Nikilite/suyu
1
1
Fork 0
forked from suyu/suyu
Code Pull requests Activity

Merge pull request #3975 from ReinUsesLisp/fast-bufcache

Browse source 
buffer_cache: Replace boost::icl::interval_map with boost::intrusive::set
This commit is contained in:
bunnei 2020-05-24 00:32:44 -04:00 committed by GitHub
commit 6a5cf1473e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 221 additions and 189 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/video_core/CMakeLists.txt
View file

@ -1,6 +1,7 @@
add_library(video_core STATIC
buffer_cache/buffer_block.h
buffer_cache/buffer_cache.h
buffer_cache/map_interval.cpp
buffer_cache/map_interval.h
dirty_flags.cpp
dirty_flags.h

227
src/video_core/buffer_cache/buffer_cache.h
View file

@ -12,11 +12,12 @@
#include <utility>
#include <vector>
#include <boost/icl/interval_map.hpp>
#include <boost/container/small_vector.hpp>
#include <boost/icl/interval_set.hpp>
#include <boost/range/iterator_range.hpp>
#include <boost/intrusive/set.hpp>
#include "common/alignment.h"
#include "common/assert.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "core/core.h"
@ -29,10 +30,12 @@
namespace VideoCommon {
using MapInterval = std::shared_ptr<MapIntervalBase>;
template <typename OwnerBuffer, typename BufferType, typename StreamBuffer>
class BufferCache {
using IntervalSet = boost::icl::interval_set<VAddr>;
using IntervalType = typename IntervalSet::interval_type;
using VectorMapInterval = boost::container::small_vector<MapInterval*, 1>;
public:
using BufferInfo = std::pair<BufferType, u64>;
@ -40,14 +43,12 @@ public:
bool is_written = false, bool use_fast_cbuf = false) {
std::lock_guard lock{mutex};
const std::optional<VAddr> cpu_addr_opt =
system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr);
const auto& memory_manager = system.GPU().MemoryManager();
const std::optional<VAddr> cpu_addr_opt = memory_manager.GpuToCpuAddress(gpu_addr);
if (!cpu_addr_opt) {
return {GetEmptyBuffer(size), 0};
}
VAddr cpu_addr = *cpu_addr_opt;
const VAddr cpu_addr = *cpu_addr_opt;
// 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.
@ -77,16 +78,19 @@ public:
}
}
auto block = GetBlock(cpu_addr, size);
auto map = MapAddress(block, gpu_addr, cpu_addr, size);
OwnerBuffer block = GetBlock(cpu_addr, size);
MapInterval* const map = MapAddress(block, gpu_addr, cpu_addr, size);
if (!map) {
return {GetEmptyBuffer(size), 0};
}
if (is_written) {
map->MarkAsModified(true, GetModifiedTicks());
if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
MarkForAsyncFlush(map);
}
if (!map->IsWritten()) {
map->MarkAsWritten(true);
MarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
if (!map->is_written) {
map->is_written = true;
MarkRegionAsWritten(map->start, map->end - 1);
}
}
@ -132,12 +136,11 @@ public:
void FlushRegion(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
std::vector<MapInterval> objects = GetMapsInRange(addr, size);
std::sort(objects.begin(), objects.end(), [](const MapInterval& a, const MapInterval& b) {
return a->GetModificationTick() < b->GetModificationTick();
});
for (auto& object : objects) {
if (object->IsModified() && object->IsRegistered()) {
VectorMapInterval objects = GetMapsInRange(addr, size);
std::sort(objects.begin(), objects.end(),
[](MapInterval* lhs, MapInterval* rhs) { return lhs->ticks < rhs->ticks; });
for (MapInterval* object : objects) {
if (object->is_modified && object->is_registered) {
mutex.unlock();
FlushMap(object);
mutex.lock();
@ -148,9 +151,9 @@ public:
bool MustFlushRegion(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
const std::vector<MapInterval> objects = GetMapsInRange(addr, size);
return std::any_of(objects.cbegin(), objects.cend(), [](const MapInterval& map) {
return map->IsModified() && map->IsRegistered();
const VectorMapInterval objects = GetMapsInRange(addr, size);
return std::any_of(objects.cbegin(), objects.cend(), [](const MapInterval* map) {
return map->is_modified && map->is_registered;
});
}
@ -158,9 +161,8 @@ public:
void InvalidateRegion(VAddr addr, u64 size) {
std::lock_guard lock{mutex};
std::vector<MapInterval> objects = GetMapsInRange(addr, size);
for (auto& object : objects) {
if (object->IsRegistered()) {
for (auto& object : GetMapsInRange(addr, size)) {
if (object->is_registered) {
Unregister(object);
}
}
@ -169,10 +171,10 @@ public:
void OnCPUWrite(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
for (const auto& object : GetMapsInRange(addr, size)) {
if (object->IsMemoryMarked() && object->IsRegistered()) {
for (MapInterval* object : GetMapsInRange(addr, size)) {
if (object->is_memory_marked && object->is_registered) {
UnmarkMemory(object);
object->SetSyncPending(true);
object->is_sync_pending = true;
marked_for_unregister.emplace_back(object);
}
}
@ -181,9 +183,9 @@ public:
void SyncGuestHost() {
std::lock_guard lock{mutex};
for (const auto& object : marked_for_unregister) {
if (object->IsRegistered()) {
object->SetSyncPending(false);
for (auto& object : marked_for_unregister) {
if (object->is_registered) {
object->is_sync_pending = false;
Unregister(object);
}
}
@ -192,9 +194,9 @@ public:
void CommitAsyncFlushes() {
if (uncommitted_flushes) {
auto commit_list = std::make_shared<std::list<MapInterval>>();
for (auto& map : *uncommitted_flushes) {
if (map->IsRegistered() && map->IsModified()) {
auto commit_list = std::make_shared<std::list<MapInterval*>>();
for (MapInterval* map : *uncommitted_flushes) {
if (map->is_registered && map->is_modified) {
// TODO(Blinkhawk): Implement backend asynchronous flushing
// AsyncFlushMap(map)
commit_list->push_back(map);
@ -228,8 +230,8 @@ public:
committed_flushes.pop_front();
return;
}
for (MapInterval& map : *flush_list) {
if (map->IsRegistered()) {
for (MapInterval* map : *flush_list) {
if (map->is_registered) {
// TODO(Blinkhawk): Replace this for reading the asynchronous flush
FlushMap(map);
}
@ -265,61 +267,60 @@ protected:
}
/// Register an object into the cache
void Register(const MapInterval& new_map, bool inherit_written = false) {
const VAddr cpu_addr = new_map->GetStart();
MapInterval* Register(MapInterval new_map, bool inherit_written = false) {
const VAddr cpu_addr = new_map.start;
if (!cpu_addr) {
LOG_CRITICAL(HW_GPU, "Failed to register buffer with unmapped gpu_address 0x{:016x}",
new_map->GetGpuAddress());
return;
new_map.gpu_addr);
return nullptr;
}
const std::size_t size = new_map->GetEnd() - new_map->GetStart();
new_map->MarkAsRegistered(true);
const IntervalType interval{new_map->GetStart(), new_map->GetEnd()};
mapped_addresses.insert({interval, new_map});
const std::size_t size = new_map.end - new_map.start;
new_map.is_registered = true;
rasterizer.UpdatePagesCachedCount(cpu_addr, size, 1);
new_map->SetMemoryMarked(true);
new_map.is_memory_marked = true;
if (inherit_written) {
MarkRegionAsWritten(new_map->GetStart(), new_map->GetEnd() - 1);
new_map->MarkAsWritten(true);
MarkRegionAsWritten(new_map.start, new_map.end - 1);
new_map.is_written = true;
}
MapInterval* const storage = mapped_addresses_allocator.Allocate();
*storage = new_map;
mapped_addresses.insert(*storage);
return storage;
}
void UnmarkMemory(const MapInterval& map) {
if (!map->IsMemoryMarked()) {
void UnmarkMemory(MapInterval* map) {
if (!map->is_memory_marked) {
return;
}
const std::size_t size = map->GetEnd() - map->GetStart();
rasterizer.UpdatePagesCachedCount(map->GetStart(), size, -1);
map->SetMemoryMarked(false);
const std::size_t size = map->end - map->start;
rasterizer.UpdatePagesCachedCount(map->start, size, -1);
map->is_memory_marked = false;
}
/// Unregisters an object from the cache
void Unregister(const MapInterval& map) {
void Unregister(MapInterval* map) {
UnmarkMemory(map);
map->MarkAsRegistered(false);
if (map->IsSyncPending()) {
map->is_registered = false;
if (map->is_sync_pending) {
map->is_sync_pending = false;
marked_for_unregister.remove(map);
map->SetSyncPending(false);
}
if (map->IsWritten()) {
UnmarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
if (map->is_written) {
UnmarkRegionAsWritten(map->start, map->end - 1);
}
const IntervalType delete_interval{map->GetStart(), map->GetEnd()};
mapped_addresses.erase(delete_interval);
const auto it = mapped_addresses.find(*map);
ASSERT(it != mapped_addresses.end());
mapped_addresses.erase(it);
mapped_addresses_allocator.Release(map);
}
private:
MapInterval CreateMap(const VAddr start, const VAddr end, const GPUVAddr gpu_addr) {
return std::make_shared<MapIntervalBase>(start, end, gpu_addr);
}
MapInterval MapAddress(const OwnerBuffer& block, const GPUVAddr gpu_addr, const VAddr cpu_addr,
const std::size_t size) {
std::vector<MapInterval> overlaps = GetMapsInRange(cpu_addr, size);
MapInterval* MapAddress(const OwnerBuffer& block, GPUVAddr gpu_addr, VAddr cpu_addr,
std::size_t size) {
const VectorMapInterval overlaps = GetMapsInRange(cpu_addr, size);
if (overlaps.empty()) {
auto& memory_manager = system.GPU().MemoryManager();
const VAddr cpu_addr_end = cpu_addr + size;
MapInterval new_map = CreateMap(cpu_addr, cpu_addr_end, gpu_addr);
if (memory_manager.IsGranularRange(gpu_addr, size)) {
u8* host_ptr = memory_manager.GetPointer(gpu_addr);
UploadBlockData(block, block->GetOffset(cpu_addr), size, host_ptr);
@ -328,13 +329,12 @@ private:
memory_manager.ReadBlockUnsafe(gpu_addr, staging_buffer.data(), size);
UploadBlockData(block, block->GetOffset(cpu_addr), size, staging_buffer.data());
}
Register(new_map);
return new_map;
return Register(MapInterval(cpu_addr, cpu_addr_end, gpu_addr));
}
const VAddr cpu_addr_end = cpu_addr + size;
if (overlaps.size() == 1) {
MapInterval& current_map = overlaps[0];
MapInterval* const current_map = overlaps[0];
if (current_map->IsInside(cpu_addr, cpu_addr_end)) {
return current_map;
}
@ -344,35 +344,39 @@ private:
bool write_inheritance = false;
bool modified_inheritance = false;
// Calculate new buffer parameters
for (auto& overlap : overlaps) {
new_start = std::min(overlap->GetStart(), new_start);
new_end = std::max(overlap->GetEnd(), new_end);
write_inheritance |= overlap->IsWritten();
modified_inheritance |= overlap->IsModified();
for (MapInterval* overlap : overlaps) {
new_start = std::min(overlap->start, new_start);
new_end = std::max(overlap->end, new_end);
write_inheritance |= overlap->is_written;
modified_inheritance |= overlap->is_modified;
}
GPUVAddr new_gpu_addr = gpu_addr + new_start - cpu_addr;
for (auto& overlap : overlaps) {
Unregister(overlap);
}
UpdateBlock(block, new_start, new_end, overlaps);
MapInterval new_map = CreateMap(new_start, new_end, new_gpu_addr);
const MapInterval new_map{new_start, new_end, new_gpu_addr};
MapInterval* const map = Register(new_map, write_inheritance);
if (!map) {
return nullptr;
}
if (modified_inheritance) {
new_map->MarkAsModified(true, GetModifiedTicks());
map->MarkAsModified(true, GetModifiedTicks());
if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
MarkForAsyncFlush(new_map);
MarkForAsyncFlush(map);
}
}
Register(new_map, write_inheritance);
return new_map;
return map;
}
void UpdateBlock(const OwnerBuffer& block, VAddr start, VAddr end,
std::vector<MapInterval>& overlaps) {
const VectorMapInterval& overlaps) {
const IntervalType base_interval{start, end};
IntervalSet interval_set{};
interval_set.add(base_interval);
for (auto& overlap : overlaps) {
const IntervalType subtract{overlap->GetStart(), overlap->GetEnd()};
const IntervalType subtract{overlap->start, overlap->end};
interval_set.subtract(subtract);
}
for (auto& interval : interval_set) {
@ -386,18 +390,24 @@ private:
}
}
std::vector<MapInterval> GetMapsInRange(VAddr addr, std::size_t size) {
VectorMapInterval GetMapsInRange(VAddr addr, std::size_t size) {
VectorMapInterval result;
if (size == 0) {
return {};
return result;
}
std::vector<MapInterval> objects{};
const IntervalType interval{addr, addr + size};
for (auto& pair : boost::make_iterator_range(mapped_addresses.equal_range(interval))) {
objects.push_back(pair.second);
const VAddr addr_end = addr + size;
auto it = mapped_addresses.lower_bound(addr);
if (it != mapped_addresses.begin()) {
--it;
}
return objects;
while (it != mapped_addresses.end() && it->start < addr_end) {
if (it->Overlaps(addr, addr_end)) {
result.push_back(&*it);
}
++it;
}
return result;
}
/// Returns a ticks counter used for tracking when cached objects were last modified
@ -405,12 +415,12 @@ private:
return ++modified_ticks;
}
void FlushMap(MapInterval map) {
std::size_t size = map->GetEnd() - map->GetStart();
OwnerBuffer block = blocks[map->GetStart() >> block_page_bits];
void FlushMap(MapInterval* map) {
const std::size_t size = map->end - map->start;
OwnerBuffer block = blocks[map->start >> block_page_bits];
staging_buffer.resize(size);
DownloadBlockData(block, block->GetOffset(map->GetStart()), size, staging_buffer.data());
system.Memory().WriteBlockUnsafe(map->GetStart(), staging_buffer.data(), size);
DownloadBlockData(block, block->GetOffset(map->start), size, staging_buffer.data());
system.Memory().WriteBlockUnsafe(map->start, staging_buffer.data(), size);
map->MarkAsModified(false, 0);
}
@ -515,7 +525,7 @@ private:
} else {
written_pages[page_start] = 1;
}
page_start++;
++page_start;
}
}
@ -531,7 +541,7 @@ private:
written_pages.erase(it);
}
}
page_start++;
++page_start;
}
}
@ -542,14 +552,14 @@ private:
if (written_pages.count(page_start) > 0) {
return true;
}
page_start++;
++page_start;
}
return false;
}
void MarkForAsyncFlush(MapInterval& map) {
void MarkForAsyncFlush(MapInterval* map) {
if (!uncommitted_flushes) {
uncommitted_flushes = std::make_shared<std::unordered_set<MapInterval>>();
uncommitted_flushes = std::make_shared<std::unordered_set<MapInterval*>>();
}
uncommitted_flushes->insert(map);
}
@ -566,10 +576,9 @@ private:
u64 buffer_offset = 0;
u64 buffer_offset_base = 0;
using IntervalSet = boost::icl::interval_set<VAddr>;
using IntervalCache = boost::icl::interval_map<VAddr, MapInterval>;
using IntervalType = typename IntervalCache::interval_type;
IntervalCache mapped_addresses;
MapIntervalAllocator mapped_addresses_allocator;
boost::intrusive::set<MapInterval, boost::intrusive::compare<MapIntervalCompare>>
mapped_addresses;
static constexpr u64 write_page_bit = 11;
std::unordered_map<u64, u32> written_pages;
@ -583,10 +592,10 @@ private:
u64 modified_ticks = 0;
std::vector<u8> staging_buffer;
std::list<MapInterval> marked_for_unregister;
std::list<MapInterval*> marked_for_unregister;
std::shared_ptr<std::unordered_set<MapInterval>> uncommitted_flushes{};
std::list<std::shared_ptr<std::list<MapInterval>>> committed_flushes;
std::shared_ptr<std::unordered_set<MapInterval*>> uncommitted_flushes;
std::list<std::shared_ptr<std::list<MapInterval*>>> committed_flushes;
std::recursive_mutex mutex;
};

33
src/video_core/buffer_cache/map_interval.cpp Normal file
View file

@ -0,0 +1,33 @@
// Copyright 2020 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <array>
#include <cstddef>
#include <memory>
#include "video_core/buffer_cache/map_interval.h"
namespace VideoCommon {
MapIntervalAllocator::MapIntervalAllocator() {
FillFreeList(first_chunk);
}
MapIntervalAllocator::~MapIntervalAllocator() = default;
void MapIntervalAllocator::AllocateNewChunk() {
*new_chunk = std::make_unique<Chunk>();
FillFreeList(**new_chunk);
new_chunk = &(*new_chunk)->next;
}
void MapIntervalAllocator::FillFreeList(Chunk& chunk) {
const std::size_t old_size = free_list.size();
free_list.resize(old_size + chunk.data.size());
std::transform(chunk.data.rbegin(), chunk.data.rend(), free_list.begin() + old_size,
[](MapInterval& interval) { return &interval; });
}
} // namespace VideoCommon

145
src/video_core/buffer_cache/map_interval.h
View file

@ -4,104 +4,89 @@
#pragma once
#include <array>
#include <cstddef>
#include <memory>
#include <vector>
#include <boost/intrusive/set_hook.hpp>
#include "common/common_types.h"
#include "video_core/gpu.h"
namespace VideoCommon {
class MapIntervalBase {
public:
MapIntervalBase(const VAddr start, const VAddr end, const GPUVAddr gpu_addr)
: start{start}, end{end}, gpu_addr{gpu_addr} {}
struct MapInterval : public boost::intrusive::set_base_hook<boost::intrusive::optimize_size<true>> {
MapInterval() = default;
void SetCpuAddress(VAddr new_cpu_addr) {
cpu_addr = new_cpu_addr;
/*implicit*/ MapInterval(VAddr start_) noexcept : start{start_} {}
explicit MapInterval(VAddr start_, VAddr end_, GPUVAddr gpu_addr_) noexcept
: start{start_}, end{end_}, gpu_addr{gpu_addr_} {}
bool IsInside(VAddr other_start, VAddr other_end) const noexcept {
return start <= other_start && other_end <= end;
}
VAddr GetCpuAddress() const {
return cpu_addr;
bool Overlaps(VAddr other_start, VAddr other_end) const noexcept {
return start < other_end && other_start < end;
}
GPUVAddr GetGpuAddress() const {
return gpu_addr;
}
bool IsInside(const VAddr other_start, const VAddr other_end) const {
return (start <= other_start && other_end <= end);
}
bool operator==(const MapIntervalBase& rhs) const {
return std::tie(start, end) == std::tie(rhs.start, rhs.end);
}
bool operator!=(const MapIntervalBase& rhs) const {
return !operator==(rhs);
}
void MarkAsRegistered(const bool registered) {
is_registered = registered;
}
bool IsRegistered() const {
return is_registered;
}
void SetMemoryMarked(bool is_memory_marked_) {
is_memory_marked = is_memory_marked_;
}
bool IsMemoryMarked() const {
return is_memory_marked;
}
void SetSyncPending(bool is_sync_pending_) {
is_sync_pending = is_sync_pending_;
}
bool IsSyncPending() const {
return is_sync_pending;
}
VAddr GetStart() const {
return start;
}
VAddr GetEnd() const {
return end;
}
void MarkAsModified(const bool is_modified_, const u64 tick) {
void MarkAsModified(bool is_modified_, u64 ticks_) noexcept {
is_modified = is_modified_;
ticks = tick;
ticks = ticks_;
}
bool IsModified() const {
return is_modified;
boost::intrusive::set_member_hook<> member_hook_;
VAddr start = 0;
VAddr end = 0;
GPUVAddr gpu_addr = 0;
u64 ticks = 0;
bool is_written = false;
bool is_modified = false;
bool is_registered = false;
bool is_memory_marked = false;
bool is_sync_pending = false;
};
struct MapIntervalCompare {
constexpr bool operator()(const MapInterval& lhs, const MapInterval& rhs) const noexcept {
return lhs.start < rhs.start;
}
};
class MapIntervalAllocator {
public:
MapIntervalAllocator();
~MapIntervalAllocator();
MapInterval* Allocate() {
if (free_list.empty()) {
AllocateNewChunk();
}
MapInterval* const interval = free_list.back();
free_list.pop_back();
return interval;
}
u64 GetModificationTick() const {
return ticks;
}
void MarkAsWritten(const bool is_written_) {
is_written = is_written_;
}
bool IsWritten() const {
return is_written;
void Release(MapInterval* interval) {
free_list.push_back(interval);
}
private:
VAddr start;
VAddr end;
GPUVAddr gpu_addr;
VAddr cpu_addr{};
bool is_written{};
bool is_modified{};
bool is_registered{};
bool is_memory_marked{};
bool is_sync_pending{};
u64 ticks{};
struct Chunk {
std::unique_ptr<Chunk> next;
std::array<MapInterval, 0x8000> data;
};
void AllocateNewChunk();
void FillFreeList(Chunk& chunk);
std::vector<MapInterval*> free_list;
std::unique_ptr<Chunk>* new_chunk = &first_chunk.next;
Chunk first_chunk;
};
} // namespace VideoCommon

1
src/video_core/renderer_opengl/gl_buffer_cache.cpp
View file

@ -8,6 +8,7 @@
#include "common/assert.h"
#include "common/microprofile.h"
#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"

1
src/video_core/renderer_opengl/gl_fence_manager.cpp
View file

@ -4,6 +4,7 @@
#include "common/assert.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_fence_manager.h"
namespace OpenGL {

1
src/video_core/renderer_vulkan/vk_buffer_cache.cpp
View file

@ -7,6 +7,7 @@
#include <memory>
#include "core/core.h"
#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/renderer_vulkan/vk_buffer_cache.h"
#include "video_core/renderer_vulkan/vk_device.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"

1
src/video_core/renderer_vulkan/vk_fence_manager.h
View file

@ -7,6 +7,7 @@
#include <memory>
#include "video_core/fence_manager.h"
#include "video_core/renderer_vulkan/vk_buffer_cache.h"
#include "video_core/renderer_vulkan/wrapper.h"
namespace Core {