Merge pull request #11225 from FernandoS27/no-laxatives-in-santas-cookies

Y.F.C: Rework the Query Cache.

5 files changed, 26 insertions, 77 deletions

diff --git a/src/video_core/engines/draw_manager.h b/src/video_core/engines/draw_manager.h
index 7c22c49f1..18d959143 100644
--- a/src/video_core/engines/draw_manager.h
+++ b/src/video_core/engines/draw_manager.h
@@ -46,6 +46,7 @@ public:
     };
 
     struct IndirectParams {
+        bool is_byte_count;
         bool is_indexed;
         bool include_count;
         GPUVAddr count_start_address;
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 06e349e43..32d767d85 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -20,8 +20,6 @@
 
 namespace Tegra::Engines {
 
-using VideoCore::QueryType;
-
 /// First register id that is actually a Macro call.
 constexpr u32 MacroRegistersStart = 0xE00;
 
@@ -500,27 +498,21 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
 }
 
 void Maxwell3D::ProcessQueryGet() {
+    VideoCommon::QueryPropertiesFlags flags{};
+    if (regs.report_semaphore.query.short_query == 0) {
+        flags |= VideoCommon::QueryPropertiesFlags::HasTimeout;
+    }
+    const GPUVAddr sequence_address{regs.report_semaphore.Address()};
+    const VideoCommon::QueryType query_type =
+        static_cast<VideoCommon::QueryType>(regs.report_semaphore.query.report.Value());
+    const u32 payload = regs.report_semaphore.payload;
+    const u32 subreport = regs.report_semaphore.query.sub_report;
     switch (regs.report_semaphore.query.operation) {
     case Regs::ReportSemaphore::Operation::Release:
         if (regs.report_semaphore.query.short_query != 0) {
-            const GPUVAddr sequence_address{regs.report_semaphore.Address()};
-            const u32 payload = regs.report_semaphore.payload;
-            std::function<void()> operation([this, sequence_address, payload] {
-                memory_manager.Write<u32>(sequence_address, payload);
-            });
-            rasterizer->SignalFence(std::move(operation));
-        } else {
-            struct LongQueryResult {
-                u64_le value;
-                u64_le timestamp;
-            };
-            const GPUVAddr sequence_address{regs.report_semaphore.Address()};
-            const u32 payload = regs.report_semaphore.payload;
-            [this, sequence_address, payload] {
-                memory_manager.Write<u64>(sequence_address + sizeof(u64), system.GPU().GetTicks());
-                memory_manager.Write<u64>(sequence_address, payload);
-            }();
+            flags |= VideoCommon::QueryPropertiesFlags::IsAFence;
         }
+        rasterizer->Query(sequence_address, query_type, flags, payload, subreport);
         break;
     case Regs::ReportSemaphore::Operation::Acquire:
         // TODO(Blinkhawk): Under this operation, the GPU waits for the CPU to write a value that
@@ -528,11 +520,7 @@ void Maxwell3D::ProcessQueryGet() {
         UNIMPLEMENTED_MSG("Unimplemented query operation ACQUIRE");
         break;
     case Regs::ReportSemaphore::Operation::ReportOnly:
-        if (const std::optional<u64> result = GetQueryResult()) {
-            // If the query returns an empty optional it means it's cached and deferred.
-            // In this case we have a non-empty result, so we stamp it immediately.
-            StampQueryResult(*result, regs.report_semaphore.query.short_query == 0);
-        }
+        rasterizer->Query(sequence_address, query_type, flags, payload, subreport);
         break;
     case Regs::ReportSemaphore::Operation::Trap:
         UNIMPLEMENTED_MSG("Unimplemented query operation TRAP");
@@ -544,6 +532,10 @@ void Maxwell3D::ProcessQueryGet() {
 }
 
 void Maxwell3D::ProcessQueryCondition() {
+    if (rasterizer->AccelerateConditionalRendering()) {
+        execute_on = true;
+        return;
+    }
     const GPUVAddr condition_address{regs.render_enable.Address()};
     switch (regs.render_enable_override) {
     case Regs::RenderEnable::Override::AlwaysRender:
@@ -553,10 +545,6 @@ void Maxwell3D::ProcessQueryCondition() {
         execute_on = false;
         break;
     case Regs::RenderEnable::Override::UseRenderEnable: {
-        if (rasterizer->AccelerateConditionalRendering()) {
-            execute_on = true;
-            return;
-        }
         switch (regs.render_enable.mode) {
         case Regs::RenderEnable::Mode::True: {
             execute_on = true;
@@ -598,15 +586,9 @@ void Maxwell3D::ProcessQueryCondition() {
 }
 
 void Maxwell3D::ProcessCounterReset() {
-#if ANDROID
-    if (!Settings::IsGPULevelHigh()) {
-        // This is problematic on Android, disable on GPU Normal.
-        return;
-    }
-#endif
     switch (regs.clear_report_value) {
     case Regs::ClearReport::ZPassPixelCount:
-        rasterizer->ResetCounter(QueryType::SamplesPassed);
+        rasterizer->ResetCounter(VideoCommon::QueryType::ZPassPixelCount64);
         break;
     default:
         LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}", regs.clear_report_value);
@@ -620,28 +602,6 @@ void Maxwell3D::ProcessSyncPoint() {
     rasterizer->SignalSyncPoint(sync_point);
 }
 
-std::optional<u64> Maxwell3D::GetQueryResult() {
-    switch (regs.report_semaphore.query.report) {
-    case Regs::ReportSemaphore::Report::Payload:
-        return regs.report_semaphore.payload;
-    case Regs::ReportSemaphore::Report::ZPassPixelCount64:
-#if ANDROID
-        if (!Settings::IsGPULevelHigh()) {
-            // This is problematic on Android, disable on GPU Normal.
-            return 120;
-        }
-#endif
-        // Deferred.
-        rasterizer->Query(regs.report_semaphore.Address(), QueryType::SamplesPassed,
-                          system.GPU().GetTicks());
-        return std::nullopt;
-    default:
-        LOG_DEBUG(HW_GPU, "Unimplemented query report type {}",
-                  regs.report_semaphore.query.report.Value());
-        return 1;
-    }
-}
-
 void Maxwell3D::ProcessCBBind(size_t stage_index) {
     // Bind the buffer currently in CB_ADDRESS to the specified index in the desired shader
     // stage.
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 6c19354e1..17faacc37 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -3182,9 +3182,6 @@ private:
     /// Handles writes to syncing register.
     void ProcessSyncPoint();
 
-    /// Returns a query's value or an empty object if the value will be deferred through a cache.
-    std::optional<u64> GetQueryResult();
-
     void RefreshParametersImpl();
 
     bool IsMethodExecutable(u32 method);
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 279f0daa1..422d4d859 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -362,21 +362,17 @@ void MaxwellDMA::ReleaseSemaphore() {
     const auto type = regs.launch_dma.semaphore_type;
     const GPUVAddr address = regs.semaphore.address;
     const u32 payload = regs.semaphore.payload;
+    VideoCommon::QueryPropertiesFlags flags{VideoCommon::QueryPropertiesFlags::IsAFence};
     switch (type) {
     case LaunchDMA::SemaphoreType::NONE:
         break;
     case LaunchDMA::SemaphoreType::RELEASE_ONE_WORD_SEMAPHORE: {
-        std::function<void()> operation(
-            [this, address, payload] { memory_manager.Write<u32>(address, payload); });
-        rasterizer->SignalFence(std::move(operation));
+        rasterizer->Query(address, VideoCommon::QueryType::Payload, flags, payload, 0);
         break;
     }
     case LaunchDMA::SemaphoreType::RELEASE_FOUR_WORD_SEMAPHORE: {
-        std::function<void()> operation([this, address, payload] {
-            memory_manager.Write<u64>(address + sizeof(u64), system.GPU().GetTicks());
-            memory_manager.Write<u64>(address, payload);
-        });
-        rasterizer->SignalFence(std::move(operation));
+        rasterizer->Query(address, VideoCommon::QueryType::Payload,
+                          flags | VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
         break;
     }
     default:
diff --git a/src/video_core/engines/puller.cpp b/src/video_core/engines/puller.cpp
index 6de2543b7..8dd34c04a 100644
--- a/src/video_core/engines/puller.cpp
+++ b/src/video_core/engines/puller.cpp
@@ -82,10 +82,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
     if (op == GpuSemaphoreOperation::WriteLong) {
         const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
         const u32 payload = regs.semaphore_sequence;
-        [this, sequence_address, payload] {
-            memory_manager.Write<u64>(sequence_address + sizeof(u64), gpu.GetTicks());
-            memory_manager.Write<u64>(sequence_address, payload);
-        }();
+        rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
+                          VideoCommon::QueryPropertiesFlags::HasTimeout, payload, 0);
     } else {
         do {
             const u32 word{memory_manager.Read<u32>(regs.semaphore_address.SemaphoreAddress())};
@@ -120,10 +118,8 @@ void Puller::ProcessSemaphoreTriggerMethod() {
 void Puller::ProcessSemaphoreRelease() {
     const GPUVAddr sequence_address{regs.semaphore_address.SemaphoreAddress()};
     const u32 payload = regs.semaphore_release;
-    std::function<void()> operation([this, sequence_address, payload] {
-        memory_manager.Write<u32>(sequence_address, payload);
-    });
-    rasterizer->SignalFence(std::move(operation));
+    rasterizer->Query(sequence_address, VideoCommon::QueryType::Payload,
+                      VideoCommon::QueryPropertiesFlags::IsAFence, payload, 0);
 }
 
 void Puller::ProcessSemaphoreAcquire() {
@@ -132,7 +128,6 @@ void Puller::ProcessSemaphoreAcquire() {
     while (word != value) {
         regs.acquire_active = true;
         regs.acquire_value = value;
-        std::this_thread::sleep_for(std::chrono::milliseconds(1));
         rasterizer->ReleaseFences();
         word = memory_manager.Read<u32>(regs.semaphore_address.SemaphoreAddress());
         // TODO(kemathe73) figure out how to do the acquire_timeout