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

src/core/hle/service/sockets/bsd.cpp

@@ -40,12 +40,6 @@ bool IsConnectionBased(Type type) {
     }
 }
 
-// Size of a real BSD sockaddr_in. The guest SockAddrIn struct is padded to 0x100 so that longer
-// input addresses can be read, but an address handed back to the game is always this long.
-// Reporting 0x100 lets a game that trusts the length copy 240 bytes past its 16-byte sockaddr
-// (suspected cause of Crysis Remastered crashing ~30 s after its getsockname() call).
-constexpr size_t GuestSockAddrInSize = 16;
-
 template <typename T>
 T GetValue(std::span<const u8> buffer) {
     T t{};
@@ -661,9 +655,6 @@ std::pair<s32, Errno> BSD_USA::AcceptImpl(s32 fd, std::vector<u8>& write_buffer)
     new_descriptor.is_connection_based = descriptor.is_connection_based;
 
     const SockAddrIn guest_addr_in = Translate(result.sockaddr_in);
-    if (write_buffer.size() > GuestSockAddrInSize) {
-        write_buffer.resize(GuestSockAddrInSize);
-    }
     PutValue(write_buffer, guest_addr_in);
 
     return {new_fd, Errno::SUCCESS};
@@ -723,8 +714,8 @@ Errno BSD_USA::GetPeerNameImpl(s32 fd, std::vector<u8>& write_buffer) {
     }
     const SockAddrIn guest_addrin = Translate(addr_in);
 
-    // The reported length must be that of a real sockaddr_in, not of the padded guest struct.
-    write_buffer.resize(GuestSockAddrInSize);
+    ASSERT(write_buffer.size() >= sizeof(guest_addrin));
+    write_buffer.resize(sizeof(guest_addrin));
     PutValue(write_buffer, guest_addrin);
     return Translate(bsd_errno);
 }
@@ -745,8 +736,8 @@ Errno BSD_USA::GetSockNameImpl(s32 fd, std::vector<u8>& write_buffer) {
     }
     const SockAddrIn guest_addrin = Translate(addr_in);
 
-    // The reported length must be that of a real sockaddr_in, not of the padded guest struct.
-    write_buffer.resize(GuestSockAddrInSize);
+    ASSERT(write_buffer.size() >= sizeof(guest_addrin));
+    write_buffer.resize(sizeof(guest_addrin));
     PutValue(write_buffer, guest_addrin);
     return Translate(bsd_errno);
 }
@@ -958,9 +949,8 @@ std::pair<s32, Errno> BSD_USA::RecvFromImpl(s32 fd, u32 flags, std::vector<u8>&
         if (ret < 0) {
             addr.clear();
         } else {
-            ASSERT(addr.size() >= GuestSockAddrInSize);
+            ASSERT(addr.size() >= 16);
             const SockAddrIn result = Translate(addr_in);
-            addr.resize(GuestSockAddrInSize);
             PutValue(addr, result);
         }
     }

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/blit_image.cpp

@@ -563,8 +563,10 @@ void RecordShaderReadBarrier(Scheduler& scheduler, const ImageView& image_view)
     });
 }
 
-void BeginRenderPass(vk::CommandBuffer& cmdbuf, VkRenderPass render_pass,
-                     VkFramebuffer framebuffer_handle, VkExtent2D render_area) {
+void BeginRenderPass(vk::CommandBuffer& cmdbuf, const Framebuffer* framebuffer) {
+    const VkRenderPass render_pass = framebuffer->RenderPass();
+    const VkFramebuffer framebuffer_handle = framebuffer->Handle();
+    const VkExtent2D render_area = framebuffer->RenderArea();
     const VkRenderPassBeginInfo renderpass_bi{
         .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
         .pNext = nullptr,
@@ -677,67 +679,27 @@ void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, const ImageV
 void BlitImageHelper::BlitColor(const Framebuffer* dst_framebuffer, VkImageView src_image_view,
                                 VkImage src_image, VkSampler src_sampler,
                                 const Region2D& dst_region, const Region2D& src_region,
-                                const Extent3D& src_size,
-                                const std::optional<VkRect2D>& scissor) {
+                                const Extent3D& src_size) {
     const BlitImagePipelineKey key{
         .renderpass = dst_framebuffer->RenderPass(),
         .operation = Tegra::Engines::Fermi2D::Operation::SrcCopy,
     };
     const VkPipelineLayout layout = *one_texture_pipeline_layout;
     const VkPipeline pipeline = FindOrEmplaceColorPipeline(key);
-
-    // Clip the destination rectangle to the scissor; the viewport keeps the full rectangle so
-    // the texture coordinates are unchanged.
-    std::optional<VkRect2D> clip;
-    if (scissor) {
-        const s64 x0 = (std::max)(s64{(std::min)(dst_region.start.x, dst_region.end.x)},
-                                  s64{scissor->offset.x});
-        const s64 y0 = (std::max)(s64{(std::min)(dst_region.start.y, dst_region.end.y)},
-                                  s64{scissor->offset.y});
-        const s64 x1 = (std::min)(s64{(std::max)(dst_region.start.x, dst_region.end.x)},
-                                  s64{scissor->offset.x} + scissor->extent.width);
-        const s64 y1 = (std::min)(s64{(std::max)(dst_region.start.y, dst_region.end.y)},
-                                  s64{scissor->offset.y} + scissor->extent.height);
-        if (x1 <= x0 || y1 <= y0) {
-            return; // Fully scissored out.
-        }
-        clip = VkRect2D{
-            .offset = {static_cast<s32>(x0), static_cast<s32>(y0)},
-            .extent = {static_cast<u32>(x1 - x0), static_cast<u32>(y1 - y0)},
-        };
-    }
-
-    // Read the framebuffer now: by the time the worker runs this, the texture cache may have
-    // retired it (moved into its delayed-destruction ring), leaving a null handle behind - Turnip
-    // then crashed in vkCmdBeginRenderPass (Crysis Remastered, DrawTexture).
-    const VkRenderPass render_pass = dst_framebuffer->RenderPass();
-    const VkFramebuffer framebuffer_handle = dst_framebuffer->Handle();
-    const VkExtent2D render_area = dst_framebuffer->RenderArea();
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([this, render_pass, framebuffer_handle, render_area, src_image_view,
-                      src_image, src_sampler, dst_region, src_region, src_size, pipeline, layout,
-                      clip](vk::CommandBuffer cmdbuf) {
-        // Images stay in GENERAL everywhere (the descriptor below says so too): only make the
-        // earlier writes visible. Moving the source to SHADER_READ_ONLY_OPTIMAL left it there
-        // for good, so every later render pass used it in the wrong layout.
-        TransitionImageLayout(cmdbuf, src_image, VK_IMAGE_LAYOUT_GENERAL);
-        BeginRenderPass(cmdbuf, render_pass, framebuffer_handle, render_area);
+    scheduler.Record([this, dst_framebuffer, src_image_view, src_image, src_sampler, dst_region,
+                      src_region, src_size, pipeline, layout](vk::CommandBuffer cmdbuf) {
+        TransitionImageLayout(cmdbuf, src_image, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+        BeginRenderPass(cmdbuf, dst_framebuffer);
         const VkDescriptorSet descriptor_set = one_texture_descriptor_allocator.Commit();
         UpdateOneTextureDescriptorSet(device, descriptor_set, src_sampler, src_image_view);
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
         cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, layout, 0, descriptor_set,
                                   nullptr);
         BindBlitState(cmdbuf, layout, dst_region, src_region, src_size);
-        if (clip) {
-            cmdbuf.SetScissor(0, *clip);
-        }
         cmdbuf.Draw(3, 1, 0, 0);
         cmdbuf.EndRenderPass();
     });
-    // The blit set its own pipeline, viewport and scissor: make the next draw set its own again,
-    // as every other helper here does. Without this the next draw kept the blit's viewport and
-    // scissor: Crysis Remastered drew its 256x256 minimap map over the whole screen.
-    scheduler.InvalidateState();
 }
 
 void BlitImageHelper::BlitImpl(const Framebuffer* dst_framebuffer,

src/video_core/renderer_vulkan/blit_image.h

@@ -7,7 +7,6 @@
 #pragma once
 
 #include <deque>
-#include <optional>
 #include <span>
 
 #include "video_core/engines/fermi_2d.h"
@@ -79,12 +78,9 @@ public:
                    Tegra::Engines::Fermi2D::Filter filter,
                    Tegra::Engines::Fermi2D::Operation operation);
 
-    /// @param scissor Optional clip rectangle in framebuffer pixels (DrawTexture honours the
-    ///                guest scissor like any other draw).
     void BlitColor(const Framebuffer* dst_framebuffer, VkImageView src_image_view,
                    VkImage src_image, VkSampler src_sampler, const Region2D& dst_region,
-                   const Region2D& src_region, const Extent3D& src_size,
-                   const std::optional<VkRect2D>& scissor = std::nullopt);
+                   const Region2D& src_region, const Extent3D& src_size);
 
     void BlitColorMSAA(const Framebuffer* dst_framebuffer, const ImageView& src_image_view,
                        const Region2D& dst_region, const Region2D& src_region);

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() &&
@@ -393,35 +490,8 @@ void RasterizerVulkan::DrawTexture() {
                                     .y = ScaleSrc(draw_texture_state.src_y1)}};
     Extent3D src_size = {static_cast<u32>(ScaleSrc(texture.size.width)),
                          static_cast<u32>(ScaleSrc(texture.size.height)), texture.size.depth};
-    // DrawTexture rectangles go through the fragment pipeline, scissor included.
-    std::optional<VkRect2D> scissor;
-    if (maxwell3d->regs.scissor_test[0].enable) {
-        u32 up_scale = 1;
-        u32 down_shift = 0;
-        if (dst_rescaling) {
-            up_scale = Settings::values.resolution_info.up_scale;
-            down_shift = Settings::values.resolution_info.down_shift;
-        }
-        scissor = GetScissorState(maxwell3d->regs, 0, up_scale, down_shift);
-    }
-
-    // Drawing a texture 1:1 onto itself without blending changes nothing on hardware, but here
-    // it samples the image the render pass is writing to: undefined in Vulkan, and on tiled GPUs
-    // it reads whatever the image memory last held. Crysis Remastered does this every frame for
-    // its post-processing buffers (suspected cause of its map covering the whole screen).
-    const auto& fb_images = framebuffer->Images();
-    const bool src_is_target =
-        std::find(fb_images.begin(), fb_images.begin() + framebuffer->NumImages(),
-                  texture.ImageHandle()) != fb_images.begin() + framebuffer->NumImages();
-    if (src_is_target && src_rescaling == dst_rescaling &&
-        src_region.start.x == dst_region.start.x && src_region.start.y == dst_region.start.y &&
-        src_region.end.x == dst_region.end.x && src_region.end.y == dst_region.end.y &&
-        maxwell3d->regs.blend.enable[0] == 0) {
-        return;
-    }
-
     blit_image.BlitColor(framebuffer, texture.RenderTarget(), texture.ImageHandle(),
-                         sampler->Handle(), dst_region, src_region, src_size, scissor);
+                         sampler->Handle(), dst_region, src_region, src_size);
 }
 
 void RasterizerVulkan::Clear(u32 layer_count) {

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();