diff options
author | Markus Wick <markus@selfnet.de> | 2021-04-07 11:41:31 +0200 |
---|---|---|
committer | Markus Wick <markus@selfnet.de> | 2021-04-07 22:38:52 +0200 |
commit | e8bd9aed8bf0f60455d0ae6a8f6f3abf92dd8305 (patch) | |
tree | 042497b8c297e874b382f4b695bb601b6ea2d1cf /src/video_core/gpu_thread.cpp | |
parent | e6fb49fa4bb2864702abcefc14f6bb62eaba7a7e (diff) |
video_core: Use a CV for blocking commands.
There is no need for a busy loop here. Let's just use a condition variable to save some power.
Diffstat (limited to 'src/video_core/gpu_thread.cpp')
-rw-r--r-- | src/video_core/gpu_thread.cpp | 43 |
1 files changed, 26 insertions, 17 deletions
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp index 9488bf5444..7addfbc7bb 100644 --- a/src/video_core/gpu_thread.cpp +++ b/src/video_core/gpu_thread.cpp @@ -56,11 +56,17 @@ static void RunThread(Core::System& system, VideoCore::RendererBase& renderer, } else if (const auto* invalidate = std::get_if<InvalidateRegionCommand>(&next.data)) { rasterizer->OnCPUWrite(invalidate->addr, invalidate->size); } else if (std::holds_alternative<EndProcessingCommand>(next.data)) { - return; + ASSERT(state.is_running == false); } else { UNREACHABLE(); } state.signaled_fence.store(next.fence); + if (next.block) { + // We have to lock the write_lock to ensure that the condition_variable wait not get a + // race between the check and the lock itself. + std::lock_guard lk(state.write_lock); + state.cv.notify_all(); + } } } @@ -105,9 +111,8 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) { case Settings::GPUAccuracy::Extreme: { auto& gpu = system.GPU(); u64 fence = gpu.RequestFlush(addr, size); - PushCommand(GPUTickCommand()); - while (fence > gpu.CurrentFlushRequestFence()) { - } + PushCommand(GPUTickCommand(), true); + ASSERT(fence <= gpu.CurrentFlushRequestFence()); break; } default: @@ -124,18 +129,16 @@ void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) { rasterizer->OnCPUWrite(addr, size); } -void ThreadManager::WaitIdle() const { - while (state.last_fence > state.signaled_fence.load(std::memory_order_relaxed) && - state.is_running) { - } -} - void ThreadManager::ShutDown() { if (!state.is_running) { return; } - state.is_running = false; + { + std::lock_guard lk(state.write_lock); + state.is_running = false; + state.cv.notify_all(); + } if (!thread.joinable()) { return; @@ -150,15 +153,21 @@ void ThreadManager::OnCommandListEnd() { PushCommand(OnCommandListEndCommand()); } -u64 ThreadManager::PushCommand(CommandData&& command_data) { +u64 ThreadManager::PushCommand(CommandData&& command_data, bool block) { + if (!is_async) { + // In synchronous GPU mode, block the caller until the command has executed + block = true; + } + std::unique_lock lk(state.write_lock); const u64 fence{++state.last_fence}; - state.queue.Push(CommandDataContainer(std::move(command_data), fence)); + state.queue.Push(CommandDataContainer(std::move(command_data), fence, block)); - if (!is_async) { - // In synchronous GPU mode, block the caller until the command has executed - lk.unlock(); - WaitIdle(); + if (block) { + state.cv.wait(lk, [this, fence] { + return fence <= state.signaled_fence.load(std::memory_order_relaxed) || + !state.is_running; + }); } return fence; |