1
0
Fork 0
forked from suyu/suyu

Fence Manager: Force ordering on WFI.

This commit is contained in:
Fernando Sahmkow 2021-07-07 16:42:26 +02:00
parent 73638ca593
commit cf38faee9b
4 changed files with 71 additions and 38 deletions

View file

@ -152,6 +152,8 @@ public:
/// Return true when there are uncommitted buffers to be downloaded /// Return true when there are uncommitted buffers to be downloaded
[[nodiscard]] bool HasUncommittedFlushes() const noexcept; [[nodiscard]] bool HasUncommittedFlushes() const noexcept;
void AccumulateFlushes();
/// Return true when the caller should wait for async downloads /// Return true when the caller should wait for async downloads
[[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept; [[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept;
@ -334,6 +336,7 @@ private:
std::vector<BufferId> cached_write_buffer_ids; std::vector<BufferId> cached_write_buffer_ids;
IntervalSet uncommitted_ranges; IntervalSet uncommitted_ranges;
std::deque<IntervalSet> committed_ranges;
size_t immediate_buffer_capacity = 0; size_t immediate_buffer_capacity = 0;
std::unique_ptr<u8[]> immediate_buffer_alloc; std::unique_ptr<u8[]> immediate_buffer_alloc;
@ -551,7 +554,19 @@ void BufferCache<P>::FlushCachedWrites() {
template <class P> template <class P>
bool BufferCache<P>::HasUncommittedFlushes() const noexcept { bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
return !uncommitted_ranges.empty(); return !uncommitted_ranges.empty() || !committed_ranges.empty();
}
template <class P>
void BufferCache<P>::AccumulateFlushes() {
if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {
uncommitted_ranges.clear();
return;
}
if (uncommitted_ranges.empty()) {
return;
}
committed_ranges.emplace_back(std::move(uncommitted_ranges));
} }
template <class P> template <class P>
@ -561,8 +576,8 @@ bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
template <class P> template <class P>
void BufferCache<P>::CommitAsyncFlushesHigh() { void BufferCache<P>::CommitAsyncFlushesHigh() {
const IntervalSet& intervals = uncommitted_ranges; AccumulateFlushes();
if (intervals.empty()) { if (committed_ranges.empty()) {
return; return;
} }
MICROPROFILE_SCOPE(GPU_DownloadMemory); MICROPROFILE_SCOPE(GPU_DownloadMemory);
@ -570,6 +585,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads; boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads;
u64 total_size_bytes = 0; u64 total_size_bytes = 0;
u64 largest_copy = 0; u64 largest_copy = 0;
for (const IntervalSet& intervals : committed_ranges) {
for (auto& interval : intervals) { for (auto& interval : intervals) {
const std::size_t size = interval.upper() - interval.lower(); const std::size_t size = interval.upper() - interval.lower();
const VAddr cpu_addr = interval.lower(); const VAddr cpu_addr = interval.lower();
@ -607,6 +623,8 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
}); });
}); });
} }
}
committed_ranges.clear();
if (downloads.empty()) { if (downloads.empty()) {
return; return;
} }
@ -644,6 +662,7 @@ void BufferCache<P>::CommitAsyncFlushes() {
CommitAsyncFlushesHigh(); CommitAsyncFlushesHigh();
} else { } else {
uncommitted_ranges.clear(); uncommitted_ranges.clear();
committed_ranges.clear();
} }
} }

View file

@ -54,6 +54,12 @@ public:
delayed_destruction_ring.Tick(); delayed_destruction_ring.Tick();
} }
// Unlike other fences, this one doesn't
void SignalOrdering() {
std::scoped_lock lock{buffer_cache.mutex};
buffer_cache.AccumulateFlushes();
}
void SignalReference() { void SignalReference() {
// Only sync references on High // Only sync references on High
if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) { if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {

View file

@ -657,6 +657,10 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
void RasterizerOpenGL::WaitForIdle() { void RasterizerOpenGL::WaitForIdle() {
glMemoryBarrier(GL_ALL_BARRIER_BITS); glMemoryBarrier(GL_ALL_BARRIER_BITS);
if (!gpu.IsAsync()) {
return;
}
fence_manager.SignalOrdering();
} }
void RasterizerOpenGL::FragmentBarrier() { void RasterizerOpenGL::FragmentBarrier() {

View file

@ -619,6 +619,10 @@ void RasterizerVulkan::WaitForIdle() {
cmdbuf.SetEvent(event, flags); cmdbuf.SetEvent(event, flags);
cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {}); cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {});
}); });
if (!gpu.IsAsync()) {
return;
}
fence_manager.SignalOrdering();
} }
void RasterizerVulkan::FragmentBarrier() { void RasterizerVulkan::FragmentBarrier() {