Implement Async downloads in normal and fix a few issues.

This commit is contained in:
Fernando Sahmkow 2023-04-22 20:10:40 +02:00
parent f2d3212de9
commit ed4553806a
3 changed files with 61 additions and 39 deletions

View file

@ -22,6 +22,8 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
void(slot_buffers.insert(runtime, NullBufferParams{})); void(slot_buffers.insert(runtime, NullBufferParams{}));
common_ranges.clear(); common_ranges.clear();
active_async_buffers = IMPLEMENTS_ASYNC_DOWNLOADS && !Settings::IsGPULevelHigh();
if (!runtime.CanReportMemoryUsage()) { if (!runtime.CanReportMemoryUsage()) {
minimum_memory = DEFAULT_EXPECTED_MEMORY; minimum_memory = DEFAULT_EXPECTED_MEMORY;
critical_memory = DEFAULT_CRITICAL_MEMORY; critical_memory = DEFAULT_CRITICAL_MEMORY;
@ -72,6 +74,8 @@ void BufferCache<P>::TickFrame() {
uniform_cache_hits[0] = 0; uniform_cache_hits[0] = 0;
uniform_cache_shots[0] = 0; uniform_cache_shots[0] = 0;
active_async_buffers = IMPLEMENTS_ASYNC_DOWNLOADS && !Settings::IsGPULevelHigh();
const bool skip_preferred = hits * 256 < shots * 251; const bool skip_preferred = hits * 256 < shots * 251;
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0; uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
@ -130,7 +134,7 @@ void BufferCache<P>::WaitOnAsyncFlushes(VAddr cpu_addr, u64 size) {
template <class P> template <class P>
void BufferCache<P>::ClearDownload(IntervalType subtract_interval) { void BufferCache<P>::ClearDownload(IntervalType subtract_interval) {
async_downloads -= std::make_pair(subtract_interval, std::numeric_limits<int>::max()); RemoveEachInOverlapCounter(async_downloads, subtract_interval, -1024);
uncommitted_ranges.subtract(subtract_interval); uncommitted_ranges.subtract(subtract_interval);
pending_ranges.subtract(subtract_interval); pending_ranges.subtract(subtract_interval);
for (auto& interval_set : committed_ranges) { for (auto& interval_set : committed_ranges) {
@ -173,18 +177,14 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
}}; }};
boost::container::small_vector<IntervalType, 4> tmp_intervals; boost::container::small_vector<IntervalType, 4> tmp_intervals;
const bool is_high_accuracy =
Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High;
auto mirror = [&](VAddr base_address, VAddr base_address_end) { auto mirror = [&](VAddr base_address, VAddr base_address_end) {
const u64 size = base_address_end - base_address; const u64 size = base_address_end - base_address;
const VAddr diff = base_address - *cpu_src_address; const VAddr diff = base_address - *cpu_src_address;
const VAddr new_base_address = *cpu_dest_address + diff; const VAddr new_base_address = *cpu_dest_address + diff;
const IntervalType add_interval{new_base_address, new_base_address + size}; const IntervalType add_interval{new_base_address, new_base_address + size};
tmp_intervals.push_back(add_interval); tmp_intervals.push_back(add_interval);
if (is_high_accuracy) { uncommitted_ranges.add(add_interval);
uncommitted_ranges.add(add_interval); pending_ranges.add(add_interval);
pending_ranges.add(add_interval);
}
}; };
ForEachInRangeSet(common_ranges, *cpu_src_address, amount, mirror); ForEachInRangeSet(common_ranges, *cpu_src_address, amount, mirror);
// This subtraction in this order is important for overlapping copies. // This subtraction in this order is important for overlapping copies.
@ -468,7 +468,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
AccumulateFlushes(); AccumulateFlushes();
if (committed_ranges.empty()) { if (committed_ranges.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { if (active_async_buffers) {
async_buffers.emplace_back(std::optional<Async_Buffer>{}); async_buffers.emplace_back(std::optional<Async_Buffer>{});
} }
return; return;
@ -529,31 +529,33 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
} }
committed_ranges.clear(); committed_ranges.clear();
if (downloads.empty()) { if (downloads.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { if (active_async_buffers) {
async_buffers.emplace_back(std::optional<Async_Buffer>{}); async_buffers.emplace_back(std::optional<Async_Buffer>{});
} }
return; return;
} }
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { if (active_async_buffers) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true); if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
boost::container::small_vector<BufferCopy, 4> normalized_copies; auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
IntervalSet new_async_range{}; boost::container::small_vector<BufferCopy, 4> normalized_copies;
runtime.PreCopyBarrier(); IntervalSet new_async_range{};
for (auto& [copy, buffer_id] : downloads) { runtime.PreCopyBarrier();
copy.dst_offset += download_staging.offset; for (auto& [copy, buffer_id] : downloads) {
const std::array copies{copy}; copy.dst_offset += download_staging.offset;
BufferCopy second_copy{copy}; const std::array copies{copy};
Buffer& buffer = slot_buffers[buffer_id]; BufferCopy second_copy{copy};
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset; Buffer& buffer = slot_buffers[buffer_id];
VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset); second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size}; VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
async_downloads += std::make_pair(base_interval, 1); const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false); async_downloads += std::make_pair(base_interval, 1);
normalized_copies.push_back(second_copy); runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
normalized_copies.push_back(second_copy);
}
runtime.PostCopyBarrier();
pending_downloads.emplace_back(std::move(normalized_copies));
async_buffers.emplace_back(download_staging);
} }
runtime.PostCopyBarrier();
pending_downloads.emplace_back(std::move(normalized_copies));
async_buffers.emplace_back(download_staging);
} else { } else {
if constexpr (USE_MEMORY_MAPS) { if constexpr (USE_MEMORY_MAPS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes); auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
@ -624,7 +626,8 @@ void BufferCache<P>::PopAsyncBuffers() {
common_ranges.subtract(base_interval); common_ranges.subtract(base_interval);
} }
}); });
async_downloads -= std::make_pair(IntervalType(cpu_addr, cpu_addr + copy.size), 1); const IntervalType subtract_interval{cpu_addr, cpu_addr + copy.size};
RemoveEachInOverlapCounter(async_downloads, subtract_interval, -1);
} }
runtime.FreeDeferredStagingBuffer(*async_buffer); runtime.FreeDeferredStagingBuffer(*async_buffer);
async_buffers.pop_front(); async_buffers.pop_front();
@ -1198,10 +1201,8 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
const IntervalType base_interval{cpu_addr, cpu_addr + size}; const IntervalType base_interval{cpu_addr, cpu_addr + size};
common_ranges.add(base_interval); common_ranges.add(base_interval);
if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) { uncommitted_ranges.add(base_interval);
uncommitted_ranges.add(base_interval); pending_ranges.add(base_interval);
pending_ranges.add(base_interval);
}
} }
template <class P> template <class P>
@ -1542,7 +1543,7 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
.size = new_size, .size = new_size,
}); });
// Align up to avoid cache conflicts // Align up to avoid cache conflicts
constexpr u64 align = 8ULL; constexpr u64 align = 64ULL;
constexpr u64 mask = ~(align - 1ULL); constexpr u64 mask = ~(align - 1ULL);
total_size_bytes += (new_size + align - 1) & mask; total_size_bytes += (new_size + align - 1) & mask;
largest_copy = std::max(largest_copy, new_size); largest_copy = std::max(largest_copy, new_size);

View file

@ -345,13 +345,30 @@ private:
if (inter_addr < start_address) { if (inter_addr < start_address) {
inter_addr = start_address; inter_addr = start_address;
} }
if (it->second <= 0) {
__debugbreak();
}
func(inter_addr, inter_addr_end, it->second); func(inter_addr, inter_addr_end, it->second);
} }
} }
void RemoveEachInOverlapCounter(OverlapCounter& current_range, const IntervalType search_interval, int subtract_value) {
bool any_removals = false;
current_range.add(std::make_pair(search_interval, subtract_value));
do {
any_removals = false;
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++) {
if (it->second <= 0) {
any_removals = true;
current_range.erase(it);
break;
}
}
} while (any_removals);
}
static bool IsRangeGranular(VAddr cpu_addr, size_t size) { static bool IsRangeGranular(VAddr cpu_addr, size_t size) {
return (cpu_addr & ~Core::Memory::YUZU_PAGEMASK) == return (cpu_addr & ~Core::Memory::YUZU_PAGEMASK) ==
((cpu_addr + size) & ~Core::Memory::YUZU_PAGEMASK); ((cpu_addr + size) & ~Core::Memory::YUZU_PAGEMASK);
@ -554,6 +571,8 @@ private:
u64 minimum_memory = 0; u64 minimum_memory = 0;
u64 critical_memory = 0; u64 critical_memory = 0;
bool active_async_buffers = false;
std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table; std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
}; };

View file

@ -273,7 +273,7 @@ public:
untracked_words[word_index] &= ~bits; untracked_words[word_index] &= ~bits;
NotifyRasterizer<true>(word_index, current_bits, ~u64{0}); NotifyRasterizer<true>(word_index, current_bits, ~u64{0});
} }
const u64 word = current_word; const u64 word = current_word & ~(type == Type::GPU ? untracked_words[word_index] : 0);
u64 page = page_begin; u64 page = page_begin;
page_begin = 0; page_begin = 0;
@ -321,6 +321,7 @@ public:
[[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept { [[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept {
static_assert(type != Type::Untracked); static_assert(type != Type::Untracked);
const u64* const untracked_words = Array<Type::Untracked>();
const u64* const state_words = Array<type>(); const u64* const state_words = Array<type>();
const u64 num_query_words = size / BYTES_PER_WORD + 1; const u64 num_query_words = size / BYTES_PER_WORD + 1;
const u64 word_begin = offset / BYTES_PER_WORD; const u64 word_begin = offset / BYTES_PER_WORD;
@ -328,7 +329,8 @@ public:
const u64 page_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE); const u64 page_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE);
u64 page_index = (offset / BYTES_PER_PAGE) % PAGES_PER_WORD; u64 page_index = (offset / BYTES_PER_PAGE) % PAGES_PER_WORD;
for (u64 word_index = word_begin; word_index < word_end; ++word_index, page_index = 0) { for (u64 word_index = word_begin; word_index < word_end; ++word_index, page_index = 0) {
const u64 word = state_words[word_index]; const u64 off_word = type == Type::GPU ? untracked_words[word_index] : 0;
const u64 word = state_words[word_index] & ~off_word;
if (word == 0) { if (word == 0) {
continue; continue;
} }