Merge pull request #2178 from ReinUsesLisp/vk-buffer-cache
vk_buffer_cache: Implement a buffer cache
This commit is contained in:
commit
e22670fbc3
@ -104,6 +104,8 @@ add_library(video_core STATIC
|
||||
if (ENABLE_VULKAN)
|
||||
target_sources(video_core PRIVATE
|
||||
renderer_vulkan/declarations.h
|
||||
renderer_vulkan/vk_buffer_cache.cpp
|
||||
renderer_vulkan/vk_buffer_cache.h
|
||||
renderer_vulkan/vk_device.cpp
|
||||
renderer_vulkan/vk_device.h
|
||||
renderer_vulkan/vk_memory_manager.cpp
|
||||
|
116
src/video_core/renderer_vulkan/vk_buffer_cache.cpp
Normal file
116
src/video_core/renderer_vulkan/vk_buffer_cache.cpp
Normal file
@ -0,0 +1,116 @@
|
||||
// Copyright 2019 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#include <cstring>
|
||||
#include <memory>
|
||||
#include <optional>
|
||||
#include <tuple>
|
||||
|
||||
#include "common/alignment.h"
|
||||
#include "core/core.h"
|
||||
#include "core/memory.h"
|
||||
#include "video_core/renderer_vulkan/declarations.h"
|
||||
#include "video_core/renderer_vulkan/vk_buffer_cache.h"
|
||||
#include "video_core/renderer_vulkan/vk_scheduler.h"
|
||||
#include "video_core/renderer_vulkan/vk_stream_buffer.h"
|
||||
|
||||
namespace Vulkan {
|
||||
|
||||
VKBufferCache::VKBufferCache(Tegra::MemoryManager& tegra_memory_manager,
|
||||
VideoCore::RasterizerInterface& rasterizer, const VKDevice& device,
|
||||
VKMemoryManager& memory_manager, VKScheduler& scheduler, u64 size)
|
||||
: RasterizerCache{rasterizer}, tegra_memory_manager{tegra_memory_manager} {
|
||||
const auto usage = vk::BufferUsageFlagBits::eVertexBuffer |
|
||||
vk::BufferUsageFlagBits::eIndexBuffer |
|
||||
vk::BufferUsageFlagBits::eUniformBuffer;
|
||||
const auto access = vk::AccessFlagBits::eVertexAttributeRead | vk::AccessFlagBits::eIndexRead |
|
||||
vk::AccessFlagBits::eUniformRead;
|
||||
stream_buffer =
|
||||
std::make_unique<VKStreamBuffer>(device, memory_manager, scheduler, size, usage, access,
|
||||
vk::PipelineStageFlagBits::eAllCommands);
|
||||
buffer_handle = stream_buffer->GetBuffer();
|
||||
}
|
||||
|
||||
VKBufferCache::~VKBufferCache() = default;
|
||||
|
||||
u64 VKBufferCache::UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, u64 alignment,
|
||||
bool cache) {
|
||||
const auto cpu_addr{tegra_memory_manager.GpuToCpuAddress(gpu_addr)};
|
||||
ASSERT(cpu_addr);
|
||||
|
||||
// Cache management is a big overhead, so only cache entries with a given size.
|
||||
// TODO: Figure out which size is the best for given games.
|
||||
cache &= size >= 2048;
|
||||
|
||||
if (cache) {
|
||||
if (auto entry = TryGet(*cpu_addr); entry) {
|
||||
if (entry->size >= size && entry->alignment == alignment) {
|
||||
return entry->offset;
|
||||
}
|
||||
Unregister(entry);
|
||||
}
|
||||
}
|
||||
|
||||
AlignBuffer(alignment);
|
||||
const u64 uploaded_offset = buffer_offset;
|
||||
|
||||
Memory::ReadBlock(*cpu_addr, buffer_ptr, size);
|
||||
|
||||
buffer_ptr += size;
|
||||
buffer_offset += size;
|
||||
|
||||
if (cache) {
|
||||
auto entry = std::make_shared<CachedBufferEntry>();
|
||||
entry->offset = uploaded_offset;
|
||||
entry->size = size;
|
||||
entry->alignment = alignment;
|
||||
entry->addr = *cpu_addr;
|
||||
Register(entry);
|
||||
}
|
||||
|
||||
return uploaded_offset;
|
||||
}
|
||||
|
||||
u64 VKBufferCache::UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment) {
|
||||
AlignBuffer(alignment);
|
||||
std::memcpy(buffer_ptr, raw_pointer, size);
|
||||
const u64 uploaded_offset = buffer_offset;
|
||||
|
||||
buffer_ptr += size;
|
||||
buffer_offset += size;
|
||||
return uploaded_offset;
|
||||
}
|
||||
|
||||
std::tuple<u8*, u64> VKBufferCache::ReserveMemory(std::size_t size, u64 alignment) {
|
||||
AlignBuffer(alignment);
|
||||
u8* const uploaded_ptr = buffer_ptr;
|
||||
const u64 uploaded_offset = buffer_offset;
|
||||
|
||||
buffer_ptr += size;
|
||||
buffer_offset += size;
|
||||
return {uploaded_ptr, uploaded_offset};
|
||||
}
|
||||
|
||||
void VKBufferCache::Reserve(std::size_t max_size) {
|
||||
bool invalidate;
|
||||
std::tie(buffer_ptr, buffer_offset_base, invalidate) = stream_buffer->Reserve(max_size);
|
||||
buffer_offset = buffer_offset_base;
|
||||
|
||||
if (invalidate) {
|
||||
InvalidateAll();
|
||||
}
|
||||
}
|
||||
|
||||
VKExecutionContext VKBufferCache::Send(VKExecutionContext exctx) {
|
||||
return stream_buffer->Send(exctx, buffer_offset - buffer_offset_base);
|
||||
}
|
||||
|
||||
void VKBufferCache::AlignBuffer(std::size_t alignment) {
|
||||
// Align the offset, not the mapped pointer
|
||||
const u64 offset_aligned = Common::AlignUp(buffer_offset, alignment);
|
||||
buffer_ptr += offset_aligned - buffer_offset;
|
||||
buffer_offset = offset_aligned;
|
||||
}
|
||||
|
||||
} // namespace Vulkan
|
87
src/video_core/renderer_vulkan/vk_buffer_cache.h
Normal file
87
src/video_core/renderer_vulkan/vk_buffer_cache.h
Normal file
@ -0,0 +1,87 @@
|
||||
// Copyright 2019 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <memory>
|
||||
#include <tuple>
|
||||
|
||||
#include "common/common_types.h"
|
||||
#include "video_core/gpu.h"
|
||||
#include "video_core/rasterizer_cache.h"
|
||||
#include "video_core/renderer_vulkan/declarations.h"
|
||||
#include "video_core/renderer_vulkan/vk_scheduler.h"
|
||||
|
||||
namespace Tegra {
|
||||
class MemoryManager;
|
||||
}
|
||||
|
||||
namespace Vulkan {
|
||||
|
||||
class VKDevice;
|
||||
class VKFence;
|
||||
class VKMemoryManager;
|
||||
class VKStreamBuffer;
|
||||
|
||||
struct CachedBufferEntry final : public RasterizerCacheObject {
|
||||
VAddr GetAddr() const override {
|
||||
return addr;
|
||||
}
|
||||
|
||||
std::size_t GetSizeInBytes() const override {
|
||||
return size;
|
||||
}
|
||||
|
||||
// We do not have to flush this cache as things in it are never modified by us.
|
||||
void Flush() override {}
|
||||
|
||||
VAddr addr;
|
||||
std::size_t size;
|
||||
u64 offset;
|
||||
std::size_t alignment;
|
||||
};
|
||||
|
||||
class VKBufferCache final : public RasterizerCache<std::shared_ptr<CachedBufferEntry>> {
|
||||
public:
|
||||
explicit VKBufferCache(Tegra::MemoryManager& tegra_memory_manager, VideoCore::RasterizerInterface& rasterizer,
|
||||
const VKDevice& device, VKMemoryManager& memory_manager,
|
||||
VKScheduler& scheduler, u64 size);
|
||||
~VKBufferCache();
|
||||
|
||||
/// Uploads data from a guest GPU address. Returns host's buffer offset where it's been
|
||||
/// allocated.
|
||||
u64 UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, u64 alignment = 4,
|
||||
bool cache = true);
|
||||
|
||||
/// Uploads from a host memory. Returns host's buffer offset where it's been allocated.
|
||||
u64 UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment = 4);
|
||||
|
||||
/// Reserves memory to be used by host's CPU. Returns mapped address and offset.
|
||||
std::tuple<u8*, u64> ReserveMemory(std::size_t size, u64 alignment = 4);
|
||||
|
||||
/// Reserves a region of memory to be used in subsequent upload/reserve operations.
|
||||
void Reserve(std::size_t max_size);
|
||||
|
||||
/// Ensures that the set data is sent to the device.
|
||||
[[nodiscard]] VKExecutionContext Send(VKExecutionContext exctx);
|
||||
|
||||
/// Returns the buffer cache handle.
|
||||
vk::Buffer GetBuffer() const {
|
||||
return buffer_handle;
|
||||
}
|
||||
|
||||
private:
|
||||
void AlignBuffer(std::size_t alignment);
|
||||
|
||||
Tegra::MemoryManager& tegra_memory_manager;
|
||||
|
||||
std::unique_ptr<VKStreamBuffer> stream_buffer;
|
||||
vk::Buffer buffer_handle;
|
||||
|
||||
u8* buffer_ptr = nullptr;
|
||||
u64 buffer_offset = 0;
|
||||
u64 buffer_offset_base = 0;
|
||||
};
|
||||
|
||||
} // namespace Vulkan
|
Loading…
Reference in New Issue
Block a user