octo/yuzu-android
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Buffer cache: always use async buffer downloads and fix regression.

Browse Source
This commit is contained in:
Fernando Sahmkow
2023-05-04 13:23:36 +02:00
parent 6f90dff293
commit 36c302fa32
4 changed files with 70 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/tests/video_core/memory_tracker.cpp
View File

@@ -535,12 +535,12 @@ TEST_CASE("MemoryTracker: Cached write downloads") {
memory_track->MarkRegionAsGpuModified(c + PAGE, PAGE); memory_track->MarkRegionAsGpuModified(c + PAGE, PAGE);
int num = 0; int num = 0;
memory_track->ForEachDownloadRangeAndClear(c, WORD, [&](u64 offset, u64 size) { ++num; }); memory_track->ForEachDownloadRangeAndClear(c, WORD, [&](u64 offset, u64 size) { ++num; });
REQUIRE(num == 1); REQUIRE(num == 0);
num = 0; num = 0;
memory_track->ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; }); memory_track->ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
REQUIRE(num == 0); REQUIRE(num == 0);
REQUIRE(!memory_track->IsRegionCpuModified(c + PAGE, PAGE)); REQUIRE(!memory_track->IsRegionCpuModified(c + PAGE, PAGE));
REQUIRE(!memory_track->IsRegionGpuModified(c + PAGE, PAGE)); REQUIRE(memory_track->IsRegionGpuModified(c + PAGE, PAGE));
memory_track->FlushCachedWrites(); memory_track->FlushCachedWrites();
REQUIRE(memory_track->IsRegionCpuModified(c + PAGE, PAGE)); REQUIRE(memory_track->IsRegionCpuModified(c + PAGE, PAGE));
REQUIRE(!memory_track->IsRegionGpuModified(c + PAGE, PAGE)); REQUIRE(!memory_track->IsRegionGpuModified(c + PAGE, PAGE));

114
src/video_core/buffer_cache/buffer_cache.h
View File

@@ -23,8 +23,6 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
common_ranges.clear(); common_ranges.clear();
inline_buffer_id = NULL_BUFFER_ID; inline_buffer_id = NULL_BUFFER_ID;
active_async_buffers = !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;
@@ -75,8 +73,6 @@ 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 = !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;
@@ -491,9 +487,8 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
if (committed_ranges.empty()) { if (committed_ranges.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { 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;
} }
@@ -554,64 +549,65 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
committed_ranges.clear(); committed_ranges.clear();
if (downloads.empty()) { if (downloads.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { 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 (active_async_buffers) { if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true); boost::container::small_vector<BufferCopy, 4> normalized_copies;
boost::container::small_vector<BufferCopy, 4> normalized_copies; IntervalSet new_async_range{};
IntervalSet new_async_range{}; runtime.PreCopyBarrier();
runtime.PreCopyBarrier(); for (auto& [copy, buffer_id] : downloads) {
for (auto& [copy, buffer_id] : downloads) { copy.dst_offset += download_staging.offset;
copy.dst_offset += download_staging.offset; const std::array copies{copy};
const std::array copies{copy}; BufferCopy second_copy{copy};
BufferCopy second_copy{copy}; Buffer& buffer = slot_buffers[buffer_id];
Buffer& buffer = slot_buffers[buffer_id]; second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset; VAddr orig_cpu_addr = static_cast<VAddr>(second_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};
const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size}; async_downloads += std::make_pair(base_interval, 1);
async_downloads += std::make_pair(base_interval, 1); runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false); normalized_copies.push_back(second_copy);
normalized_copies.push_back(second_copy); }
} runtime.PostCopyBarrier();
runtime.PostCopyBarrier(); pending_downloads.emplace_back(std::move(normalized_copies));
pending_downloads.emplace_back(std::move(normalized_copies)); async_buffers.emplace_back(download_staging);
async_buffers.emplace_back(download_staging); } else {
} else { if (!Settings::IsGPULevelHigh()) {
committed_ranges.clear(); committed_ranges.clear();
uncommitted_ranges.clear(); uncommitted_ranges.clear();
}
} 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 { } else {
const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy); if constexpr (USE_MEMORY_MAPS) {
for (const auto& [copy, buffer_id] : downloads) { auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
Buffer& buffer = slot_buffers[buffer_id]; runtime.PreCopyBarrier();
buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size)); for (auto& [copy, buffer_id] : downloads) {
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset; // Have in mind the staging buffer offset for the copy
cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size); 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) {
Buffer& buffer = slot_buffers[buffer_id];
buffer.ImmediateDownload(copy.src_offset,
immediate_buffer.subspan(0, copy.size));
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
}
} }
} }
} }

2
src/video_core/buffer_cache/buffer_cache_base.h
View File

@@ -572,8 +572,6 @@ private:
u64 critical_memory = 0; u64 critical_memory = 0;
BufferId inline_buffer_id; BufferId inline_buffer_id;
bool active_async_buffers = false;
std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table; std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
std::vector<u8> tmp_buffer; std::vector<u8> tmp_buffer;
}; };

13
src/video_core/buffer_cache/word_manager.h
View File

@@ -302,6 +302,9 @@ public:
(pending_pointer - pending_offset) * BYTES_PER_PAGE); (pending_pointer - pending_offset) * BYTES_PER_PAGE);
}; };
IterateWords(offset, size, [&](size_t index, u64 mask) { IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask; const u64 word = state_words[index] & mask;
if constexpr (clear) { if constexpr (clear) {
if constexpr (type == Type::CPU || type == Type::CachedCPU) { if constexpr (type == Type::CPU || type == Type::CachedCPU) {
@@ -350,8 +353,13 @@ public:
static_assert(type != Type::Untracked); static_assert(type != Type::Untracked);
const std::span<const u64> state_words = words.template Span<type>(); const std::span<const u64> state_words = words.template Span<type>();
[[maybe_unused]] const std::span<const u64> untracked_words =
words.template Span<Type::Untracked>();
bool result = false; bool result = false;
IterateWords(offset, size, [&](size_t index, u64 mask) { IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask; const u64 word = state_words[index] & mask;
if (word != 0) { if (word != 0) {
result = true; result = true;
@@ -372,9 +380,14 @@ public:
[[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept { [[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept {
static_assert(type != Type::Untracked); static_assert(type != Type::Untracked);
const std::span<const u64> state_words = words.template Span<type>(); const std::span<const u64> state_words = words.template Span<type>();
[[maybe_unused]] const std::span<const u64> untracked_words =
words.template Span<Type::Untracked>();
u64 begin = std::numeric_limits<u64>::max(); u64 begin = std::numeric_limits<u64>::max();
u64 end = 0; u64 end = 0;
IterateWords(offset, size, [&](size_t index, u64 mask) { IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask; const u64 word = state_words[index] & mask;
if (word == 0) { if (word == 0) {
return; return;