[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/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/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/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() &&

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -157,6 +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();