDescriptorUpdateTemplate(), {}, VULKAN_UINT8_COMP_SPV), BuildInputOutputDescriptorUpdateTemplate(), {}, VULKAN_UINT8_COMP_SPV),
@ -224,15 +224,15 @@ Uint8Pass::~Uint8Pass() = default;
std::pair<VkBuffer, u64> Uint8Pass::Assemble(u32 num_vertices, VkBuffer src_buffer, std::pair<VkBuffer, u64> Uint8Pass::Assemble(u32 num_vertices, VkBuffer src_buffer,
u64 src_offset) { u64 src_offset) {
const u32 staging_size = static_cast<u32>(num_vertices * sizeof(u16)); const u32 staging_size = static_cast<u32>(num_vertices * sizeof(u16));
auto& buffer = staging_buffer_pool.GetUnusedBuffer(staging_size, false); const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
update_descriptor_queue.Acquire(); update_descriptor_queue.Acquire();
update_descriptor_queue.AddBuffer(src_buffer, src_offset, num_vertices); update_descriptor_queue.AddBuffer(src_buffer, src_offset, num_vertices);
update_descriptor_queue.AddBuffer(*buffer.handle, 0, staging_size); update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue); const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = *buffer.handle, set, scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer, set,
num_vertices](vk::CommandBuffer cmdbuf) { num_vertices](vk::CommandBuffer cmdbuf) {
constexpr u32 dispatch_size = 1024; constexpr u32 dispatch_size = 1024;
cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline); cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
@ -252,12 +252,12 @@ std::pair<VkBuffer, u64> Uint8Pass::Assemble(u32 num_vertices, VkBuffer src_buff
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {}); VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {});
}); });
return {*buffer.handle, 0}; return {staging_ref.buffer, 0};
} }
QuadIndexedPass::QuadIndexedPass(const Device& device_, VKScheduler& scheduler_, QuadIndexedPass::QuadIndexedPass(const Device& device_, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool_, VKDescriptorPool& descriptor_pool_,
VKStagingBufferPool& staging_buffer_pool_, StagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_) VKUpdateDescriptorQueue& update_descriptor_queue_)
: VKComputePass(device_, descriptor_pool_, BuildInputOutputDescriptorSetBindings(), : VKComputePass(device_, descriptor_pool_, BuildInputOutputDescriptorSetBindings(),
BuildInputOutputDescriptorUpdateTemplate(), BuildInputOutputDescriptorUpdateTemplate(),
@ -286,15 +286,15 @@ std::pair<VkBuffer, u64> QuadIndexedPass::Assemble(
const u32 num_tri_vertices = (num_vertices / 4) * 6; const u32 num_tri_vertices = (num_vertices / 4) * 6;
const std::size_t staging_size = num_tri_vertices * sizeof(u32); const std::size_t staging_size = num_tri_vertices * sizeof(u32);
auto& buffer = staging_buffer_pool.GetUnusedBuffer(staging_size, false); const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
update_descriptor_queue.Acquire(); update_descriptor_queue.Acquire();
update_descriptor_queue.AddBuffer(src_buffer, src_offset, input_size); update_descriptor_queue.AddBuffer(src_buffer, src_offset, input_size);
update_descriptor_queue.AddBuffer(*buffer.handle, 0, staging_size); update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue); const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = *buffer.handle, set, scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer, set,
num_tri_vertices, base_vertex, index_shift](vk::CommandBuffer cmdbuf) { num_tri_vertices, base_vertex, index_shift](vk::CommandBuffer cmdbuf) {
static constexpr u32 dispatch_size = 1024; static constexpr u32 dispatch_size = 1024;
const std::array push_constants = {base_vertex, index_shift}; const std::array push_constants = {base_vertex, index_shift};
@ -317,7 +317,7 @@ std::pair<VkBuffer, u64> QuadIndexedPass::Assemble(
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {}); VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {});
}); });
return {*buffer.handle, 0}; return {staging_ref.buffer, 0};
} }
} // namespace Vulkan } // namespace Vulkan

View file

@ -16,8 +16,8 @@
namespace Vulkan { namespace Vulkan {
class Device; class Device;
class StagingBufferPool;
class VKScheduler; class VKScheduler;
class VKStagingBufferPool;
class VKUpdateDescriptorQueue; class VKUpdateDescriptorQueue;
class VKComputePass { class VKComputePass {
@ -45,7 +45,7 @@ class QuadArrayPass final : public VKComputePass {
public: public:
explicit QuadArrayPass(const Device& device_, VKScheduler& scheduler_, explicit QuadArrayPass(const Device& device_, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool_, VKDescriptorPool& descriptor_pool_,
VKStagingBufferPool& staging_buffer_pool_, StagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_); VKUpdateDescriptorQueue& update_descriptor_queue_);
~QuadArrayPass(); ~QuadArrayPass();
@ -53,15 +53,14 @@ public:
private: private:
VKScheduler& scheduler; VKScheduler& scheduler;
VKStagingBufferPool& staging_buffer_pool; StagingBufferPool& staging_buffer_pool;
VKUpdateDescriptorQueue& update_descriptor_queue; VKUpdateDescriptorQueue& update_descriptor_queue;
}; };
class Uint8Pass final : public VKComputePass { class Uint8Pass final : public VKComputePass {
public: public:
explicit Uint8Pass(const Device& device_, VKScheduler& scheduler_, explicit Uint8Pass(const Device& device_, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool_, VKDescriptorPool& descriptor_pool_, StagingBufferPool& staging_buffer_pool_,
VKStagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_); VKUpdateDescriptorQueue& update_descriptor_queue_);
~Uint8Pass(); ~Uint8Pass();
@ -69,7 +68,7 @@ public:
private: private:
VKScheduler& scheduler; VKScheduler& scheduler;
VKStagingBufferPool& staging_buffer_pool; StagingBufferPool& staging_buffer_pool;
VKUpdateDescriptorQueue& update_descriptor_queue; VKUpdateDescriptorQueue& update_descriptor_queue;
}; };
@ -77,7 +76,7 @@ class QuadIndexedPass final : public VKComputePass {
public: public:
explicit QuadIndexedPass(const Device& device_, VKScheduler& scheduler_, explicit QuadIndexedPass(const Device& device_, VKScheduler& scheduler_,
VKDescriptorPool& descriptor_pool_, VKDescriptorPool& descriptor_pool_,
VKStagingBufferPool& staging_buffer_pool_, StagingBufferPool& staging_buffer_pool_,
VKUpdateDescriptorQueue& update_descriptor_queue_); VKUpdateDescriptorQueue& update_descriptor_queue_);
~QuadIndexedPass(); ~QuadIndexedPass();
@ -87,7 +86,7 @@ public:
private: private:
VKScheduler& scheduler; VKScheduler& scheduler;
VKStagingBufferPool& staging_buffer_pool; StagingBufferPool& staging_buffer_pool;
VKUpdateDescriptorQueue& update_descriptor_queue; VKUpdateDescriptorQueue& update_descriptor_queue;
}; };

View file

@ -1,230 +0,0 @@
// Copyright 2018 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <optional>
#include <tuple>
#include <vector>
#include "common/alignment.h"
#include "common/assert.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan {
namespace {
u64 GetAllocationChunkSize(u64 required_size) {
static constexpr u64 sizes[] = {16ULL << 20, 32ULL << 20, 64ULL << 20, 128ULL << 20};
auto it = std::lower_bound(std::begin(sizes), std::end(sizes), required_size);
return it != std::end(sizes) ? *it : Common::AlignUp(required_size, 256ULL << 20);
}
} // Anonymous namespace
class VKMemoryAllocation final {
public:
explicit VKMemoryAllocation(const Device& device_, vk::DeviceMemory memory_,
VkMemoryPropertyFlags properties_, u64 allocation_size_, u32 type_)
: device{device_}, memory{std::move(memory_)}, properties{properties_},
allocation_size{allocation_size_}, shifted_type{ShiftType(type_)} {}
VKMemoryCommit Commit(VkDeviceSize commit_size, VkDeviceSize alignment) {
auto found = TryFindFreeSection(free_iterator, allocation_size,
static_cast<u64>(commit_size), static_cast<u64>(alignment));
if (!found) {
found = TryFindFreeSection(0, free_iterator, static_cast<u64>(commit_size),
static_cast<u64>(alignment));
if (!found) {
// Signal out of memory, it'll try to do more allocations.
return nullptr;
}
}
auto commit = std::make_unique<VKMemoryCommitImpl>(device, this, memory, *found,
*found + commit_size);
commits.push_back(commit.get());
// Last commit's address is highly probable to be free.
free_iterator = *found + commit_size;
return commit;
}
void Free(const VKMemoryCommitImpl* commit) {
ASSERT(commit);
const auto it = std::find(std::begin(commits), std::end(commits), commit);
if (it == commits.end()) {
UNREACHABLE_MSG("Freeing unallocated commit!");
return;
}
commits.erase(it);
}
/// Returns whether this allocation is compatible with the arguments.
bool IsCompatible(VkMemoryPropertyFlags wanted_properties, u32 type_mask) const {
return (wanted_properties & properties) && (type_mask & shifted_type) != 0;
}
private:
static constexpr u32 ShiftType(u32 type) {
return 1U << type;
}
/// A memory allocator, it may return a free region between "start" and "end" with the solicited
/// requirements.
std::optional<u64> TryFindFreeSection(u64 start, u64 end, u64 size, u64 alignment) const {
u64 iterator = Common::AlignUp(start, alignment);
while (iterator + size <= end) {
const u64 try_left = iterator;
const u64 try_right = try_left + size;
bool overlap = false;
for (const auto& commit : commits) {
const auto [commit_left, commit_right] = commit->interval;
if (try_left < commit_right && commit_left < try_right) {
// There's an overlap, continue the search where the overlapping commit ends.
iterator = Common::AlignUp(commit_right, alignment);
overlap = true;
break;
}
}
if (!overlap) {
// A free address has been found.
return try_left;
}
}
// No free regions where found, return an empty optional.
return std::nullopt;
}
const Device& device; ///< Vulkan device.
const vk::DeviceMemory memory; ///< Vulkan memory allocation handler.
const VkMemoryPropertyFlags properties; ///< Vulkan properties.
const u64 allocation_size; ///< Size of this allocation.
const u32 shifted_type; ///< Stored Vulkan type of this allocation, shifted.
/// Hints where the next free region is likely going to be.
u64 free_iterator{};
/// Stores all commits done from this allocation.
std::vector<const VKMemoryCommitImpl*> commits;
};
VKMemoryManager::VKMemoryManager(const Device& device_)
: device{device_}, properties{device_.GetPhysical().GetMemoryProperties()} {}
VKMemoryManager::~VKMemoryManager() = default;
VKMemoryCommit VKMemoryManager::Commit(const VkMemoryRequirements& requirements,
bool host_visible) {
const u64 chunk_size = GetAllocationChunkSize(requirements.size);
// When a host visible commit is asked, search for host visible and coherent, otherwise search
// for a fast device local type.
const VkMemoryPropertyFlags wanted_properties =
host_visible ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
: VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
if (auto commit = TryAllocCommit(requirements, wanted_properties)) {
return commit;
}
// Commit has failed, allocate more memory.
if (!AllocMemory(wanted_properties, requirements.memoryTypeBits, chunk_size)) {
// TODO(Rodrigo): Handle these situations in some way like flushing to guest memory.
// Allocation has failed, panic.
UNREACHABLE_MSG("Ran out of VRAM!");
return {};
}
// Commit again, this time it won't fail since there's a fresh allocation above. If it does,
// there's a bug.
auto commit = TryAllocCommit(requirements, wanted_properties);
ASSERT(commit);
return commit;
}
VKMemoryCommit VKMemoryManager::Commit(const vk::Buffer& buffer, bool host_visible) {
auto commit = Commit(device.GetLogical().GetBufferMemoryRequirements(*buffer), host_visible);
buffer.BindMemory(commit->GetMemory(), commit->GetOffset());
return commit;
}
VKMemoryCommit VKMemoryManager::Commit(const vk::Image& image, bool host_visible) {
auto commit = Commit(device.GetLogical().GetImageMemoryRequirements(*image), host_visible);
image.BindMemory(commit->GetMemory(), commit->GetOffset());
return commit;
}
bool VKMemoryManager::AllocMemory(VkMemoryPropertyFlags wanted_properties, u32 type_mask,
u64 size) {
const u32 type = [&] {
for (u32 type_index = 0; type_index < properties.memoryTypeCount; ++type_index) {
const auto flags = properties.memoryTypes[type_index].propertyFlags;
if ((type_mask & (1U << type_index)) && (flags & wanted_properties)) {
// The type matches in type and in the wanted properties.
return type_index;
}
}
UNREACHABLE_MSG("Couldn't find a compatible memory type!");
return 0U;
}();
// Try to allocate found type.
vk::DeviceMemory memory = device.GetLogical().TryAllocateMemory({
.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
.pNext = nullptr,
.allocationSize = size,
.memoryTypeIndex = type,
});
if (!memory) {
LOG_CRITICAL(Render_Vulkan, "Device allocation failed!");
return false;
}
allocations.push_back(std::make_unique<VKMemoryAllocation>(device, std::move(memory),
wanted_properties, size, type));
return true;
}
VKMemoryCommit VKMemoryManager::TryAllocCommit(const VkMemoryRequirements& requirements,
VkMemoryPropertyFlags wanted_properties) {
for (auto& allocation : allocations) {
if (!allocation->IsCompatible(wanted_properties, requirements.memoryTypeBits)) {
continue;
}
if (auto commit = allocation->Commit(requirements.size, requirements.alignment)) {
return commit;
}
}
return {};
}
VKMemoryCommitImpl::VKMemoryCommitImpl(const Device& device_, VKMemoryAllocation* allocation_,
const vk::DeviceMemory& memory_, u64 begin_, u64 end_)
: device{device_}, memory{memory_}, interval{begin_, end_}, allocation{allocation_} {}
VKMemoryCommitImpl::~VKMemoryCommitImpl() {
allocation->Free(this);
}
MemoryMap VKMemoryCommitImpl::Map(u64 size, u64 offset_) const {
return MemoryMap(this, std::span<u8>(memory.Map(interval.first + offset_, size), size));
}
void VKMemoryCommitImpl::Unmap() const {
memory.Unmap();
}
MemoryMap VKMemoryCommitImpl::Map() const {
return Map(interval.second - interval.first);
}
} // namespace Vulkan

View file

@ -1,132 +0,0 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <memory>
#include <span>
#include <utility>
#include <vector>
#include "common/common_types.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan {
class Device;
class MemoryMap;
class VKMemoryAllocation;
class VKMemoryCommitImpl;
using VKMemoryCommit = std::unique_ptr<VKMemoryCommitImpl>;
class VKMemoryManager final {
public:
explicit VKMemoryManager(const Device& device_);
VKMemoryManager(const VKMemoryManager&) = delete;
~VKMemoryManager();
/**
* Commits a memory with the specified requeriments.
* @param requirements Requirements returned from a Vulkan call.
* @param host_visible Signals the allocator that it *must* use host visible and coherent
* memory. When passing false, it will try to allocate device local memory.
* @returns A memory commit.
*/
VKMemoryCommit Commit(const VkMemoryRequirements& requirements, bool host_visible);
/// Commits memory required by the buffer and binds it.
VKMemoryCommit Commit(const vk::Buffer& buffer, bool host_visible);
/// Commits memory required by the image and binds it.
VKMemoryCommit Commit(const vk::Image& image, bool host_visible);
private:
/// Allocates a chunk of memory.
bool AllocMemory(VkMemoryPropertyFlags wanted_properties, u32 type_mask, u64 size);
/// Tries to allocate a memory commit.
VKMemoryCommit TryAllocCommit(const VkMemoryRequirements& requirements,
VkMemoryPropertyFlags wanted_properties);
const Device& device; ///< Device handler.
const VkPhysicalDeviceMemoryProperties properties; ///< Physical device properties.
std::vector<std::unique_ptr<VKMemoryAllocation>> allocations; ///< Current allocations.
};
class VKMemoryCommitImpl final {
friend VKMemoryAllocation;
friend MemoryMap;
public:
explicit VKMemoryCommitImpl(const Device& device_, VKMemoryAllocation* allocation_,
const vk::DeviceMemory& memory_, u64 begin_, u64 end_);
~VKMemoryCommitImpl();
/// Maps a memory region and returns a pointer to it.
/// It's illegal to have more than one memory map at the same time.
MemoryMap Map(u64 size, u64 offset = 0) const;
/// Maps the whole commit and returns a pointer to it.
/// It's illegal to have more than one memory map at the same time.
MemoryMap Map() const;
/// Returns the Vulkan memory handler.
VkDeviceMemory GetMemory() const {
return *memory;
}
/// Returns the start position of the commit relative to the allocation.
VkDeviceSize GetOffset() const {
return static_cast<VkDeviceSize>(interval.first);
}
private:
/// Unmaps memory.
void Unmap() const;
const Device& device; ///< Vulkan device.
const vk::DeviceMemory& memory; ///< Vulkan device memory handler.
std::pair<u64, u64> interval{}; ///< Interval where the commit exists.
VKMemoryAllocation* allocation{}; ///< Pointer to the large memory allocation.
};
/// Holds ownership of a memory map.
class MemoryMap final {
public:
explicit MemoryMap(const VKMemoryCommitImpl* commit_, std::span<u8> span_)
: commit{commit_}, span{span_} {}
~MemoryMap() {
if (commit) {
commit->Unmap();
}
}
/// Prematurely releases the memory map.
void Release() {
commit->Unmap();
commit = nullptr;
}
/// Returns a span to the memory map.
[[nodiscard]] std::span<u8> Span() const noexcept {
return span;
}
/// Returns the address of the memory map.
[[nodiscard]] u8* Address() const noexcept {
return span.data();
}
/// Returns the address of the memory map;
[[nodiscard]] operator u8*() const noexcept {
return span.data();
}
private:
const VKMemoryCommitImpl* commit{}; ///< Mapped memory commit.
std::span<u8> span; ///< Address to the mapped memory.
};
} // namespace Vulkan

View file

@ -409,24 +409,24 @@ void RasterizerVulkan::DrawParameters::Draw(vk::CommandBuffer cmdbuf) const {
RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_, RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_,
Tegra::MemoryManager& gpu_memory_, Tegra::MemoryManager& gpu_memory_,
Core::Memory::Memory& cpu_memory_, VKScreenInfo& screen_info_, Core::Memory::Memory& cpu_memory_, VKScreenInfo& screen_info_,
const Device& device_, VKMemoryManager& memory_manager_, const Device& device_, MemoryAllocator& memory_allocator_,
StateTracker& state_tracker_, VKScheduler& scheduler_) StateTracker& state_tracker_, VKScheduler& scheduler_)
: RasterizerAccelerated{cpu_memory_}, gpu{gpu_}, : RasterizerAccelerated{cpu_memory_}, gpu{gpu_},
gpu_memory{gpu_memory_}, maxwell3d{gpu.Maxwell3D()}, kepler_compute{gpu.KeplerCompute()}, gpu_memory{gpu_memory_}, maxwell3d{gpu.Maxwell3D()}, kepler_compute{gpu.KeplerCompute()},
screen_info{screen_info_}, device{device_}, memory_manager{memory_manager_}, screen_info{screen_info_}, device{device_}, memory_allocator{memory_allocator_},
state_tracker{state_tracker_}, scheduler{scheduler_}, stream_buffer(device, scheduler), state_tracker{state_tracker_}, scheduler{scheduler_}, stream_buffer(device, scheduler),
staging_pool(device, memory_manager, scheduler), descriptor_pool(device, scheduler), staging_pool(device, memory_allocator, scheduler), descriptor_pool(device, scheduler),
update_descriptor_queue(device, scheduler), update_descriptor_queue(device, scheduler),
blit_image(device, scheduler, state_tracker, descriptor_pool), blit_image(device, scheduler, state_tracker, descriptor_pool),
quad_array_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue), quad_array_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
quad_indexed_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue), quad_indexed_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
uint8_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue), uint8_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
texture_cache_runtime{device, scheduler, memory_manager, staging_pool, blit_image}, texture_cache_runtime{device, scheduler, memory_allocator, staging_pool, blit_image},
texture_cache(texture_cache_runtime, *this, maxwell3d, kepler_compute, gpu_memory), texture_cache(texture_cache_runtime, *this, maxwell3d, kepler_compute, gpu_memory),
pipeline_cache(*this, gpu, maxwell3d, kepler_compute, gpu_memory, device, scheduler, pipeline_cache(*this, gpu, maxwell3d, kepler_compute, gpu_memory, device, scheduler,
descriptor_pool, update_descriptor_queue), descriptor_pool, update_descriptor_queue),
buffer_cache(*this, gpu_memory, cpu_memory_, device, memory_manager, scheduler, stream_buffer, buffer_cache(*this, gpu_memory, cpu_memory_, device, memory_allocator, scheduler,
staging_pool), stream_buffer, staging_pool),
query_cache{*this, maxwell3d, gpu_memory, device, scheduler}, query_cache{*this, maxwell3d, gpu_memory, device, scheduler},
fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, scheduler), fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, scheduler),
wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) { wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) {
@ -1445,7 +1445,7 @@ VkBuffer RasterizerVulkan::DefaultBuffer() {
.queueFamilyIndexCount = 0, .queueFamilyIndexCount = 0,
.pQueueFamilyIndices = nullptr, .pQueueFamilyIndices = nullptr,
}); });
default_buffer_commit = memory_manager.Commit(default_buffer, false); default_buffer_commit = memory_allocator.Commit(default_buffer, MemoryUsage::DeviceLocal);
scheduler.RequestOutsideRenderPassOperationContext(); scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([buffer = *default_buffer](vk::CommandBuffer cmdbuf) { scheduler.Record([buffer = *default_buffer](vk::CommandBuffer cmdbuf) {

View file

@ -21,7 +21,6 @@
#include "video_core/renderer_vulkan/vk_compute_pass.h" #include "video_core/renderer_vulkan/vk_compute_pass.h"
#include "video_core/renderer_vulkan/vk_descriptor_pool.h" #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
#include "video_core/renderer_vulkan/vk_fence_manager.h" #include "video_core/renderer_vulkan/vk_fence_manager.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_pipeline_cache.h" #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
#include "video_core/renderer_vulkan/vk_query_cache.h" #include "video_core/renderer_vulkan/vk_query_cache.h"
#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_scheduler.h"
@ -30,6 +29,7 @@
#include "video_core/renderer_vulkan/vk_texture_cache.h" #include "video_core/renderer_vulkan/vk_texture_cache.h"
#include "video_core/renderer_vulkan/vk_update_descriptor.h" #include "video_core/renderer_vulkan/vk_update_descriptor.h"
#include "video_core/shader/async_shaders.h" #include "video_core/shader/async_shaders.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Core { namespace Core {
@ -56,7 +56,7 @@ public:
explicit RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_, explicit RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_,
Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_, Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
VKScreenInfo& screen_info_, const Device& device_, VKScreenInfo& screen_info_, const Device& device_,
VKMemoryManager& memory_manager_, StateTracker& state_tracker_, MemoryAllocator& memory_allocator_, StateTracker& state_tracker_,
VKScheduler& scheduler_); VKScheduler& scheduler_);
~RasterizerVulkan() override; ~RasterizerVulkan() override;
@ -213,12 +213,12 @@ private:
VKScreenInfo& screen_info; VKScreenInfo& screen_info;
const Device& device; const Device& device;
VKMemoryManager& memory_manager; MemoryAllocator& memory_allocator;
StateTracker& state_tracker; StateTracker& state_tracker;
VKScheduler& scheduler; VKScheduler& scheduler;
VKStreamBuffer stream_buffer; VKStreamBuffer stream_buffer;
VKStagingBufferPool staging_pool; StagingBufferPool staging_pool;
VKDescriptorPool descriptor_pool; VKDescriptorPool descriptor_pool;
VKUpdateDescriptorQueue update_descriptor_queue; VKUpdateDescriptorQueue update_descriptor_queue;
BlitImageHelper blit_image; BlitImageHelper blit_image;
@ -234,7 +234,7 @@ private:
VKFenceManager fence_manager; VKFenceManager fence_manager;
vk::Buffer default_buffer; vk::Buffer default_buffer;
VKMemoryCommit default_buffer_commit; MemoryCommit default_buffer_commit;
vk::Event wfi_event; vk::Event wfi_event;
VideoCommon::Shader::AsyncShaders async_shaders; VideoCommon::Shader::AsyncShaders async_shaders;

View file

@ -3,10 +3,12 @@
// Refer to the license.txt file included. // Refer to the license.txt file included.
#include <algorithm> #include <algorithm>
#include <unordered_map>
#include <utility> #include <utility>
#include <vector> #include <vector>
#include <fmt/format.h>
#include "common/assert.h"
#include "common/bit_util.h" #include "common/bit_util.h"
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_scheduler.h"
@ -16,45 +18,51 @@
namespace Vulkan { namespace Vulkan {
VKStagingBufferPool::StagingBuffer::StagingBuffer(std::unique_ptr<VKBuffer> buffer_) StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& memory_allocator_,
: buffer{std::move(buffer_)} {}
VKStagingBufferPool::VKStagingBufferPool(const Device& device_, VKMemoryManager& memory_manager_,
VKScheduler& scheduler_) VKScheduler& scheduler_)
: device{device_}, memory_manager{memory_manager_}, scheduler{scheduler_} {} : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_} {}
VKStagingBufferPool::~VKStagingBufferPool() = default; StagingBufferPool::~StagingBufferPool() = default;
VKBuffer& VKStagingBufferPool::GetUnusedBuffer(std::size_t size, bool host_visible) { StagingBufferRef StagingBufferPool::Request(size_t size, MemoryUsage usage) {
if (const auto buffer = TryGetReservedBuffer(size, host_visible)) { if (const std::optional<StagingBufferRef> ref = TryGetReservedBuffer(size, usage)) {
return *buffer; return *ref;
} }
return CreateStagingBuffer(size, host_visible); return CreateStagingBuffer(size, usage);
} }
void VKStagingBufferPool::TickFrame() { void StagingBufferPool::TickFrame() {
current_delete_level = (current_delete_level + 1) % NumLevels; current_delete_level = (current_delete_level + 1) % NUM_LEVELS;
ReleaseCache(true); ReleaseCache(MemoryUsage::DeviceLocal);
ReleaseCache(false); ReleaseCache(MemoryUsage::Upload);
ReleaseCache(MemoryUsage::Download);
} }
VKBuffer* VKStagingBufferPool::TryGetReservedBuffer(std::size_t size, bool host_visible) { std::optional<StagingBufferRef> StagingBufferPool::TryGetReservedBuffer(size_t size,
for (StagingBuffer& entry : GetCache(host_visible)[Common::Log2Ceil64(size)].entries) { MemoryUsage usage) {
if (!scheduler.IsFree(entry.tick)) { StagingBuffers& cache_level = GetCache(usage)[Common::Log2Ceil64(size)];
continue;
const auto is_free = [this](const StagingBuffer& entry) {
return scheduler.IsFree(entry.tick);
};
auto& entries = cache_level.entries;
const auto hint_it = entries.begin() + cache_level.iterate_index;
auto it = std::find_if(entries.begin() + cache_level.iterate_index, entries.end(), is_free);
if (it == entries.end()) {
it = std::find_if(entries.begin(), hint_it, is_free);
if (it == hint_it) {
return std::nullopt;
} }
entry.tick = scheduler.CurrentTick();
return &*entry.buffer;
} }
return nullptr; cache_level.iterate_index = std::distance(entries.begin(), it) + 1;
it->tick = scheduler.CurrentTick();
return it->Ref();
} }
VKBuffer& VKStagingBufferPool::CreateStagingBuffer(std::size_t size, bool host_visible) { StagingBufferRef StagingBufferPool::CreateStagingBuffer(size_t size, MemoryUsage usage) {
const u32 log2 = Common::Log2Ceil64(size); const u32 log2 = Common::Log2Ceil64(size);
vk::Buffer buffer = device.GetLogical().CreateBuffer({
auto buffer = std::make_unique<VKBuffer>();
buffer->handle = device.GetLogical().CreateBuffer({
.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO, .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
.pNext = nullptr, .pNext = nullptr,
.flags = 0, .flags = 0,
@ -66,49 +74,63 @@ VKBuffer& VKStagingBufferPool::CreateStagingBuffer(std::size_t size, bool host_v
.queueFamilyIndexCount = 0, .queueFamilyIndexCount = 0,
.pQueueFamilyIndices = nullptr, .pQueueFamilyIndices = nullptr,
}); });
buffer->commit = memory_manager.Commit(buffer->handle, host_visible); if (device.HasDebuggingToolAttached()) {
++buffer_index;
buffer.SetObjectNameEXT(fmt::format("Staging Buffer {}", buffer_index).c_str());
}
MemoryCommit commit = memory_allocator.Commit(buffer, usage);
const std::span<u8> mapped_span = IsHostVisible(usage) ? commit.Map() : std::span<u8>{};
std::vector<StagingBuffer>& entries = GetCache(host_visible)[log2].entries; StagingBuffer& entry = GetCache(usage)[log2].entries.emplace_back(StagingBuffer{
StagingBuffer& entry = entries.emplace_back(std::move(buffer)); .buffer = std::move(buffer),
entry.tick = scheduler.CurrentTick(); .commit = std::move(commit),
return *entry.buffer; .mapped_span = mapped_span,
.tick = scheduler.CurrentTick(),
});
return entry.Ref();
} }
VKStagingBufferPool::StagingBuffersCache& VKStagingBufferPool::GetCache(bool host_visible) { StagingBufferPool::StagingBuffersCache& StagingBufferPool::GetCache(MemoryUsage usage) {
return host_visible ? host_staging_buffers : device_staging_buffers; switch (usage) {
} case MemoryUsage::DeviceLocal:
return device_local_cache;
void VKStagingBufferPool::ReleaseCache(bool host_visible) { case MemoryUsage::Upload:
auto& cache = GetCache(host_visible); return upload_cache;
const u64 size = ReleaseLevel(cache, current_delete_level); case MemoryUsage::Download:
if (size == 0) { return download_cache;
return; default:
UNREACHABLE_MSG("Invalid memory usage={}", usage);
return upload_cache;
} }
} }
u64 VKStagingBufferPool::ReleaseLevel(StagingBuffersCache& cache, std::size_t log2) { void StagingBufferPool::ReleaseCache(MemoryUsage usage) {
static constexpr std::size_t deletions_per_tick = 16; ReleaseLevel(GetCache(usage), current_delete_level);
}
void StagingBufferPool::ReleaseLevel(StagingBuffersCache& cache, size_t log2) {
constexpr size_t deletions_per_tick = 16;
auto& staging = cache[log2]; auto& staging = cache[log2];
auto& entries = staging.entries; auto& entries = staging.entries;
const std::size_t old_size = entries.size(); const size_t old_size = entries.size();
const auto is_deleteable = [this](const StagingBuffer& entry) { const auto is_deleteable = [this](const StagingBuffer& entry) {
return scheduler.IsFree(entry.tick); return scheduler.IsFree(entry.tick);
}; };
const std::size_t begin_offset = staging.delete_index; const size_t begin_offset = staging.delete_index;
const std::size_t end_offset = std::min(begin_offset + deletions_per_tick, old_size); const size_t end_offset = std::min(begin_offset + deletions_per_tick, old_size);
const auto begin = std::begin(entries) + begin_offset; const auto begin = entries.begin() + begin_offset;
const auto end = std::begin(entries) + end_offset; const auto end = entries.begin() + end_offset;
entries.erase(std::remove_if(begin, end, is_deleteable), end); entries.erase(std::remove_if(begin, end, is_deleteable), end);
const std::size_t new_size = entries.size(); const size_t new_size = entries.size();
staging.delete_index += deletions_per_tick; staging.delete_index += deletions_per_tick;
if (staging.delete_index >= new_size) { if (staging.delete_index >= new_size) {
staging.delete_index = 0; staging.delete_index = 0;
} }
if (staging.iterate_index > new_size) {
return (1ULL << log2) * (old_size - new_size); staging.iterate_index = 0;
}
} }
} // namespace Vulkan } // namespace Vulkan

View file

@ -9,7 +9,7 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h" #include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan { namespace Vulkan {
@ -17,55 +17,65 @@ namespace Vulkan {
class Device; class Device;
class VKScheduler; class VKScheduler;
struct VKBuffer final { struct StagingBufferRef {
vk::Buffer handle; VkBuffer buffer;
VKMemoryCommit commit; std::span<u8> mapped_span;
}; };
class VKStagingBufferPool final { class StagingBufferPool {
public: public:
explicit VKStagingBufferPool(const Device& device, VKMemoryManager& memory_manager, explicit StagingBufferPool(const Device& device, MemoryAllocator& memory_allocator,
VKScheduler& scheduler); VKScheduler& scheduler);
~VKStagingBufferPool(); ~StagingBufferPool();
VKBuffer& GetUnusedBuffer(std::size_t size, bool host_visible); StagingBufferRef Request(size_t size, MemoryUsage usage);
void TickFrame(); void TickFrame();
private: private:
struct StagingBuffer final { struct StagingBuffer {
explicit StagingBuffer(std::unique_ptr<VKBuffer> buffer); vk::Buffer buffer;
MemoryCommit commit;
std::unique_ptr<VKBuffer> buffer; std::span<u8> mapped_span;
u64 tick = 0; u64 tick = 0;
StagingBufferRef Ref() const noexcept {
return {
.buffer = *buffer,
.mapped_span = mapped_span,
};
}
}; };
struct StagingBuffers final { struct StagingBuffers {
std::vector<StagingBuffer> entries; std::vector<StagingBuffer> entries;
std::size_t delete_index = 0; size_t delete_index = 0;
size_t iterate_index = 0;
}; };
static constexpr std::size_t NumLevels = sizeof(std::size_t) * CHAR_BIT; static constexpr size_t NUM_LEVELS = sizeof(size_t) * CHAR_BIT;
using StagingBuffersCache = std::array<StagingBuffers, NumLevels>; using StagingBuffersCache = std::array<StagingBuffers, NUM_LEVELS>;
VKBuffer* TryGetReservedBuffer(std::size_t size, bool host_visible); std::optional<StagingBufferRef> TryGetReservedBuffer(size_t size, MemoryUsage usage);
VKBuffer& CreateStagingBuffer(std::size_t size, bool host_visible); StagingBufferRef CreateStagingBuffer(size_t size, MemoryUsage usage);
StagingBuffersCache& GetCache(bool host_visible); StagingBuffersCache& GetCache(MemoryUsage usage);
void ReleaseCache(bool host_visible); void ReleaseCache(MemoryUsage usage);
u64 ReleaseLevel(StagingBuffersCache& cache, std::size_t log2); void ReleaseLevel(StagingBuffersCache& cache, size_t log2);
const Device& device; const Device& device;
VKMemoryManager& memory_manager; MemoryAllocator& memory_allocator;
VKScheduler& scheduler; VKScheduler& scheduler;
StagingBuffersCache host_staging_buffers; StagingBuffersCache device_local_cache;
StagingBuffersCache device_staging_buffers; StagingBuffersCache upload_cache;
StagingBuffersCache download_cache;
std::size_t current_delete_level = 0; size_t current_delete_level = 0;
u64 buffer_index = 0;
}; };
} // namespace Vulkan } // namespace Vulkan

View file

@ -10,12 +10,12 @@
#include "video_core/engines/fermi_2d.h" #include "video_core/engines/fermi_2d.h"
#include "video_core/renderer_vulkan/blit_image.h" #include "video_core/renderer_vulkan/blit_image.h"
#include "video_core/renderer_vulkan/maxwell_to_vk.h" #include "video_core/renderer_vulkan/maxwell_to_vk.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_rasterizer.h" #include "video_core/renderer_vulkan/vk_rasterizer.h"
#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h" #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
#include "video_core/renderer_vulkan/vk_texture_cache.h" #include "video_core/renderer_vulkan/vk_texture_cache.h"
#include "video_core/vulkan_common/vulkan_device.h" #include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan { namespace Vulkan {
@ -554,10 +554,18 @@ void TextureCacheRuntime::Finish() {
} }
ImageBufferMap TextureCacheRuntime::MapUploadBuffer(size_t size) { ImageBufferMap TextureCacheRuntime::MapUploadBuffer(size_t size) {
const auto& buffer = staging_buffer_pool.GetUnusedBuffer(size, true); const auto staging_ref = staging_buffer_pool.Request(size, MemoryUsage::Upload);
return ImageBufferMap{ return {
.handle = *buffer.handle, .handle = staging_ref.buffer,
.map = buffer.commit->Map(size), .span = staging_ref.mapped_span,
};
}
ImageBufferMap TextureCacheRuntime::MapDownloadBuffer(size_t size) {
const auto staging_ref = staging_buffer_pool.Request(size, MemoryUsage::Download);
return {
.handle = staging_ref.buffer,
.span = staging_ref.mapped_span,
}; };
} }
@ -788,9 +796,9 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
image(MakeImage(runtime.device, info)), buffer(MakeBuffer(runtime.device, info)), image(MakeImage(runtime.device, info)), buffer(MakeBuffer(runtime.device, info)),
aspect_mask(ImageAspectMask(info.format)) { aspect_mask(ImageAspectMask(info.format)) {
if (image) { if (image) {
commit = runtime.memory_manager.Commit(image, false); commit = runtime.memory_allocator.Commit(image, MemoryUsage::DeviceLocal);
} else { } else {
commit = runtime.memory_manager.Commit(buffer, false); commit = runtime.memory_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
} }
if (IsPixelFormatASTC(info.format) && !runtime.device.IsOptimalAstcSupported()) { if (IsPixelFormatASTC(info.format) && !runtime.device.IsOptimalAstcSupported()) {
flags |= VideoCommon::ImageFlagBits::Converted; flags |= VideoCommon::ImageFlagBits::Converted;

View file

@ -7,8 +7,8 @@
#include <compare> #include <compare>
#include <span> #include <span>
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/texture_cache/texture_cache.h" #include "video_core/texture_cache/texture_cache.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h" #include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan { namespace Vulkan {
@ -19,14 +19,13 @@ using VideoCommon::Offset2D;
using VideoCommon::RenderTargets; using VideoCommon::RenderTargets;
using VideoCore::Surface::PixelFormat; using VideoCore::Surface::PixelFormat;
class VKScheduler;
class VKStagingBufferPool;
class BlitImageHelper; class BlitImageHelper;
class Device; class Device;
class Image; class Image;
class ImageView; class ImageView;
class Framebuffer; class Framebuffer;
class StagingBufferPool;
class VKScheduler;
struct RenderPassKey { struct RenderPassKey {
constexpr auto operator<=>(const RenderPassKey&) const noexcept = default; constexpr auto operator<=>(const RenderPassKey&) const noexcept = default;
@ -60,18 +59,18 @@ struct ImageBufferMap {
} }
[[nodiscard]] std::span<u8> Span() const noexcept { [[nodiscard]] std::span<u8> Span() const noexcept {
return map.Span(); return span;
} }
VkBuffer handle; VkBuffer handle;
MemoryMap map; std::span<u8> span;
}; };
struct TextureCacheRuntime { struct TextureCacheRuntime {
const Device& device; const Device& device;
VKScheduler& scheduler; VKScheduler& scheduler;
VKMemoryManager& memory_manager; MemoryAllocator& memory_allocator;
VKStagingBufferPool& staging_buffer_pool; StagingBufferPool& staging_buffer_pool;
BlitImageHelper& blit_image_helper; BlitImageHelper& blit_image_helper;
std::unordered_map<RenderPassKey, vk::RenderPass> renderpass_cache; std::unordered_map<RenderPassKey, vk::RenderPass> renderpass_cache;
@ -79,10 +78,7 @@ struct TextureCacheRuntime {
[[nodiscard]] ImageBufferMap MapUploadBuffer(size_t size); [[nodiscard]] ImageBufferMap MapUploadBuffer(size_t size);
[[nodiscard]] ImageBufferMap MapDownloadBuffer(size_t size) { [[nodiscard]] ImageBufferMap MapDownloadBuffer(size_t size);
// TODO: Have a special function for this
return MapUploadBuffer(size);
}
void BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src, void BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src,
const std::array<Offset2D, 2>& dst_region, const std::array<Offset2D, 2>& dst_region,
@ -141,7 +137,7 @@ private:
VKScheduler* scheduler; VKScheduler* scheduler;
vk::Image image; vk::Image image;
vk::Buffer buffer; vk::Buffer buffer;
VKMemoryCommit commit; MemoryCommit commit;
VkImageAspectFlags aspect_mask = 0; VkImageAspectFlags aspect_mask = 0;
bool initialized = false; bool initialized = false;
}; };

View file

@ -0,0 +1,268 @@
// Copyright 2018 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <bit>
#include <optional>
#include <vector>
#include "common/alignment.h"
#include "common/assert.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan {
namespace {
struct Range {
u64 begin;
u64 end;
[[nodiscard]] bool Contains(u64 iterator, u64 size) const noexcept {
return iterator < end && begin < iterator + size;
}
};
[[nodiscard]] u64 AllocationChunkSize(u64 required_size) {
static constexpr std::array sizes{
0x1000ULL << 10, 0x1400ULL << 10, 0x1800ULL << 10, 0x1c00ULL << 10, 0x2000ULL << 10,
0x3200ULL << 10, 0x4000ULL << 10, 0x6000ULL << 10, 0x8000ULL << 10, 0xA000ULL << 10,
0x10000ULL << 10, 0x18000ULL << 10, 0x20000ULL << 10,
};
static_assert(std::is_sorted(sizes.begin(), sizes.end()));
const auto it = std::ranges::lower_bound(sizes, required_size);
return it != sizes.end() ? *it : Common::AlignUp(required_size, 4ULL << 20);
}
[[nodiscard]] VkMemoryPropertyFlags MemoryUsagePropertyFlags(MemoryUsage usage) {
switch (usage) {
case MemoryUsage::DeviceLocal:
return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
case MemoryUsage::Upload:
return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
case MemoryUsage::Download:
return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
}
UNREACHABLE_MSG("Invalid memory usage={}", usage);
return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
}
} // Anonymous namespace
class MemoryAllocation {
public:
explicit MemoryAllocation(const Device& device_, vk::DeviceMemory memory_,
VkMemoryPropertyFlags properties, u64 allocation_size_, u32 type)
: device{device_}, memory{std::move(memory_)}, allocation_size{allocation_size_},
property_flags{properties}, shifted_memory_type{1U << type} {}
[[nodiscard]] std::optional<MemoryCommit> Commit(VkDeviceSize size, VkDeviceSize alignment) {
const std::optional<u64> alloc = FindFreeRegion(size, alignment);
if (!alloc) {
// Signal out of memory, it'll try to do more allocations.
return std::nullopt;
}
const Range range{
.begin = *alloc,
.end = *alloc + size,
};
commits.insert(std::ranges::upper_bound(commits, *alloc, {}, &Range::begin), range);
return std::make_optional<MemoryCommit>(this, *memory, *alloc, *alloc + size);
}
void Free(u64 begin) {
const auto it = std::ranges::find(commits, begin, &Range::begin);
ASSERT_MSG(it != commits.end(), "Invalid commit");
commits.erase(it);
}
[[nodiscard]] std::span<u8> Map() {
if (memory_mapped_span.empty()) {
u8* const raw_pointer = memory.Map(0, allocation_size);
memory_mapped_span = std::span<u8>(raw_pointer, allocation_size);
}
return memory_mapped_span;
}
/// Returns whether this allocation is compatible with the arguments.
[[nodiscard]] bool IsCompatible(VkMemoryPropertyFlags flags, u32 type_mask) const {
return (flags & property_flags) && (type_mask & shifted_memory_type) != 0;
}
private:
[[nodiscard]] static constexpr u32 ShiftType(u32 type) {
return 1U << type;
}
[[nodiscard]] std::optional<u64> FindFreeRegion(u64 size, u64 alignment) noexcept {
ASSERT(std::has_single_bit(alignment));
const u64 alignment_log2 = std::countr_zero(alignment);
std::optional<u64> candidate;
u64 iterator = 0;
auto commit = commits.begin();
while (iterator + size <= allocation_size) {
candidate = candidate.value_or(iterator);
if (commit == commits.end()) {
break;
}
if (commit->Contains(*candidate, size)) {
candidate = std::nullopt;
}
iterator = Common::AlignUpLog2(commit->end, alignment_log2);
++commit;
}
return candidate;
}
const Device& device; ///< Vulkan device.
const vk::DeviceMemory memory; ///< Vulkan memory allocation handler.
const u64 allocation_size; ///< Size of this allocation.
const VkMemoryPropertyFlags property_flags; ///< Vulkan memory property flags.
const u32 shifted_memory_type; ///< Shifted Vulkan memory type.
std::vector<Range> commits; ///< All commit ranges done from this allocation.
std::span<u8> memory_mapped_span; ///< Memory mapped span. Empty if not queried before.
};
MemoryCommit::MemoryCommit(MemoryAllocation* allocation_, VkDeviceMemory memory_, u64 begin_,
u64 end_) noexcept
: allocation{allocation_}, memory{memory_}, begin{begin_}, end{end_} {}
MemoryCommit::~MemoryCommit() {
Release();
}
MemoryCommit& MemoryCommit::operator=(MemoryCommit&& rhs) noexcept {
Release();
allocation = std::exchange(rhs.allocation, nullptr);
memory = rhs.memory;
begin = rhs.begin;
end = rhs.end;
span = std::exchange(rhs.span, std::span<u8>{});
return *this;
}
MemoryCommit::MemoryCommit(MemoryCommit&& rhs) noexcept
: allocation{std::exchange(rhs.allocation, nullptr)}, memory{rhs.memory}, begin{rhs.begin},
end{rhs.end}, span{std::exchange(rhs.span, std::span<u8>{})} {}
std::span<u8> MemoryCommit::Map() {
if (span.empty()) {
span = allocation->Map().subspan(begin, end - begin);
}
return span;
}
void MemoryCommit::Release() {
if (allocation) {
allocation->Free(begin);
}
}
MemoryAllocator::MemoryAllocator(const Device& device_)
: device{device_}, properties{device_.GetPhysical().GetMemoryProperties()} {}
MemoryAllocator::~MemoryAllocator() = default;
MemoryCommit MemoryAllocator::Commit(const VkMemoryRequirements& requirements, MemoryUsage usage) {
// Find the fastest memory flags we can afford with the current requirements
const VkMemoryPropertyFlags flags = MemoryPropertyFlags(requirements.memoryTypeBits, usage);
if (std::optional<MemoryCommit> commit = TryCommit(requirements, flags)) {
return std::move(*commit);
}
// Commit has failed, allocate more memory.
// TODO(Rodrigo): Handle out of memory situations in some way like flushing to guest memory.
AllocMemory(flags, requirements.memoryTypeBits, AllocationChunkSize(requirements.size));
// Commit again, this time it won't fail since there's a fresh allocation above.
// If it does, there's a bug.
return TryCommit(requirements, flags).value();
}
MemoryCommit MemoryAllocator::Commit(const vk::Buffer& buffer, MemoryUsage usage) {
auto commit = Commit(device.GetLogical().GetBufferMemoryRequirements(*buffer), usage);
buffer.BindMemory(commit.Memory(), commit.Offset());
return commit;
}
MemoryCommit MemoryAllocator::Commit(const vk::Image& image, MemoryUsage usage) {
auto commit = Commit(device.GetLogical().GetImageMemoryRequirements(*image), usage);
image.BindMemory(commit.Memory(), commit.Offset());
return commit;
}
void MemoryAllocator::AllocMemory(VkMemoryPropertyFlags flags, u32 type_mask, u64 size) {
const u32 type = FindType(flags, type_mask).value();
vk::DeviceMemory memory = device.GetLogical().AllocateMemory({
.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
.pNext = nullptr,
.allocationSize = size,
.memoryTypeIndex = type,
});
allocations.push_back(
std::make_unique<MemoryAllocation>(device, std::move(memory), flags, size, type));
}
std::optional<MemoryCommit> MemoryAllocator::TryCommit(const VkMemoryRequirements& requirements,
VkMemoryPropertyFlags flags) {
for (auto& allocation : allocations) {
if (!allocation->IsCompatible(flags, requirements.memoryTypeBits)) {
continue;
}
if (auto commit = allocation->Commit(requirements.size, requirements.alignment)) {
return commit;
}
}
return std::nullopt;
}
VkMemoryPropertyFlags MemoryAllocator::MemoryPropertyFlags(u32 type_mask, MemoryUsage usage) const {
return MemoryPropertyFlags(type_mask, MemoryUsagePropertyFlags(usage));
}
VkMemoryPropertyFlags MemoryAllocator::MemoryPropertyFlags(u32 type_mask,
VkMemoryPropertyFlags flags) const {
if (FindType(flags, type_mask)) {
// Found a memory type with those requirements
return flags;
}
if (flags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT) {
// Remove host cached bit in case it's not supported
return MemoryPropertyFlags(type_mask, flags & ~VK_MEMORY_PROPERTY_HOST_CACHED_BIT);
}
if (flags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) {
// Remove device local, if it's not supported by the requested resource
return MemoryPropertyFlags(type_mask, flags & ~VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
}
UNREACHABLE_MSG("No compatible memory types found");
return 0;
}
std::optional<u32> MemoryAllocator::FindType(VkMemoryPropertyFlags flags, u32 type_mask) const {
for (u32 type_index = 0; type_index < properties.memoryTypeCount; ++type_index) {
const VkMemoryPropertyFlags type_flags = properties.memoryTypes[type_index].propertyFlags;
if ((type_mask & (1U << type_index)) && (type_flags & flags)) {
// The type matches in type and in the wanted properties.
return type_index;
}
}
// Failed to find index
return std::nullopt;
}
bool IsHostVisible(MemoryUsage usage) noexcept {
switch (usage) {
case MemoryUsage::DeviceLocal:
return false;
case MemoryUsage::Upload:
case MemoryUsage::Download:
return true;
}
UNREACHABLE_MSG("Invalid memory usage={}", usage);
return false;
}
} // namespace Vulkan

View file

@ -0,0 +1,118 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <memory>
#include <span>
#include <utility>
#include <vector>
#include "common/common_types.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan {
class Device;
class MemoryMap;
class MemoryAllocation;
/// Hints and requirements for the backing memory type of a commit
enum class MemoryUsage {
DeviceLocal, ///< Hints device local usages, fastest memory type to read and write from the GPU
Upload, ///< Requires a host visible memory type optimized for CPU to GPU uploads
Download, ///< Requires a host visible memory type optimized for GPU to CPU readbacks
};
/// Ownership handle of a memory commitment.
/// Points to a subregion of a memory allocation.
class MemoryCommit {
public:
explicit MemoryCommit() noexcept = default;
explicit MemoryCommit(MemoryAllocation* allocation_, VkDeviceMemory memory_, u64 begin_,
u64 end_) noexcept;
~MemoryCommit();
MemoryCommit& operator=(MemoryCommit&&) noexcept;
MemoryCommit(MemoryCommit&&) noexcept;
MemoryCommit& operator=(const MemoryCommit&) = delete;
MemoryCommit(const MemoryCommit&) = delete;
/// Returns a host visible memory map.
/// It will map the backing allocation if it hasn't been mapped before.
std::span<u8> Map();
/// Returns the Vulkan memory handler.
VkDeviceMemory Memory() const {
return memory;
}
/// Returns the start position of the commit relative to the allocation.
VkDeviceSize Offset() const {
return static_cast<VkDeviceSize>(begin);
}
private:
void Release();
MemoryAllocation* allocation{}; ///< Pointer to the large memory allocation.
VkDeviceMemory memory{}; ///< Vulkan device memory handler.
u64 begin{}; ///< Beginning offset in bytes to where the commit exists.
u64 end{}; ///< Offset in bytes where the commit ends.
std::span<u8> span; ///< Host visible memory span. Empty if not queried before.
};
/// Memory allocator container.
/// Allocates and releases memory allocations on demand.
class MemoryAllocator {
public:
explicit MemoryAllocator(const Device& device_);
~MemoryAllocator();
MemoryAllocator& operator=(const MemoryAllocator&) = delete;
MemoryAllocator(const MemoryAllocator&) = delete;
/**
* Commits a memory with the specified requeriments.
*
* @param requirements Requirements returned from a Vulkan call.
* @param host_visible Signals the allocator that it *must* use host visible and coherent
* memory. When passing false, it will try to allocate device local memory.
*
* @returns A memory commit.
*/
MemoryCommit Commit(const VkMemoryRequirements& requirements, MemoryUsage usage);
/// Commits memory required by the buffer and binds it.
MemoryCommit Commit(const vk::Buffer& buffer, MemoryUsage usage);
/// Commits memory required by the image and binds it.
MemoryCommit Commit(const vk::Image& image, MemoryUsage usage);
private:
/// Allocates a chunk of memory.
void AllocMemory(VkMemoryPropertyFlags flags, u32 type_mask, u64 size);
/// Tries to allocate a memory commit.
std::optional<MemoryCommit> TryCommit(const VkMemoryRequirements& requirements,
VkMemoryPropertyFlags flags);
/// Returns the fastest compatible memory property flags from a wanted usage.
VkMemoryPropertyFlags MemoryPropertyFlags(u32 type_mask, MemoryUsage usage) const;
/// Returns the fastest compatible memory property flags from the wanted flags.
VkMemoryPropertyFlags MemoryPropertyFlags(u32 type_mask, VkMemoryPropertyFlags flags) const;
/// Returns index to the fastest memory type compatible with the passed requirements.
std::optional<u32> FindType(VkMemoryPropertyFlags flags, u32 type_mask) const;
const Device& device; ///< Device handle.
const VkPhysicalDeviceMemoryProperties properties; ///< Physical device properties.
std::vector<std::unique_ptr<MemoryAllocation>> allocations; ///< Current allocations.
};
/// Returns true when a memory usage is guaranteed to be host visible.
bool IsHostVisible(MemoryUsage usage) noexcept;
} // namespace Vulkan