Buffer Cache rework: Setup async downloads.

This commit is contained in:
Fernando Sahmkow 2023-04-22 13:36:18 +02:00
parent a16c261131
commit f2d3212de9
2 changed files with 155 additions and 141 deletions

View file

@ -11,6 +11,8 @@
namespace VideoCommon {
using Core::Memory::YUZU_PAGESIZE;
template <class P>
BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
Core::Memory::Memory& cpu_memory_, Runtime& runtime_)
@ -87,10 +89,12 @@ void BufferCache<P>::TickFrame() {
template <class P>
void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
memory_tracker.MarkRegionAsCpuModified(cpu_addr, size);
if (memory_tracker.IsRegionGpuModified(cpu_addr, size)) {
const IntervalType subtract_interval{cpu_addr, cpu_addr + size};
ClearDownload(subtract_interval);
common_ranges.subtract(subtract_interval);
}
}
template <class P>
void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
@ -102,17 +106,33 @@ void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
template <class P>
void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
WaitOnAsyncFlushes(cpu_addr, size);
ForEachBufferInRange(cpu_addr, size, [&](BufferId, Buffer& buffer) {
DownloadBufferMemory(buffer, cpu_addr, size);
});
}
template <class P>
void BufferCache<P>::ClearDownload(IntervalType subtract_interval) {
uncommitted_ranges.subtract(subtract_interval);
for (auto& interval_set : async_downloads) {
interval_set.subtract(subtract_interval);
void BufferCache<P>::WaitOnAsyncFlushes(VAddr cpu_addr, u64 size) {
bool must_wait = false;
ForEachInOverlapCounter(async_downloads, cpu_addr, size,
[&](VAddr, VAddr, int) { must_wait = true; });
bool must_release = false;
ForEachInRangeSet(pending_ranges, cpu_addr, size, [&](VAddr, VAddr) { must_release = true; });
if (must_release) {
std::function<void()> tmp([]() {});
rasterizer.SignalFence(std::move(tmp));
}
if (must_wait || must_release) {
rasterizer.ReleaseFences();
}
}
template <class P>
void BufferCache<P>::ClearDownload(IntervalType subtract_interval) {
async_downloads -= std::make_pair(subtract_interval, std::numeric_limits<int>::max());
uncommitted_ranges.subtract(subtract_interval);
pending_ranges.subtract(subtract_interval);
for (auto& interval_set : committed_ranges) {
interval_set.subtract(subtract_interval);
}
@ -132,6 +152,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
}
const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount};
WaitOnAsyncFlushes(*cpu_src_address, static_cast<u32>(amount));
ClearDownload(subtract_interval);
BufferId buffer_a;
@ -162,6 +183,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
tmp_intervals.push_back(add_interval);
if (is_high_accuracy) {
uncommitted_ranges.add(add_interval);
pending_ranges.add(add_interval);
}
};
ForEachInRangeSet(common_ranges, *cpu_src_address, amount, mirror);
@ -413,18 +435,15 @@ template <class P>
void BufferCache<P>::FlushCachedWrites() {
cached_write_buffer_ids.clear();
memory_tracker.FlushCachedWrites();
/*for (auto& interval : cached_ranges) {
VAddr cpu_addr = interval.lower();
const std::size_t size = interval.upper() - interval.lower();
memory_tracker.FlushCachedWrites(cpu_addr, size);
// common_ranges.subtract(interval);
}*/
for (auto& interval : cached_ranges) {
ClearDownload(interval);
}
cached_ranges.clear();
}
template <class P>
bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
return !uncommitted_ranges.empty() || !committed_ranges.empty() || !pending_queries.empty();
return !uncommitted_ranges.empty() || !committed_ranges.empty();
}
template <class P>
@ -437,8 +456,11 @@ void BufferCache<P>::AccumulateFlushes() {
template <class P>
bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
return (!async_buffers.empty() && async_buffers.front().has_value()) ||
(!query_async_buffers.empty() && query_async_buffers.front().has_value());
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
return (!async_buffers.empty() && async_buffers.front().has_value());
} else {
return false;
}
}
template <class P>
@ -446,11 +468,14 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
AccumulateFlushes();
if (committed_ranges.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
return;
}
MICROPROFILE_SCOPE(GPU_DownloadMemory);
pending_ranges.clear();
auto it = committed_ranges.begin();
while (it != committed_ranges.end()) {
auto& current_intervals = *it;
@ -491,7 +516,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
buffer_id,
});
// Align up to avoid cache conflicts
constexpr u64 align = 8ULL;
constexpr u64 align = 64ULL;
constexpr u64 mask = ~(align - 1ULL);
total_size_bytes += (new_size + align - 1) & mask;
largest_copy = std::max(largest_copy, new_size);
@ -504,7 +529,9 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
}
committed_ranges.clear();
if (downloads.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
return;
}
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
@ -520,13 +547,33 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
new_async_range.add(base_interval);
async_downloads += std::make_pair(base_interval, 1);
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
normalized_copies.push_back(second_copy);
}
async_downloads.emplace_back(std::move(new_async_range));
runtime.PostCopyBarrier();
pending_downloads.emplace_back(std::move(normalized_copies));
async_buffers.emplace_back(download_staging);
} else {
if constexpr (USE_MEMORY_MAPS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
// Have in mind the staging buffer offset for the copy
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies, false);
}
runtime.PostCopyBarrier();
runtime.Finish();
for (const auto& [copy, buffer_id] : downloads) {
const Buffer& buffer = slot_buffers[buffer_id];
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
// Undo the modified offset
const u64 dst_offset = copy.dst_offset - download_staging.offset;
const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
}
} else {
const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
for (const auto& [copy, buffer_id] : downloads) {
@ -537,82 +584,17 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
}
}
}
template <class P>
void BufferCache<P>::CommitAsyncQueries() {
if (pending_queries.empty()) {
query_async_buffers.emplace_back(std::optional<Async_Buffer>{});
return;
}
MICROPROFILE_SCOPE(GPU_DownloadMemory);
boost::container::small_vector<std::pair<BufferCopy, BufferId>, 8> downloads;
u64 total_size_bytes = 0;
u64 largest_copy = 0;
do {
has_deleted_buffers = false;
downloads.clear();
total_size_bytes = 0;
largest_copy = 0;
for (const auto& query_info : pending_queries) {
const std::size_t size = query_info.second;
const VAddr cpu_addr = query_info.first;
const BufferId buffer_id = FindBuffer(cpu_addr, static_cast<u32>(size));
Buffer& buffer = slot_buffers[buffer_id];
if (has_deleted_buffers) {
break;
}
downloads.push_back({
BufferCopy{
.src_offset = buffer.Offset(cpu_addr),
.dst_offset = total_size_bytes,
.size = size,
},
buffer_id,
});
constexpr u64 align = 8ULL;
constexpr u64 mask = ~(align - 1ULL);
total_size_bytes += (size + align - 1) & mask;
largest_copy = std::max(largest_copy, size);
}
} while (has_deleted_buffers);
pending_queries.clear();
if (downloads.empty()) {
query_async_buffers.push_back(std::optional<Async_Buffer>{});
return;
}
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
boost::container::small_vector<BufferCopy, 8> normalized_copies;
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
// Have in mind the staging buffer offset for the copy
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
const Buffer& buffer = slot_buffers[buffer_id];
BufferCopy second_copy{copy};
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + second_copy.src_offset;
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
normalized_copies.push_back(second_copy);
}
committed_queries.emplace_back(std::move(normalized_copies));
query_async_buffers.emplace_back(download_staging);
} else {
query_async_buffers.push_back(std::optional<Async_Buffer>{});
}
}
template <class P>
void BufferCache<P>::CommitAsyncFlushes() {
CommitAsyncFlushesHigh();
CommitAsyncQueries();
}
template <class P>
void BufferCache<P>::PopAsyncFlushes() {
MICROPROFILE_SCOPE(GPU_DownloadMemory);
PopAsyncBuffers();
PopAsyncQueries();
}
template <class P>
@ -627,59 +609,34 @@ void BufferCache<P>::PopAsyncBuffers() {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
auto& downloads = pending_downloads.front();
auto& async_buffer = async_buffers.front();
auto& async_range = async_downloads.front();
u8* base = async_buffer->mapped_span.data();
const size_t base_offset = async_buffer->offset;
for (const auto& copy : downloads) {
const VAddr cpu_addr = static_cast<VAddr>(copy.src_offset);
const u64 dst_offset = copy.dst_offset - base_offset;
const u8* read_mapped_memory = base + dst_offset;
ForEachInRangeSet(async_range, cpu_addr, copy.size, [&](VAddr start, VAddr end) {
const size_t diff = start - cpu_addr;
const size_t new_size = end - start;
cpu_memory.WriteBlockUnsafe(start, &read_mapped_memory[diff], new_size);
ForEachInOverlapCounter(
async_downloads, cpu_addr, copy.size, [&](VAddr start, VAddr end, int count) {
cpu_memory.WriteBlockUnsafe(start, &read_mapped_memory[start - cpu_addr],
end - start);
if (count == 1) {
const IntervalType base_interval{start, end};
common_ranges.subtract(base_interval);
}
});
async_downloads -= std::make_pair(IntervalType(cpu_addr, cpu_addr + copy.size), 1);
}
runtime.FreeDeferredStagingBuffer(*async_buffer);
async_buffers.pop_front();
pending_downloads.pop_front();
async_downloads.pop_front();
}
}
template <class P>
void BufferCache<P>::PopAsyncQueries() {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
if (query_async_buffers.empty()) {
return;
}
if (!query_async_buffers.front().has_value()) {
query_async_buffers.pop_front();
return;
}
auto& downloads = committed_queries.front();
auto& async_buffer = query_async_buffers.front();
flushed_queries.clear();
u8* base = async_buffer->mapped_span.data();
const size_t base_offset = async_buffer->offset;
for (const auto& copy : downloads) {
const size_t dst_offset = copy.dst_offset - base_offset;
const u8* read_mapped_memory = base + dst_offset;
u64 new_value{};
std::memcpy(&new_value, read_mapped_memory, copy.size);
flushed_queries.push_back(new_value);
}
runtime.FreeDeferredStagingBuffer(*async_buffer);
committed_queries.pop_front();
query_async_buffers.pop_front();
}
}
template <class P>
bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
return memory_tracker.IsRegionGpuModified(addr, size);
bool is_dirty = false;
ForEachInRangeSet(common_ranges, addr, size, [&](VAddr, VAddr) { is_dirty = true; });
return is_dirty;
}
template <class P>
@ -1232,16 +1189,18 @@ void BufferCache<P>::UpdateComputeTextureBuffers() {
}
template <class P>
void BufferCache<P>::MarkWrittenBuffer(BufferId, VAddr cpu_addr, u32 size) {
void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 size) {
memory_tracker.MarkRegionAsGpuModified(cpu_addr, size);
if (memory_tracker.IsRegionCpuModified(cpu_addr, size)) {
SynchronizeBuffer(slot_buffers[buffer_id], cpu_addr, size);
}
const IntervalType base_interval{cpu_addr, cpu_addr + size};
common_ranges.add(base_interval);
for (auto& interval_set : async_downloads) {
interval_set.subtract(base_interval);
}
if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) {
uncommitted_ranges.add(base_interval);
pending_ranges.add(base_interval);
}
}
@ -1530,7 +1489,9 @@ bool BufferCache<P>::InlineMemory(VAddr dest_address, size_t copy_size,
if (!is_dirty) {
return false;
}
if (!IsRegionGpuModified(dest_address, copy_size)) {
VAddr aligned_start = Common::AlignDown(dest_address, YUZU_PAGESIZE);
VAddr aligned_end = Common::AlignUp(dest_address + copy_size, YUZU_PAGESIZE);
if (!IsRegionGpuModified(aligned_start, aligned_end - aligned_start)) {
return false;
}

View file

@ -17,6 +17,7 @@
#include <boost/pool/detail/mutex.hpp>
#undef BOOST_NO_MT
#include <boost/icl/interval_set.hpp>
#include <boost/icl/split_interval_map.hpp>
#include <boost/pool/pool.hpp>
#include <boost/pool/pool_alloc.hpp>
@ -44,8 +45,7 @@
namespace boost {
template <typename T>
class fast_pool_allocator<T, default_user_allocator_new_delete, details::pool::default_mutex, 4096,
0>;
class fast_pool_allocator<T, default_user_allocator_new_delete, details::pool::null_mutex, 4096, 0>;
}
namespace VideoCommon {
@ -123,6 +123,31 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelI
boost::icl::interval_set<VAddr, IntervalCompare, IntervalInstance, IntervalAllocator>;
using IntervalType = typename IntervalSet::interval_type;
template <typename Type>
struct counter_add_functor : public boost::icl::identity_based_inplace_combine<Type> {
// types
typedef counter_add_functor<Type> type;
typedef boost::icl::identity_based_inplace_combine<Type> base_type;
// public member functions
void operator()(Type& current, const Type& added) const {
current += added;
if (current < base_type::identity_element()) {
current = base_type::identity_element();
}
}
// public static functions
static void version(Type&){};
};
using OverlapCombine = ICL_COMBINE_INSTANCE(counter_add_functor, int);
using OverlapSection = ICL_SECTION_INSTANCE(boost::icl::inter_section, int);
using OverlapCounter =
boost::icl::split_interval_map<VAddr, int, boost::icl::partial_absorber, IntervalCompare,
OverlapCombine, OverlapSection, IntervalInstance,
IntervalAllocator>;
struct Empty {};
struct OverlapResult {
@ -219,12 +244,9 @@ public:
/// Commit asynchronous downloads
void CommitAsyncFlushes();
void CommitAsyncFlushesHigh();
void CommitAsyncQueries();
/// Pop asynchronous downloads
void PopAsyncFlushes();
void PopAsyncQueries();
void PopAsyncBuffers();
bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount);
@ -302,6 +324,34 @@ private:
}
}
template <typename Func>
void ForEachInOverlapCounter(OverlapCounter& current_range, VAddr cpu_addr, u64 size,
Func&& func) {
const VAddr start_address = cpu_addr;
const VAddr end_address = start_address + size;
const IntervalType search_interval{start_address, end_address};
auto it = current_range.lower_bound(search_interval);
if (it == current_range.end()) {
return;
}
auto end_it = current_range.upper_bound(search_interval);
for (; it != end_it; it++) {
auto& inter = it->first;
VAddr inter_addr_end = inter.upper();
VAddr inter_addr = inter.lower();
if (inter_addr_end > end_address) {
inter_addr_end = end_address;
}
if (inter_addr < start_address) {
inter_addr = start_address;
}
if (it->second <= 0) {
__debugbreak();
}
func(inter_addr, inter_addr_end, it->second);
}
}
static bool IsRangeGranular(VAddr cpu_addr, size_t size) {
return (cpu_addr & ~Core::Memory::YUZU_PAGEMASK) ==
((cpu_addr + size) & ~Core::Memory::YUZU_PAGEMASK);
@ -309,6 +359,8 @@ private:
void RunGarbageCollector();
void WaitOnAsyncFlushes(VAddr cpu_addr, u64 size);
void BindHostIndexBuffer();
void BindHostVertexBuffers();
@ -474,10 +526,11 @@ private:
IntervalSet uncommitted_ranges;
IntervalSet common_ranges;
IntervalSet cached_ranges;
IntervalSet pending_ranges;
std::deque<IntervalSet> committed_ranges;
// Async Buffers
std::deque<IntervalSet> async_downloads;
OverlapCounter async_downloads;
std::deque<std::optional<Async_Buffer>> async_buffers;
std::deque<boost::container::small_vector<BufferCopy, 4>> pending_downloads;
std::optional<Async_Buffer> current_buffer;