Merge pull request #12761 from liamwhite/mp-composite

video_core: rewrite presentation for layer composition

src/core/CMakeLists.txt

@@ -775,6 +775,9 @@ add_library(core STATIC
     hle/service/nvnflinger/graphic_buffer_producer.h
     hle/service/nvnflinger/hos_binder_driver_server.cpp
     hle/service/nvnflinger/hos_binder_driver_server.h
+    hle/service/nvnflinger/hardware_composer.cpp
+    hle/service/nvnflinger/hardware_composer.h
+    hle/service/nvnflinger/hwc_layer.h
     hle/service/nvnflinger/nvnflinger.cpp
     hle/service/nvnflinger/nvnflinger.h
     hle/service/nvnflinger/parcel.h

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -1,6 +1,8 @@
 // SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include <boost/container/small_vector.hpp>
+
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/core.h"
@@ -38,19 +40,30 @@ NvResult nvdisp_disp0::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> in
 void nvdisp_disp0::OnOpen(NvCore::SessionId session_id, DeviceFD fd) {}
 void nvdisp_disp0::OnClose(DeviceFD fd) {}
 
-void nvdisp_disp0::flip(u32 buffer_handle, u32 offset, android::PixelFormat format, u32 width,
-                        u32 height, u32 stride, android::BufferTransformFlags transform,
-                        const Common::Rectangle<int>& crop_rect,
-                        std::array<Service::Nvidia::NvFence, 4>& fences, u32 num_fences) {
-    const DAddr addr = nvmap.GetHandleAddress(buffer_handle);
-    LOG_TRACE(Service,
-              "Drawing from address {:X} offset {:08X} Width {} Height {} Stride {} Format {}",
-              addr, offset, width, height, stride, format);
+void nvdisp_disp0::Composite(std::span<const Nvnflinger::HwcLayer> sorted_layers) {
+    std::vector<Tegra::FramebufferConfig> output_layers;
+    std::vector<Service::Nvidia::NvFence> output_fences;
+    output_layers.reserve(sorted_layers.size());
+    output_fences.reserve(sorted_layers.size());
 
-    const Tegra::FramebufferConfig framebuffer{addr,   offset, width,     height,
-                                               stride, format, transform, crop_rect};
+    for (auto& layer : sorted_layers) {
+        output_layers.emplace_back(Tegra::FramebufferConfig{
+            .address = nvmap.GetHandleAddress(layer.buffer_handle),
+            .offset = layer.offset,
+            .width = layer.width,
+            .height = layer.height,
+            .stride = layer.stride,
+            .pixel_format = layer.format,
+            .transform_flags = layer.transform,
+            .crop_rect = layer.crop_rect,
+        });
 
-    system.GPU().RequestSwapBuffers(&framebuffer, fences, num_fences);
+        for (size_t i = 0; i < layer.acquire_fence.num_fences; i++) {
+            output_fences.push_back(layer.acquire_fence.fences[i]);
+        }
+    }
+
+    system.GPU().RequestComposite(std::move(output_layers), std::move(output_fences));
     system.SpeedLimiter().DoSpeedLimiting(system.CoreTiming().GetGlobalTimeUs());
     system.GetPerfStats().EndSystemFrame();
     system.GetPerfStats().BeginSystemFrame();

src/core/hle/service/nvdrv/devices/nvdisp_disp0.h

@@ -8,8 +8,7 @@
 #include "common/common_types.h"
 #include "common/math_util.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
-#include "core/hle/service/nvnflinger/buffer_transform_flags.h"
-#include "core/hle/service/nvnflinger/pixel_format.h"
+#include "core/hle/service/nvnflinger/hwc_layer.h"
 
 namespace Service::Nvidia::NvCore {
 class Container;
@@ -35,11 +34,8 @@ public:
     void OnOpen(NvCore::SessionId session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
-    /// Performs a screen flip, drawing the buffer pointed to by the handle.
-    void flip(u32 buffer_handle, u32 offset, android::PixelFormat format, u32 width, u32 height,
-              u32 stride, android::BufferTransformFlags transform,
-              const Common::Rectangle<int>& crop_rect,
-              std::array<Service::Nvidia::NvFence, 4>& fences, u32 num_fences);
+    /// Performs a screen flip, compositing each buffer.
+    void Composite(std::span<const Nvnflinger::HwcLayer> sorted_layers);
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;
 

src/core/hle/service/nvnflinger/hardware_composer.cpp

@@ -0,0 +1,215 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <boost/container/small_vector.hpp>
+
+#include "common/microprofile.h"
+#include "core/hle/service/nvdrv/devices/nvdisp_disp0.h"
+#include "core/hle/service/nvnflinger/buffer_item.h"
+#include "core/hle/service/nvnflinger/buffer_item_consumer.h"
+#include "core/hle/service/nvnflinger/buffer_queue_producer.h"
+#include "core/hle/service/nvnflinger/hardware_composer.h"
+#include "core/hle/service/nvnflinger/hwc_layer.h"
+#include "core/hle/service/nvnflinger/ui/graphic_buffer.h"
+#include "core/hle/service/vi/display/vi_display.h"
+#include "core/hle/service/vi/layer/vi_layer.h"
+
+namespace Service::Nvnflinger {
+
+namespace {
+
+s32 NormalizeSwapInterval(f32* out_speed_scale, s32 swap_interval) {
+    if (swap_interval <= 0) {
+        // As an extension, treat nonpositive swap interval as speed multiplier.
+        if (out_speed_scale) {
+            *out_speed_scale = 2.f * static_cast<f32>(1 - swap_interval);
+        }
+
+        swap_interval = 1;
+    }
+
+    if (swap_interval >= 5) {
+        // As an extension, treat high swap interval as precise speed control.
+        if (out_speed_scale) {
+            *out_speed_scale = static_cast<f32>(swap_interval) / 100.f;
+        }
+
+        swap_interval = 1;
+    }
+
+    return swap_interval;
+}
+
+} // namespace
+
+HardwareComposer::HardwareComposer() = default;
+HardwareComposer::~HardwareComposer() = default;
+
+u32 HardwareComposer::ComposeLocked(f32* out_speed_scale, VI::Display& display,
+                                    Nvidia::Devices::nvdisp_disp0& nvdisp, u32 frame_advance) {
+    boost::container::small_vector<HwcLayer, 2> composition_stack;
+
+    m_frame_number += frame_advance;
+
+    // Release any necessary framebuffers.
+    for (auto& [layer_id, framebuffer] : m_framebuffers) {
+        if (framebuffer.release_frame_number > m_frame_number) {
+            // Not yet ready to release this framebuffer.
+            continue;
+        }
+
+        if (!framebuffer.is_acquired) {
+            // Already released.
+            continue;
+        }
+
+        if (auto* layer = display.FindLayer(layer_id); layer != nullptr) {
+            // TODO: support release fence
+            // This is needed to prevent screen tearing
+            layer->GetConsumer().ReleaseBuffer(framebuffer.item, android::Fence::NoFence());
+            framebuffer.is_acquired = false;
+        }
+    }
+
+    // Set default speed limit to 100%.
+    *out_speed_scale = 1.0f;
+
+    // Determine the number of vsync periods to wait before composing again.
+    std::optional<s32> swap_interval{};
+    bool has_acquired_buffer{};
+
+    // Acquire all necessary framebuffers.
+    for (size_t i = 0; i < display.GetNumLayers(); i++) {
+        auto& layer = display.GetLayer(i);
+        auto layer_id = layer.GetLayerId();
+
+        // Try to fetch the framebuffer (either new or stale).
+        const auto result = this->CacheFramebufferLocked(layer, layer_id);
+
+        // If we failed, skip this layer.
+        if (result == CacheStatus::NoBufferAvailable) {
+            continue;
+        }
+
+        // If we acquired a new buffer, we need to present.
+        if (result == CacheStatus::BufferAcquired) {
+            has_acquired_buffer = true;
+        }
+
+        const auto& buffer = m_framebuffers[layer_id];
+        const auto& item = buffer.item;
+        const auto& igbp_buffer = *item.graphic_buffer;
+
+        // TODO: get proper Z-index from layer
+        composition_stack.emplace_back(HwcLayer{
+            .buffer_handle = igbp_buffer.BufferId(),
+            .offset = igbp_buffer.Offset(),
+            .format = igbp_buffer.ExternalFormat(),
+            .width = igbp_buffer.Width(),
+            .height = igbp_buffer.Height(),
+            .stride = igbp_buffer.Stride(),
+            .z_index = 0,
+            .transform = static_cast<android::BufferTransformFlags>(item.transform),
+            .crop_rect = item.crop,
+            .acquire_fence = item.fence,
+        });
+
+        // We need to compose again either before this frame is supposed to
+        // be released, or exactly on the vsync period it should be released.
+        const s32 item_swap_interval = NormalizeSwapInterval(out_speed_scale, item.swap_interval);
+
+        // TODO: handle cases where swap intervals are relatively prime. So far,
+        // only swap intervals of 0, 1 and 2 have been observed, but if 3 were
+        // to be introduced, this would cause an issue.
+        if (swap_interval) {
+            swap_interval = std::min(*swap_interval, item_swap_interval);
+        } else {
+            swap_interval = item_swap_interval;
+        }
+    }
+
+    // If any new buffers were acquired, we can present.
+    if (has_acquired_buffer) {
+        // Sort by Z-index.
+        std::stable_sort(composition_stack.begin(), composition_stack.end(),
+                         [&](auto& l, auto& r) { return l.z_index < r.z_index; });
+
+        // Composite.
+        nvdisp.Composite(composition_stack);
+    }
+
+    // Render MicroProfile.
+    MicroProfileFlip();
+
+    // Advance by at least one frame.
+    return swap_interval.value_or(1);
+}
+
+void HardwareComposer::RemoveLayerLocked(VI::Display& display, LayerId layer_id) {
+    // Check if we are tracking a slot with this layer_id.
+    const auto it = m_framebuffers.find(layer_id);
+    if (it == m_framebuffers.end()) {
+        return;
+    }
+
+    // Try to release the buffer item.
+    auto* const layer = display.FindLayer(layer_id);
+    if (layer && it->second.is_acquired) {
+        layer->GetConsumer().ReleaseBuffer(it->second.item, android::Fence::NoFence());
+    }
+
+    // Erase the slot.
+    m_framebuffers.erase(it);
+}
+
+bool HardwareComposer::TryAcquireFramebufferLocked(VI::Layer& layer, Framebuffer& framebuffer) {
+    // Attempt the update.
+    const auto status = layer.GetConsumer().AcquireBuffer(&framebuffer.item, {}, false);
+    if (status != android::Status::NoError) {
+        return false;
+    }
+
+    // We succeeded, so set the new release frame info.
+    framebuffer.release_frame_number =
+        NormalizeSwapInterval(nullptr, framebuffer.item.swap_interval);
+    framebuffer.is_acquired = true;
+
+    return true;
+}
+
+HardwareComposer::CacheStatus HardwareComposer::CacheFramebufferLocked(VI::Layer& layer,
+                                                                       LayerId layer_id) {
+    // Check if this framebuffer is already present.
+    const auto it = m_framebuffers.find(layer_id);
+    if (it != m_framebuffers.end()) {
+        // If it's currently still acquired, we are done.
+        if (it->second.is_acquired) {
+            return CacheStatus::CachedBufferReused;
+        }
+
+        // Try to acquire a new item.
+        if (this->TryAcquireFramebufferLocked(layer, it->second)) {
+            // We got a new item.
+            return CacheStatus::BufferAcquired;
+        } else {
+            // We didn't acquire a new item, but we can reuse the slot.
+            return CacheStatus::CachedBufferReused;
+        }
+    }
+
+    // Framebuffer is not present, so try to create it.
+    Framebuffer framebuffer{};
+
+    if (this->TryAcquireFramebufferLocked(layer, framebuffer)) {
+        // Move the buffer item into a new slot.
+        m_framebuffers.emplace(layer_id, std::move(framebuffer));
+
+        // We succeeded.
+        return CacheStatus::BufferAcquired;
+    }
+
+    // We couldn't acquire the buffer item, so don't create a slot.
+    return CacheStatus::NoBufferAvailable;
+}
+
+} // namespace Service::Nvnflinger

src/core/hle/service/nvnflinger/hardware_composer.h

@@ -0,0 +1,59 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <memory>
+#include <boost/container/flat_map.hpp>
+
+#include "core/hle/service/nvnflinger/buffer_item.h"
+
+namespace Service::Nvidia::Devices {
+class nvdisp_disp0;
+}
+
+namespace Service::VI {
+class Display;
+class Layer;
+} // namespace Service::VI
+
+namespace Service::Nvnflinger {
+
+using LayerId = u64;
+
+class HardwareComposer {
+public:
+    explicit HardwareComposer();
+    ~HardwareComposer();
+
+    u32 ComposeLocked(f32* out_speed_scale, VI::Display& display,
+                      Nvidia::Devices::nvdisp_disp0& nvdisp, u32 frame_advance);
+    void RemoveLayerLocked(VI::Display& display, LayerId layer_id);
+
+private:
+    // TODO: do we want to track frame number in vi instead?
+    u64 m_frame_number{0};
+
+private:
+    using ReleaseFrameNumber = u64;
+
+    struct Framebuffer {
+        android::BufferItem item{};
+        ReleaseFrameNumber release_frame_number{};
+        bool is_acquired{false};
+    };
+
+    enum class CacheStatus : u32 {
+        NoBufferAvailable,
+        BufferAcquired,
+        CachedBufferReused,
+    };
+
+    boost::container::flat_map<LayerId, Framebuffer> m_framebuffers{};
+
+private:
+    bool TryAcquireFramebufferLocked(VI::Layer& layer, Framebuffer& framebuffer);
+    CacheStatus CacheFramebufferLocked(VI::Layer& layer, LayerId layer_id);
+};
+
+} // namespace Service::Nvnflinger

src/core/hle/service/nvnflinger/hwc_layer.h

@@ -0,0 +1,27 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/math_util.h"
+#include "core/hle/service/nvdrv/nvdata.h"
+#include "core/hle/service/nvnflinger/buffer_transform_flags.h"
+#include "core/hle/service/nvnflinger/pixel_format.h"
+#include "core/hle/service/nvnflinger/ui/fence.h"
+
+namespace Service::Nvnflinger {
+
+struct HwcLayer {
+    u32 buffer_handle;
+    u32 offset;
+    android::PixelFormat format;
+    u32 width;
+    u32 height;
+    u32 stride;
+    s32 z_index;
+    android::BufferTransformFlags transform;
+    Common::Rectangle<int> crop_rect;
+    android::Fence acquire_fence;
+};
+
+} // namespace Service::Nvnflinger

src/core/hle/service/nvnflinger/nvnflinger.cpp

@@ -18,6 +18,7 @@
 #include "core/hle/service/nvnflinger/buffer_item_consumer.h"
 #include "core/hle/service/nvnflinger/buffer_queue_core.h"
 #include "core/hle/service/nvnflinger/fb_share_buffer_manager.h"
+#include "core/hle/service/nvnflinger/hardware_composer.h"
 #include "core/hle/service/nvnflinger/hos_binder_driver_server.h"
 #include "core/hle/service/nvnflinger/nvnflinger.h"
 #include "core/hle/service/nvnflinger/ui/graphic_buffer.h"
@@ -279,45 +280,19 @@ void Nvnflinger::Compose() {
         SCOPE_EXIT({ display.SignalVSyncEvent(); });
 
         // Don't do anything for displays without layers.
-        if (!display.HasLayers())
-            continue;
-
-        // TODO(Subv): Support more than 1 layer.
-        VI::Layer& layer = display.GetLayer(0);
-
-        android::BufferItem buffer{};
-        const auto status = layer.GetConsumer().AcquireBuffer(&buffer, {}, false);
-
-        if (status != android::Status::NoError) {
+        if (!display.HasLayers()) {
             continue;
         }
 
-        const auto& igbp_buffer = *buffer.graphic_buffer;
-
         if (!system.IsPoweredOn()) {
             return; // We are likely shutting down
         }
 
-        // Now send the buffer to the GPU for drawing.
-        // TODO(Subv): Support more than just disp0. The display device selection is probably based
-        // on which display we're drawing (Default, Internal, External, etc)
         auto nvdisp = nvdrv->GetDevice<Nvidia::Devices::nvdisp_disp0>(disp_fd);
         ASSERT(nvdisp);
 
-        Common::Rectangle<int> crop_rect{
-            static_cast<int>(buffer.crop.Left()), static_cast<int>(buffer.crop.Top()),
-            static_cast<int>(buffer.crop.Right()), static_cast<int>(buffer.crop.Bottom())};
-
-        nvdisp->flip(igbp_buffer.BufferId(), igbp_buffer.Offset(), igbp_buffer.ExternalFormat(),
-                     igbp_buffer.Width(), igbp_buffer.Height(), igbp_buffer.Stride(),
-                     static_cast<android::BufferTransformFlags>(buffer.transform), crop_rect,
-                     buffer.fence.fences, buffer.fence.num_fences);
-
-        MicroProfileFlip();
-
-        swap_interval = buffer.swap_interval;
-
-        layer.GetConsumer().ReleaseBuffer(buffer, android::Fence::NoFence());
+        swap_interval = display.GetComposer().ComposeLocked(&compose_speed_scale, display, *nvdisp,
+                                                            swap_interval);
     }
 }
 
@@ -334,15 +309,16 @@ s64 Nvnflinger::GetNextTicks() const {
             speed_scale = 0.01f;
         }
     }
+
+    // Adjust by speed limit determined during composition.
+    speed_scale /= compose_speed_scale;
+
     if (system.GetNVDECActive() && settings.use_video_framerate.GetValue()) {
         // Run at intended presentation rate during video playback.
         speed_scale = 1.f;
     }
 
-    // As an extension, treat nonpositive swap interval as framerate multiplier.
-    const f32 effective_fps = swap_interval <= 0 ? 120.f * static_cast<f32>(1 - swap_interval)
-                                                 : 60.f / static_cast<f32>(swap_interval);
-
+    const f32 effective_fps = 60.f / static_cast<f32>(swap_interval);
     return static_cast<s64>(speed_scale * (1000000000.f / effective_fps));
 }
 

src/core/hle/service/nvnflinger/nvnflinger.h

@@ -46,6 +46,7 @@ class BufferQueueProducer;
 namespace Service::Nvnflinger {
 
 class FbShareBufferManager;
+class HardwareComposer;
 class HosBinderDriverServer;
 
 class Nvnflinger final {
@@ -143,6 +144,7 @@ private:
     u32 next_buffer_queue_id = 1;
 
     s32 swap_interval = 1;
+    f32 compose_speed_scale = 1.0f;
 
     bool is_abandoned = false;
 

src/core/hle/service/vi/display/vi_display.cpp

@@ -16,6 +16,7 @@
 #include "core/hle/service/nvnflinger/buffer_queue_consumer.h"
 #include "core/hle/service/nvnflinger/buffer_queue_core.h"
 #include "core/hle/service/nvnflinger/buffer_queue_producer.h"
+#include "core/hle/service/nvnflinger/hardware_composer.h"
 #include "core/hle/service/nvnflinger/hos_binder_driver_server.h"
 #include "core/hle/service/vi/display/vi_display.h"
 #include "core/hle/service/vi/layer/vi_layer.h"
@@ -43,6 +44,7 @@ Display::Display(u64 id, std::string name_,
                  KernelHelpers::ServiceContext& service_context_, Core::System& system_)
     : display_id{id}, name{std::move(name_)}, hos_binder_driver_server{hos_binder_driver_server_},
       service_context{service_context_} {
+    hardware_composer = std::make_unique<Nvnflinger::HardwareComposer>();
     vsync_event = service_context.CreateEvent(fmt::format("Display VSync Event {}", id));
 }
 
@@ -81,8 +83,6 @@ void Display::SignalVSyncEvent() {
 
 void Display::CreateLayer(u64 layer_id, u32 binder_id,
                           Service::Nvidia::NvCore::Container& nv_core) {
-    ASSERT_MSG(layers.empty(), "Only one layer is supported per display at the moment");
-
     auto [core, producer, consumer] = CreateBufferQueue(service_context, nv_core.GetNvMapFile());
 
     auto buffer_item_consumer = std::make_shared<android::BufferItemConsumer>(std::move(consumer));

src/core/hle/service/vi/display/vi_display.h

@@ -11,9 +11,14 @@
 #include "common/common_types.h"
 #include "core/hle/result.h"
 
+namespace Core {
+class System;
+}
+
 namespace Kernel {
 class KEvent;
-}
+class KReadableEvent;
+} // namespace Kernel
 
 namespace Service::android {
 class BufferQueueProducer;
@@ -24,8 +29,9 @@ class ServiceContext;
 }
 
 namespace Service::Nvnflinger {
+class HardwareComposer;
 class HosBinderDriverServer;
-}
+} // namespace Service::Nvnflinger
 
 namespace Service::Nvidia::NvCore {
 class Container;
@@ -118,6 +124,10 @@ public:
     ///
     const Layer* FindLayer(u64 layer_id) const;
 
+    Nvnflinger::HardwareComposer& GetComposer() const {
+        return *hardware_composer;
+    }
+
 private:
     u64 display_id;
     std::string name;
@@ -125,6 +135,7 @@ private:
     KernelHelpers::ServiceContext& service_context;
 
     std::vector<std::unique_ptr<Layer>> layers;
+    std::unique_ptr<Nvnflinger::HardwareComposer> hardware_composer;
     Kernel::KEvent* vsync_event{};
     bool is_abandoned{};
 };

src/core/hle/service/vi/vi.cpp

@@ -195,8 +195,9 @@ private:
     void GetSharedBufferMemoryHandleId(HLERequestContext& ctx) {
         IPC::RequestParser rp{ctx};
         const u64 buffer_id = rp.PopRaw<u64>();
+        const u64 aruid = ctx.GetPID();
 
-        LOG_INFO(Service_VI, "called. buffer_id={:#x}", buffer_id);
+        LOG_INFO(Service_VI, "called. buffer_id={:#x}, aruid={:#x}", buffer_id, aruid);
 
         struct OutputParameters {
             s32 nvmap_handle;
@@ -206,7 +207,7 @@ private:
         OutputParameters out{};
         Nvnflinger::SharedMemoryPoolLayout layout{};
         const auto result = nvnflinger.GetSystemBufferManager().GetSharedBufferMemoryHandleId(
-            &out.size, &out.nvmap_handle, &layout, buffer_id, 0);
+            &out.size, &out.nvmap_handle, &layout, buffer_id, aruid);
 
         ctx.WriteBuffer(&layout, sizeof(layout));