N-archive/suyu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

gpu_asynch: Simplify synchronization to a simpler consumer->producer scheme.

Browse source
This commit is contained in:
Fernando Sahmkow 2019-06-10 08:36:22 -04:00 committed by FernandoS27
parent 0706d633bf
commit f2e026a1d8
2 changed files with 3 additions and 47 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/video_core/gpu_thread.cpp
View file

@ -21,7 +21,7 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
MicroProfileOnThreadCreate("GpuThread");
// Wait for first GPU command before acquiring the window context
state.WaitForCommands();
while (state.queue.Empty());
// If emulation was stopped during disk shader loading, abort before trying to acquire context
if (!state.is_running) {
@ -32,7 +32,6 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
CommandDataContainer next;
while (state.is_running) {
state.WaitForCommands();
while (!state.queue.Empty()) {
state.queue.Pop(next);
if (const auto submit_list = std::get_if<SubmitListCommand>(&next.data)) {
@ -49,8 +48,7 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
} else {
UNREACHABLE();
}
state.signaled_fence = next.fence;
state.TrySynchronize();
state.signaled_fence.store(next.fence);
}
}
}
@ -100,22 +98,12 @@ void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
u64 ThreadManager::PushCommand(CommandData&& command_data) {
const u64 fence{++state.last_fence};
state.queue.Push(CommandDataContainer(std::move(command_data), fence));
state.SignalCommands();
return fence;
}
MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
void SynchState::WaitForSynchronization(u64 fence) {
if (signaled_fence >= fence) {
return;
}
// Wait for the GPU to be idle (all commands to be executed)
{
MICROPROFILE_SCOPE(GPU_wait);
std::unique_lock lock{synchronization_mutex};
synchronization_condition.wait(lock, [this, fence] { return signaled_fence >= fence; });
}
while (signaled_fence.load() < fence);
}
} // namespace VideoCommon::GPUThread

32
src/video_core/gpu_thread.h
View file

@ -88,41 +88,9 @@ struct CommandDataContainer {
/// Struct used to synchronize the GPU thread
struct SynchState final {
std::atomic_bool is_running{true};
std::atomic_int queued_frame_count{};
std::mutex synchronization_mutex;
std::mutex commands_mutex;
std::condition_variable commands_condition;
std::condition_variable synchronization_condition;
/// Returns true if the gap in GPU commands is small enough that we can consider the CPU and GPU
/// synchronized. This is entirely empirical.
bool IsSynchronized() const {
constexpr std::size_t max_queue_gap{5};
return queue.Size() <= max_queue_gap;
}
void TrySynchronize() {
if (IsSynchronized()) {
std::lock_guard lock{synchronization_mutex};
synchronization_condition.notify_one();
}
}
void WaitForSynchronization(u64 fence);
void SignalCommands() {
if (queue.Empty()) {
return;
}
commands_condition.notify_one();
}
void WaitForCommands() {
std::unique_lock lock{commands_mutex};
commands_condition.wait(lock, [this] { return !queue.Empty(); });
}
using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
CommandQueue queue;
u64 last_fence{};