From 084ceb925acad470b69467d64e4dfbb3bd7ef3f1 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 18 Feb 2020 16:51:42 -0400
Subject: [PATCH 01/31] UI: Replasce accurate GPU option for GPU Accuracy Level
---
src/core/settings.cpp | 2 +-
src/core/settings.h | 16 ++++++++-
src/core/telemetry_session.cpp | 16 +++++++--
src/video_core/engines/maxwell_dma.cpp | 2 +-
.../renderer_opengl/gl_rasterizer.cpp | 2 +-
src/video_core/texture_cache/texture_cache.h | 8 ++---
src/yuzu/configuration/config.cpp | 7 ++--
.../configure_graphics_advanced.cpp | 5 +--
.../configure_graphics_advanced.ui | 33 ++++++++++++++++---
src/yuzu_cmd/config.cpp | 4 +--
src/yuzu_cmd/default_ini.h | 6 ++--
src/yuzu_tester/config.cpp | 4 +--
12 files changed, 77 insertions(+), 28 deletions(-)
diff --git a/src/core/settings.cpp b/src/core/settings.cpp
index c1282cb804..445047469e 100644
--- a/src/core/settings.cpp
+++ b/src/core/settings.cpp
@@ -92,7 +92,7 @@ void LogSettings() {
LogSetting("Renderer_UseFrameLimit", Settings::values.use_frame_limit);
LogSetting("Renderer_FrameLimit", Settings::values.frame_limit);
LogSetting("Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
- LogSetting("Renderer_UseAccurateGpuEmulation", Settings::values.use_accurate_gpu_emulation);
+ LogSetting("Renderer_GPUAccuracyLevel", Settings::values.gpu_accuracy);
LogSetting("Renderer_UseAsynchronousGpuEmulation",
Settings::values.use_asynchronous_gpu_emulation);
LogSetting("Renderer_UseVsync", Settings::values.use_vsync);
diff --git a/src/core/settings.h b/src/core/settings.h
index c73d1c5967..b54a0d4ea9 100644
--- a/src/core/settings.h
+++ b/src/core/settings.h
@@ -376,6 +376,12 @@ enum class RendererBackend {
Vulkan = 1,
};
+enum class GPUAccuracy : u32 {
+ Normal = 0,
+ High = 1,
+ Extreme = 2,
+};
+
struct Values {
// System
bool use_docked_mode;
@@ -436,7 +442,7 @@ struct Values {
bool use_frame_limit;
u16 frame_limit;
bool use_disk_shader_cache;
- bool use_accurate_gpu_emulation;
+ GPUAccuracy gpu_accuracy;
bool use_asynchronous_gpu_emulation;
bool use_vsync;
bool force_30fps_mode;
@@ -480,6 +486,14 @@ struct Values {
std::map<u64, std::vector<std::string>> disabled_addons;
} extern values;
+constexpr bool IsGPULevelExtreme() {
+ return values.gpu_accuracy == GPUAccuracy::Extreme;
+}
+
+constexpr bool IsGPULevelHigh() {
+ return values.gpu_accuracy == GPUAccuracy::Extreme || values.gpu_accuracy == GPUAccuracy::High;
+}
+
void Apply();
void LogSettings();
} // namespace Settings
diff --git a/src/core/telemetry_session.cpp b/src/core/telemetry_session.cpp
index fd5a3ee9f1..1c3b03a1c2 100644
--- a/src/core/telemetry_session.cpp
+++ b/src/core/telemetry_session.cpp
@@ -56,6 +56,18 @@ static const char* TranslateRenderer(Settings::RendererBackend backend) {
return "Unknown";
}
+static const char* TranslateGPUAccuracyLevel(Settings::GPUAccuracy backend) {
+ switch (backend) {
+ case Settings::GPUAccuracy::Normal:
+ return "Normal";
+ case Settings::GPUAccuracy::High:
+ return "High";
+ case Settings::GPUAccuracy::Extreme:
+ return "Extreme";
+ }
+ return "Unknown";
+}
+
u64 GetTelemetryId() {
u64 telemetry_id{};
const std::string filename{FileUtil::GetUserPath(FileUtil::UserPath::ConfigDir) +
@@ -184,8 +196,8 @@ void TelemetrySession::AddInitialInfo(Loader::AppLoader& app_loader) {
AddField(field_type, "Renderer_UseFrameLimit", Settings::values.use_frame_limit);
AddField(field_type, "Renderer_FrameLimit", Settings::values.frame_limit);
AddField(field_type, "Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
- AddField(field_type, "Renderer_UseAccurateGpuEmulation",
- Settings::values.use_accurate_gpu_emulation);
+ AddField(field_type, "Renderer_GPUAccuracyLevel",
+ TranslateGPUAccuracyLevel(Settings::values.gpu_accuracy));
AddField(field_type, "Renderer_UseAsynchronousGpuEmulation",
Settings::values.use_asynchronous_gpu_emulation);
AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync);
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index c2610f992a..32b04e31ea 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -136,7 +136,7 @@ void MaxwellDMA::HandleCopy() {
write_buffer.resize(dst_size);
}
- if (Settings::values.use_accurate_gpu_emulation) {
+ if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(source, read_buffer.data(), src_size);
memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
} else {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 175374f0df..ac4485a18a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -661,7 +661,7 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
}
void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
- if (Settings::values.use_accurate_gpu_emulation) {
+ if (Settings::IsGPULevelExtreme()) {
FlushRegion(addr, size);
}
InvalidateRegion(addr, size);
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 69ca08fd10..7432691d1f 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -417,7 +417,7 @@ private:
**/
RecycleStrategy PickStrategy(std::vector<TSurface>& overlaps, const SurfaceParams& params,
const GPUVAddr gpu_addr, const MatchTopologyResult untopological) {
- if (Settings::values.use_accurate_gpu_emulation) {
+ if (Settings::IsGPULevelExtreme()) {
return RecycleStrategy::Flush;
}
// 3D Textures decision
@@ -461,7 +461,7 @@ private:
}
switch (PickStrategy(overlaps, params, gpu_addr, untopological)) {
case RecycleStrategy::Ignore: {
- return InitializeSurface(gpu_addr, params, Settings::values.use_accurate_gpu_emulation);
+ return InitializeSurface(gpu_addr, params, Settings::IsGPULevelExtreme());
}
case RecycleStrategy::Flush: {
std::sort(overlaps.begin(), overlaps.end(),
@@ -598,7 +598,7 @@ private:
if (passed_tests == 0) {
return {};
// In Accurate GPU all tests should pass, else we recycle
- } else if (Settings::values.use_accurate_gpu_emulation && passed_tests != overlaps.size()) {
+ } else if (Settings::IsGPULevelExtreme() && passed_tests != overlaps.size()) {
return {};
}
for (const auto& surface : overlaps) {
@@ -668,7 +668,7 @@ private:
for (const auto& surface : overlaps) {
if (!surface->MatchTarget(params.target)) {
if (overlaps.size() == 1 && surface->GetCpuAddr() == cpu_addr) {
- if (Settings::values.use_accurate_gpu_emulation) {
+ if (Settings::IsGPULevelExtreme()) {
return std::nullopt;
}
Unregister(surface);
diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp
index 7f6dfac849..da1fa4e029 100644
--- a/src/yuzu/configuration/config.cpp
+++ b/src/yuzu/configuration/config.cpp
@@ -639,8 +639,8 @@ void Config::ReadRendererValues() {
Settings::values.frame_limit = ReadSetting(QStringLiteral("frame_limit"), 100).toInt();
Settings::values.use_disk_shader_cache =
ReadSetting(QStringLiteral("use_disk_shader_cache"), true).toBool();
- Settings::values.use_accurate_gpu_emulation =
- ReadSetting(QStringLiteral("use_accurate_gpu_emulation"), false).toBool();
+ const int gpu_accuracy_level = ReadSetting(QStringLiteral("gpu_accuracy"), 0).toInt();
+ Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
Settings::values.use_asynchronous_gpu_emulation =
ReadSetting(QStringLiteral("use_asynchronous_gpu_emulation"), false).toBool();
Settings::values.use_vsync = ReadSetting(QStringLiteral("use_vsync"), true).toBool();
@@ -1080,8 +1080,7 @@ void Config::SaveRendererValues() {
WriteSetting(QStringLiteral("frame_limit"), Settings::values.frame_limit, 100);
WriteSetting(QStringLiteral("use_disk_shader_cache"), Settings::values.use_disk_shader_cache,
true);
- WriteSetting(QStringLiteral("use_accurate_gpu_emulation"),
- Settings::values.use_accurate_gpu_emulation, false);
+ WriteSetting(QStringLiteral("gpu_accuracy"), static_cast<int>(Settings::values.gpu_accuracy), 0);
WriteSetting(QStringLiteral("use_asynchronous_gpu_emulation"),
Settings::values.use_asynchronous_gpu_emulation, false);
WriteSetting(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
diff --git a/src/yuzu/configuration/configure_graphics_advanced.cpp b/src/yuzu/configuration/configure_graphics_advanced.cpp
index b9f429f840..0a3f47339d 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.cpp
+++ b/src/yuzu/configuration/configure_graphics_advanced.cpp
@@ -19,7 +19,7 @@ ConfigureGraphicsAdvanced::~ConfigureGraphicsAdvanced() = default;
void ConfigureGraphicsAdvanced::SetConfiguration() {
const bool runtime_lock = !Core::System::GetInstance().IsPoweredOn();
- ui->use_accurate_gpu_emulation->setChecked(Settings::values.use_accurate_gpu_emulation);
+ ui->gpu_accuracy->setCurrentIndex(static_cast<int>(Settings::values.gpu_accuracy));
ui->use_vsync->setEnabled(runtime_lock);
ui->use_vsync->setChecked(Settings::values.use_vsync);
ui->force_30fps_mode->setEnabled(runtime_lock);
@@ -29,7 +29,8 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
}
void ConfigureGraphicsAdvanced::ApplyConfiguration() {
- Settings::values.use_accurate_gpu_emulation = ui->use_accurate_gpu_emulation->isChecked();
+ auto gpu_accuracy = static_cast<Settings::GPUAccuracy>(ui->gpu_accuracy->currentIndex());
+ Settings::values.gpu_accuracy = gpu_accuracy;
Settings::values.use_vsync = ui->use_vsync->isChecked();
Settings::values.force_30fps_mode = ui->force_30fps_mode->isChecked();
Settings::values.max_anisotropy = ui->anisotropic_filtering_combobox->currentIndex();
diff --git a/src/yuzu/configuration/configure_graphics_advanced.ui b/src/yuzu/configuration/configure_graphics_advanced.ui
index 42eec278ed..0c7b383e00 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.ui
+++ b/src/yuzu/configuration/configure_graphics_advanced.ui
@@ -23,11 +23,34 @@
</property>
<layout class="QVBoxLayout" name="verticalLayout_3">
<item>
- <widget class="QCheckBox" name="use_accurate_gpu_emulation">
- <property name="text">
- <string>Use accurate GPU emulation (slow)</string>
- </property>
- </widget>
+ <layout class="QHBoxLayout" name="horizontalLayout_2">
+ <item>
+ <widget class="QLabel" name="label_gpu_accuracy">
+ <property name="text">
+ <string>Accuracy Level:</string>
+ </property>
+ </widget>
+ </item>
+ <item>
+ <widget class="QComboBox" name="gpu_accuracy">
+ <item>
+ <property name="text">
+ <string notr="true">Normal</string>
+ </property>
+ </item>
+ <item>
+ <property name="text">
+ <string notr="true">High</string>
+ </property>
+ </item>
+ <item>
+ <property name="text">
+ <string notr="true">Extreme(very slow)</string>
+ </property>
+ </item>
+ </widget>
+ </item>
+ </layout>
</item>
<item>
<widget class="QCheckBox" name="use_vsync">
diff --git a/src/yuzu_cmd/config.cpp b/src/yuzu_cmd/config.cpp
index 80341747f9..d1ac354bf1 100644
--- a/src/yuzu_cmd/config.cpp
+++ b/src/yuzu_cmd/config.cpp
@@ -388,8 +388,8 @@ void Config::ReadValues() {
static_cast<u16>(sdl2_config->GetInteger("Renderer", "frame_limit", 100));
Settings::values.use_disk_shader_cache =
sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false);
- Settings::values.use_accurate_gpu_emulation =
- sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
+ const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 0);
+ Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
Settings::values.use_asynchronous_gpu_emulation =
sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
Settings::values.use_vsync =
diff --git a/src/yuzu_cmd/default_ini.h b/src/yuzu_cmd/default_ini.h
index 171d16fa00..60b1a62fad 100644
--- a/src/yuzu_cmd/default_ini.h
+++ b/src/yuzu_cmd/default_ini.h
@@ -146,9 +146,9 @@ frame_limit =
# 0 (default): Off, 1 : On
use_disk_shader_cache =
-# Whether to use accurate GPU emulation
-# 0 (default): Off (fast), 1 : On (slow)
-use_accurate_gpu_emulation =
+# Which gpu accuracy level to use
+# 0 (Normal), 1 (High), 2 (Extreme)
+gpu_accuracy =
# Whether to use asynchronous GPU emulation
# 0 : Off (slow), 1 (default): On (fast)
diff --git a/src/yuzu_tester/config.cpp b/src/yuzu_tester/config.cpp
index ee2591c8fb..c0325cc3ca 100644
--- a/src/yuzu_tester/config.cpp
+++ b/src/yuzu_tester/config.cpp
@@ -126,8 +126,8 @@ void Config::ReadValues() {
Settings::values.frame_limit = 100;
Settings::values.use_disk_shader_cache =
sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false);
- Settings::values.use_accurate_gpu_emulation =
- sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
+ const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 0);
+ Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
Settings::values.use_asynchronous_gpu_emulation =
sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
From a60a22d9c284537be2e3a6a89b8afb37e6f0c510 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Feb 2020 09:33:03 -0400
Subject: [PATCH 02/31] Texture Cache: Implement OnCPUWrite and SyncGuestHost
---
src/video_core/texture_cache/surface_base.h | 18 ++++++++
src/video_core/texture_cache/texture_cache.h | 48 ++++++++++++++++++--
2 files changed, 63 insertions(+), 3 deletions(-)
diff --git a/src/video_core/texture_cache/surface_base.h b/src/video_core/texture_cache/surface_base.h
index c5ab21f567..79e10ffbb1 100644
--- a/src/video_core/texture_cache/surface_base.h
+++ b/src/video_core/texture_cache/surface_base.h
@@ -192,6 +192,22 @@ public:
index = index_;
}
+ void SetMemoryMarked(bool is_memory_marked_) {
+ is_memory_marked = is_memory_marked_;
+ }
+
+ bool IsMemoryMarked() const {
+ return is_memory_marked;
+ }
+
+ void SetSyncPending(bool is_sync_pending_) {
+ is_sync_pending = is_sync_pending_;
+ }
+
+ bool IsSyncPending() const {
+ return is_sync_pending;
+ }
+
void MarkAsPicked(bool is_picked_) {
is_picked = is_picked_;
}
@@ -303,6 +319,8 @@ private:
bool is_target{};
bool is_registered{};
bool is_picked{};
+ bool is_memory_marked{};
+ bool is_sync_pending{};
u32 index{NO_RT};
u64 modification_tick{};
};
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 7432691d1f..441fda53d8 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -6,6 +6,7 @@
#include <algorithm>
#include <array>
+#include <list>
#include <memory>
#include <mutex>
#include <set>
@@ -62,6 +63,30 @@ public:
}
}
+ void OnCPUWrite(CacheAddr addr, std::size_t size) {
+ std::lock_guard lock{mutex};
+
+ for (const auto& surface : GetSurfacesInRegion(addr, size)) {
+ if (surface->IsMemoryMarked()) {
+ Unmark(surface);
+ surface->SetSyncPending(true);
+ marked_for_unregister.emplace_back(surface);
+ }
+ }
+ }
+
+ void SyncGuestHost() {
+ std::lock_guard lock{mutex};
+
+ for (const auto& surface : marked_for_unregister) {
+ if (surface->IsRegistered()) {
+ surface->SetSyncPending(false);
+ Unregister(surface);
+ }
+ }
+ marked_for_unregister.clear();
+ }
+
/**
* Guarantees that rendertargets don't unregister themselves if the
* collide. Protection is currently only done on 3D slices.
@@ -85,7 +110,9 @@ public:
return a->GetModificationTick() < b->GetModificationTick();
});
for (const auto& surface : surfaces) {
+ mutex.unlock();
FlushSurface(surface);
+ mutex.lock();
}
}
@@ -345,9 +372,20 @@ protected:
surface->SetCpuAddr(*cpu_addr);
RegisterInnerCache(surface);
surface->MarkAsRegistered(true);
+ surface->SetMemoryMarked(true);
rasterizer.UpdatePagesCachedCount(*cpu_addr, size, 1);
}
+ void Unmark(TSurface surface) {
+ if (!surface->IsMemoryMarked()) {
+ return;
+ }
+ const std::size_t size = surface->GetSizeInBytes();
+ const VAddr cpu_addr = surface->GetCpuAddr();
+ rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1);
+ surface->SetMemoryMarked(false);
+ }
+
void Unregister(TSurface surface) {
if (guard_render_targets && surface->IsProtected()) {
return;
@@ -355,9 +393,11 @@ protected:
if (!guard_render_targets && surface->IsRenderTarget()) {
ManageRenderTargetUnregister(surface);
}
- const std::size_t size = surface->GetSizeInBytes();
- const VAddr cpu_addr = surface->GetCpuAddr();
- rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1);
+ Unmark(surface);
+ if (surface->IsSyncPending()) {
+ marked_for_unregister.remove(surface);
+ surface->SetSyncPending(false);
+ }
UnregisterInnerCache(surface);
surface->MarkAsRegistered(false);
ReserveSurface(surface->GetSurfaceParams(), surface);
@@ -1150,6 +1190,8 @@ private:
std::unordered_map<u32, TSurface> invalid_cache;
std::vector<u8> invalid_memory;
+ std::list<TSurface> marked_for_unregister;
+
StagingCache staging_cache;
std::recursive_mutex mutex;
};
From da8f17715dbdc7eec92f5f0c11c968a51b86cab4 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Feb 2020 09:51:37 -0400
Subject: [PATCH 03/31] GPU: Refactor synchronization on Async GPU
---
src/video_core/dma_pusher.cpp | 1 +
src/video_core/engines/maxwell_3d.cpp | 8 ++++++--
src/video_core/gpu.cpp | 3 +++
src/video_core/gpu.h | 1 +
src/video_core/gpu_thread.cpp | 6 +++---
src/video_core/rasterizer_interface.h | 6 ++++++
src/video_core/renderer_opengl/gl_rasterizer.cpp | 16 ++++++++++++++++
src/video_core/renderer_opengl/gl_rasterizer.h | 2 ++
src/video_core/renderer_vulkan/vk_rasterizer.cpp | 14 ++++++++++++++
src/video_core/renderer_vulkan/vk_rasterizer.h | 2 ++
src/video_core/texture_cache/texture_cache.h | 4 ++--
11 files changed, 56 insertions(+), 7 deletions(-)
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 0b77afc717..2516ea9935 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -32,6 +32,7 @@ void DmaPusher::DispatchCalls() {
}
}
gpu.FlushCommands();
+ gpu.SyncGuestHost();
}
bool DmaPusher::Step() {
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index baa74ad4ca..2298a62738 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -403,9 +403,13 @@ void Maxwell3D::ProcessQueryGet() {
"Units other than CROP are unimplemented");
switch (regs.query.query_get.operation) {
- case Regs::QueryOperation::Release:
- StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
+ case Regs::QueryOperation::Release: {
+ rasterizer.FlushCommands();
+ rasterizer.SyncGuestHost();
+ const u64 result = regs.query.query_sequence;
+ StampQueryResult(result, regs.query.query_get.short_query == 0);
break;
+ }
case Regs::QueryOperation::Acquire:
// TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
// matches the current payload.
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index a606f4abdb..13bca5a78d 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -142,6 +142,9 @@ void GPU::FlushCommands() {
renderer->Rasterizer().FlushCommands();
}
+void GPU::SyncGuestHost() {
+ renderer->Rasterizer().SyncGuestHost();
+}
// Note that, traditionally, methods are treated as 4-byte addressable locations, and hence
// their numbers are written down multiplied by 4 in Docs. Here we are not multiply by 4.
// So the values you see in docs might be multiplied by 4.
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 1a2d747bef..99ed190bcd 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -156,6 +156,7 @@ public:
void CallMethod(const MethodCall& method_call);
void FlushCommands();
+ void SyncGuestHost();
/// Returns a reference to the Maxwell3D GPU engine.
Engines::Maxwell3D& Maxwell3D();
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 10cda686b9..1994d3bb45 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -40,7 +40,7 @@ static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::Graphic
} else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
renderer.Rasterizer().FlushRegion(data->addr, data->size);
} else if (const auto data = std::get_if<InvalidateRegionCommand>(&next.data)) {
- renderer.Rasterizer().InvalidateRegion(data->addr, data->size);
+ renderer.Rasterizer().OnCPUWrite(data->addr, data->size);
} else if (std::holds_alternative<EndProcessingCommand>(next.data)) {
return;
} else {
@@ -82,12 +82,12 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) {
}
void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
- system.Renderer().Rasterizer().InvalidateRegion(addr, size);
+ system.Renderer().Rasterizer().OnCPUWrite(addr, size);
}
void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
// Skip flush on asynch mode, as FlushAndInvalidateRegion is not used for anything too important
- InvalidateRegion(addr, size);
+ system.Renderer().Rasterizer().OnCPUWrite(addr, size);
}
void ThreadManager::WaitIdle() const {
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 8ae5b9c4e9..0d05a3fc71 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -58,6 +58,12 @@ public:
/// Notify rasterizer that any caches of the specified region should be invalidated
virtual void InvalidateRegion(VAddr addr, u64 size) = 0;
+ /// Notify rasterizer that any caches of the specified region are desync with guest
+ virtual void OnCPUWrite(VAddr addr, u64 size) = 0;
+
+ /// Sync memory between guest and host.
+ virtual void SyncGuestHost() = 0;
+
/// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
/// and invalidated
virtual void FlushAndInvalidateRegion(VAddr addr, u64 size) = 0;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index ac4485a18a..5379127454 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -660,6 +660,22 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
query_cache.InvalidateRegion(addr, size);
}
+void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
+ MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+ if (!addr || !size) {
+ return;
+ }
+ texture_cache.OnCPUWrite(addr, size);
+ shader_cache.InvalidateRegion(addr, size);
+ buffer_cache.InvalidateRegion(addr, size);
+}
+
+void RasterizerOpenGL::SyncGuestHost() {
+ MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+ texture_cache.SyncGuestHost();
+ // buffer_cache.SyncGuestHost();
+}
+
void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
if (Settings::IsGPULevelExtreme()) {
FlushRegion(addr, size);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index caea174d2c..a870024c63 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -67,6 +67,8 @@ public:
void FlushAll() override;
void FlushRegion(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size) override;
+ void OnCPUWrite(VAddr addr, u64 size) override;
+ void SyncGuestHost() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
void TickFrame() override;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 71007bbe85..ad59f558de 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -524,6 +524,20 @@ void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
query_cache.InvalidateRegion(addr, size);
}
+void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
+ if (!addr || !size) {
+ return;
+ }
+ texture_cache.OnCPUWrite(addr, size);
+ pipeline_cache.InvalidateRegion(addr, size);
+ buffer_cache.InvalidateRegion(addr, size);
+}
+
+void RasterizerVulkan::SyncGuestHost() {
+ texture_cache.SyncGuestHost();
+ // buffer_cache.SyncGuestHost();
+}
+
void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
FlushRegion(addr, size);
InvalidateRegion(addr, size);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index d9108f862a..285f731bce 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -119,6 +119,8 @@ public:
void FlushAll() override;
void FlushRegion(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size) override;
+ void OnCPUWrite(VAddr addr, u64 size) override;
+ void SyncGuestHost() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
void TickFrame() override;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 441fda53d8..c23b9f9b9f 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -63,7 +63,7 @@ public:
}
}
- void OnCPUWrite(CacheAddr addr, std::size_t size) {
+ void OnCPUWrite(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
for (const auto& surface : GetSurfacesInRegion(addr, size)) {
@@ -549,7 +549,7 @@ private:
}
const auto& final_params = new_surface->GetSurfaceParams();
if (cr_params.type != final_params.type) {
- if (Settings::values.use_accurate_gpu_emulation) {
+ if (Settings::IsGPULevelExtreme()) {
BufferCopy(current_surface, new_surface);
}
} else {
From 8b1eb44b3ed5483071dc6754662a277b45e4a809 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Feb 2020 10:08:07 -0400
Subject: [PATCH 04/31] BufferCache: Implement OnCPUWrite and SyncGuestHost
---
src/video_core/buffer_cache/buffer_cache.h | 45 ++++++++++++++++++-
src/video_core/buffer_cache/map_interval.h | 18 ++++++++
src/video_core/dma_pusher.cpp | 1 +
src/video_core/gpu_thread.cpp | 2 +-
.../renderer_opengl/gl_rasterizer.cpp | 4 +-
.../renderer_vulkan/vk_rasterizer.cpp | 4 +-
6 files changed, 67 insertions(+), 7 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 83e7a1cde1..5b14d52e2b 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -5,6 +5,7 @@
#pragma once
#include <array>
+#include <list>
#include <memory>
#include <mutex>
#include <unordered_map>
@@ -137,7 +138,9 @@ public:
});
for (auto& object : objects) {
if (object->IsModified() && object->IsRegistered()) {
+ mutex.unlock();
FlushMap(object);
+ mutex.lock();
}
}
}
@@ -154,6 +157,30 @@ public:
}
}
+ void OnCPUWrite(VAddr addr, std::size_t size) {
+ std::lock_guard lock{mutex};
+
+ for (const auto& object : GetMapsInRange(addr, size)) {
+ if (object->IsMemoryMarked() && object->IsRegistered()) {
+ Unmark(object);
+ object->SetSyncPending(true);
+ marked_for_unregister.emplace_back(object);
+ }
+ }
+ }
+
+ void SyncGuestHost() {
+ std::lock_guard lock{mutex};
+
+ for (const auto& object : marked_for_unregister) {
+ if (object->IsRegistered()) {
+ object->SetSyncPending(false);
+ Unregister(object);
+ }
+ }
+ marked_for_unregister.clear();
+ }
+
virtual BufferType GetEmptyBuffer(std::size_t size) = 0;
protected:
@@ -196,17 +223,30 @@ protected:
const IntervalType interval{new_map->GetStart(), new_map->GetEnd()};
mapped_addresses.insert({interval, new_map});
rasterizer.UpdatePagesCachedCount(cpu_addr, size, 1);
+ new_map->SetMemoryMarked(true);
if (inherit_written) {
MarkRegionAsWritten(new_map->GetStart(), new_map->GetEnd() - 1);
new_map->MarkAsWritten(true);
}
}
- /// Unregisters an object from the cache
- void Unregister(MapInterval& map) {
+ void Unmark(const MapInterval& map) {
+ if (!map->IsMemoryMarked()) {
+ return;
+ }
const std::size_t size = map->GetEnd() - map->GetStart();
rasterizer.UpdatePagesCachedCount(map->GetStart(), size, -1);
+ map->SetMemoryMarked(false);
+ }
+
+ /// Unregisters an object from the cache
+ void Unregister(const MapInterval& map) {
+ Unmark(map);
map->MarkAsRegistered(false);
+ if (map->IsSyncPending()) {
+ marked_for_unregister.remove(map);
+ map->SetSyncPending(false);
+ }
if (map->IsWritten()) {
UnmarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
}
@@ -479,6 +519,7 @@ private:
u64 modified_ticks = 0;
std::vector<u8> staging_buffer;
+ std::list<MapInterval> marked_for_unregister;
std::recursive_mutex mutex;
};
diff --git a/src/video_core/buffer_cache/map_interval.h b/src/video_core/buffer_cache/map_interval.h
index b0956029d3..29d8b26f3d 100644
--- a/src/video_core/buffer_cache/map_interval.h
+++ b/src/video_core/buffer_cache/map_interval.h
@@ -46,6 +46,22 @@ public:
return is_registered;
}
+ void SetMemoryMarked(bool is_memory_marked_) {
+ is_memory_marked = is_memory_marked_;
+ }
+
+ bool IsMemoryMarked() const {
+ return is_memory_marked;
+ }
+
+ void SetSyncPending(bool is_sync_pending_) {
+ is_sync_pending = is_sync_pending_;
+ }
+
+ bool IsSyncPending() const {
+ return is_sync_pending;
+ }
+
VAddr GetStart() const {
return start;
}
@@ -83,6 +99,8 @@ private:
bool is_written{};
bool is_modified{};
bool is_registered{};
+ bool is_memory_marked{};
+ bool is_sync_pending{};
u64 ticks{};
};
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 2516ea9935..31627b8121 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -21,6 +21,7 @@ MICROPROFILE_DEFINE(DispatchCalls, "GPU", "Execute command buffer", MP_RGB(128,
void DmaPusher::DispatchCalls() {
MICROPROFILE_SCOPE(DispatchCalls);
+ gpu.SyncGuestHost();
// On entering GPU code, assume all memory may be touched by the ARM core.
gpu.Maxwell3D().OnMemoryWrite();
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 1994d3bb45..0a8123cfef 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -78,7 +78,7 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
- PushCommand(FlushRegionCommand(addr, size));
+ system.Renderer().Rasterizer().FlushRegion(addr, size);
}
void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 5379127454..988eaeaa5a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -667,13 +667,13 @@ void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
}
texture_cache.OnCPUWrite(addr, size);
shader_cache.InvalidateRegion(addr, size);
- buffer_cache.InvalidateRegion(addr, size);
+ buffer_cache.OnCPUWrite(addr, size);
}
void RasterizerOpenGL::SyncGuestHost() {
MICROPROFILE_SCOPE(OpenGL_CacheManagement);
texture_cache.SyncGuestHost();
- // buffer_cache.SyncGuestHost();
+ buffer_cache.SyncGuestHost();
}
void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index ad59f558de..4d0c90aa39 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -530,12 +530,12 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
}
texture_cache.OnCPUWrite(addr, size);
pipeline_cache.InvalidateRegion(addr, size);
- buffer_cache.InvalidateRegion(addr, size);
+ buffer_cache.OnCPUWrite(addr, size);
}
void RasterizerVulkan::SyncGuestHost() {
texture_cache.SyncGuestHost();
- // buffer_cache.SyncGuestHost();
+ buffer_cache.SyncGuestHost();
}
void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
From 339d0d9d6c02cf79d6025dae7c60d8635fa4ea3b Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Feb 2020 16:24:37 -0400
Subject: [PATCH 05/31] GPU: Delay Fences.
---
src/video_core/dma_pusher.cpp | 1 +
src/video_core/engines/maxwell_3d.cpp | 10 +++++++++-
src/video_core/engines/maxwell_3d.h | 4 ++++
src/video_core/gpu.cpp | 4 ++++
src/video_core/gpu.h | 1 +
src/video_core/gpu_thread.cpp | 2 +-
6 files changed, 20 insertions(+), 2 deletions(-)
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 31627b8121..324dafdcda 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -34,6 +34,7 @@ void DmaPusher::DispatchCalls() {
}
gpu.FlushCommands();
gpu.SyncGuestHost();
+ gpu.OnCommandListEnd();
}
bool DmaPusher::Step() {
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 2298a62738..2605c3b429 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -397,6 +397,14 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
}
}
+void Maxwell3D::ReleaseFences() {
+ for (const auto pair : delay_fences) {
+ const auto [addr, payload] = pair;
+ memory_manager.Write<u32>(addr, static_cast<u32>(payload));
+ }
+ delay_fences.clear();
+}
+
void Maxwell3D::ProcessQueryGet() {
// TODO(Subv): Support the other query units.
ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop,
@@ -407,7 +415,7 @@ void Maxwell3D::ProcessQueryGet() {
rasterizer.FlushCommands();
rasterizer.SyncGuestHost();
const u64 result = regs.query.query_sequence;
- StampQueryResult(result, regs.query.query_get.short_query == 0);
+ delay_fences.emplace_back(regs.query.QueryAddress(), result);
break;
}
case Regs::QueryOperation::Acquire:
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 59d5752d26..0a93827ec8 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -1427,6 +1427,8 @@ public:
Tables tables{};
} dirty;
+ void ReleaseFences();
+
private:
void InitializeRegisterDefaults();
@@ -1467,6 +1469,8 @@ private:
std::array<u8, Regs::NUM_REGS> dirty_pointers{};
+ std::vector<std::pair<GPUVAddr, u64>> delay_fences;
+
/// Retrieves information about a specific TIC entry from the TIC buffer.
Texture::TICEntry GetTICEntry(u32 tic_index) const;
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 13bca5a78d..71ddfbd26e 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -145,6 +145,10 @@ void GPU::FlushCommands() {
void GPU::SyncGuestHost() {
renderer->Rasterizer().SyncGuestHost();
}
+
+void GPU::OnCommandListEnd() {
+ maxwell_3d->ReleaseFences();
+}
// Note that, traditionally, methods are treated as 4-byte addressable locations, and hence
// their numbers are written down multiplied by 4 in Docs. Here we are not multiply by 4.
// So the values you see in docs might be multiplied by 4.
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 99ed190bcd..b884456349 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -157,6 +157,7 @@ public:
void FlushCommands();
void SyncGuestHost();
+ void OnCommandListEnd();
/// Returns a reference to the Maxwell3D GPU engine.
Engines::Maxwell3D& Maxwell3D();
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 0a8123cfef..1994d3bb45 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -78,7 +78,7 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
- system.Renderer().Rasterizer().FlushRegion(addr, size);
+ PushCommand(FlushRegionCommand(addr, size));
}
void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
From ed7e9657120faea849af2933e539c72bc961c2a9 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Feb 2020 17:13:19 -0400
Subject: [PATCH 06/31] TextureCache: Flush linear textures after finishing
rendering.
---
src/video_core/texture_cache/texture_cache.h | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index c23b9f9b9f..d8c8390bb9 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -233,8 +233,14 @@ public:
auto surface_view = GetSurface(gpu_addr, *cpu_addr,
SurfaceParams::CreateForFramebuffer(system, index), true);
- if (render_targets[index].target)
- render_targets[index].target->MarkAsRenderTarget(false, NO_RT);
+ if (render_targets[index].target) {
+ auto& surface = render_targets[index].target;
+ surface->MarkAsRenderTarget(false, NO_RT);
+ const auto& cr_params = surface->GetSurfaceParams();
+ if (!cr_params.is_tiled) {
+ FlushSurface(surface);
+ }
+ }
render_targets[index].target = surface_view.first;
render_targets[index].view = surface_view.second;
if (render_targets[index].target)
From 487379c593bcaf3787ede187c5d44f7923b54dc9 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 17 Feb 2020 18:10:23 -0400
Subject: [PATCH 07/31] OpenGL: Implement Fencing backend.
---
src/video_core/engines/maxwell_3d.cpp | 16 +++-------
src/video_core/engines/maxwell_3d.h | 4 ---
src/video_core/gpu.cpp | 2 +-
src/video_core/gpu.h | 2 +-
src/video_core/gpu_asynch.cpp | 4 +++
src/video_core/gpu_asynch.h | 2 ++
src/video_core/gpu_thread.cpp | 6 ++++
src/video_core/gpu_thread.h | 7 +++-
src/video_core/rasterizer_interface.h | 8 +++++
.../renderer_opengl/gl_rasterizer.cpp | 28 ++++++++++++++++
.../renderer_opengl/gl_rasterizer.h | 2 ++
src/video_core/texture_cache/texture_cache.h | 32 ++++++++++++++++++-
12 files changed, 94 insertions(+), 19 deletions(-)
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 2605c3b429..c297bc31ba 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -397,14 +397,6 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
}
}
-void Maxwell3D::ReleaseFences() {
- for (const auto pair : delay_fences) {
- const auto [addr, payload] = pair;
- memory_manager.Write<u32>(addr, static_cast<u32>(payload));
- }
- delay_fences.clear();
-}
-
void Maxwell3D::ProcessQueryGet() {
// TODO(Subv): Support the other query units.
ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop,
@@ -412,10 +404,12 @@ void Maxwell3D::ProcessQueryGet() {
switch (regs.query.query_get.operation) {
case Regs::QueryOperation::Release: {
- rasterizer.FlushCommands();
- rasterizer.SyncGuestHost();
const u64 result = regs.query.query_sequence;
- delay_fences.emplace_back(regs.query.QueryAddress(), result);
+ if (regs.query.query_get.fence == 1) {
+ rasterizer.SignalFence(regs.query.QueryAddress(), static_cast<u32>(result));
+ } else {
+ StampQueryResult(result, regs.query.query_get.short_query == 0);
+ }
break;
}
case Regs::QueryOperation::Acquire:
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 0a93827ec8..59d5752d26 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -1427,8 +1427,6 @@ public:
Tables tables{};
} dirty;
- void ReleaseFences();
-
private:
void InitializeRegisterDefaults();
@@ -1469,8 +1467,6 @@ private:
std::array<u8, Regs::NUM_REGS> dirty_pointers{};
- std::vector<std::pair<GPUVAddr, u64>> delay_fences;
-
/// Retrieves information about a specific TIC entry from the TIC buffer.
Texture::TICEntry GetTICEntry(u32 tic_index) const;
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 71ddfbd26e..d05b6a9d2f 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -147,7 +147,7 @@ void GPU::SyncGuestHost() {
}
void GPU::OnCommandListEnd() {
- maxwell_3d->ReleaseFences();
+ renderer.Rasterizer().ReleaseFences();
}
// Note that, traditionally, methods are treated as 4-byte addressable locations, and hence
// their numbers are written down multiplied by 4 in Docs. Here we are not multiply by 4.
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index b884456349..fa9991c871 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -157,7 +157,7 @@ public:
void FlushCommands();
void SyncGuestHost();
- void OnCommandListEnd();
+ virtual void OnCommandListEnd();
/// Returns a reference to the Maxwell3D GPU engine.
Engines::Maxwell3D& Maxwell3D();
diff --git a/src/video_core/gpu_asynch.cpp b/src/video_core/gpu_asynch.cpp
index 20e73a37e4..53305ab436 100644
--- a/src/video_core/gpu_asynch.cpp
+++ b/src/video_core/gpu_asynch.cpp
@@ -52,4 +52,8 @@ void GPUAsynch::WaitIdle() const {
gpu_thread.WaitIdle();
}
+void GPUAsynch::OnCommandListEnd() {
+ gpu_thread.OnCommandListEnd();
+}
+
} // namespace VideoCommon
diff --git a/src/video_core/gpu_asynch.h b/src/video_core/gpu_asynch.h
index 03fd0eef0a..5176586127 100644
--- a/src/video_core/gpu_asynch.h
+++ b/src/video_core/gpu_asynch.h
@@ -32,6 +32,8 @@ public:
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void WaitIdle() const override;
+ void OnCommandListEnd() override;
+
protected:
void TriggerCpuInterrupt(u32 syncpoint_id, u32 value) const override;
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 1994d3bb45..251a9d911f 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -37,6 +37,8 @@ static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::Graphic
dma_pusher.DispatchCalls();
} else if (const auto data = std::get_if<SwapBuffersCommand>(&next.data)) {
renderer.SwapBuffers(data->framebuffer ? &*data->framebuffer : nullptr);
+ } else if (const auto data = std::get_if<OnCommandListEndCommand>(&next.data)) {
+ renderer.Rasterizer().ReleaseFences();
} else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
renderer.Rasterizer().FlushRegion(data->addr, data->size);
} else if (const auto data = std::get_if<InvalidateRegionCommand>(&next.data)) {
@@ -95,6 +97,10 @@ void ThreadManager::WaitIdle() const {
}
}
+void ThreadManager::OnCommandListEnd() {
+ PushCommand(OnCommandListEndCommand());
+}
+
u64 ThreadManager::PushCommand(CommandData&& command_data) {
const u64 fence{++state.last_fence};
state.queue.Push(CommandDataContainer(std::move(command_data), fence));
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index cd74ad330a..9d08779216 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -70,9 +70,12 @@ struct FlushAndInvalidateRegionCommand final {
u64 size;
};
+/// Command to signal to the GPU thread that processing has ended
+struct OnCommandListEndCommand final {};
+
using CommandData =
std::variant<EndProcessingCommand, SubmitListCommand, SwapBuffersCommand, FlushRegionCommand,
- InvalidateRegionCommand, FlushAndInvalidateRegionCommand>;
+ InvalidateRegionCommand, FlushAndInvalidateRegionCommand, OnCommandListEndCommand>;
struct CommandDataContainer {
CommandDataContainer() = default;
@@ -122,6 +125,8 @@ public:
// Wait until the gpu thread is idle.
void WaitIdle() const;
+ void OnCommandListEnd();
+
private:
/// Pushes a command to be executed by the GPU thread
u64 PushCommand(CommandData&& command_data);
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 0d05a3fc71..72f65b166c 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -49,6 +49,14 @@ public:
/// Records a GPU query and caches it
virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
+ virtual void SignalFence(GPUVAddr addr, u32 value) {
+
+ }
+
+ virtual void ReleaseFences() {
+
+ }
+
/// Notify rasterizer that all caches should be flushed to Switch memory
virtual void FlushAll() = 0;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 988eaeaa5a..93bb33e8cc 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -676,6 +676,34 @@ void RasterizerOpenGL::SyncGuestHost() {
buffer_cache.SyncGuestHost();
}
+void RasterizerOpenGL::SignalFence(GPUVAddr addr, u32 value) {
+ if (!fences.empty()) {
+ const std::pair<GPUVAddr, u32>& current_fence = fences.front();
+ const auto [address, payload] = current_fence;
+ texture_cache.PopAsyncFlushes();
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(address, payload);
+ fences.pop_front();
+ }
+ fences.emplace_back(addr, value);
+ texture_cache.CommitAsyncFlushes();
+ FlushCommands();
+ SyncGuestHost();
+}
+
+void RasterizerOpenGL::ReleaseFences() {
+ while (!fences.empty()) {
+ const std::pair<GPUVAddr, u32>& current_fence = fences.front();
+ const auto [address, payload] = current_fence;
+ texture_cache.PopAsyncFlushes();
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(address, payload);
+ fences.pop_front();
+ }
+}
+
void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
if (Settings::IsGPULevelExtreme()) {
FlushRegion(addr, size);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index a870024c63..486a154ad6 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -69,6 +69,8 @@ public:
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
+ void SignalFence(GPUVAddr addr, u32 value) override;
+ void ReleaseFences() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
void TickFrame() override;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index d8c8390bb9..6629c59ed2 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -238,7 +238,7 @@ public:
surface->MarkAsRenderTarget(false, NO_RT);
const auto& cr_params = surface->GetSurfaceParams();
if (!cr_params.is_tiled) {
- FlushSurface(surface);
+ AsyncFlushSurface(surface);
}
}
render_targets[index].target = surface_view.first;
@@ -317,6 +317,26 @@ public:
return ++ticks;
}
+ void CommitAsyncFlushes() {
+ commited_flushes.push_back(uncommited_flushes);
+ uncommited_flushes.reset();
+ }
+
+ void PopAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ commited_flushes.pop_front();
+ return;
+ }
+ for (TSurface& surface : *flush_list) {
+ FlushSurface(surface);
+ }
+ commited_flushes.pop_front();
+ }
+
protected:
explicit TextureCache(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
bool is_astc_supported)
@@ -1152,6 +1172,13 @@ private:
TView view;
};
+ void AsyncFlushSurface(TSurface& surface) {
+ if (!uncommited_flushes) {
+ uncommited_flushes = std::make_shared<std::list<TSurface>>();
+ }
+ uncommited_flushes->push_back(surface);
+ }
+
VideoCore::RasterizerInterface& rasterizer;
FormatLookupTable format_lookup_table;
@@ -1198,6 +1225,9 @@ private:
std::list<TSurface> marked_for_unregister;
+ std::shared_ptr<std::list<TSurface>> uncommited_flushes{};
+ std::list<std::shared_ptr<std::list<TSurface>>> commited_flushes;
+
StagingCache staging_cache;
std::recursive_mutex mutex;
};
From 1f345ebe3a5501b50f26f0c5c21cac5d55dd79c1 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 17 Feb 2020 20:19:26 -0400
Subject: [PATCH 08/31] GPU: Implement a Fence Manager.
---
src/video_core/CMakeLists.txt | 3 +
src/video_core/fence_manager.h | 97 +++++++++++++++++++
.../renderer_opengl/gl_fence_manager.cpp | 55 +++++++++++
.../renderer_opengl/gl_fence_manager.h | 47 +++++++++
.../renderer_opengl/gl_rasterizer.cpp | 27 +-----
.../renderer_opengl/gl_rasterizer.h | 2 +
6 files changed, 208 insertions(+), 23 deletions(-)
create mode 100644 src/video_core/fence_manager.h
create mode 100644 src/video_core/renderer_opengl/gl_fence_manager.cpp
create mode 100644 src/video_core/renderer_opengl/gl_fence_manager.h
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 258d58eba7..9a3f568f9c 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -23,6 +23,7 @@ add_library(video_core STATIC
engines/shader_bytecode.h
engines/shader_header.h
engines/shader_type.h
+ fence_manager.h
gpu.cpp
gpu.h
gpu_asynch.cpp
@@ -51,6 +52,8 @@ add_library(video_core STATIC
renderer_opengl/gl_buffer_cache.h
renderer_opengl/gl_device.cpp
renderer_opengl/gl_device.h
+ renderer_opengl/gl_fence_manager.cpp
+ renderer_opengl/gl_fence_manager.h
renderer_opengl/gl_framebuffer_cache.cpp
renderer_opengl/gl_framebuffer_cache.h
renderer_opengl/gl_rasterizer.cpp
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
new file mode 100644
index 0000000000..19cec0f66a
--- /dev/null
+++ b/src/video_core/fence_manager.h
@@ -0,0 +1,97 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <queue>
+#include <memory>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "core/core.h"
+#include "core/memory.h"
+#include "core/settings.h"
+#include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
+#include "video_core/rasterizer_interface.h"
+
+namespace VideoCommon {
+
+class FenceBase {
+public:
+ FenceBase(GPUVAddr address, u32 payload) : address{address}, payload{payload} {}
+
+ constexpr GPUVAddr GetAddress() const {
+ return address;
+ }
+
+ constexpr u32 GetPayload() const {
+ return payload;
+ }
+
+private:
+ GPUVAddr address;
+ u32 payload;
+};
+
+template <typename TFence, typename TTextureCache>
+class FenceManager {
+public:
+ void SignalFence(GPUVAddr addr, u32 value) {
+ TryReleasePendingFences();
+ TFence new_fence = CreateFence(addr, value);
+ QueueFence(new_fence);
+ fences.push(new_fence);
+ texture_cache.CommitAsyncFlushes();
+ rasterizer.FlushCommands();
+ rasterizer.SyncGuestHost();
+ }
+
+ void WaitPendingFences() {
+ while (!fences.empty()) {
+ TFence& current_fence = fences.front();
+ WaitFence(current_fence);
+ texture_cache.PopAsyncFlushes();
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ fences.pop();
+ }
+ }
+
+protected:
+ FenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+ TTextureCache& texture_cache)
+ : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache} {}
+
+ virtual TFence CreateFence(GPUVAddr addr, u32 value) = 0;
+ virtual void QueueFence(TFence& fence) = 0;
+ virtual bool IsFenceSignaled(TFence& fence) = 0;
+ virtual void WaitFence(TFence& fence) = 0;
+
+ Core::System& system;
+ VideoCore::RasterizerInterface& rasterizer;
+ TTextureCache& texture_cache;
+
+private:
+ void TryReleasePendingFences() {
+ while (!fences.empty()) {
+ TFence& current_fence = fences.front();
+ if (!IsFenceSignaled(current_fence)) {
+ return;
+ }
+ texture_cache.PopAsyncFlushes();
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ fences.pop();
+ }
+ }
+
+ std::queue<TFence> fences;
+};
+
+} // namespace VideoCommon
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
new file mode 100644
index 0000000000..4517ef1507
--- /dev/null
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -0,0 +1,55 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+
+#include "video_core/renderer_opengl/gl_fence_manager.h"
+
+namespace OpenGL {
+
+GLInnerFence::GLInnerFence(GPUVAddr address, u32 payload)
+ : VideoCommon::FenceBase(address, payload), sync_object{} {}
+
+GLInnerFence::~GLInnerFence() = default;
+
+void GLInnerFence::Queue() {
+ ASSERT(sync_object.handle == 0);
+ sync_object.Create();
+}
+
+bool GLInnerFence::IsSignaled() const {
+ ASSERT(sync_object.handle != 0);
+ GLsizei length;
+ GLint sync_status;
+ glGetSynciv(sync_object.handle, GL_SYNC_STATUS, sizeof(GLint), &length, &sync_status);
+ return sync_status == GL_SIGNALED;
+}
+
+void GLInnerFence::Wait() {
+ ASSERT(sync_object.handle != 0);
+ while (glClientWaitSync(sync_object.handle, 0, 1000) == GL_TIMEOUT_EXPIRED)
+ ;
+}
+
+FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+ TextureCacheOpenGL& texture_cache)
+ : GenericFenceManager(system, rasterizer, texture_cache) {}
+
+Fence FenceManagerOpenGL::CreateFence(GPUVAddr addr, u32 value) {
+ return std::make_shared<GLInnerFence>(addr, value);
+}
+
+void FenceManagerOpenGL::QueueFence(Fence& fence) {
+ fence->Queue();
+}
+
+bool FenceManagerOpenGL::IsFenceSignaled(Fence& fence) {
+ return fence->IsSignaled();
+}
+
+void FenceManagerOpenGL::WaitFence(Fence& fence) {
+ fence->Wait();
+}
+
+} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.h b/src/video_core/renderer_opengl/gl_fence_manager.h
new file mode 100644
index 0000000000..3cfa8b1d01
--- /dev/null
+++ b/src/video_core/renderer_opengl/gl_fence_manager.h
@@ -0,0 +1,47 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <glad/glad.h>
+
+#include "common/common_types.h"
+#include "video_core/fence_manager.h"
+#include "video_core/renderer_opengl/gl_resource_manager.h"
+#include "video_core/renderer_opengl/gl_texture_cache.h"
+
+namespace OpenGL {
+
+class GLInnerFence : public VideoCommon::FenceBase {
+public:
+ GLInnerFence(GPUVAddr address, u32 payload);
+ ~GLInnerFence();
+
+ void Queue();
+
+ bool IsSignaled() const;
+
+ void Wait();
+
+private:
+ OGLSync sync_object;
+};
+
+using Fence = std::shared_ptr<GLInnerFence>;
+using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCacheOpenGL>;
+
+class FenceManagerOpenGL final : public GenericFenceManager {
+public:
+ FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+ TextureCacheOpenGL& texture_cache);
+
+protected:
+ Fence CreateFence(GPUVAddr addr, u32 value) override;
+ void QueueFence(Fence& fence) override;
+ bool IsFenceSignaled(Fence& fence) override;
+ void WaitFence(Fence& fence) override;
+};
+
+} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 93bb33e8cc..35bed444f8 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -101,7 +101,8 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
: RasterizerAccelerated{system.Memory()}, texture_cache{system, *this, device, state_tracker},
shader_cache{*this, system, emu_window, device}, query_cache{system, *this}, system{system},
screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker},
- buffer_cache{*this, system, device, STREAM_BUFFER_SIZE} {
+ buffer_cache{*this, system, device, STREAM_BUFFER_SIZE}, fence_manager{system, *this,
+ texture_cache} {
CheckExtensions();
}
@@ -677,31 +678,11 @@ void RasterizerOpenGL::SyncGuestHost() {
}
void RasterizerOpenGL::SignalFence(GPUVAddr addr, u32 value) {
- if (!fences.empty()) {
- const std::pair<GPUVAddr, u32>& current_fence = fences.front();
- const auto [address, payload] = current_fence;
- texture_cache.PopAsyncFlushes();
- auto& gpu{system.GPU()};
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(address, payload);
- fences.pop_front();
- }
- fences.emplace_back(addr, value);
- texture_cache.CommitAsyncFlushes();
- FlushCommands();
- SyncGuestHost();
+ fence_manager.SignalFence(addr, value);
}
void RasterizerOpenGL::ReleaseFences() {
- while (!fences.empty()) {
- const std::pair<GPUVAddr, u32>& current_fence = fences.front();
- const auto [address, payload] = current_fence;
- texture_cache.PopAsyncFlushes();
- auto& gpu{system.GPU()};
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(address, payload);
- fences.pop_front();
- }
+ fence_manager.WaitPendingFences();
}
void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 486a154ad6..6d173a9220 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -23,6 +23,7 @@
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_device.h"
+#include "video_core/renderer_opengl/gl_fence_manager.h"
#include "video_core/renderer_opengl/gl_framebuffer_cache.h"
#include "video_core/renderer_opengl/gl_query_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
@@ -226,6 +227,7 @@ private:
SamplerCacheOpenGL sampler_cache;
FramebufferCacheOpenGL framebuffer_cache;
QueryCache query_cache;
+ FenceManagerOpenGL fence_manager;
Core::System& system;
ScreenInfo& screen_info;
From 57fdbd9b8992de4eaf2b262e6a2cece43c141894 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 17 Feb 2020 22:15:43 -0400
Subject: [PATCH 09/31] FenceManager: Implement should wait.
---
src/video_core/fence_manager.h | 8 ++++++--
src/video_core/texture_cache/texture_cache.h | 11 +++++++++++
2 files changed, 17 insertions(+), 2 deletions(-)
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 19cec0f66a..036f3996c3 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -53,7 +53,10 @@ public:
void WaitPendingFences() {
while (!fences.empty()) {
TFence& current_fence = fences.front();
- WaitFence(current_fence);
+ bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
+ if (should_wait) {
+ WaitFence(current_fence);
+ }
texture_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
@@ -80,7 +83,8 @@ private:
void TryReleasePendingFences() {
while (!fences.empty()) {
TFence& current_fence = fences.front();
- if (!IsFenceSignaled(current_fence)) {
+ bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
+ if (should_wait && !IsFenceSignaled(current_fence)) {
return;
}
texture_cache.PopAsyncFlushes();
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 6629c59ed2..04fe69c11d 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -322,6 +322,17 @@ public:
uncommited_flushes.reset();
}
+ bool ShouldWaitAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return false;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ return false;
+ }
+ return true;
+ }
+
void PopAsyncFlushes() {
if (commited_flushes.empty()) {
return;
From 165ae823f522aa981129927f42e76763a9fa6006 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 17 Feb 2020 22:29:04 -0400
Subject: [PATCH 10/31] ThreadManager: Sync async reads on accurate gpu.
---
src/video_core/buffer_cache/buffer_cache.h | 12 ++++++++++++
src/video_core/gpu.cpp | 2 +-
src/video_core/gpu_thread.cpp | 7 ++++++-
src/video_core/rasterizer_interface.h | 10 ++++------
src/video_core/renderer_opengl/gl_rasterizer.cpp | 4 ++++
src/video_core/renderer_opengl/gl_rasterizer.h | 1 +
src/video_core/renderer_vulkan/vk_rasterizer.cpp | 4 ++++
src/video_core/renderer_vulkan/vk_rasterizer.h | 1 +
src/video_core/texture_cache/texture_cache.h | 15 +++++++++++++++
9 files changed, 48 insertions(+), 8 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 5b14d52e2b..df4c0211e3 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -145,6 +145,18 @@ public:
}
}
+ bool MustFlushRegion(VAddr addr, std::size_t size) {
+ std::lock_guard lock{mutex};
+
+ std::vector<MapInterval> objects = GetMapsInRange(addr, size);
+ for (auto& object : objects) {
+ if (object->IsModified() && object->IsRegistered()) {
+ return true;
+ }
+ }
+ return false;
+ }
+
/// Mark the specified region as being invalidated
void InvalidateRegion(VAddr addr, u64 size) {
std::lock_guard lock{mutex};
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index d05b6a9d2f..19d3bd3052 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -147,7 +147,7 @@ void GPU::SyncGuestHost() {
}
void GPU::OnCommandListEnd() {
- renderer.Rasterizer().ReleaseFences();
+ renderer->Rasterizer().ReleaseFences();
}
// Note that, traditionally, methods are treated as 4-byte addressable locations, and hence
// their numbers are written down multiplied by 4 in Docs. Here we are not multiply by 4.
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 251a9d911f..672f8d2fab 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -6,6 +6,7 @@
#include "common/microprofile.h"
#include "core/core.h"
#include "core/frontend/emu_window.h"
+#include "core/settings.h"
#include "video_core/dma_pusher.h"
#include "video_core/gpu.h"
#include "video_core/gpu_thread.h"
@@ -80,7 +81,11 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
- PushCommand(FlushRegionCommand(addr, size));
+ if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
+ u64 fence = PushCommand(FlushRegionCommand(addr, size));
+ while (fence < state.signaled_fence.load(std::memory_order_relaxed)) {
+ }
+ }
}
void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 72f65b166c..2287521314 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -49,13 +49,9 @@ public:
/// Records a GPU query and caches it
virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
- virtual void SignalFence(GPUVAddr addr, u32 value) {
+ virtual void SignalFence(GPUVAddr addr, u32 value) {}
- }
-
- virtual void ReleaseFences() {
-
- }
+ virtual void ReleaseFences() {}
/// Notify rasterizer that all caches should be flushed to Switch memory
virtual void FlushAll() = 0;
@@ -63,6 +59,8 @@ public:
/// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
virtual void FlushRegion(VAddr addr, u64 size) = 0;
+ virtual bool MustFlushRegion(VAddr addr, u64 size) = 0;
+
/// Notify rasterizer that any caches of the specified region should be invalidated
virtual void InvalidateRegion(VAddr addr, u64 size) = 0;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 35bed444f8..bbf37a00d5 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -650,6 +650,10 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
query_cache.FlushRegion(addr, size);
}
+bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+ return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
+}
+
void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
MICROPROFILE_SCOPE(OpenGL_CacheManagement);
if (addr == 0 || size == 0) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 6d173a9220..5c0f88e6f1 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -67,6 +67,7 @@ public:
void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
void FlushAll() override;
void FlushRegion(VAddr addr, u64 size) override;
+ bool MustFlushRegion(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 4d0c90aa39..9437a4aa11 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -514,6 +514,10 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
query_cache.FlushRegion(addr, size);
}
+bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size) {
+ return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
+}
+
void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
if (addr == 0 || size == 0) {
return;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 285f731bce..7002a4fa39 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -118,6 +118,7 @@ public:
void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
void FlushAll() override;
void FlushRegion(VAddr addr, u64 size) override;
+ bool MustFlushRegion(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 04fe69c11d..e251a30c30 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -116,6 +116,21 @@ public:
}
}
+ bool MustFlushRegion(VAddr addr, std::size_t size) {
+ std::lock_guard lock{mutex};
+
+ auto surfaces = GetSurfacesInRegion(addr, size);
+ if (surfaces.empty()) {
+ return false;
+ }
+ for (const auto& surface : surfaces) {
+ if (surface->IsModified()) {
+ return true;
+ }
+ }
+ return false;
+ }
+
TView GetTextureSurface(const Tegra::Texture::TICEntry& tic,
const VideoCommon::Shader::Sampler& entry) {
std::lock_guard lock{mutex};
From e84eb64e511cd4699cc2371744fccd24628c0749 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 18 Feb 2020 11:26:31 -0400
Subject: [PATCH 11/31] Rasterizer: Disable fence managing in synchronous gpu.
---
src/video_core/renderer_opengl/gl_rasterizer.cpp | 10 ++++++++++
src/video_core/texture_cache/texture_cache.h | 2 +-
2 files changed, 11 insertions(+), 1 deletion(-)
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index bbf37a00d5..ffa0a04292 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -682,10 +682,20 @@ void RasterizerOpenGL::SyncGuestHost() {
}
void RasterizerOpenGL::SignalFence(GPUVAddr addr, u32 value) {
+ auto& gpu{system.GPU()};
+ if (!gpu.IsAsync()) {
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(addr, value);
+ return;
+ }
fence_manager.SignalFence(addr, value);
}
void RasterizerOpenGL::ReleaseFences() {
+ auto& gpu{system.GPU()};
+ if (!gpu.IsAsync()) {
+ return;
+ }
fence_manager.WaitPendingFences();
}
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index e251a30c30..e1a1edbd24 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -252,7 +252,7 @@ public:
auto& surface = render_targets[index].target;
surface->MarkAsRenderTarget(false, NO_RT);
const auto& cr_params = surface->GetSurfaceParams();
- if (!cr_params.is_tiled) {
+ if (!cr_params.is_tiled && Settings::values.use_asynchronous_gpu_emulation) {
AsyncFlushSurface(surface);
}
}
From a081a7c8558500249f4871f3a1b8c9046cb1b6f1 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 18 Feb 2020 13:19:24 -0400
Subject: [PATCH 12/31] GPU: Fix rebase errors.
---
src/video_core/engines/maxwell_3d.cpp | 7 +++----
src/video_core/renderer_opengl/gl_rasterizer.cpp | 1 +
2 files changed, 4 insertions(+), 4 deletions(-)
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index c297bc31ba..cff90bc1d9 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -403,12 +403,11 @@ void Maxwell3D::ProcessQueryGet() {
"Units other than CROP are unimplemented");
switch (regs.query.query_get.operation) {
- case Regs::QueryOperation::Release: {
- const u64 result = regs.query.query_sequence;
+ case Regs::QueryOperation::Release:
if (regs.query.query_get.fence == 1) {
- rasterizer.SignalFence(regs.query.QueryAddress(), static_cast<u32>(result));
+ rasterizer.SignalFence(regs.query.QueryAddress(), regs.query.query_sequence);
} else {
- StampQueryResult(result, regs.query.query_get.short_query == 0);
+ StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
}
break;
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index ffa0a04292..db7eae065c 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -673,6 +673,7 @@ void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
texture_cache.OnCPUWrite(addr, size);
shader_cache.InvalidateRegion(addr, size);
buffer_cache.OnCPUWrite(addr, size);
+ query_cache.InvalidateRegion(addr, size);
}
void RasterizerOpenGL::SyncGuestHost() {
From 4adfc9bb0870296b372dc96296436538d6aa6c32 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 18 Feb 2020 13:24:38 -0400
Subject: [PATCH 13/31] Rasterizer: Document SignalFence & ReleaseFences and
setup skeletons on Vulkan.
---
src/video_core/engines/maxwell_3d.cpp | 1 -
src/video_core/gpu_thread.cpp | 5 +++-
src/video_core/rasterizer_interface.h | 6 +++--
.../renderer_vulkan/vk_rasterizer.cpp | 25 +++++++++++++++++++
.../renderer_vulkan/vk_rasterizer.h | 2 ++
5 files changed, 35 insertions(+), 4 deletions(-)
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index cff90bc1d9..a7e9514330 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -410,7 +410,6 @@ void Maxwell3D::ProcessQueryGet() {
StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
}
break;
- }
case Regs::QueryOperation::Acquire:
// TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
// matches the current payload.
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 672f8d2fab..1c3ab2145d 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -81,9 +81,12 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
+ if (!Settings::IsGPULevelExtreme()) {
+ return;
+ }
if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
u64 fence = PushCommand(FlushRegionCommand(addr, size));
- while (fence < state.signaled_fence.load(std::memory_order_relaxed)) {
+ while (fence > state.signaled_fence.load(std::memory_order_relaxed)) {
}
}
}
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 2287521314..b49f15df2e 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -49,9 +49,11 @@ public:
/// Records a GPU query and caches it
virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
- virtual void SignalFence(GPUVAddr addr, u32 value) {}
+ /// Signal a GPU based fence
+ virtual void SignalFence(GPUVAddr addr, u32 value) = 0;
- virtual void ReleaseFences() {}
+ /// Release all pending fences.
+ virtual void ReleaseFences() = 0;
/// Notify rasterizer that all caches should be flushed to Switch memory
virtual void FlushAll() = 0;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 9437a4aa11..1d75a4766a 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -535,6 +535,7 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
texture_cache.OnCPUWrite(addr, size);
pipeline_cache.InvalidateRegion(addr, size);
buffer_cache.OnCPUWrite(addr, size);
+ query_cache.InvalidateRegion(addr, size);
}
void RasterizerVulkan::SyncGuestHost() {
@@ -542,6 +543,30 @@ void RasterizerVulkan::SyncGuestHost() {
buffer_cache.SyncGuestHost();
}
+void RasterizerVulkan::SignalFence(GPUVAddr addr, u32 value) {
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(addr, value);
+ /*
+ if (!gpu.IsAsync()) {
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(addr, value);
+ return;
+ }
+ fence_manager.SignalFence(addr, value);
+ */
+}
+
+void RasterizerVulkan::ReleaseFences() {
+ /*
+ auto& gpu{system.GPU()};
+ if (!gpu.IsAsync()) {
+ return;
+ }
+ fence_manager.WaitPendingFences();
+ */
+}
+
void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
FlushRegion(addr, size);
InvalidateRegion(addr, size);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 7002a4fa39..08a9af401c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -122,6 +122,8 @@ public:
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
+ void SignalFence(GPUVAddr addr, u32 value) override;
+ void ReleaseFences() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
void TickFrame() override;
From b10db7e4a5f43414679b7969ea309b1829937a37 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 18 Feb 2020 17:20:39 -0400
Subject: [PATCH 14/31] FenceManager: Implement async buffer cache flushes on
High settings
---
src/video_core/buffer_cache/buffer_cache.h | 50 +++++++++++++++++++
src/video_core/fence_manager.h | 15 ++++--
src/video_core/gpu_thread.cpp | 2 +-
.../renderer_opengl/gl_fence_manager.cpp | 4 +-
.../renderer_opengl/gl_fence_manager.h | 5 +-
.../renderer_opengl/gl_rasterizer.cpp | 3 +-
6 files changed, 69 insertions(+), 10 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index df4c0211e3..d72df90eff 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -21,6 +21,7 @@
#include "common/common_types.h"
#include "core/core.h"
#include "core/memory.h"
+#include "core/settings.h"
#include "video_core/buffer_cache/buffer_block.h"
#include "video_core/buffer_cache/map_interval.h"
#include "video_core/memory_manager.h"
@@ -80,6 +81,9 @@ public:
auto map = MapAddress(block, gpu_addr, cpu_addr, size);
if (is_written) {
map->MarkAsModified(true, GetModifiedTicks());
+ if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+ AsyncFlushMap(map);
+ }
if (!map->IsWritten()) {
map->MarkAsWritten(true);
MarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
@@ -193,6 +197,39 @@ public:
marked_for_unregister.clear();
}
+ void CommitAsyncFlushes() {
+ commited_flushes.push_back(uncommited_flushes);
+ uncommited_flushes.reset();
+ }
+
+ bool ShouldWaitAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return false;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ return false;
+ }
+ return true;
+ }
+
+ void PopAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ commited_flushes.pop_front();
+ return;
+ }
+ for (MapInterval& map : *flush_list) {
+ if (map->IsRegistered()) {
+ FlushMap(map);
+ }
+ }
+ commited_flushes.pop_front();
+ }
+
virtual BufferType GetEmptyBuffer(std::size_t size) = 0;
protected:
@@ -316,6 +353,9 @@ private:
MapInterval new_map = CreateMap(new_start, new_end, new_gpu_addr);
if (modified_inheritance) {
new_map->MarkAsModified(true, GetModifiedTicks());
+ if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+ AsyncFlushMap(new_map);
+ }
}
Register(new_map, write_inheritance);
return new_map;
@@ -502,6 +542,13 @@ private:
return false;
}
+ void AsyncFlushMap(MapInterval& map) {
+ if (!uncommited_flushes) {
+ uncommited_flushes = std::make_shared<std::list<MapInterval>>();
+ }
+ uncommited_flushes->push_back(map);
+ }
+
VideoCore::RasterizerInterface& rasterizer;
Core::System& system;
@@ -533,6 +580,9 @@ private:
std::vector<u8> staging_buffer;
std::list<MapInterval> marked_for_unregister;
+ std::shared_ptr<std::list<MapInterval>> uncommited_flushes{};
+ std::list<std::shared_ptr<std::list<MapInterval>>> commited_flushes;
+
std::recursive_mutex mutex;
};
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 036f3996c3..c4b190503f 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -6,8 +6,8 @@
#include <algorithm>
#include <array>
-#include <queue>
#include <memory>
+#include <queue>
#include "common/assert.h"
#include "common/common_types.h"
@@ -37,7 +37,7 @@ private:
u32 payload;
};
-template <typename TFence, typename TTextureCache>
+template <typename TFence, typename TTextureCache, typename TTBufferCache>
class FenceManager {
public:
void SignalFence(GPUVAddr addr, u32 value) {
@@ -46,6 +46,7 @@ public:
QueueFence(new_fence);
fences.push(new_fence);
texture_cache.CommitAsyncFlushes();
+ buffer_cache.CommitAsyncFlushes();
rasterizer.FlushCommands();
rasterizer.SyncGuestHost();
}
@@ -54,10 +55,12 @@ public:
while (!fences.empty()) {
TFence& current_fence = fences.front();
bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
+ should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
if (should_wait) {
WaitFence(current_fence);
}
texture_cache.PopAsyncFlushes();
+ buffer_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
@@ -67,8 +70,9 @@ public:
protected:
FenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TTextureCache& texture_cache)
- : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache} {}
+ TTextureCache& texture_cache, TTBufferCache& buffer_cache)
+ : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache}, buffer_cache{
+ buffer_cache} {}
virtual TFence CreateFence(GPUVAddr addr, u32 value) = 0;
virtual void QueueFence(TFence& fence) = 0;
@@ -78,16 +82,19 @@ protected:
Core::System& system;
VideoCore::RasterizerInterface& rasterizer;
TTextureCache& texture_cache;
+ TTBufferCache& buffer_cache;
private:
void TryReleasePendingFences() {
while (!fences.empty()) {
TFence& current_fence = fences.front();
bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
+ should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
if (should_wait && !IsFenceSignaled(current_fence)) {
return;
}
texture_cache.PopAsyncFlushes();
+ buffer_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 1c3ab2145d..3e2be00e9d 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -81,7 +81,7 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
- if (!Settings::IsGPULevelExtreme()) {
+ if (!Settings::IsGPULevelHigh()) {
return;
}
if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
index 4517ef1507..69dd3211b0 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -33,8 +33,8 @@ void GLInnerFence::Wait() {
}
FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TextureCacheOpenGL& texture_cache)
- : GenericFenceManager(system, rasterizer, texture_cache) {}
+ TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache)
+ : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache) {}
Fence FenceManagerOpenGL::CreateFence(GPUVAddr addr, u32 value) {
return std::make_shared<GLInnerFence>(addr, value);
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.h b/src/video_core/renderer_opengl/gl_fence_manager.h
index 3cfa8b1d01..b48d5eaa0d 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.h
+++ b/src/video_core/renderer_opengl/gl_fence_manager.h
@@ -9,6 +9,7 @@
#include "common/common_types.h"
#include "video_core/fence_manager.h"
+#include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_texture_cache.h"
@@ -30,12 +31,12 @@ private:
};
using Fence = std::shared_ptr<GLInnerFence>;
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCacheOpenGL>;
+using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCacheOpenGL, OGLBufferCache>;
class FenceManagerOpenGL final : public GenericFenceManager {
public:
FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TextureCacheOpenGL& texture_cache);
+ TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache);
protected:
Fence CreateFence(GPUVAddr addr, u32 value) override;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index db7eae065c..88914828ca 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -102,7 +102,8 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
shader_cache{*this, system, emu_window, device}, query_cache{system, *this}, system{system},
screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker},
buffer_cache{*this, system, device, STREAM_BUFFER_SIZE}, fence_manager{system, *this,
- texture_cache} {
+ texture_cache,
+ buffer_cache} {
CheckExtensions();
}
From 96bb961a6485c2f5c8b7fb91aa0dd7eb24fa5e5d Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 19 Feb 2020 10:49:07 -0400
Subject: [PATCH 15/31] BufferCache: Refactor async managing.
---
src/video_core/buffer_cache/buffer_cache.h | 31 +++++++++++++++++-----
src/video_core/fence_manager.h | 6 ++---
2 files changed, 27 insertions(+), 10 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index d72df90eff..06fb931d75 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -82,7 +82,7 @@ public:
if (is_written) {
map->MarkAsModified(true, GetModifiedTicks());
if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
- AsyncFlushMap(map);
+ MarkForAsyncFlush(map);
}
if (!map->IsWritten()) {
map->MarkAsWritten(true);
@@ -198,7 +198,23 @@ public:
}
void CommitAsyncFlushes() {
- commited_flushes.push_back(uncommited_flushes);
+ if (uncommited_flushes) {
+ auto commit_list = std::make_shared<std::list<MapInterval>>();
+ for (auto& map : *uncommited_flushes) {
+ if (map->IsRegistered() && map->IsModified()) {
+ // TODO(Blinkhawk): Implement backend asynchronous flushing
+ // AsyncFlushMap(map)
+ commit_list->push_back(map);
+ }
+ }
+ if (!commit_list->empty()) {
+ commited_flushes.push_back(commit_list);
+ } else {
+ commited_flushes.emplace_back();
+ }
+ } else {
+ commited_flushes.emplace_back();
+ }
uncommited_flushes.reset();
}
@@ -224,6 +240,7 @@ public:
}
for (MapInterval& map : *flush_list) {
if (map->IsRegistered()) {
+ // TODO(Blinkhawk): Replace this for reading the asynchronous flush
FlushMap(map);
}
}
@@ -354,7 +371,7 @@ private:
if (modified_inheritance) {
new_map->MarkAsModified(true, GetModifiedTicks());
if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
- AsyncFlushMap(new_map);
+ MarkForAsyncFlush(new_map);
}
}
Register(new_map, write_inheritance);
@@ -542,11 +559,11 @@ private:
return false;
}
- void AsyncFlushMap(MapInterval& map) {
+ void MarkForAsyncFlush(MapInterval& map) {
if (!uncommited_flushes) {
- uncommited_flushes = std::make_shared<std::list<MapInterval>>();
+ uncommited_flushes = std::make_shared<std::unordered_set<MapInterval>>();
}
- uncommited_flushes->push_back(map);
+ uncommited_flushes->insert(map);
}
VideoCore::RasterizerInterface& rasterizer;
@@ -580,7 +597,7 @@ private:
std::vector<u8> staging_buffer;
std::list<MapInterval> marked_for_unregister;
- std::shared_ptr<std::list<MapInterval>> uncommited_flushes{};
+ std::shared_ptr<std::unordered_set<MapInterval>> uncommited_flushes{};
std::list<std::shared_ptr<std::list<MapInterval>>> commited_flushes;
std::recursive_mutex mutex;
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index c4b190503f..72ee509550 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -42,11 +42,11 @@ class FenceManager {
public:
void SignalFence(GPUVAddr addr, u32 value) {
TryReleasePendingFences();
- TFence new_fence = CreateFence(addr, value);
- QueueFence(new_fence);
- fences.push(new_fence);
texture_cache.CommitAsyncFlushes();
buffer_cache.CommitAsyncFlushes();
+ TFence new_fence = CreateFence(addr, value);
+ fences.push(new_fence);
+ QueueFence(new_fence);
rasterizer.FlushCommands();
rasterizer.SyncGuestHost();
}
From b7bc3c25496849661846f2fe42f591f2a81fbc87 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 19 Feb 2020 13:40:37 -0400
Subject: [PATCH 16/31] FenceManager: Manage syncpoints and rename fences to
semaphores.
---
src/video_core/buffer_cache/buffer_cache.h | 7 +++
src/video_core/engines/maxwell_3d.cpp | 4 +-
src/video_core/fence_manager.h | 58 ++++++++++++++++---
src/video_core/rasterizer_interface.h | 7 ++-
.../renderer_opengl/gl_fence_manager.cpp | 24 ++++++--
.../renderer_opengl/gl_fence_manager.h | 6 +-
.../renderer_opengl/gl_rasterizer.cpp | 13 ++++-
.../renderer_opengl/gl_rasterizer.h | 3 +-
.../renderer_vulkan/vk_rasterizer.cpp | 16 ++++-
.../renderer_vulkan/vk_rasterizer.h | 3 +-
src/video_core/texture_cache/texture_cache.h | 7 +++
11 files changed, 123 insertions(+), 25 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 06fb931d75..54c75ca4e1 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -229,6 +229,13 @@ public:
return true;
}
+ bool HasUncommitedFlushes() {
+ if (uncommited_flushes) {
+ return true;
+ }
+ return false;
+ }
+
void PopAsyncFlushes() {
if (commited_flushes.empty()) {
return;
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index a7e9514330..2824ed707d 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -405,7 +405,7 @@ void Maxwell3D::ProcessQueryGet() {
switch (regs.query.query_get.operation) {
case Regs::QueryOperation::Release:
if (regs.query.query_get.fence == 1) {
- rasterizer.SignalFence(regs.query.QueryAddress(), regs.query.query_sequence);
+ rasterizer.SignalSemaphore(regs.query.QueryAddress(), regs.query.query_sequence);
} else {
StampQueryResult(regs.query.query_sequence, regs.query.query_get.short_query == 0);
}
@@ -487,7 +487,7 @@ void Maxwell3D::ProcessSyncPoint() {
const u32 increment = regs.sync_info.increment.Value();
[[maybe_unused]] const u32 cache_flush = regs.sync_info.unknown.Value();
if (increment) {
- system.GPU().IncrementSyncPoint(sync_point);
+ rasterizer.SignalSyncPoint(sync_point);
}
}
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 72ee509550..417cb113fb 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -22,7 +22,11 @@ namespace VideoCommon {
class FenceBase {
public:
- FenceBase(GPUVAddr address, u32 payload) : address{address}, payload{payload} {}
+ FenceBase(u32 payload, bool is_stubbed)
+ : address{}, payload{payload}, is_semaphore{false}, is_stubbed{is_stubbed} {}
+
+ FenceBase(GPUVAddr address, u32 payload, bool is_stubbed)
+ : address{address}, payload{payload}, is_semaphore{true}, is_stubbed{is_stubbed} {}
constexpr GPUVAddr GetAddress() const {
return address;
@@ -32,22 +36,49 @@ public:
return payload;
}
+ constexpr bool IsSemaphore() const {
+ return is_semaphore;
+ }
+
private:
GPUVAddr address;
u32 payload;
+ bool is_semaphore;
+
+protected:
+ bool is_stubbed;
};
template <typename TFence, typename TTextureCache, typename TTBufferCache>
class FenceManager {
public:
- void SignalFence(GPUVAddr addr, u32 value) {
+ void SignalSemaphore(GPUVAddr addr, u32 value) {
TryReleasePendingFences();
+ bool should_flush = texture_cache.HasUncommitedFlushes();
+ should_flush |= buffer_cache.HasUncommitedFlushes();
texture_cache.CommitAsyncFlushes();
buffer_cache.CommitAsyncFlushes();
- TFence new_fence = CreateFence(addr, value);
+ TFence new_fence = CreateFence(addr, value, !should_flush);
fences.push(new_fence);
QueueFence(new_fence);
- rasterizer.FlushCommands();
+ if (should_flush) {
+ rasterizer.FlushCommands();
+ }
+ rasterizer.SyncGuestHost();
+ }
+
+ void SignalSyncPoint(u32 value) {
+ TryReleasePendingFences();
+ bool should_flush = texture_cache.HasUncommitedFlushes();
+ should_flush |= buffer_cache.HasUncommitedFlushes();
+ texture_cache.CommitAsyncFlushes();
+ buffer_cache.CommitAsyncFlushes();
+ TFence new_fence = CreateFence(value, !should_flush);
+ fences.push(new_fence);
+ QueueFence(new_fence);
+ if (should_flush) {
+ rasterizer.FlushCommands();
+ }
rasterizer.SyncGuestHost();
}
@@ -62,8 +93,12 @@ public:
texture_cache.PopAsyncFlushes();
buffer_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ if (current_fence->IsSemaphore()) {
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ } else {
+ gpu.IncrementSyncPoint(current_fence->GetPayload());
+ }
fences.pop();
}
}
@@ -74,7 +109,8 @@ protected:
: system{system}, rasterizer{rasterizer}, texture_cache{texture_cache}, buffer_cache{
buffer_cache} {}
- virtual TFence CreateFence(GPUVAddr addr, u32 value) = 0;
+ virtual TFence CreateFence(u32 value, bool is_stubbed) = 0;
+ virtual TFence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) = 0;
virtual void QueueFence(TFence& fence) = 0;
virtual bool IsFenceSignaled(TFence& fence) = 0;
virtual void WaitFence(TFence& fence) = 0;
@@ -96,8 +132,12 @@ private:
texture_cache.PopAsyncFlushes();
buffer_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ if (current_fence->IsSemaphore()) {
+ auto& memory_manager{gpu.MemoryManager()};
+ memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
+ } else {
+ gpu.IncrementSyncPoint(current_fence->GetPayload());
+ }
fences.pop();
}
}
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index b49f15df2e..4e9c8fb595 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -49,8 +49,11 @@ public:
/// Records a GPU query and caches it
virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
- /// Signal a GPU based fence
- virtual void SignalFence(GPUVAddr addr, u32 value) = 0;
+ /// Signal a GPU based semaphore as a fence
+ virtual void SignalSemaphore(GPUVAddr addr, u32 value) = 0;
+
+ /// Signal a GPU based syncpoint as a fence
+ virtual void SignalSyncPoint(u32 value) = 0;
/// Release all pending fences.
virtual void ReleaseFences() = 0;
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
index 69dd3211b0..579c03a1ea 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -8,17 +8,26 @@
namespace OpenGL {
-GLInnerFence::GLInnerFence(GPUVAddr address, u32 payload)
- : VideoCommon::FenceBase(address, payload), sync_object{} {}
+GLInnerFence::GLInnerFence(u32 payload, bool is_stubbed)
+ : VideoCommon::FenceBase(payload, is_stubbed), sync_object{} {}
+
+GLInnerFence::GLInnerFence(GPUVAddr address, u32 payload, bool is_stubbed)
+ : VideoCommon::FenceBase(address, payload, is_stubbed), sync_object{} {}
GLInnerFence::~GLInnerFence() = default;
void GLInnerFence::Queue() {
+ if (is_stubbed) {
+ return;
+ }
ASSERT(sync_object.handle == 0);
sync_object.Create();
}
bool GLInnerFence::IsSignaled() const {
+ if (is_stubbed) {
+ return true;
+ }
ASSERT(sync_object.handle != 0);
GLsizei length;
GLint sync_status;
@@ -27,6 +36,9 @@ bool GLInnerFence::IsSignaled() const {
}
void GLInnerFence::Wait() {
+ if (is_stubbed) {
+ return;
+ }
ASSERT(sync_object.handle != 0);
while (glClientWaitSync(sync_object.handle, 0, 1000) == GL_TIMEOUT_EXPIRED)
;
@@ -36,8 +48,12 @@ FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system, VideoCore::Rasteriz
TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache)
: GenericFenceManager(system, rasterizer, texture_cache, buffer_cache) {}
-Fence FenceManagerOpenGL::CreateFence(GPUVAddr addr, u32 value) {
- return std::make_shared<GLInnerFence>(addr, value);
+Fence FenceManagerOpenGL::CreateFence(u32 value, bool is_stubbed) {
+ return std::make_shared<GLInnerFence>(value, is_stubbed);
+}
+
+Fence FenceManagerOpenGL::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) {
+ return std::make_shared<GLInnerFence>(addr, value, is_stubbed);
}
void FenceManagerOpenGL::QueueFence(Fence& fence) {
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.h b/src/video_core/renderer_opengl/gl_fence_manager.h
index b48d5eaa0d..ba48d2f840 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.h
+++ b/src/video_core/renderer_opengl/gl_fence_manager.h
@@ -17,7 +17,8 @@ namespace OpenGL {
class GLInnerFence : public VideoCommon::FenceBase {
public:
- GLInnerFence(GPUVAddr address, u32 payload);
+ GLInnerFence(u32 payload, bool is_stubbed);
+ GLInnerFence(GPUVAddr address, u32 payload, bool is_stubbed);
~GLInnerFence();
void Queue();
@@ -39,7 +40,8 @@ public:
TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache);
protected:
- Fence CreateFence(GPUVAddr addr, u32 value) override;
+ Fence CreateFence(u32 value, bool is_stubbed) override;
+ Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
void QueueFence(Fence& fence) override;
bool IsFenceSignaled(Fence& fence) override;
void WaitFence(Fence& fence) override;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 88914828ca..e52e5961f1 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -683,14 +683,23 @@ void RasterizerOpenGL::SyncGuestHost() {
buffer_cache.SyncGuestHost();
}
-void RasterizerOpenGL::SignalFence(GPUVAddr addr, u32 value) {
+void RasterizerOpenGL::SignalSemaphore(GPUVAddr addr, u32 value) {
auto& gpu{system.GPU()};
if (!gpu.IsAsync()) {
auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(addr, value);
return;
}
- fence_manager.SignalFence(addr, value);
+ fence_manager.SignalSemaphore(addr, value);
+}
+
+void RasterizerOpenGL::SignalSyncPoint(u32 value) {
+ auto& gpu{system.GPU()};
+ if (!gpu.IsAsync()) {
+ gpu.IncrementSyncPoint(value);
+ return;
+ }
+ fence_manager.SignalSyncPoint(value);
}
void RasterizerOpenGL::ReleaseFences() {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 5c0f88e6f1..15e9ff7d79 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -71,7 +71,8 @@ public:
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
- void SignalFence(GPUVAddr addr, u32 value) override;
+ void SignalSemaphore(GPUVAddr addr, u32 value) override;
+ void SignalSyncPoint(u32 value) override;
void ReleaseFences() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 1d75a4766a..507262c8fe 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -543,7 +543,7 @@ void RasterizerVulkan::SyncGuestHost() {
buffer_cache.SyncGuestHost();
}
-void RasterizerVulkan::SignalFence(GPUVAddr addr, u32 value) {
+void RasterizerVulkan::SignalSemaphore(GPUVAddr addr, u32 value) {
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(addr, value);
@@ -553,7 +553,19 @@ void RasterizerVulkan::SignalFence(GPUVAddr addr, u32 value) {
memory_manager.Write<u32>(addr, value);
return;
}
- fence_manager.SignalFence(addr, value);
+ fence_manager.SignalSemaphore(addr, value);
+ */
+}
+
+void RasterizerVulkan::SignalSyncPoint(u32 value) {
+ auto& gpu{system.GPU()};
+ gpu.IncrementSyncPoint(value);
+ /*
+ if (!gpu.IsAsync()) {
+ gpu.IncrementSyncPoint(value);
+ return;
+ }
+ fence_manager.SignalSyncPoint(value);
*/
}
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 08a9af401c..145bdf899a 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -122,7 +122,8 @@ public:
void InvalidateRegion(VAddr addr, u64 size) override;
void OnCPUWrite(VAddr addr, u64 size) override;
void SyncGuestHost() override;
- void SignalFence(GPUVAddr addr, u32 value) override;
+ void SignalSemaphore(GPUVAddr addr, u32 value) override;
+ void SignalSyncPoint(u32 value) override;
void ReleaseFences() override;
void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
void FlushCommands() override;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index e1a1edbd24..f3ca1ffd15 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -337,6 +337,13 @@ public:
uncommited_flushes.reset();
}
+ bool HasUncommitedFlushes() {
+ if (uncommited_flushes) {
+ return true;
+ }
+ return false;
+ }
+
bool ShouldWaitAsyncFlushes() {
if (commited_flushes.empty()) {
return false;
From 1fb516cd979ed0dbf8fa7cb4f6a334932dfb6434 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Thu, 20 Feb 2020 11:55:32 -0400
Subject: [PATCH 17/31] GPU: Implement Flush Requests for Async mode.
---
src/video_core/gpu.cpp | 22 +++++++++++++++++++
src/video_core/gpu.h | 21 ++++++++++++++++++
src/video_core/gpu_thread.cpp | 17 +++++++++-----
src/video_core/gpu_thread.h | 8 +++++--
.../renderer_opengl/gl_rasterizer.cpp | 6 +++++
.../renderer_vulkan/vk_rasterizer.cpp | 4 ++++
6 files changed, 70 insertions(+), 8 deletions(-)
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 19d3bd3052..85a6c7bb5a 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -125,6 +125,28 @@ bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
return true;
}
+u64 GPU::RequestFlush(CacheAddr addr, std::size_t size) {
+ std::unique_lock lck{flush_request_mutex};
+ const u64 fence = ++last_flush_fence;
+ flush_requests.emplace_back(fence, addr, size);
+ return fence;
+}
+
+void GPU::TickWork() {
+ std::unique_lock lck{flush_request_mutex};
+ while (!flush_requests.empty()) {
+ auto& request = flush_requests.front();
+ const u64 fence = request.fence;
+ const CacheAddr addr = request.addr;
+ const std::size_t size = request.size;
+ flush_requests.pop_front();
+ flush_request_mutex.unlock();
+ renderer->Rasterizer().FlushRegion(addr, size);
+ current_flush_fence.store(fence);
+ flush_request_mutex.lock();
+ }
+}
+
u64 GPU::GetTicks() const {
// This values were reversed engineered by fincs from NVN
// The gpu clock is reported in units of 385/625 nanoseconds
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index fa9991c871..943a5b1100 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -159,6 +159,14 @@ public:
void SyncGuestHost();
virtual void OnCommandListEnd();
+ u64 RequestFlush(CacheAddr addr, std::size_t size);
+
+ u64 CurrentFlushRequestFence() const {
+ return current_flush_fence.load(std::memory_order_relaxed);
+ }
+
+ void TickWork();
+
/// Returns a reference to the Maxwell3D GPU engine.
Engines::Maxwell3D& Maxwell3D();
@@ -327,6 +335,19 @@ private:
std::condition_variable sync_cv;
+ struct FlushRequest {
+ FlushRequest(u64 fence, CacheAddr addr, std::size_t size)
+ : fence{fence}, addr{addr}, size{size} {}
+ u64 fence;
+ CacheAddr addr;
+ std::size_t size;
+ };
+
+ std::list<FlushRequest> flush_requests;
+ std::atomic<u64> current_flush_fence{};
+ u64 last_flush_fence{};
+ std::mutex flush_request_mutex;
+
const bool is_async;
};
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 3e2be00e9d..9460364a3e 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -15,8 +15,9 @@
namespace VideoCommon::GPUThread {
/// Runs the GPU thread
-static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::GraphicsContext& context,
- Tegra::DmaPusher& dma_pusher, SynchState& state) {
+static void RunThread(Core::System& system, VideoCore::RendererBase& renderer,
+ Core::Frontend::GraphicsContext& context, Tegra::DmaPusher& dma_pusher,
+ SynchState& state) {
MicroProfileOnThreadCreate("GpuThread");
// Wait for first GPU command before acquiring the window context
@@ -40,6 +41,8 @@ static void RunThread(VideoCore::RendererBase& renderer, Core::Frontend::Graphic
renderer.SwapBuffers(data->framebuffer ? &*data->framebuffer : nullptr);
} else if (const auto data = std::get_if<OnCommandListEndCommand>(&next.data)) {
renderer.Rasterizer().ReleaseFences();
+ } else if (const auto data = std::get_if<GPUTickCommand>(&next.data)) {
+ system.GPU().TickWork();
} else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
renderer.Rasterizer().FlushRegion(data->addr, data->size);
} else if (const auto data = std::get_if<InvalidateRegionCommand>(&next.data)) {
@@ -68,8 +71,8 @@ ThreadManager::~ThreadManager() {
void ThreadManager::StartThread(VideoCore::RendererBase& renderer,
Core::Frontend::GraphicsContext& context,
Tegra::DmaPusher& dma_pusher) {
- thread = std::thread{RunThread, std::ref(renderer), std::ref(context), std::ref(dma_pusher),
- std::ref(state)};
+ thread = std::thread{RunThread, std::ref(system), std::ref(renderer),
+ std::ref(context), std::ref(dma_pusher), std::ref(state)};
}
void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
@@ -85,8 +88,10 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) {
return;
}
if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
- u64 fence = PushCommand(FlushRegionCommand(addr, size));
- while (fence > state.signaled_fence.load(std::memory_order_relaxed)) {
+ auto& gpu = system.GPU();
+ u64 fence = gpu.RequestFlush(addr, size);
+ PushCommand(GPUTickCommand());
+ while (fence > gpu.CurrentFlushRequestFence()) {
}
}
}
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index 9d08779216..5a28335d61 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -70,12 +70,16 @@ struct FlushAndInvalidateRegionCommand final {
u64 size;
};
-/// Command to signal to the GPU thread that processing has ended
+/// Command called within the gpu, to schedule actions after a command list end
struct OnCommandListEndCommand final {};
+/// Command to make the gpu look into pending requests
+struct GPUTickCommand final {};
+
using CommandData =
std::variant<EndProcessingCommand, SubmitListCommand, SwapBuffersCommand, FlushRegionCommand,
- InvalidateRegionCommand, FlushAndInvalidateRegionCommand, OnCommandListEndCommand>;
+ InvalidateRegionCommand, FlushAndInvalidateRegionCommand, OnCommandListEndCommand,
+ GPUTickCommand>;
struct CommandDataContainer {
CommandDataContainer() = default;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index e52e5961f1..fbd81b895f 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -601,6 +601,8 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
EndTransformFeedback();
++num_queued_commands;
+
+ system.GPU().TickWork();
}
void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
@@ -628,6 +630,7 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
glDispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
++num_queued_commands;
+ system.GPU().TickWork();
}
void RasterizerOpenGL::ResetCounter(VideoCore::QueryType type) {
@@ -652,6 +655,9 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
}
bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+ if (!Settings::IsGPULevelExtreme()) {
+ return buffer_cache.MustFlushRegion(addr, size);
+ }
return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
}
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 507262c8fe..926ecf38ea 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -365,6 +365,8 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
});
EndTransformFeedback();
+
+ system.GPU().TickWork();
}
void RasterizerVulkan::Clear() {
@@ -492,6 +494,8 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
descriptor_set, {});
cmdbuf.Dispatch(grid_x, grid_y, grid_z);
});
+
+ system.GPU().TickWork();
}
void RasterizerVulkan::ResetCounter(VideoCore::QueryType type) {
From 131b342130f21cf66ca64ece8034951ad6cce1f7 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Thu, 20 Feb 2020 13:14:03 -0400
Subject: [PATCH 18/31] OpenGL: Guarantee writes to Buffers.
---
src/video_core/renderer_opengl/gl_buffer_cache.cpp | 3 ++-
src/video_core/renderer_opengl/gl_rasterizer.cpp | 1 -
src/video_core/renderer_vulkan/vk_rasterizer.cpp | 2 --
3 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.cpp b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
index cb5792407e..4efce0de77 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
@@ -52,7 +52,7 @@ Buffer OGLBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
}
void OGLBufferCache::WriteBarrier() {
- glMemoryBarrier(GL_ALL_BARRIER_BITS);
+ glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
}
GLuint OGLBufferCache::ToHandle(const Buffer& buffer) {
@@ -72,6 +72,7 @@ void OGLBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, s
void OGLBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
u8* data) {
MICROPROFILE_SCOPE(OpenGL_Buffer_Download);
+ glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
glGetNamedBufferSubData(buffer->GetHandle(), static_cast<GLintptr>(offset),
static_cast<GLsizeiptr>(size), data);
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index fbd81b895f..bc57d396ec 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -630,7 +630,6 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
glDispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
++num_queued_commands;
- system.GPU().TickWork();
}
void RasterizerOpenGL::ResetCounter(VideoCore::QueryType type) {
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 926ecf38ea..f8b5a5a929 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -494,8 +494,6 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
descriptor_set, {});
cmdbuf.Dispatch(grid_x, grid_y, grid_z);
});
-
- system.GPU().TickWork();
}
void RasterizerVulkan::ResetCounter(VideoCore::QueryType type) {
From 0649f0590047e8cc0b16a10dec5eb74938fef718 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 16:36:14 -0400
Subject: [PATCH 19/31] QueryCache: Implement Async Flushes.
---
src/video_core/fence_manager.h | 18 +++++--
src/video_core/query_cache.h | 50 +++++++++++++++++++
.../renderer_opengl/gl_fence_manager.cpp | 8 +--
.../renderer_opengl/gl_fence_manager.h | 7 ++-
.../renderer_opengl/gl_rasterizer.cpp | 6 +--
5 files changed, 77 insertions(+), 12 deletions(-)
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 417cb113fb..99a138b5b7 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -49,15 +49,17 @@ protected:
bool is_stubbed;
};
-template <typename TFence, typename TTextureCache, typename TTBufferCache>
+template <typename TFence, typename TTextureCache, typename TTBufferCache, typename TQueryCache>
class FenceManager {
public:
void SignalSemaphore(GPUVAddr addr, u32 value) {
TryReleasePendingFences();
bool should_flush = texture_cache.HasUncommitedFlushes();
should_flush |= buffer_cache.HasUncommitedFlushes();
+ should_flush |= query_cache.HasUncommitedFlushes();
texture_cache.CommitAsyncFlushes();
buffer_cache.CommitAsyncFlushes();
+ query_cache.CommitAsyncFlushes();
TFence new_fence = CreateFence(addr, value, !should_flush);
fences.push(new_fence);
QueueFence(new_fence);
@@ -71,8 +73,10 @@ public:
TryReleasePendingFences();
bool should_flush = texture_cache.HasUncommitedFlushes();
should_flush |= buffer_cache.HasUncommitedFlushes();
+ should_flush |= query_cache.HasUncommitedFlushes();
texture_cache.CommitAsyncFlushes();
buffer_cache.CommitAsyncFlushes();
+ query_cache.CommitAsyncFlushes();
TFence new_fence = CreateFence(value, !should_flush);
fences.push(new_fence);
QueueFence(new_fence);
@@ -87,11 +91,13 @@ public:
TFence& current_fence = fences.front();
bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
+ should_wait |= query_cache.ShouldWaitAsyncFlushes();
if (should_wait) {
WaitFence(current_fence);
}
texture_cache.PopAsyncFlushes();
buffer_cache.PopAsyncFlushes();
+ query_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
auto& memory_manager{gpu.MemoryManager()};
@@ -105,9 +111,10 @@ public:
protected:
FenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TTextureCache& texture_cache, TTBufferCache& buffer_cache)
- : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache}, buffer_cache{
- buffer_cache} {}
+ TTextureCache& texture_cache, TTBufferCache& buffer_cache,
+ TQueryCache& query_cache)
+ : system{system}, rasterizer{rasterizer}, texture_cache{texture_cache},
+ buffer_cache{buffer_cache}, query_cache{query_cache} {}
virtual TFence CreateFence(u32 value, bool is_stubbed) = 0;
virtual TFence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) = 0;
@@ -119,6 +126,7 @@ protected:
VideoCore::RasterizerInterface& rasterizer;
TTextureCache& texture_cache;
TTBufferCache& buffer_cache;
+ TQueryCache& query_cache;
private:
void TryReleasePendingFences() {
@@ -126,11 +134,13 @@ private:
TFence& current_fence = fences.front();
bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
+ should_wait |= query_cache.ShouldWaitAsyncFlushes();
if (should_wait && !IsFenceSignaled(current_fence)) {
return;
}
texture_cache.PopAsyncFlushes();
buffer_cache.PopAsyncFlushes();
+ query_cache.PopAsyncFlushes();
auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
auto& memory_manager{gpu.MemoryManager()};
diff --git a/src/video_core/query_cache.h b/src/video_core/query_cache.h
index 5ea2b01f2a..1b1c23995f 100644
--- a/src/video_core/query_cache.h
+++ b/src/video_core/query_cache.h
@@ -12,6 +12,7 @@
#include <mutex>
#include <optional>
#include <unordered_map>
+#include <unordered_set>
#include <vector>
#include "common/assert.h"
@@ -130,6 +131,7 @@ public:
}
query->BindCounter(Stream(type).Current(), timestamp);
+ AsyncFlushQuery(cpu_addr);
}
/// Updates counters from GPU state. Expected to be called once per draw, clear or dispatch.
@@ -170,6 +172,44 @@ public:
return streams[static_cast<std::size_t>(type)];
}
+ void CommitAsyncFlushes() {
+ commited_flushes.push_back(uncommited_flushes);
+ uncommited_flushes.reset();
+ }
+
+ bool HasUncommitedFlushes() {
+ if (uncommited_flushes) {
+ return true;
+ }
+ return false;
+ }
+
+ bool ShouldWaitAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return false;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ return false;
+ }
+ return true;
+ }
+
+ void PopAsyncFlushes() {
+ if (commited_flushes.empty()) {
+ return;
+ }
+ auto& flush_list = commited_flushes.front();
+ if (!flush_list) {
+ commited_flushes.pop_front();
+ return;
+ }
+ for (VAddr query_address : *flush_list) {
+ FlushAndRemoveRegion(query_address, 4);
+ }
+ commited_flushes.pop_front();
+ }
+
protected:
std::array<QueryPool, VideoCore::NumQueryTypes> query_pools;
@@ -224,6 +264,13 @@ private:
return found != std::end(contents) ? &*found : nullptr;
}
+ void AsyncFlushQuery(VAddr addr) {
+ if (!uncommited_flushes) {
+ uncommited_flushes = std::make_shared<std::unordered_set<VAddr>>();
+ }
+ uncommited_flushes->insert(addr);
+ }
+
static constexpr std::uintptr_t PAGE_SIZE = 4096;
static constexpr unsigned PAGE_SHIFT = 12;
@@ -235,6 +282,9 @@ private:
std::unordered_map<u64, std::vector<CachedQuery>> cached_queries;
std::array<CounterStream, VideoCore::NumQueryTypes> streams;
+
+ std::shared_ptr<std::unordered_set<VAddr>> uncommited_flushes{};
+ std::list<std::shared_ptr<std::unordered_set<VAddr>>> commited_flushes;
};
template <class QueryCache, class HostCounter>
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
index 579c03a1ea..aa57a0ae01 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -44,9 +44,11 @@ void GLInnerFence::Wait() {
;
}
-FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache)
- : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache) {}
+FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system,
+ VideoCore::RasterizerInterface& rasterizer,
+ TextureCacheOpenGL& texture_cache,
+ OGLBufferCache& buffer_cache, QueryCache& query_cache)
+ : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache, query_cache) {}
Fence FenceManagerOpenGL::CreateFence(u32 value, bool is_stubbed) {
return std::make_shared<GLInnerFence>(value, is_stubbed);
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.h b/src/video_core/renderer_opengl/gl_fence_manager.h
index ba48d2f840..c76e69cb8d 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.h
+++ b/src/video_core/renderer_opengl/gl_fence_manager.h
@@ -10,6 +10,7 @@
#include "common/common_types.h"
#include "video_core/fence_manager.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
+#include "video_core/renderer_opengl/gl_query_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_texture_cache.h"
@@ -32,12 +33,14 @@ private:
};
using Fence = std::shared_ptr<GLInnerFence>;
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCacheOpenGL, OGLBufferCache>;
+using GenericFenceManager =
+ VideoCommon::FenceManager<Fence, TextureCacheOpenGL, OGLBufferCache, QueryCache>;
class FenceManagerOpenGL final : public GenericFenceManager {
public:
FenceManagerOpenGL(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
- TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache);
+ TextureCacheOpenGL& texture_cache, OGLBufferCache& buffer_cache,
+ QueryCache& query_cache);
protected:
Fence CreateFence(u32 value, bool is_stubbed) override;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index bc57d396ec..6d3b5f3f40 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -101,9 +101,9 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
: RasterizerAccelerated{system.Memory()}, texture_cache{system, *this, device, state_tracker},
shader_cache{*this, system, emu_window, device}, query_cache{system, *this}, system{system},
screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker},
- buffer_cache{*this, system, device, STREAM_BUFFER_SIZE}, fence_manager{system, *this,
- texture_cache,
- buffer_cache} {
+ buffer_cache{*this, system, device, STREAM_BUFFER_SIZE}, fence_manager{
+ system, *this, texture_cache,
+ buffer_cache, query_cache} {
CheckExtensions();
}
From b752faf2d3aae882a1a35a3aec393ef5765c035f Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 16 Mar 2020 21:43:05 -0300
Subject: [PATCH 20/31] vk_fence_manager: Initial implementation
---
src/video_core/CMakeLists.txt | 2 +
src/video_core/buffer_cache/buffer_cache.h | 1 +
.../renderer_vulkan/vk_fence_manager.cpp | 101 ++++++++++++++++++
.../renderer_vulkan/vk_fence_manager.h | 74 +++++++++++++
.../renderer_vulkan/vk_rasterizer.cpp | 16 +--
.../renderer_vulkan/vk_rasterizer.h | 2 +
src/video_core/renderer_vulkan/wrapper.cpp | 18 ++++
src/video_core/renderer_vulkan/wrapper.h | 20 ++++
8 files changed, 222 insertions(+), 12 deletions(-)
create mode 100644 src/video_core/renderer_vulkan/vk_fence_manager.cpp
create mode 100644 src/video_core/renderer_vulkan/vk_fence_manager.h
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 9a3f568f9c..55047dde41 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -177,6 +177,8 @@ if (ENABLE_VULKAN)
renderer_vulkan/vk_descriptor_pool.h
renderer_vulkan/vk_device.cpp
renderer_vulkan/vk_device.h
+ renderer_vulkan/vk_fence_manager.cpp
+ renderer_vulkan/vk_fence_manager.h
renderer_vulkan/vk_graphics_pipeline.cpp
renderer_vulkan/vk_graphics_pipeline.h
renderer_vulkan/vk_image.cpp
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 54c75ca4e1..3725450803 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -19,6 +19,7 @@
#include "common/alignment.h"
#include "common/common_types.h"
+#include "common/logging/log.h"
#include "core/core.h"
#include "core/memory.h"
#include "core/settings.h"
diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.cpp b/src/video_core/renderer_vulkan/vk_fence_manager.cpp
new file mode 100644
index 0000000000..a2b2bc408f
--- /dev/null
+++ b/src/video_core/renderer_vulkan/vk_fence_manager.cpp
@@ -0,0 +1,101 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <memory>
+#include <thread>
+
+#include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_device.h"
+#include "video_core/renderer_vulkan/vk_fence_manager.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_texture_cache.h"
+#include "video_core/renderer_vulkan/wrapper.h"
+
+namespace Vulkan {
+
+InnerFence::InnerFence(const VKDevice& device, VKScheduler& scheduler, u32 payload, bool is_stubbed)
+ : VideoCommon::FenceBase(payload, is_stubbed), device{device}, scheduler{scheduler} {}
+
+InnerFence::InnerFence(const VKDevice& device, VKScheduler& scheduler, GPUVAddr address,
+ u32 payload, bool is_stubbed)
+ : VideoCommon::FenceBase(address, payload, is_stubbed), device{device}, scheduler{scheduler} {}
+
+InnerFence::~InnerFence() = default;
+
+void InnerFence::Queue() {
+ if (is_stubbed) {
+ return;
+ }
+ ASSERT(!event);
+
+ event = device.GetLogical().CreateEvent();
+ ticks = scheduler.Ticks();
+
+ scheduler.RequestOutsideRenderPassOperationContext();
+ scheduler.Record([event = *event](vk::CommandBuffer cmdbuf) {
+ cmdbuf.SetEvent(event, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
+ });
+}
+
+bool InnerFence::IsSignaled() const {
+ if (is_stubbed) {
+ return true;
+ }
+ ASSERT(event);
+ return IsEventSignalled();
+}
+
+void InnerFence::Wait() {
+ if (is_stubbed) {
+ return;
+ }
+ ASSERT(event);
+
+ if (ticks >= scheduler.Ticks()) {
+ scheduler.Flush();
+ }
+ while (!IsEventSignalled()) {
+ std::this_thread::yield();
+ }
+}
+
+bool InnerFence::IsEventSignalled() const {
+ switch (const VkResult result = event.GetStatus()) {
+ case VK_EVENT_SET:
+ return true;
+ case VK_EVENT_RESET:
+ return false;
+ default:
+ throw vk::Exception(result);
+ }
+}
+
+VKFenceManager::VKFenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+ const VKDevice& device, VKScheduler& scheduler,
+ VKTextureCache& texture_cache, VKBufferCache& buffer_cache,
+ VKQueryCache& query_cache)
+ : GenericFenceManager(system, rasterizer, texture_cache, buffer_cache, query_cache),
+ device{device}, scheduler{scheduler} {}
+
+Fence VKFenceManager::CreateFence(u32 value, bool is_stubbed) {
+ return std::make_shared<InnerFence>(device, scheduler, value, is_stubbed);
+}
+
+Fence VKFenceManager::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) {
+ return std::make_shared<InnerFence>(device, scheduler, addr, value, is_stubbed);
+}
+
+void VKFenceManager::QueueFence(Fence& fence) {
+ fence->Queue();
+}
+
+bool VKFenceManager::IsFenceSignaled(Fence& fence) {
+ return fence->IsSignaled();
+}
+
+void VKFenceManager::WaitFence(Fence& fence) {
+ fence->Wait();
+}
+
+} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.h b/src/video_core/renderer_vulkan/vk_fence_manager.h
new file mode 100644
index 0000000000..30651e9c72
--- /dev/null
+++ b/src/video_core/renderer_vulkan/vk_fence_manager.h
@@ -0,0 +1,74 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+
+#include "video_core/fence_manager.h"
+#include "video_core/renderer_vulkan/wrapper.h"
+
+namespace Core {
+class System;
+}
+
+namespace VideoCore {
+class RasterizerInterface;
+}
+
+namespace Vulkan {
+
+class VKBufferCache;
+class VKDevice;
+class VKQueryCache;
+class VKScheduler;
+class VKTextureCache;
+
+class InnerFence : public VideoCommon::FenceBase {
+public:
+ explicit InnerFence(const VKDevice& device, VKScheduler& scheduler, u32 payload,
+ bool is_stubbed);
+ explicit InnerFence(const VKDevice& device, VKScheduler& scheduler, GPUVAddr address,
+ u32 payload, bool is_stubbed);
+ ~InnerFence();
+
+ void Queue();
+
+ bool IsSignaled() const;
+
+ void Wait();
+
+private:
+ bool IsEventSignalled() const;
+
+ const VKDevice& device;
+ VKScheduler& scheduler;
+ vk::Event event;
+ u64 ticks = 0;
+};
+using Fence = std::shared_ptr<InnerFence>;
+
+using GenericFenceManager =
+ VideoCommon::FenceManager<Fence, VKTextureCache, VKBufferCache, VKQueryCache>;
+
+class VKFenceManager final : public GenericFenceManager {
+public:
+ explicit VKFenceManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
+ const VKDevice& device, VKScheduler& scheduler,
+ VKTextureCache& texture_cache, VKBufferCache& buffer_cache,
+ VKQueryCache& query_cache);
+
+protected:
+ Fence CreateFence(u32 value, bool is_stubbed) override;
+ Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
+ void QueueFence(Fence& fence) override;
+ bool IsFenceSignaled(Fence& fence) override;
+ void WaitFence(Fence& fence) override;
+
+private:
+ const VKDevice& device;
+ VKScheduler& scheduler;
+};
+
+} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index f8b5a5a929..4dc7555aa0 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -299,7 +299,9 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind
pipeline_cache(system, *this, device, scheduler, descriptor_pool, update_descriptor_queue,
renderpass_cache),
buffer_cache(*this, system, device, memory_manager, scheduler, staging_pool),
- sampler_cache(device), query_cache(system, *this, device, scheduler) {
+ sampler_cache(device),
+ fence_manager(system, *this, device, scheduler, texture_cache, buffer_cache, query_cache),
+ query_cache(system, *this, device, scheduler) {
scheduler.SetQueryCache(query_cache);
}
@@ -547,38 +549,28 @@ void RasterizerVulkan::SyncGuestHost() {
void RasterizerVulkan::SignalSemaphore(GPUVAddr addr, u32 value) {
auto& gpu{system.GPU()};
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(addr, value);
- /*
if (!gpu.IsAsync()) {
- auto& memory_manager{gpu.MemoryManager()};
- memory_manager.Write<u32>(addr, value);
+ gpu.MemoryManager().Write<u32>(addr, value);
return;
}
fence_manager.SignalSemaphore(addr, value);
- */
}
void RasterizerVulkan::SignalSyncPoint(u32 value) {
auto& gpu{system.GPU()};
- gpu.IncrementSyncPoint(value);
- /*
if (!gpu.IsAsync()) {
gpu.IncrementSyncPoint(value);
return;
}
fence_manager.SignalSyncPoint(value);
- */
}
void RasterizerVulkan::ReleaseFences() {
- /*
auto& gpu{system.GPU()};
if (!gpu.IsAsync()) {
return;
}
fence_manager.WaitPendingFences();
- */
}
void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 145bdf899a..2fa46b0cc4 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -21,6 +21,7 @@
#include "video_core/renderer_vulkan/vk_buffer_cache.h"
#include "video_core/renderer_vulkan/vk_compute_pass.h"
#include "video_core/renderer_vulkan/vk_descriptor_pool.h"
+#include "video_core/renderer_vulkan/vk_fence_manager.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_pipeline_cache.h"
#include "video_core/renderer_vulkan/vk_query_cache.h"
@@ -267,6 +268,7 @@ private:
VKPipelineCache pipeline_cache;
VKBufferCache buffer_cache;
VKSamplerCache sampler_cache;
+ VKFenceManager fence_manager;
VKQueryCache query_cache;
std::array<View, Maxwell::NumRenderTargets> color_attachments;
diff --git a/src/video_core/renderer_vulkan/wrapper.cpp b/src/video_core/renderer_vulkan/wrapper.cpp
index 9b94dfff18..5e24b8f052 100644
--- a/src/video_core/renderer_vulkan/wrapper.cpp
+++ b/src/video_core/renderer_vulkan/wrapper.cpp
@@ -64,6 +64,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
X(vkCmdSetCheckpointNV);
X(vkCmdSetDepthBias);
X(vkCmdSetDepthBounds);
+ X(vkCmdSetEvent);
X(vkCmdSetScissor);
X(vkCmdSetStencilCompareMask);
X(vkCmdSetStencilReference);
@@ -76,6 +77,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
X(vkCreateDescriptorPool);
X(vkCreateDescriptorSetLayout);
X(vkCreateDescriptorUpdateTemplateKHR);
+ X(vkCreateEvent);
X(vkCreateFence);
X(vkCreateFramebuffer);
X(vkCreateGraphicsPipelines);
@@ -94,6 +96,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
X(vkDestroyDescriptorPool);
X(vkDestroyDescriptorSetLayout);
X(vkDestroyDescriptorUpdateTemplateKHR);
+ X(vkDestroyEvent);
X(vkDestroyFence);
X(vkDestroyFramebuffer);
X(vkDestroyImage);
@@ -113,6 +116,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
X(vkFreeMemory);
X(vkGetBufferMemoryRequirements);
X(vkGetDeviceQueue);
+ X(vkGetEventStatus);
X(vkGetFenceStatus);
X(vkGetImageMemoryRequirements);
X(vkGetQueryPoolResults);
@@ -271,6 +275,10 @@ void Destroy(VkDevice device, VkDeviceMemory handle, const DeviceDispatch& dld)
dld.vkFreeMemory(device, handle, nullptr);
}
+void Destroy(VkDevice device, VkEvent handle, const DeviceDispatch& dld) noexcept {
+ dld.vkDestroyEvent(device, handle, nullptr);
+}
+
void Destroy(VkDevice device, VkFence handle, const DeviceDispatch& dld) noexcept {
dld.vkDestroyFence(device, handle, nullptr);
}
@@ -613,6 +621,16 @@ ShaderModule Device::CreateShaderModule(const VkShaderModuleCreateInfo& ci) cons
return ShaderModule(object, handle, *dld);
}
+Event Device::CreateEvent() const {
+ VkEventCreateInfo ci;
+ ci.sType = VK_STRUCTURE_TYPE_EVENT_CREATE_INFO;
+ ci.pNext = nullptr;
+ ci.flags = 0;
+ VkEvent object;
+ Check(dld->vkCreateEvent(handle, &ci, nullptr, &object));
+ return Event(object, handle, *dld);
+}
+
SwapchainKHR Device::CreateSwapchainKHR(const VkSwapchainCreateInfoKHR& ci) const {
VkSwapchainKHR object;
Check(dld->vkCreateSwapchainKHR(handle, &ci, nullptr, &object));
diff --git a/src/video_core/renderer_vulkan/wrapper.h b/src/video_core/renderer_vulkan/wrapper.h
index fb36578198..cd2b8e6c9e 100644
--- a/src/video_core/renderer_vulkan/wrapper.h
+++ b/src/video_core/renderer_vulkan/wrapper.h
@@ -200,6 +200,7 @@ struct DeviceDispatch : public InstanceDispatch {
PFN_vkCmdSetCheckpointNV vkCmdSetCheckpointNV;
PFN_vkCmdSetDepthBias vkCmdSetDepthBias;
PFN_vkCmdSetDepthBounds vkCmdSetDepthBounds;
+ PFN_vkCmdSetEvent vkCmdSetEvent;
PFN_vkCmdSetScissor vkCmdSetScissor;
PFN_vkCmdSetStencilCompareMask vkCmdSetStencilCompareMask;
PFN_vkCmdSetStencilReference vkCmdSetStencilReference;
@@ -212,6 +213,7 @@ struct DeviceDispatch : public InstanceDispatch {
PFN_vkCreateDescriptorPool vkCreateDescriptorPool;
PFN_vkCreateDescriptorSetLayout vkCreateDescriptorSetLayout;
PFN_vkCreateDescriptorUpdateTemplateKHR vkCreateDescriptorUpdateTemplateKHR;
+ PFN_vkCreateEvent vkCreateEvent;
PFN_vkCreateFence vkCreateFence;
PFN_vkCreateFramebuffer vkCreateFramebuffer;
PFN_vkCreateGraphicsPipelines vkCreateGraphicsPipelines;
@@ -230,6 +232,7 @@ struct DeviceDispatch : public InstanceDispatch {
PFN_vkDestroyDescriptorPool vkDestroyDescriptorPool;
PFN_vkDestroyDescriptorSetLayout vkDestroyDescriptorSetLayout;
PFN_vkDestroyDescriptorUpdateTemplateKHR vkDestroyDescriptorUpdateTemplateKHR;
+ PFN_vkDestroyEvent vkDestroyEvent;
PFN_vkDestroyFence vkDestroyFence;
PFN_vkDestroyFramebuffer vkDestroyFramebuffer;
PFN_vkDestroyImage vkDestroyImage;
@@ -249,6 +252,7 @@ struct DeviceDispatch : public InstanceDispatch {
PFN_vkFreeMemory vkFreeMemory;
PFN_vkGetBufferMemoryRequirements vkGetBufferMemoryRequirements;
PFN_vkGetDeviceQueue vkGetDeviceQueue;
+ PFN_vkGetEventStatus vkGetEventStatus;
PFN_vkGetFenceStatus vkGetFenceStatus;
PFN_vkGetImageMemoryRequirements vkGetImageMemoryRequirements;
PFN_vkGetQueryPoolResults vkGetQueryPoolResults;
@@ -281,6 +285,7 @@ void Destroy(VkDevice, VkDescriptorPool, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkDescriptorSetLayout, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkDescriptorUpdateTemplateKHR, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkDeviceMemory, const DeviceDispatch&) noexcept;
+void Destroy(VkDevice, VkEvent, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkFence, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkFramebuffer, const DeviceDispatch&) noexcept;
void Destroy(VkDevice, VkImage, const DeviceDispatch&) noexcept;
@@ -654,6 +659,15 @@ public:
std::vector<VkImage> GetImages() const;
};
+class Event : public Handle<VkEvent, VkDevice, DeviceDispatch> {
+ using Handle<VkEvent, VkDevice, DeviceDispatch>::Handle;
+
+public:
+ VkResult GetStatus() const noexcept {
+ return dld->vkGetEventStatus(owner, handle);
+ }
+};
+
class Device : public Handle<VkDevice, NoOwner, DeviceDispatch> {
using Handle<VkDevice, NoOwner, DeviceDispatch>::Handle;
@@ -702,6 +716,8 @@ public:
ShaderModule CreateShaderModule(const VkShaderModuleCreateInfo& ci) const;
+ Event CreateEvent() const;
+
SwapchainKHR CreateSwapchainKHR(const VkSwapchainCreateInfoKHR& ci) const;
DeviceMemory TryAllocateMemory(const VkMemoryAllocateInfo& ai) const noexcept;
@@ -956,6 +972,10 @@ public:
dld->vkCmdSetDepthBounds(handle, min_depth_bounds, max_depth_bounds);
}
+ void SetEvent(VkEvent event, VkPipelineStageFlags stage_flags) const noexcept {
+ dld->vkCmdSetEvent(handle, event, stage_flags);
+ }
+
void BindTransformFeedbackBuffersEXT(u32 first, u32 count, const VkBuffer* buffers,
const VkDeviceSize* offsets,
const VkDeviceSize* sizes) const noexcept {
From f4ab223ef0eca55666de32c7f9b9b591e6c17235 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 20:07:32 -0400
Subject: [PATCH 21/31] Async GPU: Only do reactive flushing on Extreme Level.
---
src/video_core/gpu_thread.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 9460364a3e..7df854a2f9 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -84,7 +84,7 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
- if (!Settings::IsGPULevelHigh()) {
+ if (!Settings::IsGPULevelExtreme()) {
return;
}
if (system.Renderer().Rasterizer().MustFlushRegion(addr, size)) {
From b3e5f177ba3de381c4ad4e4b20a3bd17e4577e24 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 21:03:30 -0400
Subject: [PATCH 22/31] QueryCache: Only do async flushes on async gpu.
---
src/video_core/query_cache.h | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/src/video_core/query_cache.h b/src/video_core/query_cache.h
index 1b1c23995f..98d956b68c 100644
--- a/src/video_core/query_cache.h
+++ b/src/video_core/query_cache.h
@@ -17,6 +17,7 @@
#include "common/assert.h"
#include "core/core.h"
+#include "core/settings.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/gpu.h"
#include "video_core/memory_manager.h"
@@ -131,7 +132,9 @@ public:
}
query->BindCounter(Stream(type).Current(), timestamp);
- AsyncFlushQuery(cpu_addr);
+ if (Settings::values.use_asynchronous_gpu_emulation) {
+ AsyncFlushQuery(cpu_addr);
+ }
}
/// Updates counters from GPU state. Expected to be called once per draw, clear or dispatch.
From d2d4a6cbcf089d6a56adb84358155dbc1bff5ee6 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 21:17:06 -0400
Subject: [PATCH 23/31] Clang format.
---
src/yuzu/configuration/config.cpp | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp
index da1fa4e029..196a3a1168 100644
--- a/src/yuzu/configuration/config.cpp
+++ b/src/yuzu/configuration/config.cpp
@@ -1080,7 +1080,8 @@ void Config::SaveRendererValues() {
WriteSetting(QStringLiteral("frame_limit"), Settings::values.frame_limit, 100);
WriteSetting(QStringLiteral("use_disk_shader_cache"), Settings::values.use_disk_shader_cache,
true);
- WriteSetting(QStringLiteral("gpu_accuracy"), static_cast<int>(Settings::values.gpu_accuracy), 0);
+ WriteSetting(QStringLiteral("gpu_accuracy"), static_cast<int>(Settings::values.gpu_accuracy),
+ 0);
WriteSetting(QStringLiteral("use_asynchronous_gpu_emulation"),
Settings::values.use_asynchronous_gpu_emulation, false);
WriteSetting(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
From 7f44f224517a853be945350e6f537a7977f7e41c Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 21:42:05 -0400
Subject: [PATCH 24/31] Correct Linux Compile Error.
---
src/core/settings.cpp | 8 ++++++++
src/core/settings.h | 9 ++-------
2 files changed, 10 insertions(+), 7 deletions(-)
diff --git a/src/core/settings.cpp b/src/core/settings.cpp
index 445047469e..cd6c257f5a 100644
--- a/src/core/settings.cpp
+++ b/src/core/settings.cpp
@@ -109,4 +109,12 @@ void LogSettings() {
LogSetting("Services_BCATBoxcatLocal", Settings::values.bcat_boxcat_local);
}
+bool IsGPULevelExtreme() {
+ return values.gpu_accuracy == GPUAccuracy::Extreme;
+}
+
+bool IsGPULevelHigh() {
+ return values.gpu_accuracy == GPUAccuracy::Extreme || values.gpu_accuracy == GPUAccuracy::High;
+}
+
} // namespace Settings
diff --git a/src/core/settings.h b/src/core/settings.h
index b54a0d4ea9..7d09253f58 100644
--- a/src/core/settings.h
+++ b/src/core/settings.h
@@ -486,13 +486,8 @@ struct Values {
std::map<u64, std::vector<std::string>> disabled_addons;
} extern values;
-constexpr bool IsGPULevelExtreme() {
- return values.gpu_accuracy == GPUAccuracy::Extreme;
-}
-
-constexpr bool IsGPULevelHigh() {
- return values.gpu_accuracy == GPUAccuracy::Extreme || values.gpu_accuracy == GPUAccuracy::High;
-}
+bool IsGPULevelExtreme();
+bool IsGPULevelHigh();
void Apply();
void LogSettings();
From ec2f3e48e196508bb2e777511a6d2138e325e722 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 15 Apr 2020 22:59:29 -0400
Subject: [PATCH 25/31] Fix GCC error.
---
src/video_core/renderer_opengl/gl_rasterizer.cpp | 9 ++++-----
src/video_core/renderer_opengl/gl_rasterizer.h | 2 +-
2 files changed, 5 insertions(+), 6 deletions(-)
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 6d3b5f3f40..847d67159b 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -99,11 +99,10 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
ScreenInfo& info, GLShader::ProgramManager& program_manager,
StateTracker& state_tracker)
: RasterizerAccelerated{system.Memory()}, texture_cache{system, *this, device, state_tracker},
- shader_cache{*this, system, emu_window, device}, query_cache{system, *this}, system{system},
- screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker},
- buffer_cache{*this, system, device, STREAM_BUFFER_SIZE}, fence_manager{
- system, *this, texture_cache,
- buffer_cache, query_cache} {
+ shader_cache{*this, system, emu_window, device}, query_cache{system, *this},
+ buffer_cache{*this, system, device, STREAM_BUFFER_SIZE},
+ fence_manager{system, *this, texture_cache, buffer_cache, query_cache}, system{system},
+ screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker} {
CheckExtensions();
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 15e9ff7d79..ebd2173eba 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -229,6 +229,7 @@ private:
SamplerCacheOpenGL sampler_cache;
FramebufferCacheOpenGL framebuffer_cache;
QueryCache query_cache;
+ OGLBufferCache buffer_cache;
FenceManagerOpenGL fence_manager;
Core::System& system;
@@ -237,7 +238,6 @@ private:
StateTracker& state_tracker;
static constexpr std::size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
- OGLBufferCache buffer_cache;
GLint vertex_binding = 0;
From f616dc0b591b783b3fb75ca89633f1c26cce05a9 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Thu, 16 Apr 2020 12:29:53 -0400
Subject: [PATCH 26/31] Address Feedback.
---
src/video_core/buffer_cache/buffer_cache.h | 56 ++++++---------
src/video_core/fence_manager.h | 72 +++++++++++--------
src/video_core/gpu.cpp | 4 +-
src/video_core/gpu.h | 12 +++-
src/video_core/query_cache.h | 39 +++++-----
src/video_core/rasterizer_interface.h | 1 +
.../renderer_opengl/gl_fence_manager.cpp | 2 +-
.../renderer_opengl/gl_fence_manager.h | 2 +-
.../renderer_opengl/gl_rasterizer.cpp | 5 +-
.../renderer_vulkan/vk_fence_manager.cpp | 2 +-
.../renderer_vulkan/vk_fence_manager.h | 2 +-
.../renderer_vulkan/vk_rasterizer.cpp | 2 +-
src/video_core/texture_cache/texture_cache.h | 50 +++++--------
13 files changed, 117 insertions(+), 132 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 3725450803..f3aa352959 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -154,12 +154,9 @@ public:
std::lock_guard lock{mutex};
std::vector<MapInterval> objects = GetMapsInRange(addr, size);
- for (auto& object : objects) {
- if (object->IsModified() && object->IsRegistered()) {
- return true;
- }
- }
- return false;
+ return std::any_of(objects.begin(), objects.end(), [](const MapInterval& map) {
+ return map->IsModified() && map->IsRegistered();
+ });
}
/// Mark the specified region as being invalidated
@@ -199,9 +196,9 @@ public:
}
void CommitAsyncFlushes() {
- if (uncommited_flushes) {
+ if (uncommitted_flushes) {
auto commit_list = std::make_shared<std::list<MapInterval>>();
- for (auto& map : *uncommited_flushes) {
+ for (auto& map : *uncommitted_flushes) {
if (map->IsRegistered() && map->IsModified()) {
// TODO(Blinkhawk): Implement backend asynchronous flushing
// AsyncFlushMap(map)
@@ -209,41 +206,34 @@ public:
}
}
if (!commit_list->empty()) {
- commited_flushes.push_back(commit_list);
+ committed_flushes.push_back(commit_list);
} else {
- commited_flushes.emplace_back();
+ committed_flushes.emplace_back();
}
} else {
- commited_flushes.emplace_back();
+ committed_flushes.emplace_back();
}
- uncommited_flushes.reset();
+ uncommitted_flushes.reset();
}
- bool ShouldWaitAsyncFlushes() {
- if (commited_flushes.empty()) {
+ bool ShouldWaitAsyncFlushes() const {
+ if (committed_flushes.empty()) {
return false;
}
- auto& flush_list = commited_flushes.front();
- if (!flush_list) {
- return false;
- }
- return true;
+ return committed_flushes.front() != nullptr;
}
- bool HasUncommitedFlushes() {
- if (uncommited_flushes) {
- return true;
- }
- return false;
+ bool HasUncommittedFlushes() const {
+ return uncommitted_flushes != nullptr;
}
void PopAsyncFlushes() {
- if (commited_flushes.empty()) {
+ if (committed_flushes.empty()) {
return;
}
- auto& flush_list = commited_flushes.front();
+ auto& flush_list = committed_flushes.front();
if (!flush_list) {
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
return;
}
for (MapInterval& map : *flush_list) {
@@ -252,7 +242,7 @@ public:
FlushMap(map);
}
}
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
}
virtual BufferType GetEmptyBuffer(std::size_t size) = 0;
@@ -568,10 +558,10 @@ private:
}
void MarkForAsyncFlush(MapInterval& map) {
- if (!uncommited_flushes) {
- uncommited_flushes = std::make_shared<std::unordered_set<MapInterval>>();
+ if (!uncommitted_flushes) {
+ uncommitted_flushes = std::make_shared<std::unordered_set<MapInterval>>();
}
- uncommited_flushes->insert(map);
+ uncommitted_flushes->insert(map);
}
VideoCore::RasterizerInterface& rasterizer;
@@ -605,8 +595,8 @@ private:
std::vector<u8> staging_buffer;
std::list<MapInterval> marked_for_unregister;
- std::shared_ptr<std::unordered_set<MapInterval>> uncommited_flushes{};
- std::list<std::shared_ptr<std::list<MapInterval>>> commited_flushes;
+ std::shared_ptr<std::unordered_set<MapInterval>> uncommitted_flushes{};
+ std::list<std::shared_ptr<std::list<MapInterval>>> committed_flushes;
std::recursive_mutex mutex;
};
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 99a138b5b7..9fe9c1bf25 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -28,15 +28,15 @@ public:
FenceBase(GPUVAddr address, u32 payload, bool is_stubbed)
: address{address}, payload{payload}, is_semaphore{true}, is_stubbed{is_stubbed} {}
- constexpr GPUVAddr GetAddress() const {
+ GPUVAddr GetAddress() const {
return address;
}
- constexpr u32 GetPayload() const {
+ u32 GetPayload() const {
return payload;
}
- constexpr bool IsSemaphore() const {
+ bool IsSemaphore() const {
return is_semaphore;
}
@@ -54,12 +54,8 @@ class FenceManager {
public:
void SignalSemaphore(GPUVAddr addr, u32 value) {
TryReleasePendingFences();
- bool should_flush = texture_cache.HasUncommitedFlushes();
- should_flush |= buffer_cache.HasUncommitedFlushes();
- should_flush |= query_cache.HasUncommitedFlushes();
- texture_cache.CommitAsyncFlushes();
- buffer_cache.CommitAsyncFlushes();
- query_cache.CommitAsyncFlushes();
+ bool should_flush = ShouldFlush();
+ CommitAsyncFlushes();
TFence new_fence = CreateFence(addr, value, !should_flush);
fences.push(new_fence);
QueueFence(new_fence);
@@ -71,12 +67,8 @@ public:
void SignalSyncPoint(u32 value) {
TryReleasePendingFences();
- bool should_flush = texture_cache.HasUncommitedFlushes();
- should_flush |= buffer_cache.HasUncommitedFlushes();
- should_flush |= query_cache.HasUncommitedFlushes();
- texture_cache.CommitAsyncFlushes();
- buffer_cache.CommitAsyncFlushes();
- query_cache.CommitAsyncFlushes();
+ bool should_flush = ShouldFlush();
+ CommitAsyncFlushes();
TFence new_fence = CreateFence(value, !should_flush);
fences.push(new_fence);
QueueFence(new_fence);
@@ -89,15 +81,10 @@ public:
void WaitPendingFences() {
while (!fences.empty()) {
TFence& current_fence = fences.front();
- bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
- should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
- should_wait |= query_cache.ShouldWaitAsyncFlushes();
- if (should_wait) {
+ if (ShouldWait()) {
WaitFence(current_fence);
}
- texture_cache.PopAsyncFlushes();
- buffer_cache.PopAsyncFlushes();
- query_cache.PopAsyncFlushes();
+ PopAsyncFlushes();
auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
auto& memory_manager{gpu.MemoryManager()};
@@ -116,10 +103,18 @@ protected:
: system{system}, rasterizer{rasterizer}, texture_cache{texture_cache},
buffer_cache{buffer_cache}, query_cache{query_cache} {}
+ virtual ~FenceManager() {}
+
+ /// Creates a Sync Point Fence Interface, does not create a backend fence if 'is_stubbed' is
+ /// true
virtual TFence CreateFence(u32 value, bool is_stubbed) = 0;
+ /// Creates a Semaphore Fence Interface, does not create a backend fence if 'is_stubbed' is true
virtual TFence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) = 0;
+ /// Queues a fence into the backend if the fence isn't stubbed.
virtual void QueueFence(TFence& fence) = 0;
- virtual bool IsFenceSignaled(TFence& fence) = 0;
+ /// Notifies that the backend fence has been signaled/reached in host GPU.
+ virtual bool IsFenceSignaled(TFence& fence) const = 0;
+ /// Waits until a fence has been signalled by the host GPU.
virtual void WaitFence(TFence& fence) = 0;
Core::System& system;
@@ -132,15 +127,10 @@ private:
void TryReleasePendingFences() {
while (!fences.empty()) {
TFence& current_fence = fences.front();
- bool should_wait = texture_cache.ShouldWaitAsyncFlushes();
- should_wait |= buffer_cache.ShouldWaitAsyncFlushes();
- should_wait |= query_cache.ShouldWaitAsyncFlushes();
- if (should_wait && !IsFenceSignaled(current_fence)) {
+ if (ShouldWait() && !IsFenceSignaled(current_fence)) {
return;
}
- texture_cache.PopAsyncFlushes();
- buffer_cache.PopAsyncFlushes();
- query_cache.PopAsyncFlushes();
+ PopAsyncFlushes();
auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
auto& memory_manager{gpu.MemoryManager()};
@@ -152,6 +142,28 @@ private:
}
}
+ bool ShouldWait() const {
+ return texture_cache.ShouldWaitAsyncFlushes() || buffer_cache.ShouldWaitAsyncFlushes() ||
+ query_cache.ShouldWaitAsyncFlushes();
+ }
+
+ bool ShouldFlush() const {
+ return texture_cache.HasUncommittedFlushes() || buffer_cache.HasUncommittedFlushes() ||
+ query_cache.HasUncommittedFlushes();
+ }
+
+ void PopAsyncFlushes() {
+ texture_cache.PopAsyncFlushes();
+ buffer_cache.PopAsyncFlushes();
+ query_cache.PopAsyncFlushes();
+ }
+
+ void CommitAsyncFlushes() {
+ texture_cache.CommitAsyncFlushes();
+ buffer_cache.CommitAsyncFlushes();
+ query_cache.CommitAsyncFlushes();
+ }
+
std::queue<TFence> fences;
};
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 85a6c7bb5a..3b7572d61a 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -125,7 +125,7 @@ bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
return true;
}
-u64 GPU::RequestFlush(CacheAddr addr, std::size_t size) {
+u64 GPU::RequestFlush(VAddr addr, std::size_t size) {
std::unique_lock lck{flush_request_mutex};
const u64 fence = ++last_flush_fence;
flush_requests.emplace_back(fence, addr, size);
@@ -137,7 +137,7 @@ void GPU::TickWork() {
while (!flush_requests.empty()) {
auto& request = flush_requests.front();
const u64 fence = request.fence;
- const CacheAddr addr = request.addr;
+ const VAddr addr = request.addr;
const std::size_t size = request.size;
flush_requests.pop_front();
flush_request_mutex.unlock();
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 943a5b1100..5e3eb94e95 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -155,16 +155,22 @@ public:
/// Calls a GPU method.
void CallMethod(const MethodCall& method_call);
+ /// Flush all current written commands into the host GPU for execution.
void FlushCommands();
+ /// Synchronizes CPU writes with Host GPU memory.
void SyncGuestHost();
+ /// Signal the ending of command list.
virtual void OnCommandListEnd();
- u64 RequestFlush(CacheAddr addr, std::size_t size);
+ /// Request a host GPU memory flush from the CPU.
+ u64 RequestFlush(VAddr addr, std::size_t size);
+ /// Obtains current flush request fence id.
u64 CurrentFlushRequestFence() const {
return current_flush_fence.load(std::memory_order_relaxed);
}
+ /// Tick pending requests within the GPU.
void TickWork();
/// Returns a reference to the Maxwell3D GPU engine.
@@ -336,10 +342,10 @@ private:
std::condition_variable sync_cv;
struct FlushRequest {
- FlushRequest(u64 fence, CacheAddr addr, std::size_t size)
+ FlushRequest(u64 fence, VAddr addr, std::size_t size)
: fence{fence}, addr{addr}, size{size} {}
u64 fence;
- CacheAddr addr;
+ VAddr addr;
std::size_t size;
};
diff --git a/src/video_core/query_cache.h b/src/video_core/query_cache.h
index 98d956b68c..2f75f88015 100644
--- a/src/video_core/query_cache.h
+++ b/src/video_core/query_cache.h
@@ -176,41 +176,34 @@ public:
}
void CommitAsyncFlushes() {
- commited_flushes.push_back(uncommited_flushes);
- uncommited_flushes.reset();
+ committed_flushes.push_back(uncommitted_flushes);
+ uncommitted_flushes.reset();
}
- bool HasUncommitedFlushes() {
- if (uncommited_flushes) {
- return true;
- }
- return false;
+ bool HasUncommittedFlushes() const {
+ return uncommitted_flushes != nullptr;
}
- bool ShouldWaitAsyncFlushes() {
- if (commited_flushes.empty()) {
+ bool ShouldWaitAsyncFlushes() const {
+ if (committed_flushes.empty()) {
return false;
}
- auto& flush_list = commited_flushes.front();
- if (!flush_list) {
- return false;
- }
- return true;
+ return committed_flushes.front() != nullptr;
}
void PopAsyncFlushes() {
- if (commited_flushes.empty()) {
+ if (committed_flushes.empty()) {
return;
}
- auto& flush_list = commited_flushes.front();
+ auto& flush_list = committed_flushes.front();
if (!flush_list) {
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
return;
}
for (VAddr query_address : *flush_list) {
FlushAndRemoveRegion(query_address, 4);
}
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
}
protected:
@@ -268,10 +261,10 @@ private:
}
void AsyncFlushQuery(VAddr addr) {
- if (!uncommited_flushes) {
- uncommited_flushes = std::make_shared<std::unordered_set<VAddr>>();
+ if (!uncommitted_flushes) {
+ uncommitted_flushes = std::make_shared<std::unordered_set<VAddr>>();
}
- uncommited_flushes->insert(addr);
+ uncommitted_flushes->insert(addr);
}
static constexpr std::uintptr_t PAGE_SIZE = 4096;
@@ -286,8 +279,8 @@ private:
std::array<CounterStream, VideoCore::NumQueryTypes> streams;
- std::shared_ptr<std::unordered_set<VAddr>> uncommited_flushes{};
- std::list<std::shared_ptr<std::unordered_set<VAddr>>> commited_flushes;
+ std::shared_ptr<std::unordered_set<VAddr>> uncommitted_flushes{};
+ std::list<std::shared_ptr<std::unordered_set<VAddr>>> committed_flushes;
};
template <class QueryCache, class HostCounter>
diff --git a/src/video_core/rasterizer_interface.h b/src/video_core/rasterizer_interface.h
index 4e9c8fb595..603f619523 100644
--- a/src/video_core/rasterizer_interface.h
+++ b/src/video_core/rasterizer_interface.h
@@ -64,6 +64,7 @@ public:
/// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
virtual void FlushRegion(VAddr addr, u64 size) = 0;
+ /// Check if the the specified memory area requires flushing to CPU Memory.
virtual bool MustFlushRegion(VAddr addr, u64 size) = 0;
/// Notify rasterizer that any caches of the specified region should be invalidated
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
index aa57a0ae01..476c899408 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -62,7 +62,7 @@ void FenceManagerOpenGL::QueueFence(Fence& fence) {
fence->Queue();
}
-bool FenceManagerOpenGL::IsFenceSignaled(Fence& fence) {
+bool FenceManagerOpenGL::IsFenceSignaled(Fence& fence) const {
return fence->IsSignaled();
}
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.h b/src/video_core/renderer_opengl/gl_fence_manager.h
index c76e69cb8d..c917b3343f 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.h
+++ b/src/video_core/renderer_opengl/gl_fence_manager.h
@@ -46,7 +46,7 @@ protected:
Fence CreateFence(u32 value, bool is_stubbed) override;
Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
void QueueFence(Fence& fence) override;
- bool IsFenceSignaled(Fence& fence) override;
+ bool IsFenceSignaled(Fence& fence) const override;
void WaitFence(Fence& fence) override;
};
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 847d67159b..d662657cf5 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -653,9 +653,6 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
}
bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
- if (!Settings::IsGPULevelExtreme()) {
- return buffer_cache.MustFlushRegion(addr, size);
- }
return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
}
@@ -672,7 +669,7 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
MICROPROFILE_SCOPE(OpenGL_CacheManagement);
- if (!addr || !size) {
+ if (addr == 0 || size == 0) {
return;
}
texture_cache.OnCPUWrite(addr, size);
diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.cpp b/src/video_core/renderer_vulkan/vk_fence_manager.cpp
index a2b2bc408f..a02be5487a 100644
--- a/src/video_core/renderer_vulkan/vk_fence_manager.cpp
+++ b/src/video_core/renderer_vulkan/vk_fence_manager.cpp
@@ -90,7 +90,7 @@ void VKFenceManager::QueueFence(Fence& fence) {
fence->Queue();
}
-bool VKFenceManager::IsFenceSignaled(Fence& fence) {
+bool VKFenceManager::IsFenceSignaled(Fence& fence) const {
return fence->IsSignaled();
}
diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.h b/src/video_core/renderer_vulkan/vk_fence_manager.h
index 30651e9c72..04d07fe6af 100644
--- a/src/video_core/renderer_vulkan/vk_fence_manager.h
+++ b/src/video_core/renderer_vulkan/vk_fence_manager.h
@@ -63,7 +63,7 @@ protected:
Fence CreateFence(u32 value, bool is_stubbed) override;
Fence CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) override;
void QueueFence(Fence& fence) override;
- bool IsFenceSignaled(Fence& fence) override;
+ bool IsFenceSignaled(Fence& fence) const override;
void WaitFence(Fence& fence) override;
private:
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 4dc7555aa0..2350cd5f4c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -533,7 +533,7 @@ void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
}
void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
- if (!addr || !size) {
+ if (addr == 0 || size == 0) {
return;
}
texture_cache.OnCPUWrite(addr, size);
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index f3ca1ffd15..1148c3a340 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -120,15 +120,8 @@ public:
std::lock_guard lock{mutex};
auto surfaces = GetSurfacesInRegion(addr, size);
- if (surfaces.empty()) {
- return false;
- }
- for (const auto& surface : surfaces) {
- if (surface->IsModified()) {
- return true;
- }
- }
- return false;
+ return std::any_of(surfaces.begin(), surfaces.end(),
+ [](const TSurface& surface) { return surface->IsModified(); });
}
TView GetTextureSurface(const Tegra::Texture::TICEntry& tic,
@@ -333,41 +326,34 @@ public:
}
void CommitAsyncFlushes() {
- commited_flushes.push_back(uncommited_flushes);
- uncommited_flushes.reset();
+ committed_flushes.push_back(uncommitted_flushes);
+ uncommitted_flushes.reset();
}
- bool HasUncommitedFlushes() {
- if (uncommited_flushes) {
- return true;
- }
- return false;
+ bool HasUncommittedFlushes() const {
+ return uncommitted_flushes != nullptr;
}
- bool ShouldWaitAsyncFlushes() {
- if (commited_flushes.empty()) {
+ bool ShouldWaitAsyncFlushes() const {
+ if (committed_flushes.empty()) {
return false;
}
- auto& flush_list = commited_flushes.front();
- if (!flush_list) {
- return false;
- }
- return true;
+ return committed_flushes.front() != nullptr;
}
void PopAsyncFlushes() {
- if (commited_flushes.empty()) {
+ if (committed_flushes.empty()) {
return;
}
- auto& flush_list = commited_flushes.front();
+ auto& flush_list = committed_flushes.front();
if (!flush_list) {
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
return;
}
for (TSurface& surface : *flush_list) {
FlushSurface(surface);
}
- commited_flushes.pop_front();
+ committed_flushes.pop_front();
}
protected:
@@ -1206,10 +1192,10 @@ private:
};
void AsyncFlushSurface(TSurface& surface) {
- if (!uncommited_flushes) {
- uncommited_flushes = std::make_shared<std::list<TSurface>>();
+ if (!uncommitted_flushes) {
+ uncommitted_flushes = std::make_shared<std::list<TSurface>>();
}
- uncommited_flushes->push_back(surface);
+ uncommitted_flushes->push_back(surface);
}
VideoCore::RasterizerInterface& rasterizer;
@@ -1258,8 +1244,8 @@ private:
std::list<TSurface> marked_for_unregister;
- std::shared_ptr<std::list<TSurface>> uncommited_flushes{};
- std::list<std::shared_ptr<std::list<TSurface>>> commited_flushes;
+ std::shared_ptr<std::list<TSurface>> uncommitted_flushes{};
+ std::list<std::shared_ptr<std::list<TSurface>>> committed_flushes;
StagingCache staging_cache;
std::recursive_mutex mutex;
From 644588fd883fb45bf6d1cb7895e98ec65120c7f1 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Thu, 16 Apr 2020 13:50:12 -0400
Subject: [PATCH 27/31] ShaderCache/PipelineCache: Cache null shaders.
---
.../renderer_opengl/gl_shader_cache.cpp | 17 +++++++++++++----
.../renderer_opengl/gl_shader_cache.h | 3 +++
.../renderer_vulkan/vk_pipeline_cache.cpp | 16 ++++++++++++----
.../renderer_vulkan/vk_pipeline_cache.h | 3 +++
4 files changed, 31 insertions(+), 8 deletions(-)
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp
index 6d2ff20f96..f63156b8d9 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp
@@ -448,7 +448,7 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
// Look up shader in the cache based on address
const auto cpu_addr{memory_manager.GpuToCpuAddress(address)};
- Shader shader{cpu_addr ? TryGet(*cpu_addr) : nullptr};
+ Shader shader{cpu_addr ? TryGet(*cpu_addr) : null_shader};
if (shader) {
return last_shaders[static_cast<std::size_t>(program)] = shader;
}
@@ -477,7 +477,12 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
const std::size_t size_in_bytes = code.size() * sizeof(u64);
shader = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
}
- Register(shader);
+
+ if (cpu_addr) {
+ Register(shader);
+ } else {
+ null_shader = shader;
+ }
return last_shaders[static_cast<std::size_t>(program)] = shader;
}
@@ -486,7 +491,7 @@ Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
auto& memory_manager{system.GPU().MemoryManager()};
const auto cpu_addr{memory_manager.GpuToCpuAddress(code_addr)};
- auto kernel = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+ auto kernel = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
if (kernel) {
return kernel;
}
@@ -507,7 +512,11 @@ Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
kernel = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
}
- Register(kernel);
+ if (cpu_addr) {
+ Register(kernel);
+ } else {
+ null_kernel = kernel;
+ }
return kernel;
}
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.h b/src/video_core/renderer_opengl/gl_shader_cache.h
index c836df5bd1..91690b4705 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.h
+++ b/src/video_core/renderer_opengl/gl_shader_cache.h
@@ -125,6 +125,9 @@ private:
ShaderDiskCacheOpenGL disk_cache;
std::unordered_map<u64, PrecompiledShader> runtime_cache;
+ Shader null_shader{};
+ Shader null_kernel{};
+
std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
};
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index 8fdc6400dc..c4b3bc6c87 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -207,7 +207,7 @@ std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
const GPUVAddr program_addr{GetShaderAddress(system, program)};
const std::optional cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
ASSERT(cpu_addr);
- auto shader = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+ auto shader = cpu_addr ? TryGet(*cpu_addr) : null_shader;
if (!shader) {
const auto host_ptr{memory_manager.GetPointer(program_addr)};
@@ -218,7 +218,11 @@ std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
shader = std::make_shared<CachedShader>(system, stage, program_addr, *cpu_addr,
std::move(code), stage_offset);
- Register(shader);
+ if (cpu_addr) {
+ Register(shader);
+ } else {
+ null_shader = shader;
+ }
}
shaders[index] = std::move(shader);
}
@@ -261,7 +265,7 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
const auto cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
ASSERT(cpu_addr);
- auto shader = cpu_addr ? TryGet(*cpu_addr) : nullptr;
+ auto shader = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
if (!shader) {
// No shader found - create a new one
const auto host_ptr = memory_manager.GetPointer(program_addr);
@@ -271,7 +275,11 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
shader = std::make_shared<CachedShader>(system, Tegra::Engines::ShaderType::Compute,
program_addr, *cpu_addr, std::move(code),
kernel_main_offset);
- Register(shader);
+ if (cpu_addr) {
+ Register(shader);
+ } else {
+ null_kernel = shader;
+ }
}
Specialization specialization;
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h
index 7ccdb7083d..602a0a3403 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h
@@ -182,6 +182,9 @@ private:
VKUpdateDescriptorQueue& update_descriptor_queue;
VKRenderPassCache& renderpass_cache;
+ Shader null_shader{};
+ Shader null_kernel{};
+
std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
GraphicsPipelineCacheKey last_graphics_key;
From 1b3be8a8f86f0315004a4c60ef1828ba09f48b32 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 19 Apr 2020 13:27:56 -0400
Subject: [PATCH 28/31] MaxwellDMA: Correct copying on accuracy level.
---
src/video_core/engines/maxwell_dma.cpp | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 32b04e31ea..3bfed6ab81 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -104,8 +104,13 @@ void MaxwellDMA::HandleCopy() {
write_buffer.resize(dst_size);
}
- memory_manager.ReadBlock(source, read_buffer.data(), src_size);
- memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+ if (Settings::IsGPULevelExtreme()) {
+ memory_manager.ReadBlock(source, read_buffer.data(), src_size);
+ memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+ } else {
+ memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
+ memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
+ }
Texture::UnswizzleSubrect(
regs.x_count, regs.y_count, regs.dst_pitch, regs.src_params.size_x, bytes_per_pixel,
From 39e5b7294898c45cf247b61e46ef735bd16e96ae Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 19 Apr 2020 13:47:45 -0400
Subject: [PATCH 29/31] Async GPU: Correct flushing behavior to be similar to
old async GPU behavior.
---
src/video_core/gpu_thread.cpp | 4 ++++
src/video_core/renderer_opengl/gl_rasterizer.cpp | 3 +++
src/video_core/renderer_vulkan/vk_rasterizer.cpp | 4 ++++
3 files changed, 11 insertions(+)
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 7df854a2f9..c3bb4fe062 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -84,6 +84,10 @@ void ThreadManager::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
}
void ThreadManager::FlushRegion(VAddr addr, u64 size) {
+ if (!Settings::IsGPULevelHigh()) {
+ PushCommand(FlushRegionCommand(addr, size));
+ return;
+ }
if (!Settings::IsGPULevelExtreme()) {
return;
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index d662657cf5..4c16c89d29 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -653,6 +653,9 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
}
bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+ if (!Settings::IsGPULevelHigh()) {
+ return buffer_cache.MustFlushRegion(addr, size);
+ }
return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
}
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 2350cd5f4c..0dc9979165 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -17,6 +17,7 @@
#include "common/microprofile.h"
#include "core/core.h"
#include "core/memory.h"
+#include "core/settings.h"
#include "video_core/engines/kepler_compute.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/renderer_vulkan/fixed_pipeline_state.h"
@@ -519,6 +520,9 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
}
bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size) {
+ if (!Settings::IsGPULevelHigh()) {
+ return buffer_cache.MustFlushRegion(addr, size);
+ }
return texture_cache.MustFlushRegion(addr, size) || buffer_cache.MustFlushRegion(addr, size);
}
From 4e37f1b1130b083b42f21029155e5a2e4e9a9eb3 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 22 Apr 2020 11:14:40 -0400
Subject: [PATCH 30/31] Address Feedback.
---
src/video_core/buffer_cache/buffer_cache.h | 15 ++++++---------
src/video_core/fence_manager.h | 12 ++++++------
src/video_core/texture_cache/texture_cache.h | 15 ++++++---------
3 files changed, 18 insertions(+), 24 deletions(-)
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index f3aa352959..510f11089f 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -153,8 +153,8 @@ public:
bool MustFlushRegion(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
- std::vector<MapInterval> objects = GetMapsInRange(addr, size);
- return std::any_of(objects.begin(), objects.end(), [](const MapInterval& map) {
+ const std::vector<MapInterval> objects = GetMapsInRange(addr, size);
+ return std::any_of(objects.cbegin(), objects.cend(), [](const MapInterval& map) {
return map->IsModified() && map->IsRegistered();
});
}
@@ -176,7 +176,7 @@ public:
for (const auto& object : GetMapsInRange(addr, size)) {
if (object->IsMemoryMarked() && object->IsRegistered()) {
- Unmark(object);
+ UnmarkMemory(object);
object->SetSyncPending(true);
marked_for_unregister.emplace_back(object);
}
@@ -217,10 +217,7 @@ public:
}
bool ShouldWaitAsyncFlushes() const {
- if (committed_flushes.empty()) {
- return false;
- }
- return committed_flushes.front() != nullptr;
+ return !committed_flushes.empty() && committed_flushes.front() != nullptr;
}
bool HasUncommittedFlushes() const {
@@ -294,7 +291,7 @@ protected:
}
}
- void Unmark(const MapInterval& map) {
+ void UnmarkMemory(const MapInterval& map) {
if (!map->IsMemoryMarked()) {
return;
}
@@ -305,7 +302,7 @@ protected:
/// Unregisters an object from the cache
void Unregister(const MapInterval& map) {
- Unmark(map);
+ UnmarkMemory(map);
map->MarkAsRegistered(false);
if (map->IsSyncPending()) {
marked_for_unregister.remove(map);
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index 9fe9c1bf25..dabd1588ce 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -54,7 +54,7 @@ class FenceManager {
public:
void SignalSemaphore(GPUVAddr addr, u32 value) {
TryReleasePendingFences();
- bool should_flush = ShouldFlush();
+ const bool should_flush = ShouldFlush();
CommitAsyncFlushes();
TFence new_fence = CreateFence(addr, value, !should_flush);
fences.push(new_fence);
@@ -67,7 +67,7 @@ public:
void SignalSyncPoint(u32 value) {
TryReleasePendingFences();
- bool should_flush = ShouldFlush();
+ const bool should_flush = ShouldFlush();
CommitAsyncFlushes();
TFence new_fence = CreateFence(value, !should_flush);
fences.push(new_fence);
@@ -79,15 +79,15 @@ public:
}
void WaitPendingFences() {
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
while (!fences.empty()) {
TFence& current_fence = fences.front();
if (ShouldWait()) {
WaitFence(current_fence);
}
PopAsyncFlushes();
- auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
- auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
} else {
gpu.IncrementSyncPoint(current_fence->GetPayload());
@@ -125,15 +125,15 @@ protected:
private:
void TryReleasePendingFences() {
+ auto& gpu{system.GPU()};
+ auto& memory_manager{gpu.MemoryManager()};
while (!fences.empty()) {
TFence& current_fence = fences.front();
if (ShouldWait() && !IsFenceSignaled(current_fence)) {
return;
}
PopAsyncFlushes();
- auto& gpu{system.GPU()};
if (current_fence->IsSemaphore()) {
- auto& memory_manager{gpu.MemoryManager()};
memory_manager.Write<u32>(current_fence->GetAddress(), current_fence->GetPayload());
} else {
gpu.IncrementSyncPoint(current_fence->GetPayload());
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 1148c3a340..cf6bd005aa 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -68,7 +68,7 @@ public:
for (const auto& surface : GetSurfacesInRegion(addr, size)) {
if (surface->IsMemoryMarked()) {
- Unmark(surface);
+ UnmarkMemory(surface);
surface->SetSyncPending(true);
marked_for_unregister.emplace_back(surface);
}
@@ -119,8 +119,8 @@ public:
bool MustFlushRegion(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
- auto surfaces = GetSurfacesInRegion(addr, size);
- return std::any_of(surfaces.begin(), surfaces.end(),
+ const auto surfaces = GetSurfacesInRegion(addr, size);
+ return std::any_of(surfaces.cbegin(), surfaces.cend(),
[](const TSurface& surface) { return surface->IsModified(); });
}
@@ -335,10 +335,7 @@ public:
}
bool ShouldWaitAsyncFlushes() const {
- if (committed_flushes.empty()) {
- return false;
- }
- return committed_flushes.front() != nullptr;
+ return !committed_flushes.empty() && committed_flushes.front() != nullptr;
}
void PopAsyncFlushes() {
@@ -421,7 +418,7 @@ protected:
rasterizer.UpdatePagesCachedCount(*cpu_addr, size, 1);
}
- void Unmark(TSurface surface) {
+ void UnmarkMemory(TSurface surface) {
if (!surface->IsMemoryMarked()) {
return;
}
@@ -438,7 +435,7 @@ protected:
if (!guard_render_targets && surface->IsRenderTarget()) {
ManageRenderTargetUnregister(surface);
}
- Unmark(surface);
+ UnmarkMemory(surface);
if (surface->IsSyncPending()) {
marked_for_unregister.remove(surface);
surface->SetSyncPending(false);
From c043ac4f139be7cbf5b7edbf050aa84e1896f75d Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 22 Apr 2020 20:34:32 -0400
Subject: [PATCH 31/31] GL_Fence_Manager: use GL_TIMEOUT_IGNORED instead of a
loop,
---
src/video_core/renderer_opengl/gl_fence_manager.cpp | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/src/video_core/renderer_opengl/gl_fence_manager.cpp b/src/video_core/renderer_opengl/gl_fence_manager.cpp
index 476c899408..99ddcb3f85 100644
--- a/src/video_core/renderer_opengl/gl_fence_manager.cpp
+++ b/src/video_core/renderer_opengl/gl_fence_manager.cpp
@@ -40,8 +40,7 @@ void GLInnerFence::Wait() {
return;
}
ASSERT(sync_object.handle != 0);
- while (glClientWaitSync(sync_object.handle, 0, 1000) == GL_TIMEOUT_EXPIRED)
- ;
+ glClientWaitSync(sync_object.handle, 0, GL_TIMEOUT_IGNORED);
}
FenceManagerOpenGL::FenceManagerOpenGL(Core::System& system,