[shader_recompiler, renderer_vulkan] draw sprites whose texture handle comes from a vertex attribute

Performance loss regression fixed in Link's Awakening, Castlevania, Super Mario RPG, etc

Tested by Mr. Knobs on Android and myself on Linux

Multicore ON may be fixed for some users using Intel CPUs.
Games include: FF Tactics, Shredder's Revenge, Unity Engine Games.

- [x] I have read and followed the [Contribution Guidelines](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/CONTRIBUTING.md#code-contributions).
- [x] I have read and followed the [AI Policy](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/docs/policies/AI.md).
- [x] I have read and followed the [Coding Guidelines](https://git.eden-emu.dev/eden-emu/eden/src/branch/master/docs/policies/Coding.md) to the best of my ability.

-------------------

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/4548
Reviewed-by: Maufeat <sahyno1996@gmail.com>
Reviewed-by: lizzie <lizzie@eden-emu.dev>

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/model/BooleanSetting.kt

@@ -16,6 +16,7 @@ enum class BooleanSetting(override val key: String) : AbstractBooleanSetting {
     RENDERER_USE_SPEED_LIMIT("use_speed_limit"),
     USE_CUSTOM_CPU_TICKS("use_custom_cpu_ticks"),
     SKIP_CPU_INNER_INVALIDATION("skip_cpu_inner_invalidation"),
+    NCE_INVALIDATION_GPU_READBACK("nce_invalidation_gpu_readback"),
     FIX_BLOOM_EFFECTS("fix_bloom_effects"),
     EMULATE_BGR565("emulate_bgr565"),
     RESCALE_HACK("rescale_hack"),

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/model/view/SettingsItem.kt

@@ -859,6 +859,13 @@ abstract class SettingsItem(
                     descriptionId = R.string.skip_cpu_inner_invalidation_description
                 )
             )
+            put(
+                SwitchSetting(
+                    BooleanSetting.NCE_INVALIDATION_GPU_READBACK,
+                    titleId = R.string.nce_invalidation_gpu_readback,
+                    descriptionId = R.string.nce_invalidation_gpu_readback_description
+                )
+            )
             put(
                 SwitchSetting(
                     BooleanSetting.FIX_BLOOM_EFFECTS,

src/android/app/src/main/java/org/yuzu/yuzu_emu/features/settings/ui/SettingsFragmentPresenter.kt

@@ -554,6 +554,7 @@ class SettingsFragmentPresenter(
             add(HeaderSetting(R.string.hacks))
 
             add(BooleanSetting.SKIP_CPU_INNER_INVALIDATION.key)
+            add(BooleanSetting.NCE_INVALIDATION_GPU_READBACK.key)
             add(BooleanSetting.FIX_BLOOM_EFFECTS.key)
             add(BooleanSetting.EMULATE_BGR565.key)
             add(BooleanSetting.RENDERER_ASYNCHRONOUS_SHADERS.key)

src/android/app/src/main/res/values-ar/strings.xml

@@ -587,6 +587,8 @@
     <string name="fast_gpu_time_description">يجعل اللعبة تعتقد أن عمل وحدة معالجة الرسومات ينتهي بشكل أسرع مما هو عليه في الواقع، لذلك تتوقف عن خفض الدقة ومسافة العرض لتتناسب مع سرعات جهاز Switch.</string>
     <string name="skip_cpu_inner_invalidation">تخطي إبطال صلاحية وحدة المعالجة المركزية الداخلية</string>
     <string name="skip_cpu_inner_invalidation_description">يتخطى بعض عمليات إبطال ذاكرة التخزين المؤقتة من جانب وحدة المعالجة المركزية أثناء تحديثات الذاكرة، مما يقلل من استخدام وحدة المعالجة المركزية ويحسن أداءها. قد يتسبب ذلك في حدوث أعطال أو تعطل في بعض الألعاب.</string>
+    <string name="nce_invalidation_gpu_readback">إعادة قراءة وحدة معالجة الرسومات قبل إبطال صلاحية NCE</string>
+    <string name="nce_invalidation_gpu_readback_description">إصلاح مؤقت. يعالج مشكلة تعطل لعبة Ender Magnolia (وغيرها من ألعاب UE). قد يؤدي إلى انخفاض في الأداء. يجري العمل على تحسينه.</string>
     <string name="fix_bloom_effects">إصلاح تأثيرات التوهج</string>
     <string name="fix_bloom_effects_description">يقلل من ضبابية التوهج في LA/EOW  (Adreno A6XX - A7XX/ Turnip)، ويزيل التوهج في Burnout. تحذير: قد يسبب تشوهات رسومية في ألعاب أخرى.</string>
     <string name="emulate_bgr565">محاكاة BGR565</string>

src/android/app/src/main/res/values-zh-rTW/strings.xml

@@ -577,6 +577,8 @@
     <string name="fast_gpu_time_description">讓遊戲誤以為 GPU 工作能更快完成，避免遊戲為了配合 Switch 的時脈而降低解析度與渲染距離</string>
     <string name="skip_cpu_inner_invalidation">跳過 CPU 內部失效處理</string>
     <string name="skip_cpu_inner_invalidation_description">在記憶體更新期間跳過某些 CPU 端快取的失效處理，減少 CPU 使用率並提高其性能。可能會導致某些遊戲出現故障或當機</string>
+    <string name="nce_invalidation_gpu_readback">在 NCE 無效化前進行 GPU 回讀</string>
+    <string name="nce_invalidation_gpu_readback_description">（暫時性功能）修復 \"終結者瑪格諾莉亞：霧中綻放\" 及其它使用虛幻引擎之遊戲中的當機問題，可能會降低效能。此功能正在改善中</string>
     <string name="fix_bloom_effects">修正光暈特效</string>
     <string name="fix_bloom_effects_description">在 \"薩爾達傳說：織夢島\" 和 \"薩爾達傳說：智慧的再現\" 中減少光暈模糊的問題並移除 \"橫衝直撞：狂飆樂園重製版\" 中的光暈效果（適用於 Adreno A6XX - A7XX/ Turnip）。警告：可能會在其它遊戲中導致圖形異常</string>
     <string name="emulate_bgr565">模擬 BGR565</string>

src/android/app/src/main/res/values/strings.xml

@@ -593,6 +593,8 @@
     <string name="fast_gpu_time_description">Makes the game believe GPU work finishes faster than it does, so it stops lowering resolution and render distance to fit the Switch\'s clocks.</string>
     <string name="skip_cpu_inner_invalidation">Skip CPU Inner Invalidation</string>
     <string name="skip_cpu_inner_invalidation_description">Skips certain CPU-side cache invalidations during memory updates, reducing CPU usage and improving it\'s performance. This may cause glitches or crashes on some games.</string>
+    <string name="nce_invalidation_gpu_readback">GPU Readback Before NCE Invalidation</string>
+    <string name="nce_invalidation_gpu_readback_description">Temporary. Fix Ender Magnolia (and other UE titles) crashes. May reduce performance. Improvement in progress.</string>
     <string name="fix_bloom_effects">Fix Bloom Effects</string>
     <string name="fix_bloom_effects_description">Reduces bloom blur in LA/EOW (Adreno A6XX - A7XX/ Turnip), removes bloom in Burnout. Warning: may cause graphical artifacts in other games.</string>
     <string name="emulate_bgr565">Emulate BGR565</string>

src/common/settings.h

@@ -639,6 +639,8 @@ struct Values {
                                                         Specialization::Default,
                                                         true,
                                                         true};
+    SwitchableSetting<bool> nce_invalidation_gpu_readback{
+        linkage, false, "nce_invalidation_gpu_readback", Category::RendererHacks};
     SwitchableSetting<bool> async_presentation{linkage,
 #ifdef __ANDROID__
                                                true,

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

@@ -57,8 +57,21 @@ u32 HardwareComposer::ComposeLocked(f32* out_speed_scale, Display& display,
     // Set default speed limit to 100%.
     *out_speed_scale = 1.0f;
 
+    // If no layers are visible, skip the logic. Acquiring and holding buffers here would
+    // consume the producer's queue without ever presenting.
+    bool any_visible = false;
+    for (auto& layer : display.stack.layers) {
+        if (layer->visible) {
+            any_visible = true;
+            break;
+        }
+    }
+    if (!any_visible) {
+        return 1;
+    }
+
+    // Wait for the previous composite, bounding the number of in-flight composites.
     nvdisp.WaitForComposite();
-    this->ReleaseFramebuffersLocked(display);
 
     // Determine the number of vsync periods to wait before composing again.
     std::optional<s32> swap_interval{};
@@ -152,6 +165,13 @@ u32 HardwareComposer::ComposeLocked(f32* out_speed_scale, Display& display,
     // Advance by 1 frame (60 FPS compositing)
     m_frame_number += 1;
 
+    // Release framebuffers once the frame has been handed to the compositor. This happens
+    // within the same vsync period as the acquire (release_frame_number is checked against
+    // the advanced counter), so the producer gets its buffers back immediately instead of
+    // losing a queue slot for a whole frame.
+    // TODO: support release fences; releasing before the composite has finished can tear.
+    this->ReleaseFramebuffersLocked(display);
+
     return 1;
 }
 

src/input_common/drivers/joycon.cpp

@@ -83,21 +83,20 @@ void Joycons::ScanThread(std::stop_token stop_token) {
     constexpr u16 nintendo_vendor_id = 0x057e;
     Common::SetCurrentThreadName("JoyconScanThread");
 
-    auto last_ret = 0u;
     do {
-        auto ret = SDL_hid_device_change_count();
-        if (last_ret == 0u || last_ret != ret) {
-            last_ret = ret;
-            SDL_hid_device_info* devs = SDL_hid_enumerate(nintendo_vendor_id, 0x0);
-            for (auto* cur_dev = devs; cur_dev; cur_dev = cur_dev->next) {
-                if (IsDeviceNew(cur_dev)) {
-                    LOG_DEBUG(Input, "Device Found,type : {:04X} {:04X}", cur_dev->vendor_id, cur_dev->product_id);
-                    RegisterNewDevice(cur_dev);
-                }
-                cur_dev = cur_dev->next;
+        SDL_hid_device_info* devs = SDL_hid_enumerate(nintendo_vendor_id, 0x0);
+        SDL_hid_device_info* cur_dev = devs;
+
+        while (cur_dev) {
+            if (IsDeviceNew(cur_dev)) {
+                LOG_DEBUG(Input, "Device Found,type : {:04X} {:04X}", cur_dev->vendor_id,
+                          cur_dev->product_id);
+                RegisterNewDevice(cur_dev);
             }
-            SDL_hid_free_enumeration(devs);
+            cur_dev = cur_dev->next;
         }
+
+        SDL_hid_free_enumeration(devs);
     } while (Common::StoppableTimedWait(stop_token, std::chrono::seconds{5}));
 }
 

src/qt_common/config/shared_translation.cpp

@@ -194,6 +194,10 @@ std::unique_ptr<TranslationMap> InitializeTranslations(QObject* parent) {
     INSERT(Settings, skip_cpu_inner_invalidation, tr("Skip CPU Inner Invalidation"),
            tr("Skips certain cache invalidations during memory updates, reducing CPU usage and "
               "improving latency. This may cause soft-crashes."));
+    INSERT(Settings, nce_invalidation_gpu_readback,
+           tr("GPU Readback Before NCE Invalidation"),
+           tr("Flushes GPU buffer data back to guest memory before NCE invalidates the whole "
+              "guest page, preserving it while the CPU store proceeds."));
     INSERT(Settings, vsync_mode, tr("VSync Mode:"),
            tr("FIFO (VSync) does not drop frames or exhibit tearing but is limited by the screen "
               "refresh rate.\nFIFO Relaxed allows tearing as it recovers from a slow down.\n"

src/shader_recompiler/host_translate_info.h

@@ -40,6 +40,8 @@ struct HostTranslateInfo {
     bool support_conditional_barrier{}; ///< True when the device supports barriers in conditional
                                         ///< control flow
     bool single_lane_geometry_subgroups{};
+    bool support_attribute_texture_handle{}; ///< True when the host binds texture handles that
+                                             ///< arrive as vertex attributes
 
     void ApplyDescriptorLimitPolicy() noexcept {
         if (min_ssbo_alignment == 0) {

src/shader_recompiler/ir_opt/texture_pass.cpp

@@ -268,6 +268,10 @@ static inline u32 ReadCbufCached(Environment& env, u32 index, u32 offset) {
     return v;
 }
 
+static inline bool IsAttributeHandle(const ConstBufferAddr& cbuf) {
+    return cbuf.index == ATTRIBUTE_HANDLE_CBUF_INDEX;
+}
+
 static inline u32 GetTextureHandleCached(Environment& env, const ConstBufferAddr& cbuf) {
     // Must all be uniquely different variables
     // If has secondary, then it will be cbuf.secondary_{index|offset}, else its 0.
@@ -287,14 +291,23 @@ static inline u32 GetTextureHandleCached(Environment& env, const ConstBufferAddr
 
 // Cached variants of existing helpers
 static inline TextureType ReadTextureTypeCached(Environment& env, const ConstBufferAddr& cbuf) {
+    if (IsAttributeHandle(cbuf)) {
+        return TextureType::Color2D;
+    }
     return env.ReadTextureType(GetTextureHandleCached(env, cbuf));
 }
 static inline TexturePixelFormat ReadTexturePixelFormatCached(Environment& env,
                                                                 const ConstBufferAddr& cbuf) {
+    if (IsAttributeHandle(cbuf)) {
+        return TexturePixelFormat::A8B8G8R8_UNORM;
+    }
     return env.ReadTexturePixelFormat(GetTextureHandleCached(env, cbuf));
 }
 static inline bool IsTexturePixelFormatIntegerCached(Environment& env,
                                                         const ConstBufferAddr& cbuf) {
+    if (IsAttributeHandle(cbuf)) {
+        return false;
+    }
     return env.IsTexturePixelFormatInteger(GetTextureHandleCached(env, cbuf));
 }
 
@@ -466,15 +479,43 @@ std::optional<ConstBufferAddr> TryGetConstBuffer(const IR::Inst* inst, Environme
     };
 }
 
+bool IsAttributeSourcedHandle(const IR::Value& handle) {
+    if (handle.IsImmediate()) {
+        return false;
+    }
+    const IR::Inst* inst{handle.InstRecursive()};
+    if (inst->GetOpcode() == IR::Opcode::BitCastU32F32 && !inst->Arg(0).IsImmediate()) {
+        inst = inst->Arg(0).InstRecursive();
+    }
+    return inst->GetOpcode() == IR::Opcode::GetAttribute ||
+           inst->GetOpcode() == IR::Opcode::GetAttributeU32;
+}
+
 TextureInst MakeInst(Environment& env, IR::Block* block, IR::Inst& inst, const HostTranslateInfo& host_info) {
     ConstBufferAddr addr;
     if (IsBindless(inst)) {
         const std::optional<ConstBufferAddr> track_addr{TrackCached(inst.Arg(0), env, host_info)};
 
-        if (!track_addr) {
-            throw NotImplementedException("Failed to track bindless texture constant buffer");
-        } else {
+        if (track_addr) {
             addr = *track_addr;
+        } else if (host_info.support_attribute_texture_handle &&
+                   IsAttributeSourcedHandle(inst.Arg(0))) {
+            // The handle travels with the vertices (sprite batches pick the texture per quad), so
+            // there is no constant buffer to read it from. The host binds the handle of each run
+            // of vertices that share one.
+            addr = ConstBufferAddr{
+                .index = ATTRIBUTE_HANDLE_CBUF_INDEX,
+                .offset = 0,
+                .shift_left = 0,
+                .secondary_index = 0,
+                .secondary_offset = 0,
+                .secondary_shift_left = 0,
+                .dynamic_offset = {},
+                .count = 1,
+                .has_secondary = false,
+            };
+        } else {
+            throw NotImplementedException("Failed to track bindless texture constant buffer");
         }
     } else {
         addr = ConstBufferAddr{

src/shader_recompiler/shader_info.h

@@ -38,6 +38,10 @@ enum class TextureType : u32 {
 };
 constexpr u32 NUM_TEXTURE_TYPES = 9;
 
+/// Constant buffer index marking a texture handle that the shader receives as a vertex attribute
+/// instead of reading it from a constant buffer. The host supplies the handle of each draw run.
+constexpr u32 ATTRIBUTE_HANDLE_CBUF_INDEX = 0xFFFF;
+
 enum class TexturePixelFormat {
     A8B8G8R8_UNORM,
     A8B8G8R8_SNORM,

src/video_core/gpu.cpp

@@ -227,17 +227,17 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be invalidated
     void InvalidateRegion(DAddr addr, u64 size, bool preserve_gpu_writes) {
-        if (preserve_gpu_writes) {
+        if (Settings::values.nce_invalidation_gpu_readback.GetValue()) {
             VideoCore::RasterizerInterface* rasterizer = renderer->ReadRasterizer();
-            if (!rasterizer->TryCacheInvalidation(addr, size)) {
+            if (preserve_gpu_writes && rasterizer->MustFlushRegion(addr, size, VideoCommon::CacheType::BufferCache)) {
                 const u64 fence = RequestSyncOperation([rasterizer, addr, size] {
                     rasterizer->FlushRegion(addr, size, VideoCommon::CacheType::BufferCache);
                     rasterizer->OnCacheInvalidation(addr, size);
                 });
                 gpu_thread.TickGPU(is_async);
                 WaitForSyncOperation(fence);
+                return;
             }
-            return;
         }
         gpu_thread.InvalidateRegion(addr, size);
     }

src/video_core/rasterizer_interface.h

@@ -1,6 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
 // SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -114,12 +111,6 @@ public:
     /// Notify rasterizer that any caches of the specified region are desync with guest
     virtual void OnCacheInvalidation(PAddr addr, u64 size) = 0;
 
-    virtual bool TryCacheInvalidation(DAddr addr, u64 size) {
-        if (MustFlushRegion(addr, size, VideoCommon::CacheType::BufferCache)) return false;
-        OnCacheInvalidation(addr, size);
-        return true;
-    }
-
     virtual bool OnCPUWrite(PAddr addr, u64 size) = 0;
 
     /// Sync memory between guest and host.

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -276,6 +276,9 @@ GraphicsPipeline::GraphicsPipeline(
         num_textures += Shader::NumDescriptors(info->texture_descriptors);
         num_image_elements += Shader::NumDescriptors(info->texture_descriptors);
         num_image_elements += Shader::NumDescriptors(info->image_descriptors);
+        for (const auto& desc : info->texture_descriptors) {
+            uses_attribute_handle |= desc.cbuf_index == Shader::ATTRIBUTE_HANDLE_CBUF_INDEX;
+        }
         num_descriptor_entries += NumDescriptorEntries(*info);
     }
     fragment_has_color0_output = stage_infos[NUM_STAGES - 1].stores_frag_color[0];
@@ -375,6 +378,9 @@ bool GraphicsPipeline::ConfigureImpl(bool is_indexed) {
         }
         const auto& cbufs{maxwell3d->state.shader_stages[stage].const_buffers};
         const auto read_handle{[&](const auto& desc, u32 index) {
+            if (desc.cbuf_index == Shader::ATTRIBUTE_HANDLE_CBUF_INDEX) {
+                return TexturePair(attribute_handle, via_header_index);
+            }
             ASSERT(cbufs[desc.cbuf_index].enabled);
             const u32 index_offset{index << desc.size_shift};
             const u32 offset{desc.cbuf_offset + index_offset};

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -110,6 +110,16 @@ public:
         return configure_func(this, is_indexed);
     }
 
+    /// True when a shader samples a texture whose handle arrives as a vertex attribute.
+    [[nodiscard]] bool UsesAttributeHandle() const noexcept {
+        return uses_attribute_handle;
+    }
+
+    /// Handle bound to those textures by the next Configure.
+    void SetAttributeHandle(u32 handle) noexcept {
+        attribute_handle = handle;
+    }
+
     [[nodiscard]] GraphicsPipeline* Next(const GraphicsPipelineCacheKey& current_key) noexcept {
         if (key == current_key) {
             return this;
@@ -168,6 +178,8 @@ private:
     u32 num_descriptor_entries{};
     size_t num_image_elements{};
     u32 num_textures{};
+    u32 attribute_handle{};
+    bool uses_attribute_handle{};
     bool fragment_has_color0_output{};
 
     vk::DescriptorSetLayout descriptor_set_layout;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -491,6 +491,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_conditional_barrier = device.SupportsConditionalBarriers(),
         .single_lane_geometry_subgroups = !profile.SupportsSubgroupStage(Shader::Stage::Geometry),
+        .support_attribute_texture_handle = true,
     };
     host_info.ApplyDescriptorLimitPolicy();
 

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -248,6 +248,15 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     // update engine as channel may be different.
     pipeline->SetEngine(maxwell3d, gpu_memory);
+    attribute_handle_pipeline = nullptr;
+    attribute_handle_runs.clear();
+    if (split_attribute_handles && pipeline->UsesAttributeHandle()) {
+        CollectAttributeHandleRuns(is_indexed);
+        if (!attribute_handle_runs.empty()) {
+            attribute_handle_pipeline = pipeline;
+            pipeline->SetAttributeHandle(attribute_handle_runs.front().handle);
+        }
+    }
     if (!pipeline->Configure(is_indexed))
         return;
 
@@ -259,22 +268,110 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     draw_func();
 }
 
+void RasterizerVulkan::CollectAttributeHandleRuns(bool is_indexed) {
+    using VertexAttribute = Maxwell::VertexAttribute;
+    const auto& regs = maxwell3d->regs;
+    const auto& draw_state = maxwell3d->draw_manager.draw_state;
+    const u32 total = is_indexed ? draw_state.index_buffer.count : draw_state.vertex_buffer.count;
+    if (total == 0) {
+        return;
+    }
+    // The handle is an integer attribute; sprite batches repeat it on every vertex of a quad.
+    const VertexAttribute* source{};
+    for (const auto& attr : regs.vertex_attrib_format) {
+        if (attr.constant || attr.size == VertexAttribute::Size::Invalid) {
+            continue;
+        }
+        if (attr.type == VertexAttribute::Type::UInt || attr.type == VertexAttribute::Type::SInt) {
+            source = &attr;
+            break;
+        }
+    }
+    const auto& stream = regs.vertex_streams[source ? source->buffer.Value() : 0];
+    if (!source || !stream.IsEnabled()) {
+        attribute_handle_runs.push_back({0, total, 0});
+        return;
+    }
+    // Only plain triangle lists are split; anything else keeps the handle of its first vertex.
+    const u32 primitive_size =
+        draw_state.topology == Maxwell::PrimitiveTopology::Triangles ? 3 : total;
+    const GPUVAddr base_address = stream.Address() + source->offset;
+    const auto read_handle = [&](u32 position) {
+        u32 vertex;
+        if (is_indexed) {
+            const auto& index_buffer = draw_state.index_buffer;
+            const GPUVAddr address =
+                index_buffer.IndexStart() + size_t{position} * index_buffer.FormatSizeInBytes();
+            u32 index{};
+            gpu_memory->ReadBlockUnsafe(address, &index, index_buffer.FormatSizeInBytes());
+            vertex = static_cast<u32>(draw_state.base_index + index);
+        } else {
+            vertex = draw_state.vertex_buffer.first + position;
+        }
+        return gpu_memory->Read<u32>(base_address + static_cast<GPUVAddr>(vertex) * stream.stride);
+    };
+    for (u32 position = 0; position + primitive_size <= total; position += primitive_size) {
+        const u32 handle = read_handle(position);
+        if (!attribute_handle_runs.empty() && attribute_handle_runs.back().handle == handle) {
+            attribute_handle_runs.back().count += primitive_size;
+        } else {
+            attribute_handle_runs.push_back({position, primitive_size, handle});
+        }
+    }
+    if (attribute_handle_runs.empty()) {
+        attribute_handle_runs.push_back({0, total, read_handle(0)});
+    } else {
+        const u32 covered = attribute_handle_runs.back().first + attribute_handle_runs.back().count;
+        if (covered < total) {
+            attribute_handle_runs.back().count += total - covered;
+        }
+    }
+}
+
 void RasterizerVulkan::Draw(bool is_indexed, u32 instance_count) {
+    split_attribute_handles = true;
+    SCOPE_EXIT {
+        split_attribute_handles = false;
+    };
     PrepareDraw(is_indexed, [this, is_indexed, instance_count] {
         const auto& draw_state = maxwell3d->draw_manager.draw_state;
         const u32 num_instances{instance_count};
         const DrawParams draw_params{MakeDrawParams(draw_state, num_instances, is_indexed)};
 
-        scheduler.Record([draw_params](vk::CommandBuffer cmdbuf) {
-            if (draw_params.is_indexed) {
-                cmdbuf.DrawIndexed(draw_params.num_vertices, draw_params.num_instances,
-                                   draw_params.first_index, draw_params.base_vertex,
-                                   draw_params.base_instance);
-            } else {
-                cmdbuf.Draw(draw_params.num_vertices, draw_params.num_instances,
-                            draw_params.base_vertex, draw_params.base_instance);
+        const auto record_draw = [this](const DrawParams& params) {
+            scheduler.Record([params](vk::CommandBuffer cmdbuf) {
+                if (params.is_indexed) {
+                    cmdbuf.DrawIndexed(params.num_vertices, params.num_instances,
+                                       params.first_index, params.base_vertex,
+                                       params.base_instance);
+                } else {
+                    cmdbuf.Draw(params.num_vertices, params.num_instances, params.base_vertex,
+                                params.base_instance);
+                }
+            });
+        };
+        if (attribute_handle_pipeline && attribute_handle_runs.size() > 1) {
+            // The first run is already configured, each further one binds its own texture.
+            for (size_t i = 0; i < attribute_handle_runs.size(); ++i) {
+                const AttributeHandleRun& run = attribute_handle_runs[i];
+                if (i != 0) {
+                    attribute_handle_pipeline->SetAttributeHandle(run.handle);
+                    if (!attribute_handle_pipeline->Configure(is_indexed)) {
+                        continue;
+                    }
+                }
+                DrawParams run_params = draw_params;
+                run_params.num_vertices = run.count;
+                if (is_indexed) {
+                    run_params.first_index += run.first;
+                } else {
+                    run_params.base_vertex += run.first;
+                }
+                record_draw(run_params);
             }
-        });
+        } else {
+            record_draw(draw_params);
+        }
 
         // Log draw call
         if (GPU::Logging::IsActive() &&
@@ -798,18 +895,6 @@ void RasterizerVulkan::OnCacheInvalidation(DAddr addr, u64 size) {
     pipeline_cache.InvalidateRegion(addr, size);
 }
 
-bool RasterizerVulkan::TryCacheInvalidation(DAddr addr, u64 size) {
-    if (addr == 0 || size == 0) return true;
-    {
-        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
-        if (buffer_cache.IsRegionGpuModified(addr, size)) return false;
-        texture_cache.WriteMemory(addr, size);
-        buffer_cache.WriteMemory(addr, size);
-    }
-    pipeline_cache.InvalidateRegion(addr, size);
-    return true;
-}
-
 void RasterizerVulkan::InvalidateGPUCache() {
     gpu.InvalidateGPUCache();
 }

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -106,7 +106,6 @@ public:
                           VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
     void InnerInvalidation(std::span<const std::pair<DAddr, std::size_t>> sequences) override;
     void OnCacheInvalidation(DAddr addr, u64 size) override;
-    bool TryCacheInvalidation(DAddr addr, u64 size) override;
     bool OnCPUWrite(DAddr addr, u64 size) override;
     void InvalidateGPUCache() override;
     void UnmapMemory(DAddr addr, u64 size) override;
@@ -158,6 +157,20 @@ private:
     template <typename Func>
     void PrepareDraw(bool is_indexed, Func&&);
 
+    /// A stretch of a draw whose vertices all carry the same texture handle.
+    struct AttributeHandleRun {
+        u32 first;
+        u32 count;
+        u32 handle;
+    };
+
+    /// Splits the current draw wherever the texture handle carried in the vertex data changes.
+    void CollectAttributeHandleRuns(bool is_indexed);
+
+    std::vector<AttributeHandleRun> attribute_handle_runs;
+    GraphicsPipeline* attribute_handle_pipeline{};
+    bool split_attribute_handles{};
+
     void FlushWork();
 
     void UpdateDynamicStates();