Merge pull request #132 from shadps4-emu/gpu_flip

graphics: separate IRQ for GPU driven flips

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -49,10 +49,11 @@ s32 PS4_SYSV_ABI sceGnmAddEqEvent(SceKernelEqueue eq, u64 id, void* udata) {
     kernel_event.event.udata = udata;
     eq->addEvent(kernel_event);
 
-    Platform::IrqC::Instance()->Register([=](Platform::InterruptId irq) {
+    Platform::IrqC::Instance()->Register(
+        Platform::InterruptId::GfxEop, [=](Platform::InterruptId irq) {
             ASSERT_MSG(irq == Platform::InterruptId::GfxEop,
-                   "An unexpected IRQ occured"); // We need to conver IRQ# to event id and do proper
+                       "An unexpected IRQ occured"); // We need to conver IRQ# to event id and do
-                                                 // filtering in trigger function
+                                                     // proper filtering in trigger function
             eq->triggerEvent(SceKernelEvent::Type::GfxEop, EVFILT_GRAPHICS_CORE, nullptr);
         });
     return ORBIS_OK;
@@ -164,7 +165,7 @@ s32 PS4_SYSV_ABI sceGnmDeleteEqEvent(SceKernelEqueue eq, u64 id) {
 
     eq->removeEvent(id);
 
-    Platform::IrqC::Instance()->Unregister();
+    Platform::IrqC::Instance()->Unregister(Platform::InterruptId::GfxEop);
     return ORBIS_OK;
 }
 

src/core/libraries/videoout/video_out.cpp

@@ -223,8 +223,9 @@ s32 sceVideoOutSubmitEopFlip(s32 handle, u32 buf_id, u32 mode, u32 arg, void** u
         return 0x8029000b;
     }
 
-    Platform::IrqC::Instance()->RegisterOnce([=](Platform::InterruptId irq) {
+    Platform::IrqC::Instance()->RegisterOnce(
-        ASSERT_MSG(irq == Platform::InterruptId::GfxEop, "An unexpected IRQ occured");
+        Platform::InterruptId::GfxFlip, [=](Platform::InterruptId irq) {
+            ASSERT_MSG(irq == Platform::InterruptId::GfxFlip, "An unexpected IRQ occured");
             const auto result = driver->SubmitFlip(port, buf_id, arg, true);
             ASSERT_MSG(result, "EOP flip submission failed");
         });

src/core/platform.h

@@ -24,52 +24,64 @@ enum class InterruptId : u32 {
     Compute4RelMem = 4u,
     Compute5RelMem = 5u,
     Compute6RelMem = 6u,
-    GfxEop = 0x40u
+    GfxEop = 7u,
+    GfxFlip = 8u
 };
 
 using IrqHandler = std::function<void(InterruptId)>;
 
 struct IrqController {
-    void RegisterOnce(IrqHandler handler) {
+    void RegisterOnce(InterruptId irq, IrqHandler handler) {
-        std::unique_lock lock{m_lock};
+        ASSERT_MSG(static_cast<u32>(irq) < irq_contexts.size(), "Invalid IRQ number");
-        one_time_subscribers.emplace(handler);
+        auto& ctx = irq_contexts[static_cast<u32>(irq)];
+        std::unique_lock lock{ctx.m_lock};
+        ctx.one_time_subscribers.emplace(handler);
     }
 
-    void Register(IrqHandler handler) {
+    void Register(InterruptId irq, IrqHandler handler) {
-        ASSERT_MSG(!persistent_handler.has_value(),
+        ASSERT_MSG(static_cast<u32>(irq) < irq_contexts.size(), "Invalid IRQ number");
+        auto& ctx = irq_contexts[static_cast<u32>(irq)];
+        ASSERT_MSG(!ctx.persistent_handler.has_value(),
                    "Too many persistent handlers"); // Add a slot map if so
 
-        std::unique_lock lock{m_lock};
+        std::unique_lock lock{ctx.m_lock};
-        persistent_handler.emplace(handler);
+        ctx.persistent_handler.emplace(handler);
     }
 
-    void Unregister() {
+    void Unregister(InterruptId irq) {
-        std::unique_lock lock{m_lock};
+        ASSERT_MSG(static_cast<u32>(irq) < irq_contexts.size(), "Invalid IRQ number");
-        persistent_handler.reset();
+        auto& ctx = irq_contexts[static_cast<u32>(irq)];
+        std::unique_lock lock{ctx.m_lock};
+        ctx.persistent_handler.reset();
     }
 
     void Signal(InterruptId irq) {
-        std::unique_lock lock{m_lock};
+        ASSERT_MSG(static_cast<u32>(irq) < irq_contexts.size(), "Unexpected IRQ signaled");
+        auto& ctx = irq_contexts[static_cast<u32>(irq)];
+        std::unique_lock lock{ctx.m_lock};
 
         LOG_TRACE(Core, "IRQ signaled: {}", magic_enum::enum_name(irq));
 
-        if (persistent_handler) {
+        if (ctx.persistent_handler) {
-            persistent_handler.value()(irq);
+            ctx.persistent_handler.value()(irq);
         }
 
-        while (!one_time_subscribers.empty()) {
+        while (!ctx.one_time_subscribers.empty()) {
-            const auto& h = one_time_subscribers.front();
+            const auto& h = ctx.one_time_subscribers.front();
             h(irq);
 
-            one_time_subscribers.pop();
+            ctx.one_time_subscribers.pop();
         }
     }
 
 private:
+    struct IrqContext {
         std::optional<IrqHandler> persistent_handler{};
         std::queue<IrqHandler> one_time_subscribers{};
         std::mutex m_lock{};
     };
+    std::array<IrqContext, magic_enum::enum_count<InterruptId>()> irq_contexts{};
+};
 
 using IrqC = Common::Singleton<IrqController>;
 

src/video_core/amdgpu/liverpool.cpp

@@ -25,8 +25,24 @@ void Liverpool::ProcessCmdList(u32* cmdbuf, u32 size_in_bytes) {
             const PM4ItOpcode opcode = header->type3.opcode;
             const u32 count = header->type3.NumWords();
             switch (opcode) {
-            case PM4ItOpcode::Nop:
+            case PM4ItOpcode::Nop: {
+                const auto* nop = reinterpret_cast<PM4CmdNop*>(header);
+                if (nop->header.count.Value() == 0) {
                     break;
+                }
+
+                switch (nop->data_block[0]) {
+                case PM4CmdNop::PayloadType::PatchedFlip: {
+                    // There is no evidence that GPU CP drives flip events by parsing
+                    // special NOP packets. For convenience lets assume that it does.
+                    Platform::IrqC::Instance()->Signal(Platform::InterruptId::GfxFlip);
+                    break;
+                }
+                default:
+                    break;
+                }
+                break;
+            }
             case PM4ItOpcode::SetContextReg: {
                 const auto* set_data = reinterpret_cast<PM4CmdSetData*>(header);
                 std::memcpy(&regs.reg_array[ContextRegWordOffset + set_data->reg_offset],