video_core: CPU flip relay (#415)

* video_core: cpu flip is propagated via gpu thread now

* tentative fix for cpu flips racing

* libraries: videoout: better flip status handling

src/core/libraries/videoout/driver.cpp

@@ -9,6 +9,7 @@
 #include "core/libraries/error_codes.h"
 #include "core/libraries/kernel/time_management.h"
 #include "core/libraries/videoout/driver.h"
+#include "core/platform.h"
 #include "video_core/renderer_vulkan/renderer_vulkan.h"
 
 extern std::unique_ptr<Vulkan::RendererVulkan> renderer;
@@ -173,14 +174,19 @@ std::chrono::microseconds VideoOutDriver::Flip(const Request& req) {
 
     // Update flip status.
     auto* port = req.port;
+    {
+        std::unique_lock lock{port->port_mutex};
         auto& flip_status = port->flip_status;
         flip_status.count++;
         flip_status.processTime = Libraries::Kernel::sceKernelGetProcessTime();
         flip_status.tsc = Libraries::Kernel::sceKernelReadTsc();
-    flip_status.submitTsc = Libraries::Kernel::sceKernelReadTsc();
         flip_status.flipArg = req.flip_arg;
         flip_status.currentBuffer = req.index;
-    flip_status.flipPendingNum = static_cast<int>(requests.size());
+        if (req.eop) {
+            --flip_status.gcQueueNum;
+        }
+        --flip_status.flipPendingNum;
+    }
 
     // Trigger flip events for the port.
     for (auto& event : port->flip_events) {
@@ -202,34 +208,54 @@ std::chrono::microseconds VideoOutDriver::Flip(const Request& req) {
 
 bool VideoOutDriver::SubmitFlip(VideoOutPort* port, s32 index, s64 flip_arg,
                                 bool is_eop /*= false*/) {
+    {
+        std::unique_lock lock{port->port_mutex};
+        if (index != -1 && port->flip_status.flipPendingNum >= port->NumRegisteredBuffers()) {
+            LOG_ERROR(Lib_VideoOut, "Flip queue is full");
+            return false;
+        }
+
+        if (is_eop) {
+            ++port->flip_status.gcQueueNum;
+        }
+        ++port->flip_status.flipPendingNum; // integral GPU and CPU pending flips counter
+        port->flip_status.submitTsc = Libraries::Kernel::sceKernelReadTsc();
+    }
+
+    if (!is_eop) {
+        // Before processing the flip we need to ask GPU thread to flush command list as at this
+        // point VO surface is ready to be presented, and we will need have an actual state of
+        // Vulkan image at the time of frame presentation.
+        liverpool->SendCommand([=, this]() {
+            renderer->FlushDraw();
+            SubmitFlipInternal(port, index, flip_arg, is_eop);
+        });
+    } else {
+        SubmitFlipInternal(port, index, flip_arg, is_eop);
+    }
+
+    return true;
+}
+
+void VideoOutDriver::SubmitFlipInternal(VideoOutPort* port, s32 index, s64 flip_arg,
+                                        bool is_eop /*= false*/) {
     Vulkan::Frame* frame;
     if (index == -1) {
-        frame = renderer->PrepareBlankFrame();
+        frame = renderer->PrepareBlankFrame(is_eop);
     } else {
         const auto& buffer = port->buffer_slots[index];
         const auto& group = port->groups[buffer.group_index];
         frame = renderer->PrepareFrame(group, buffer.address_left, is_eop);
     }
 
-    if (index != -1 && requests.size() >= port->NumRegisteredBuffers()) {
-        LOG_ERROR(Lib_VideoOut, "Flip queue is full");
-        return false;
-    }
-
     std::scoped_lock lock{mutex};
     requests.push({
         .frame = frame,
         .port = port,
         .index = index,
         .flip_arg = flip_arg,
-        .submit_tsc = Libraries::Kernel::sceKernelReadTsc(),
         .eop = is_eop,
     });
-
-    port->flip_status.flipPendingNum = static_cast<int>(requests.size());
-    port->flip_status.gcQueueNum = 0;
-
-    return true;
 }
 
 void VideoOutDriver::PresentThread(std::stop_token token) {

src/core/libraries/videoout/driver.h

@@ -29,6 +29,7 @@ struct VideoOutPort {
     std::vector<Kernel::SceKernelEqueue> flip_events;
     std::vector<Kernel::SceKernelEqueue> vblank_events;
     std::mutex vo_mutex;
+    std::mutex port_mutex;
     std::condition_variable vo_cv;
     std::condition_variable vblank_cv;
     int flip_rate = 0;
@@ -93,7 +94,6 @@ private:
         VideoOutPort* port;
         s32 index;
         s64 flip_arg;
-        u64 submit_tsc;
         bool eop;
 
         operator bool() const noexcept {
@@ -102,6 +102,7 @@ private:
     };
 
     std::chrono::microseconds Flip(const Request& req);
+    void SubmitFlipInternal(VideoOutPort* port, s32 index, s64 flip_arg, bool is_eop = false);
     void PresentThread(std::stop_token token);
 
     std::mutex mutex;

src/core/libraries/videoout/video_out.cpp

@@ -113,7 +113,9 @@ s32 PS4_SYSV_ABI sceVideoOutSetFlipRate(s32 handle, s32 rate) {
 
 s32 PS4_SYSV_ABI sceVideoOutIsFlipPending(s32 handle) {
     LOG_INFO(Lib_VideoOut, "called");
-    s32 pending = driver->GetPort(handle)->flip_status.flipPendingNum;
+    auto* port = driver->GetPort(handle);
+    std::unique_lock lock{port->port_mutex};
+    s32 pending = port->flip_status.flipPendingNum;
     return pending;
 }
 
@@ -161,6 +163,7 @@ s32 PS4_SYSV_ABI sceVideoOutGetFlipStatus(s32 handle, FlipStatus* status) {
         return ORBIS_VIDEO_OUT_ERROR_INVALID_HANDLE;
     }
 
+    std::unique_lock lock{port->port_mutex};
     *status = port->flip_status;
 
     LOG_INFO(Lib_VideoOut,

src/video_core/amdgpu/liverpool.cpp

@@ -35,7 +35,7 @@ void Liverpool::Process(std::stop_token stoken) {
         {
             std::unique_lock lk{submit_mutex};
             Common::CondvarWait(submit_cv, lk, stoken,
-                                [this] { return num_submits != 0 || submit_done; });
+                                [this] { return num_commands || num_submits || submit_done; });
         }
         if (stoken.stop_requested()) {
             break;
@@ -45,7 +45,23 @@ void Liverpool::Process(std::stop_token stoken) {
 
         int qid = -1;
 
-        while (num_submits) {
+        while (num_submits || num_commands) {
+
+            // Process incoming commands with high priority
+            while (num_commands) {
+
+                Common::UniqueFunction<void> callback{};
+                {
+                    std::unique_lock lk{submit_mutex};
+                    callback = std::move(command_queue.back());
+                    command_queue.pop();
+                }
+
+                callback();
+
+                --num_commands;
+            }
+
             qid = (qid + 1) % NumTotalQueues;
 
             auto& queue = mapped_queues[qid];
@@ -219,7 +235,7 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
             // In the case of HW, render target memory has alignment as color block operates on
             // tiles. There is no information of actual resource extents stored in CB context
             // regs, so any deduction of it from slices/pitch will lead to a larger surface created.
-            // The same applies to the depth targets. Fortunatelly, the guest always sends
+            // The same applies to the depth targets. Fortunately, the guest always sends
             // a trailing NOP packet right after the context regs setup, so we can use the heuristic
             // below and extract the hint to determine actual resource dims.
 

src/video_core/amdgpu/liverpool.h

@@ -11,10 +11,12 @@
 #include <span>
 #include <thread>
 #include <queue>
+
 #include "common/assert.h"
 #include "common/bit_field.h"
 #include "common/polyfill_thread.h"
 #include "common/types.h"
+#include "common/unique_function.h"
 #include "video_core/amdgpu/pixel_format.h"
 #include "video_core/amdgpu/resource.h"
 
@@ -1054,6 +1056,13 @@ public:
         rasterizer = rasterizer_;
     }
 
+    void SendCommand(Common::UniqueFunction<void>&& func) {
+        std::scoped_lock lk{submit_mutex};
+        command_queue.emplace(std::move(func));
+        ++num_commands;
+        submit_cv.notify_one();
+    }
+
 private:
     struct Task {
         struct promise_type {
@@ -1122,9 +1131,11 @@ private:
     Libraries::VideoOut::VideoOutPort* vo_port{};
     std::jthread process_thread{};
     std::atomic<u32> num_submits{};
+    std::atomic<u32> num_commands{};
     std::atomic<bool> submit_done{};
     std::mutex submit_mutex;
     std::condition_variable_any submit_cv;
+    std::queue<Common::UniqueFunction<void>> command_queue{};
 };
 
 static_assert(GFX6_3D_REG_INDEX(ps_program) == 0x2C08);

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -48,13 +48,14 @@ public:
                         VAddr cpu_address, bool is_eop) {
         const auto info = VideoCore::ImageInfo{attribute, cpu_address};
         const auto image_id = texture_cache.FindImage(info);
+        texture_cache.UpdateImage(image_id, is_eop ? nullptr : &flip_scheduler);
         auto& image = texture_cache.GetImage(image_id);
         return PrepareFrameInternal(image, is_eop);
     }
 
-    Frame* PrepareBlankFrame() {
+    Frame* PrepareBlankFrame(bool is_eop) {
         auto& image = texture_cache.GetImage(VideoCore::NULL_IMAGE_ID);
-        return PrepareFrameInternal(image, true);
+        return PrepareFrameInternal(image, is_eop);
     }
 
     VideoCore::Image& RegisterVideoOutSurface(
@@ -75,6 +76,11 @@ public:
     void Present(Frame* frame);
     void RecreateFrame(Frame* frame, u32 width, u32 height);
 
+    void FlushDraw() {
+        SubmitInfo info{};
+        draw_scheduler.Flush(info);
+    }
+
 private:
     Frame* PrepareFrameInternal(VideoCore::Image& image, bool is_eop = true);
     Frame* GetRenderFrame();

src/video_core/texture_cache/texture_cache.cpp

@@ -223,7 +223,7 @@ ImageView& TextureCache::FindDepthTarget(const ImageInfo& image_info,
     return RegisterImageView(image_id, view_info);
 }
 
-void TextureCache::RefreshImage(Image& image) {
+void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_scheduler /*= nullptr*/) {
     // Mark image as validated.
     image.flags &= ~ImageFlagBits::CpuModified;
 
@@ -269,8 +269,10 @@ void TextureCache::RefreshImage(Image& image) {
         return;
     }
 
-    scheduler.EndRendering();
+    auto* sched_ptr = custom_scheduler ? custom_scheduler : &scheduler;
-    const auto cmdbuf = scheduler.CommandBuffer();
+    sched_ptr->EndRendering();
+
+    const auto cmdbuf = sched_ptr->CommandBuffer();
     image.Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits::eTransferWrite, cmdbuf);
 
     const VAddr image_addr = image.info.guest_address;

src/video_core/texture_cache/texture_cache.h

@@ -59,17 +59,17 @@ public:
                                              const ImageViewInfo& view_info);
 
     /// Updates image contents if it was modified by CPU.
-    void UpdateImage(ImageId image_id) {
+    void UpdateImage(ImageId image_id, Vulkan::Scheduler* custom_scheduler = nullptr) {
         Image& image = slot_images[image_id];
         if (False(image.flags & ImageFlagBits::CpuModified)) {
             return;
         }
-        RefreshImage(image);
+        RefreshImage(image, custom_scheduler);
         TrackImage(image, image_id);
     }
 
     /// Reuploads image contents.
-    void RefreshImage(Image& image);
+    void RefreshImage(Image& image, Vulkan::Scheduler* custom_scheduler = nullptr);
 
     /// Retrieves the sampler that matches the provided S# descriptor.
     [[nodiscard]] vk::Sampler GetSampler(const AmdGpu::Sampler& sampler);