From 76525432317fcc4b4847c5414bed6660c85e582b Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Mon, 3 Jan 2022 20:28:54 -0500
Subject: [PATCH 1/2] Revert "Merge pull request #7668 from
ameerj/fence-stop-token"
This reverts commit e7733544779f2706d108682dd027d44e7fa5ff4b, reversing
changes made to abbbdc2bc027ed7af236625ae8427a46df63f7e7.
---
src/core/hle/service/nvflinger/nvflinger.cpp | 3 +--
src/video_core/gpu.cpp | 19 +++++++++++++------
src/video_core/gpu.h | 3 +--
3 files changed, 15 insertions(+), 10 deletions(-)
diff --git a/src/core/hle/service/nvflinger/nvflinger.cpp b/src/core/hle/service/nvflinger/nvflinger.cpp
index 396cc5afa4..a22811ec12 100644
--- a/src/core/hle/service/nvflinger/nvflinger.cpp
+++ b/src/core/hle/service/nvflinger/nvflinger.cpp
@@ -266,11 +266,10 @@ void NVFlinger::Compose() {
auto& gpu = system.GPU();
const auto& multi_fence = buffer->get().multi_fence;
- const auto stop_token = vsync_thread.get_stop_token();
guard->unlock();
for (u32 fence_id = 0; fence_id < multi_fence.num_fences; fence_id++) {
const auto& fence = multi_fence.fences[fence_id];
- gpu.WaitFence(fence.id, fence.value, stop_token);
+ gpu.WaitFence(fence.id, fence.value);
}
guard->lock();
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index d988741503..8788f5148f 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -206,7 +206,7 @@ struct GPU::Impl {
}
/// Allows the CPU/NvFlinger to wait on the GPU before presenting a frame.
- void WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token = {}) {
+ void WaitFence(u32 syncpoint_id, u32 value) {
// Synced GPU, is always in sync
if (!is_async) {
return;
@@ -218,8 +218,13 @@ struct GPU::Impl {
}
MICROPROFILE_SCOPE(GPU_wait);
std::unique_lock lock{sync_mutex};
- sync_cv.wait(lock, stop_token,
- [=, this] { return syncpoints.at(syncpoint_id).load() >= value; });
+ sync_cv.wait(lock, [=, this] {
+ if (shutting_down.load(std::memory_order_relaxed)) {
+ // We're shutting down, ensure no threads continue to wait for the next syncpoint
+ return true;
+ }
+ return syncpoints.at(syncpoint_id).load() >= value;
+ });
}
void IncrementSyncPoint(u32 syncpoint_id) {
@@ -665,6 +670,8 @@ struct GPU::Impl {
std::unique_ptr<Engines::KeplerMemory> kepler_memory;
/// Shader build notifier
std::unique_ptr<VideoCore::ShaderNotify> shader_notify;
+ /// When true, we are about to shut down emulation session, so terminate outstanding tasks
+ std::atomic_bool shutting_down{};
std::array<std::atomic<u32>, Service::Nvidia::MaxSyncPoints> syncpoints{};
@@ -673,7 +680,7 @@ struct GPU::Impl {
std::mutex sync_mutex;
std::mutex device_mutex;
- std::condition_variable_any sync_cv;
+ std::condition_variable sync_cv;
struct FlushRequest {
explicit FlushRequest(u64 fence_, VAddr addr_, std::size_t size_)
@@ -812,8 +819,8 @@ const VideoCore::ShaderNotify& GPU::ShaderNotify() const {
return impl->ShaderNotify();
}
-void GPU::WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token) {
- impl->WaitFence(syncpoint_id, value, stop_token);
+void GPU::WaitFence(u32 syncpoint_id, u32 value) {
+ impl->WaitFence(syncpoint_id, value);
}
void GPU::IncrementSyncPoint(u32 syncpoint_id) {
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index cc65a78704..500411176f 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -5,7 +5,6 @@
#pragma once
#include <memory>
-#include <stop_token>
#include "common/bit_field.h"
#include "common/common_types.h"
@@ -210,7 +209,7 @@ public:
[[nodiscard]] const VideoCore::ShaderNotify& ShaderNotify() const;
/// Allows the CPU/NvFlinger to wait on the GPU before presenting a frame.
- void WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token = {});
+ void WaitFence(u32 syncpoint_id, u32 value);
void IncrementSyncPoint(u32 syncpoint_id);
From c17938f96ba9a3f2571387b21328743db8050250 Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Mon, 3 Jan 2022 20:31:51 -0500
Subject: [PATCH 2/2] gpu: Add shut down method to synchronize threads before
destruction
---
src/core/core.cpp | 2 ++
src/video_core/gpu.cpp | 10 ++++++++++
src/video_core/gpu.h | 3 +++
3 files changed, 15 insertions(+)
diff --git a/src/core/core.cpp b/src/core/core.cpp
index aa96f709b5..3f9a7f44bc 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -317,6 +317,8 @@ struct System::Impl {
is_powered_on = false;
exit_lock = false;
+ gpu_core->NotifyShutdown();
+
services.reset();
service_manager.reset();
cheat_engine.reset();
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 8788f5148f..44fda27ef9 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -312,6 +312,12 @@ struct GPU::Impl {
cpu_context->MakeCurrent();
}
+ void NotifyShutdown() {
+ std::unique_lock lk{sync_mutex};
+ shutting_down.store(true, std::memory_order::relaxed);
+ sync_cv.notify_all();
+ }
+
/// Obtain the CPU Context
void ObtainContext() {
cpu_context->MakeCurrent();
@@ -859,6 +865,10 @@ void GPU::Start() {
impl->Start();
}
+void GPU::NotifyShutdown() {
+ impl->NotifyShutdown();
+}
+
void GPU::ObtainContext() {
impl->ObtainContext();
}
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 500411176f..3188b83edf 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -232,6 +232,9 @@ public:
/// core timing events.
void Start();
+ /// Performs any additional necessary steps to shutdown GPU emulation.
+ void NotifyShutdown();
+
/// Obtain the CPU Context
void ObtainContext();