video_core: Introduce buffer cache

CMakeLists.txt

@@ -294,6 +294,7 @@ set(COMMON src/common/logging/backend.cpp
            src/common/thread.h
            src/common/types.h
            src/common/uint128.h
+           src/common/unique_function.h
            src/common/version.h
            src/common/ntapi.h
            src/common/ntapi.cpp
@@ -451,6 +452,14 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/amdgpu/pm4_cmds.h
                src/video_core/amdgpu/pm4_opcodes.h
                src/video_core/amdgpu/resource.h
+               src/video_core/buffer_cache/buffer.cpp
+               src/video_core/buffer_cache/buffer.h
+               src/video_core/buffer_cache/buffer_cache.cpp
+               src/video_core/buffer_cache/buffer_cache.h
+               src/video_core/buffer_cache/memory_tracker_base.h
+               src/video_core/buffer_cache/range_set.cpp
+               src/video_core/buffer_cache/range_set.h
+               src/video_core/buffer_cache/word_manager.h
                src/video_core/renderer_vulkan/liverpool_to_vk.cpp
                src/video_core/renderer_vulkan/liverpool_to_vk.h
                src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -496,6 +505,8 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/texture_cache/tile_manager.cpp
                src/video_core/texture_cache/tile_manager.h
                src/video_core/texture_cache/types.h
+               src/video_core/page_manager.cpp
+               src/video_core/page_manager.h
                src/video_core/renderdoc.cpp
                src/video_core/renderdoc.h
 )

src/common/unique_function.h

@@ -0,0 +1,61 @@
+// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <memory>
+#include <utility>
+
+namespace Common {
+
+/// General purpose function wrapper similar to std::function.
+/// Unlike std::function, the captured values don't have to be copyable.
+/// This class can be moved but not copied.
+template <typename ResultType, typename... Args>
+class UniqueFunction {
+    class CallableBase {
+    public:
+        virtual ~CallableBase() = default;
+        virtual ResultType operator()(Args&&...) = 0;
+    };
+
+    template <typename Functor>
+    class Callable final : public CallableBase {
+    public:
+        Callable(Functor&& functor_) : functor{std::move(functor_)} {}
+        ~Callable() override = default;
+
+        ResultType operator()(Args&&... args) override {
+            return functor(std::forward<Args>(args)...);
+        }
+
+    private:
+        Functor functor;
+    };
+
+public:
+    UniqueFunction() = default;
+
+    template <typename Functor>
+    UniqueFunction(Functor&& functor)
+        : callable{std::make_unique<Callable<Functor>>(std::move(functor))} {}
+
+    UniqueFunction& operator=(UniqueFunction&& rhs) noexcept = default;
+    UniqueFunction(UniqueFunction&& rhs) noexcept = default;
+
+    UniqueFunction& operator=(const UniqueFunction&) = delete;
+    UniqueFunction(const UniqueFunction&) = delete;
+
+    ResultType operator()(Args&&... args) const {
+        return (*callable)(std::forward<Args>(args)...);
+    }
+
+    explicit operator bool() const noexcept {
+        return static_cast<bool>(callable);
+    }
+
+private:
+    std::unique_ptr<CallableBase> callable;
+};
+
+} // namespace Common

src/core/libraries/kernel/thread_management.cpp

@@ -1061,7 +1061,16 @@ ScePthread PThreadPool::Create() {
         }
     }
 
+#ifdef _WIN64
     auto* ret = new PthreadInternal{};
+#else
+    // TODO: Linux specific hack
+    static u8* hint_address = reinterpret_cast<u8*>(0x7FFFFC000ULL);
+    auto* ret = reinterpret_cast<PthreadInternal*>(
+        mmap(hint_address, sizeof(PthreadInternal), PROT_READ | PROT_WRITE,
+             MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0));
+    hint_address += Common::AlignUp(sizeof(PthreadInternal), 4_KB);
+#endif
     ret->is_free = false;
     ret->is_detached = false;
     ret->is_almost_done = false;

src/core/memory.cpp

@@ -7,7 +7,7 @@
 #include "core/libraries/error_codes.h"
 #include "core/libraries/kernel/memory_management.h"
 #include "core/memory.h"
-#include "video_core/renderer_vulkan/vk_instance.h"
+#include "video_core/renderer_vulkan/vk_rasterizer.h"
 
 namespace Core {
 
@@ -172,7 +172,7 @@ int MemoryManager::MapMemory(void** out_addr, VAddr virtual_addr, size_t size, M
 
     if (type == VMAType::Direct) {
         new_vma.phys_base = phys_addr;
-        MapVulkanMemory(mapped_addr, size);
+        rasterizer->MapMemory(mapped_addr, size);
     }
     if (type == VMAType::Flexible) {
         flexible_usage += size;
@@ -222,7 +222,7 @@ void MemoryManager::UnmapMemory(VAddr virtual_addr, size_t size) {
     const auto type = it->second.type;
     const bool has_backing = type == VMAType::Direct || type == VMAType::File;
     if (type == VMAType::Direct) {
-        UnmapVulkanMemory(virtual_addr, size);
+        rasterizer->UnmapMemory(virtual_addr, size);
     }
     if (type == VMAType::Flexible) {
         flexible_usage -= size;
@@ -263,7 +263,7 @@ int MemoryManager::QueryProtection(VAddr addr, void** start, void** end, u32* pr
 }
 
 int MemoryManager::VirtualQuery(VAddr addr, int flags,
-                                Libraries::Kernel::OrbisVirtualQueryInfo* info) {
+                                ::Libraries::Kernel::OrbisVirtualQueryInfo* info) {
     std::scoped_lock lk{mutex};
 
     auto it = FindVMA(addr);
@@ -293,7 +293,7 @@ int MemoryManager::VirtualQuery(VAddr addr, int flags,
 }
 
 int MemoryManager::DirectMemoryQuery(PAddr addr, bool find_next,
-                                     Libraries::Kernel::OrbisQueryInfo* out_info) {
+                                     ::Libraries::Kernel::OrbisQueryInfo* out_info) {
     std::scoped_lock lk{mutex};
 
     auto dmem_area = FindDmemArea(addr);
@@ -333,13 +333,6 @@ int MemoryManager::DirectQueryAvailable(PAddr search_start, PAddr search_end, si
     return ORBIS_OK;
 }
 
-std::pair<vk::Buffer, size_t> MemoryManager::GetVulkanBuffer(VAddr addr) {
-    auto it = mapped_memories.upper_bound(addr);
-    it = std::prev(it);
-    ASSERT(it != mapped_memories.end() && it->first <= addr);
-    return std::make_pair(*it->second.buffer, addr - it->first);
-}
-
 void MemoryManager::NameVirtualRange(VAddr virtual_addr, size_t size, std::string_view name) {
     auto it = FindVMA(virtual_addr);
 
@@ -455,85 +448,6 @@ MemoryManager::DMemHandle MemoryManager::Split(DMemHandle dmem_handle, size_t of
     return dmem_map.emplace_hint(std::next(dmem_handle), new_area.base, new_area);
 };
 
-void MemoryManager::MapVulkanMemory(VAddr addr, size_t size) {
-    return;
-    const vk::Device device = instance->GetDevice();
-    const auto memory_props = instance->GetPhysicalDevice().getMemoryProperties();
-    void* host_pointer = reinterpret_cast<void*>(addr);
-    const auto host_mem_props = device.getMemoryHostPointerPropertiesEXT(
-        vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT, host_pointer);
-    ASSERT(host_mem_props.memoryTypeBits != 0);
-
-    int mapped_memory_type = -1;
-    auto find_mem_type_with_flag = [&](const vk::MemoryPropertyFlags flags) {
-        u32 host_mem_types = host_mem_props.memoryTypeBits;
-        while (host_mem_types != 0) {
-            // Try to find a cached memory type
-            mapped_memory_type = std::countr_zero(host_mem_types);
-            host_mem_types -= (1 << mapped_memory_type);
-
-            if ((memory_props.memoryTypes[mapped_memory_type].propertyFlags & flags) == flags) {
-                return;
-            }
-        }
-
-        mapped_memory_type = -1;
-    };
-
-    // First try to find a memory that is both coherent and cached
-    find_mem_type_with_flag(vk::MemoryPropertyFlagBits::eHostCoherent |
-                            vk::MemoryPropertyFlagBits::eHostCached);
-    if (mapped_memory_type == -1)
-        // Then only coherent (lower performance)
-        find_mem_type_with_flag(vk::MemoryPropertyFlagBits::eHostCoherent);
-
-    if (mapped_memory_type == -1) {
-        LOG_CRITICAL(Render_Vulkan, "No coherent memory available for memory mapping");
-        mapped_memory_type = std::countr_zero(host_mem_props.memoryTypeBits);
-    }
-
-    const vk::StructureChain alloc_info = {
-        vk::MemoryAllocateInfo{
-            .allocationSize = size,
-            .memoryTypeIndex = static_cast<uint32_t>(mapped_memory_type),
-        },
-        vk::ImportMemoryHostPointerInfoEXT{
-            .handleType = vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT,
-            .pHostPointer = host_pointer,
-        },
-    };
-
-    const auto [it, new_memory] = mapped_memories.try_emplace(addr);
-    ASSERT_MSG(new_memory, "Attempting to remap already mapped vulkan memory");
-
-    auto& memory = it->second;
-    memory.backing = device.allocateMemoryUnique(alloc_info.get());
-
-    constexpr vk::BufferUsageFlags MapFlags =
-        vk::BufferUsageFlagBits::eIndexBuffer | vk::BufferUsageFlagBits::eVertexBuffer |
-        vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst |
-        vk::BufferUsageFlagBits::eUniformBuffer | vk::BufferUsageFlagBits::eStorageBuffer;
-
-    const vk::StructureChain buffer_info = {
-        vk::BufferCreateInfo{
-            .size = size,
-            .usage = MapFlags,
-            .sharingMode = vk::SharingMode::eExclusive,
-        },
-        vk::ExternalMemoryBufferCreateInfoKHR{
-            .handleTypes = vk::ExternalMemoryHandleTypeFlagBits::eHostAllocationEXT,
-        }};
-    memory.buffer = device.createBufferUnique(buffer_info.get());
-    device.bindBufferMemory(*memory.buffer, *memory.backing, 0);
-}
-
-void MemoryManager::UnmapVulkanMemory(VAddr addr, size_t size) {
-    return;
-    const auto it = mapped_memories.find(addr);
-    ASSERT(it != mapped_memories.end() && it->second.buffer_size == size);
-    mapped_memories.erase(it);
-}
-
 int MemoryManager::GetDirectMemoryType(PAddr addr, int* directMemoryTypeOut,
                                        void** directMemoryStartOut, void** directMemoryEndOut) {
     std::scoped_lock lk{mutex};

src/core/memory.h

@@ -3,20 +3,17 @@
 
 #pragma once
 
-#include <functional>
+#include <map>
 #include <mutex>
 #include <string_view>
-#include <vector>
-#include <boost/icl/split_interval_map.hpp>
 #include "common/enum.h"
 #include "common/singleton.h"
 #include "common/types.h"
 #include "core/address_space.h"
 #include "core/libraries/kernel/memory_management.h"
-#include "video_core/renderer_vulkan/vk_common.h"
 
 namespace Vulkan {
-class Instance;
+class Rasterizer;
 }
 
 namespace Libraries::Kernel {
@@ -128,8 +125,8 @@ public:
     explicit MemoryManager();
     ~MemoryManager();
 
-    void SetInstance(const Vulkan::Instance* instance_) {
+    void SetRasterizer(Vulkan::Rasterizer* rasterizer_) {
-        instance = instance_;
+        rasterizer = rasterizer_;
     }
 
     void SetTotalFlexibleSize(u64 size) {
@@ -140,9 +137,7 @@ public:
         return total_flexible_size - flexible_usage;
     }
 
-    /// Returns the offset of the mapped virtual system managed memory base from where it usually
+    VAddr SystemReservedVirtualBase() noexcept {
-    /// would be mapped.
-    [[nodiscard]] VAddr SystemReservedVirtualBase() noexcept {
         return impl.SystemReservedVirtualBase();
     }
 
@@ -172,8 +167,6 @@ public:
     int DirectQueryAvailable(PAddr search_start, PAddr search_end, size_t alignment,
                              PAddr* phys_addr_out, size_t* size_out);
 
-    std::pair<vk::Buffer, size_t> GetVulkanBuffer(VAddr addr);
-
     int GetDirectMemoryType(PAddr addr, int* directMemoryTypeOut, void** directMemoryStartOut,
                             void** directMemoryEndOut);
 
@@ -218,10 +211,6 @@ private:
 
     DMemHandle Split(DMemHandle dmem_handle, size_t offset_in_area);
 
-    void MapVulkanMemory(VAddr addr, size_t size);
-
-    void UnmapVulkanMemory(VAddr addr, size_t size);
-
 private:
     AddressSpace impl;
     DMemMap dmem_map;
@@ -229,14 +218,7 @@ private:
     std::recursive_mutex mutex;
     size_t total_flexible_size = 448_MB;
     size_t flexible_usage{};
-
+    Vulkan::Rasterizer* rasterizer{};
-    struct MappedMemory {
-        vk::UniqueBuffer buffer;
-        vk::UniqueDeviceMemory backing;
-        size_t buffer_size;
-    };
-    std::map<VAddr, MappedMemory> mapped_memories;
-    const Vulkan::Instance* instance{};
 };
 
 using Memory = Common::Singleton<MemoryManager>;

src/core/module.cpp

@@ -88,6 +88,7 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
                       aligned_base_size + TrampolineSize, MemoryProt::CpuReadWrite,
                       MemoryMapFlags::Fixed, VMAType::Code, name, true);
     LoadOffset += CODE_BASE_INCR * (1 + aligned_base_size / CODE_BASE_INCR);
+    LOG_INFO(Core_Linker, "Loading module {} to {}", name, fmt::ptr(*out_addr));
 
     // Initialize trampoline generator.
     void* trampoline_addr = std::bit_cast<void*>(base_virtual_addr + aligned_base_size);

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -21,6 +21,7 @@ Id VsOutputAttrPointer(EmitContext& ctx, VsOutput output) {
     case VsOutput::ClipDist7: {
         const u32 index = u32(output) - u32(VsOutput::ClipDist0);
         const Id clip_num{ctx.ConstU32(index)};
+        ASSERT_MSG(Sirit::ValidId(ctx.clip_distances), "Clip distance used but not defined");
         return ctx.OpAccessChain(ctx.output_f32, ctx.clip_distances, clip_num);
     }
     case VsOutput::CullDist0:
@@ -33,6 +34,7 @@ Id VsOutputAttrPointer(EmitContext& ctx, VsOutput output) {
     case VsOutput::CullDist7: {
         const u32 index = u32(output) - u32(VsOutput::CullDist0);
         const Id cull_num{ctx.ConstU32(index)};
+        ASSERT_MSG(Sirit::ValidId(ctx.cull_distances), "Cull distance used but not defined");
         return ctx.OpAccessChain(ctx.output_f32, ctx.cull_distances, cull_num);
     }
     default:
@@ -125,7 +127,12 @@ Id EmitReadConst(EmitContext& ctx) {
 }
 
 Id EmitReadConstBuffer(EmitContext& ctx, u32 handle, Id index) {
-    const auto& buffer = ctx.buffers[handle];
+    auto& buffer = ctx.buffers[handle];
+    if (!Sirit::ValidId(buffer.offset)) {
+        buffer.offset = ctx.GetBufferOffset(buffer.global_binding);
+    }
+    const Id offset_dwords{ctx.OpShiftRightLogical(ctx.U32[1], buffer.offset, ctx.ConstU32(2U))};
+    index = ctx.OpIAdd(ctx.U32[1], index, offset_dwords);
     const Id ptr{ctx.OpAccessChain(buffer.pointer_type, buffer.id, ctx.u32_zero_value, index)};
     return ctx.OpLoad(buffer.data_types->Get(1), ptr);
 }
@@ -137,7 +144,7 @@ Id EmitReadConstBufferU32(EmitContext& ctx, u32 handle, Id index) {
 Id EmitReadStepRate(EmitContext& ctx, int rate_idx) {
     return ctx.OpLoad(
         ctx.U32[1], ctx.OpAccessChain(ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1]),
-                                      ctx.instance_step_rates,
+                                      ctx.push_data_block,
                                       rate_idx == 0 ? ctx.u32_zero_value : ctx.u32_one_value));
 }
 
@@ -221,7 +228,11 @@ Id EmitLoadBufferU32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id address) {
 
 template <u32 N>
 static Id EmitLoadBufferF32xN(EmitContext& ctx, u32 handle, Id address) {
-    const auto& buffer = ctx.buffers[handle];
+    auto& buffer = ctx.buffers[handle];
+    if (!Sirit::ValidId(buffer.offset)) {
+        buffer.offset = ctx.GetBufferOffset(buffer.global_binding);
+    }
+    address = ctx.OpIAdd(ctx.U32[1], address, buffer.offset);
     const Id index = ctx.OpShiftRightLogical(ctx.U32[1], address, ctx.ConstU32(2u));
     if constexpr (N == 1) {
         const Id ptr{ctx.OpAccessChain(buffer.pointer_type, buffer.id, ctx.u32_zero_value, index)};
@@ -314,7 +325,7 @@ static Id ComponentOffset(EmitContext& ctx, Id address, u32 stride, u32 bit_offs
 }
 
 static Id GetBufferFormatValue(EmitContext& ctx, u32 handle, Id address, u32 comp) {
-    const auto& buffer = ctx.buffers[handle];
+    auto& buffer = ctx.buffers[handle];
     const auto format = buffer.buffer.GetDataFmt();
     switch (format) {
     case AmdGpu::DataFormat::FormatInvalid:
@@ -399,6 +410,11 @@ static Id GetBufferFormatValue(EmitContext& ctx, u32 handle, Id address, u32 com
 
 template <u32 N>
 static Id EmitLoadBufferFormatF32xN(EmitContext& ctx, IR::Inst* inst, u32 handle, Id address) {
+    auto& buffer = ctx.buffers[handle];
+    if (!Sirit::ValidId(buffer.offset)) {
+        buffer.offset = ctx.GetBufferOffset(buffer.global_binding);
+    }
+    address = ctx.OpIAdd(ctx.U32[1], address, buffer.offset);
     if constexpr (N == 1) {
         return GetBufferFormatValue(ctx, handle, address, 0);
     } else {
@@ -428,7 +444,11 @@ Id EmitLoadBufferFormatF32x4(EmitContext& ctx, IR::Inst* inst, u32 handle, Id ad
 
 template <u32 N>
 static void EmitStoreBufferF32xN(EmitContext& ctx, u32 handle, Id address, Id value) {
-    const auto& buffer = ctx.buffers[handle];
+    auto& buffer = ctx.buffers[handle];
+    if (!Sirit::ValidId(buffer.offset)) {
+        buffer.offset = ctx.GetBufferOffset(buffer.global_binding);
+    }
+    address = ctx.OpIAdd(ctx.U32[1], address, buffer.offset);
     const Id index = ctx.OpShiftRightLogical(ctx.U32[1], address, ctx.ConstU32(2u));
     if constexpr (N == 1) {
         const Id ptr{ctx.OpAccessChain(buffer.pointer_type, buffer.id, ctx.u32_zero_value, index)};

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -46,9 +46,9 @@ EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& bin
       stage{program.info.stage}, binding{binding_} {
     AddCapability(spv::Capability::Shader);
     DefineArithmeticTypes();
-    DefineInterfaces(program);
+    DefineInterfaces();
-    DefineBuffers(info);
+    DefineBuffers();
-    DefineImagesAndSamplers(info);
+    DefineImagesAndSamplers();
     DefineSharedMemory();
 }
 
@@ -117,9 +117,10 @@ void EmitContext::DefineArithmeticTypes() {
     full_result_u32x2 = Name(TypeStruct(U32[1], U32[1]), "full_result_u32x2");
 }
 
-void EmitContext::DefineInterfaces(const IR::Program& program) {
+void EmitContext::DefineInterfaces() {
-    DefineInputs(program.info);
+    DefinePushDataBlock();
-    DefineOutputs(program.info);
+    DefineInputs();
+    DefineOutputs();
 }
 
 Id GetAttributeType(EmitContext& ctx, AmdGpu::NumberFormat fmt) {
@@ -164,6 +165,16 @@ EmitContext::SpirvAttribute EmitContext::GetAttributeInfo(AmdGpu::NumberFormat f
     throw InvalidArgument("Invalid attribute type {}", fmt);
 }
 
+Id EmitContext::GetBufferOffset(u32 binding) {
+    const u32 half = Shader::PushData::BufOffsetIndex + (binding >> 4);
+    const u32 comp = (binding & 0xf) >> 2;
+    const u32 offset = (binding & 0x3) << 3;
+    const Id ptr{OpAccessChain(TypePointer(spv::StorageClass::PushConstant, U32[1]),
+                               push_data_block, ConstU32(half), ConstU32(comp))};
+    const Id value{OpLoad(U32[1], ptr)};
+    return OpBitFieldUExtract(U32[1], value, ConstU32(offset), ConstU32(8U));
+}
+
 Id MakeDefaultValue(EmitContext& ctx, u32 default_value) {
     switch (default_value) {
     case 0:
@@ -179,24 +190,13 @@ Id MakeDefaultValue(EmitContext& ctx, u32 default_value) {
     }
 }
 
-void EmitContext::DefineInputs(const Info& info) {
+void EmitContext::DefineInputs() {
     switch (stage) {
     case Stage::Vertex: {
         vertex_index = DefineVariable(U32[1], spv::BuiltIn::VertexIndex, spv::StorageClass::Input);
         base_vertex = DefineVariable(U32[1], spv::BuiltIn::BaseVertex, spv::StorageClass::Input);
         instance_id = DefineVariable(U32[1], spv::BuiltIn::InstanceIndex, spv::StorageClass::Input);
 
-        // Create push constants block for instance steps rates
-        const Id struct_type{Name(TypeStruct(U32[1], U32[1]), "instance_step_rates")};
-        Decorate(struct_type, spv::Decoration::Block);
-        MemberName(struct_type, 0, "sr0");
-        MemberName(struct_type, 1, "sr1");
-        MemberDecorate(struct_type, 0, spv::Decoration::Offset, 0U);
-        MemberDecorate(struct_type, 1, spv::Decoration::Offset, 4U);
-        instance_step_rates = DefineVar(struct_type, spv::StorageClass::PushConstant);
-        Name(instance_step_rates, "step_rates");
-        interfaces.push_back(instance_step_rates);
-
         for (const auto& input : info.vs_inputs) {
             const Id type{GetAttributeType(*this, input.fmt)};
             if (input.instance_step_rate == Info::VsInput::InstanceIdType::OverStepRate0 ||
@@ -260,19 +260,20 @@ void EmitContext::DefineInputs(const Info& info) {
     }
 }
 
-void EmitContext::DefineOutputs(const Info& info) {
+void EmitContext::DefineOutputs() {
     switch (stage) {
     case Stage::Vertex: {
         output_position = DefineVariable(F32[4], spv::BuiltIn::Position, spv::StorageClass::Output);
-        const std::array<Id, 8> zero{f32_zero_value, f32_zero_value, f32_zero_value,
+        const bool has_extra_pos_stores = info.stores.Get(IR::Attribute::Position1) ||
-                                     f32_zero_value, f32_zero_value, f32_zero_value,
+                                          info.stores.Get(IR::Attribute::Position2) ||
-                                     f32_zero_value, f32_zero_value};
+                                          info.stores.Get(IR::Attribute::Position3);
+        if (has_extra_pos_stores) {
             const Id type{TypeArray(F32[1], ConstU32(8U))};
-        const Id initializer{ConstantComposite(type, zero)};
+            clip_distances =
-        clip_distances = DefineVariable(type, spv::BuiltIn::ClipDistance, spv::StorageClass::Output,
+                DefineVariable(type, spv::BuiltIn::ClipDistance, spv::StorageClass::Output);
-                                        initializer);
+            cull_distances =
-        cull_distances = DefineVariable(type, spv::BuiltIn::CullDistance, spv::StorageClass::Output,
+                DefineVariable(type, spv::BuiltIn::CullDistance, spv::StorageClass::Output);
-                                        initializer);
+        }
         for (u32 i = 0; i < IR::NumParams; i++) {
             const IR::Attribute param{IR::Attribute::Param0 + i};
             if (!info.stores.GetAny(param)) {
@@ -304,7 +305,24 @@ void EmitContext::DefineOutputs(const Info& info) {
     }
 }
 
-void EmitContext::DefineBuffers(const Info& info) {
+void EmitContext::DefinePushDataBlock() {
+    // Create push constants block for instance steps rates
+    const Id struct_type{Name(TypeStruct(U32[1], U32[1], U32[4], U32[4]), "AuxData")};
+    Decorate(struct_type, spv::Decoration::Block);
+    MemberName(struct_type, 0, "sr0");
+    MemberName(struct_type, 1, "sr1");
+    MemberName(struct_type, 2, "buf_offsets0");
+    MemberName(struct_type, 3, "buf_offsets1");
+    MemberDecorate(struct_type, 0, spv::Decoration::Offset, 0U);
+    MemberDecorate(struct_type, 1, spv::Decoration::Offset, 4U);
+    MemberDecorate(struct_type, 2, spv::Decoration::Offset, 8U);
+    MemberDecorate(struct_type, 3, spv::Decoration::Offset, 24U);
+    push_data_block = DefineVar(struct_type, spv::StorageClass::PushConstant);
+    Name(push_data_block, "push_data");
+    interfaces.push_back(push_data_block);
+}
+
+void EmitContext::DefineBuffers() {
     boost::container::small_vector<Id, 8> type_ids;
     for (u32 i = 0; const auto& buffer : info.buffers) {
         const auto* data_types = True(buffer.used_types & IR::Type::F32) ? &F32 : &U32;
@@ -322,8 +340,8 @@ void EmitContext::DefineBuffers(const Info& info) {
             Decorate(struct_type, spv::Decoration::Block);
             MemberName(struct_type, 0, "data");
             MemberDecorate(struct_type, 0, spv::Decoration::Offset, 0U);
-        }
             type_ids.push_back(record_array_type);
+        }
 
         const auto storage_class =
             buffer.is_storage ? spv::StorageClass::StorageBuffer : spv::StorageClass::Uniform;
@@ -334,9 +352,9 @@ void EmitContext::DefineBuffers(const Info& info) {
         Decorate(id, spv::Decoration::DescriptorSet, 0U);
         Name(id, fmt::format("{}_{}", buffer.is_storage ? "ssbo" : "cbuf", buffer.sgpr_base));
 
-        binding++;
         buffers.push_back({
             .id = id,
+            .global_binding = binding++,
             .data_types = data_types,
             .pointer_type = pointer_type,
             .buffer = buffer.GetVsharp(info),
@@ -430,7 +448,7 @@ Id ImageType(EmitContext& ctx, const ImageResource& desc, Id sampled_type) {
     throw InvalidArgument("Invalid texture type {}", desc.type);
 }
 
-void EmitContext::DefineImagesAndSamplers(const Info& info) {
+void EmitContext::DefineImagesAndSamplers() {
     for (const auto& image_desc : info.images) {
         const VectorIds* data_types = [&] {
             switch (image_desc.nfmt) {

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -40,6 +40,7 @@ public:
     ~EmitContext();
 
     Id Def(const IR::Value& value);
+    Id GetBufferOffset(u32 binding);
 
     [[nodiscard]] Id DefineInput(Id type, u32 location) {
         const Id input_id{DefineVar(type, spv::StorageClass::Input)};
@@ -168,7 +169,7 @@ public:
     Id output_position{};
     Id vertex_index{};
     Id instance_id{};
-    Id instance_step_rates{};
+    Id push_data_block{};
     Id base_vertex{};
     Id frag_coord{};
     Id front_facing{};
@@ -201,14 +202,16 @@ public:
 
     struct BufferDefinition {
         Id id;
+        Id offset;
+        u32 global_binding;
         const VectorIds* data_types;
         Id pointer_type;
         AmdGpu::Buffer buffer;
     };
 
     u32& binding;
-    boost::container::small_vector<BufferDefinition, 4> buffers;
+    boost::container::small_vector<BufferDefinition, 16> buffers;
-    boost::container::small_vector<TextureDefinition, 4> images;
+    boost::container::small_vector<TextureDefinition, 8> images;
     boost::container::small_vector<Id, 4> samplers;
 
     Id sampler_type{};
@@ -227,11 +230,12 @@ public:
 
 private:
     void DefineArithmeticTypes();
-    void DefineInterfaces(const IR::Program& program);
+    void DefineInterfaces();
-    void DefineInputs(const Info& info);
+    void DefineInputs();
-    void DefineOutputs(const Info& info);
+    void DefineOutputs();
-    void DefineBuffers(const Info& info);
+    void DefinePushDataBlock();
-    void DefineImagesAndSamplers(const Info& info);
+    void DefineBuffers();
+    void DefineImagesAndSamplers();
     void DefineSharedMemory();
 
     SpirvAttribute GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id);

src/shader_recompiler/frontend/translate/translate.cpp

@@ -447,6 +447,7 @@ void Translator::EmitFetch(const GcnInst& inst) {
                 .is_instance_data = true,
             });
             instance_buf_handle = s32(info.buffers.size() - 1);
+            info.uses_step_rates = true;
         }
 
         const u32 num_components = AmdGpu::NumComponents(buffer.GetDataFmt());

src/shader_recompiler/frontend/translate/vector_memory.cpp

@@ -338,6 +338,11 @@ void Translator::BUFFER_LOAD_FORMAT(u32 num_dwords, bool is_typed, bool is_forma
     if (is_typed) {
         info.dmft.Assign(static_cast<AmdGpu::DataFormat>(mtbuf.dfmt));
         info.nfmt.Assign(static_cast<AmdGpu::NumberFormat>(mtbuf.nfmt));
+        ASSERT(info.nfmt == AmdGpu::NumberFormat::Float &&
+               (info.dmft == AmdGpu::DataFormat::Format32_32_32_32 ||
+                info.dmft == AmdGpu::DataFormat::Format32_32_32 ||
+                info.dmft == AmdGpu::DataFormat::Format32_32 ||
+                info.dmft == AmdGpu::DataFormat::Format32));
     }
 
     const IR::Value handle =

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -173,10 +173,9 @@ bool IsImageStorageInstruction(const IR::Inst& inst) {
 
 class Descriptors {
 public:
-    explicit Descriptors(BufferResourceList& buffer_resources_, ImageResourceList& image_resources_,
+    explicit Descriptors(Info& info_)
-                         SamplerResourceList& sampler_resources_)
+        : info{info_}, buffer_resources{info_.buffers}, image_resources{info_.images},
-        : buffer_resources{buffer_resources_}, image_resources{image_resources_},
+          sampler_resources{info_.samplers} {}
-          sampler_resources{sampler_resources_} {}
 
     u32 Add(const BufferResource& desc) {
         const u32 index{Add(buffer_resources, desc, [&desc](const auto& existing) {
@@ -188,6 +187,7 @@ public:
         ASSERT(buffer.length == desc.length);
         buffer.is_storage |= desc.is_storage;
         buffer.used_types |= desc.used_types;
+        buffer.is_written |= desc.is_written;
         return index;
     }
 
@@ -201,9 +201,16 @@ public:
     }
 
     u32 Add(const SamplerResource& desc) {
-        const u32 index{Add(sampler_resources, desc, [&desc](const auto& existing) {
+        const u32 index{Add(sampler_resources, desc, [this, &desc](const auto& existing) {
-            return desc.sgpr_base == existing.sgpr_base &&
+            if (desc.sgpr_base == existing.sgpr_base &&
-                   desc.dword_offset == existing.dword_offset;
+                desc.dword_offset == existing.dword_offset) {
+                return true;
+            }
+            // Samplers with different bindings might still be the same.
+            const auto old_sharp =
+                info.ReadUd<AmdGpu::Sampler>(existing.sgpr_base, existing.dword_offset);
+            const auto new_sharp = info.ReadUd<AmdGpu::Sampler>(desc.sgpr_base, desc.dword_offset);
+            return old_sharp == new_sharp;
         })};
         return index;
     }
@@ -219,6 +226,7 @@ private:
         return static_cast<u32>(descriptors.size()) - 1;
     }
 
+    const Info& info;
     BufferResourceList& buffer_resources;
     ImageResourceList& image_resources;
     SamplerResourceList& sampler_resources;
@@ -328,16 +336,6 @@ static bool IsLoadBufferFormat(const IR::Inst& inst) {
     }
 }
 
-static bool IsReadConstBuffer(const IR::Inst& inst) {
-    switch (inst.GetOpcode()) {
-    case IR::Opcode::ReadConstBuffer:
-    case IR::Opcode::ReadConstBufferU32:
-        return true;
-    default:
-        return false;
-    }
-}
-
 static u32 BufferLength(const AmdGpu::Buffer& buffer) {
     const auto stride = buffer.GetStride();
     if (stride < sizeof(f32)) {
@@ -401,43 +399,42 @@ void PatchBufferInstruction(IR::Block& block, IR::Inst& inst, Info& info,
         IR::Inst* handle = inst.Arg(0).InstRecursive();
         IR::Inst* producer = handle->Arg(0).InstRecursive();
         const auto sharp = TrackSharp(producer);
+        const bool is_store = IsBufferStore(inst);
         buffer = info.ReadUd<AmdGpu::Buffer>(sharp.sgpr_base, sharp.dword_offset);
         binding = descriptors.Add(BufferResource{
             .sgpr_base = sharp.sgpr_base,
             .dword_offset = sharp.dword_offset,
             .length = BufferLength(buffer),
             .used_types = BufferDataType(inst, buffer.GetNumberFmt()),
-            .is_storage = IsBufferStore(inst) || buffer.GetSize() > MaxUboSize,
+            .is_storage = true || is_store || buffer.GetSize() > MaxUboSize,
+            .is_written = is_store,
         });
     }
 
+    // Update buffer descriptor format.
     const auto inst_info = inst.Flags<IR::BufferInstInfo>();
-    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+    auto& buffer_desc = info.buffers[binding];
-    // Replace handle with binding index in buffer resource list.
-    inst.SetArg(0, ir.Imm32(binding));
-    ASSERT(!buffer.swizzle_enable && !buffer.add_tid_enable);
     if (inst_info.is_typed) {
-        ASSERT(inst_info.nfmt == AmdGpu::NumberFormat::Float &&
+        buffer_desc.dfmt = inst_info.dmft;
-               (inst_info.dmft == AmdGpu::DataFormat::Format32_32_32_32 ||
+        buffer_desc.nfmt = inst_info.nfmt;
-                inst_info.dmft == AmdGpu::DataFormat::Format32_32_32 ||
+    } else {
-                inst_info.dmft == AmdGpu::DataFormat::Format32_32 ||
+        buffer_desc.dfmt = buffer.GetDataFmt();
-                inst_info.dmft == AmdGpu::DataFormat::Format32));
+        buffer_desc.nfmt = buffer.GetNumberFmt();
     }
 
-    if (IsReadConstBuffer(inst)) {
+    // Replace handle with binding index in buffer resource list.
+    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+    inst.SetArg(0, ir.Imm32(binding));
+    ASSERT(!buffer.swizzle_enable && !buffer.add_tid_enable);
+
+    // Address of constant buffer reads can be calculated at IR emittion time.
+    if (inst.GetOpcode() == IR::Opcode::ReadConstBuffer ||
+        inst.GetOpcode() == IR::Opcode::ReadConstBufferU32) {
         return;
     }
 
-    if (IsLoadBufferFormat(inst)) {
+    // Compute address of the buffer using the stride.
-        if (UseFP16(buffer.GetDataFmt(), buffer.GetNumberFmt())) {
+    // Todo: What if buffer is rebound with different stride?
-            info.uses_fp16 = true;
-        }
-    } else {
-        const u32 stride = buffer.GetStride();
-        ASSERT_MSG(stride >= 4, "non-formatting load_buffer_* is not implemented for stride {}",
-                   stride);
-    }
-
     IR::U32 address = ir.Imm32(inst_info.inst_offset.Value());
     if (inst_info.index_enable) {
         const IR::U32 index = inst_info.offset_enable ? IR::U32{ir.CompositeExtract(inst.Arg(1), 0)}
@@ -587,39 +584,9 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
 }
 
 void ResourceTrackingPass(IR::Program& program) {
-    // When loading data from untyped buffer we don't have if it is float or integer.
-    // Most of the time it is float so that is the default. This pass detects float buffer loads
-    // combined with bitcasts and patches them to be integer loads.
-    for (IR::Block* const block : program.post_order_blocks) {
-        break;
-        for (IR::Inst& inst : block->Instructions()) {
-            if (inst.GetOpcode() != IR::Opcode::BitCastU32F32) {
-                continue;
-            }
-            // Replace the bitcast with a typed buffer read
-            IR::Inst* const arg_inst{inst.Arg(0).TryInstRecursive()};
-            if (!arg_inst) {
-                continue;
-            }
-            const auto replace{[&](IR::Opcode new_opcode) {
-                inst.ReplaceOpcode(new_opcode);
-                inst.SetArg(0, arg_inst->Arg(0));
-                inst.SetArg(1, arg_inst->Arg(1));
-                inst.SetFlags(arg_inst->Flags<u32>());
-                arg_inst->Invalidate();
-            }};
-            if (arg_inst->GetOpcode() == IR::Opcode::ReadConstBuffer) {
-                replace(IR::Opcode::ReadConstBufferU32);
-            }
-            if (arg_inst->GetOpcode() == IR::Opcode::LoadBufferF32) {
-                replace(IR::Opcode::LoadBufferU32);
-            }
-        }
-    }
-
     // Iterate resource instructions and patch them after finding the sharp.
     auto& info = program.info;
-    Descriptors descriptors{info.buffers, info.images, info.samplers};
+    Descriptors descriptors{info};
     for (IR::Block* const block : program.blocks) {
         for (IR::Inst& inst : block->Instructions()) {
             if (IsBufferInstruction(inst)) {

src/shader_recompiler/recompiler.cpp

@@ -61,7 +61,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     Shader::Optimization::IdentityRemovalPass(program.blocks);
     Shader::Optimization::DeadCodeEliminationPass(program);
     Shader::Optimization::CollectShaderInfoPass(program);
-    LOG_INFO(Render_Vulkan, "{}", Shader::IR::DumpProgram(program));
+    LOG_DEBUG(Render_Vulkan, "{}", Shader::IR::DumpProgram(program));
 
     return program;
 }

src/shader_recompiler/runtime_info.h

@@ -77,8 +77,11 @@ struct BufferResource {
     u32 length;
     IR::Type used_types;
     AmdGpu::Buffer inline_cbuf;
-    bool is_storage{false};
+    AmdGpu::DataFormat dfmt;
-    bool is_instance_data{false};
+    AmdGpu::NumberFormat nfmt;
+    bool is_storage{};
+    bool is_instance_data{};
+    bool is_written{};
 
     constexpr AmdGpu::Buffer GetVsharp(const Info& info) const noexcept;
 };
@@ -105,6 +108,19 @@ struct SamplerResource {
 };
 using SamplerResourceList = boost::container::static_vector<SamplerResource, 16>;
 
+struct PushData {
+    static constexpr size_t BufOffsetIndex = 2;
+
+    u32 step0;
+    u32 step1;
+    std::array<u8, 32> buf_offsets;
+
+    void AddOffset(u32 binding, u32 offset) {
+        ASSERT(offset < 64 && binding < 32);
+        buf_offsets[binding] = offset;
+    }
+};
+
 struct Info {
     struct VsInput {
         enum InstanceIdType : u8 {
@@ -182,6 +198,7 @@ struct Info {
     bool uses_shared_u8{};
     bool uses_shared_u16{};
     bool uses_fp16{};
+    bool uses_step_rates{};
     bool translation_failed{}; // indicates that shader has unsupported instructions
 
     template <typename T>

src/video_core/amdgpu/resource.h

@@ -363,6 +363,10 @@ struct Sampler {
         return raw0 != 0 || raw1 != 0;
     }
 
+    bool operator==(const Sampler& other) const noexcept {
+        return std::memcmp(this, &other, sizeof(Sampler)) == 0;
+    }
+
     float LodBias() const noexcept {
         return static_cast<float>(static_cast<int16_t>((lod_bias.Value() ^ 0x2000u) - 0x2000u)) /
                256.0f;

src/video_core/buffer_cache/buffer.cpp

@@ -0,0 +1,227 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "video_core/buffer_cache/buffer.h"
+#include "video_core/renderer_vulkan/liverpool_to_vk.h"
+#include "video_core/renderer_vulkan/vk_instance.h"
+#include "video_core/renderer_vulkan/vk_platform.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+
+#include <vk_mem_alloc.h>
+
+namespace VideoCore {
+
+constexpr vk::BufferUsageFlags AllFlags =
+    vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst |
+    vk::BufferUsageFlagBits::eUniformTexelBuffer | vk::BufferUsageFlagBits::eStorageTexelBuffer |
+    vk::BufferUsageFlagBits::eUniformBuffer | vk::BufferUsageFlagBits::eStorageBuffer |
+    vk::BufferUsageFlagBits::eIndexBuffer | vk::BufferUsageFlagBits::eVertexBuffer;
+
+std::string_view BufferTypeName(MemoryUsage type) {
+    switch (type) {
+    case MemoryUsage::Upload:
+        return "Upload";
+    case MemoryUsage::Download:
+        return "Download";
+    case MemoryUsage::Stream:
+        return "Stream";
+    case MemoryUsage::DeviceLocal:
+        return "DeviceLocal";
+    default:
+        return "Invalid";
+    }
+}
+
+[[nodiscard]] VkMemoryPropertyFlags MemoryUsagePreferredVmaFlags(MemoryUsage usage) {
+    return usage != MemoryUsage::DeviceLocal ? VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
+                                             : VkMemoryPropertyFlagBits{};
+}
+
+[[nodiscard]] VmaAllocationCreateFlags MemoryUsageVmaFlags(MemoryUsage usage) {
+    switch (usage) {
+    case MemoryUsage::Upload:
+    case MemoryUsage::Stream:
+        return VMA_ALLOCATION_CREATE_MAPPED_BIT |
+               VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
+    case MemoryUsage::Download:
+        return VMA_ALLOCATION_CREATE_MAPPED_BIT | VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
+    case MemoryUsage::DeviceLocal:
+        return {};
+    }
+    return {};
+}
+
+[[nodiscard]] VmaMemoryUsage MemoryUsageVma(MemoryUsage usage) {
+    switch (usage) {
+    case MemoryUsage::DeviceLocal:
+    case MemoryUsage::Stream:
+        return VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+    case MemoryUsage::Upload:
+    case MemoryUsage::Download:
+        return VMA_MEMORY_USAGE_AUTO_PREFER_HOST;
+    }
+    return VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+}
+
+UniqueBuffer::UniqueBuffer(vk::Device device_, VmaAllocator allocator_)
+    : device{device_}, allocator{allocator_} {}
+
+UniqueBuffer::~UniqueBuffer() {
+    if (buffer) {
+        vmaDestroyBuffer(allocator, buffer, allocation);
+    }
+}
+
+void UniqueBuffer::Create(const vk::BufferCreateInfo& buffer_ci, MemoryUsage usage,
+                          VmaAllocationInfo* out_alloc_info) {
+    const VmaAllocationCreateInfo alloc_ci = {
+        .flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | MemoryUsageVmaFlags(usage),
+        .usage = MemoryUsageVma(usage),
+        .requiredFlags = 0,
+        .preferredFlags = MemoryUsagePreferredVmaFlags(usage),
+        .pool = VK_NULL_HANDLE,
+        .pUserData = nullptr,
+    };
+
+    const VkBufferCreateInfo buffer_ci_unsafe = static_cast<VkBufferCreateInfo>(buffer_ci);
+    VkBuffer unsafe_buffer{};
+    VkResult result = vmaCreateBuffer(allocator, &buffer_ci_unsafe, &alloc_ci, &unsafe_buffer,
+                                      &allocation, out_alloc_info);
+    ASSERT_MSG(result == VK_SUCCESS, "Failed allocating buffer with error {}",
+               vk::to_string(vk::Result{result}));
+    buffer = vk::Buffer{unsafe_buffer};
+}
+
+Buffer::Buffer(const Vulkan::Instance& instance_, MemoryUsage usage_, VAddr cpu_addr_,
+               u64 size_bytes_)
+    : cpu_addr{cpu_addr_}, size_bytes{size_bytes_}, instance{&instance_}, usage{usage_},
+      buffer{instance->GetDevice(), instance->GetAllocator()} {
+    // Create buffer object.
+    const vk::BufferCreateInfo buffer_ci = {
+        .size = size_bytes,
+        .usage = AllFlags,
+    };
+    VmaAllocationInfo alloc_info{};
+    buffer.Create(buffer_ci, usage, &alloc_info);
+
+    if (instance->HasDebuggingToolAttached()) {
+        const auto device = instance->GetDevice();
+        Vulkan::SetObjectName(device, Handle(), "Buffer {:#x} {} KiB", cpu_addr, size_bytes / 1024);
+    }
+
+    // Map it if it is host visible.
+    VkMemoryPropertyFlags property_flags{};
+    vmaGetAllocationMemoryProperties(instance->GetAllocator(), buffer.allocation, &property_flags);
+    if (alloc_info.pMappedData) {
+        mapped_data = std::span<u8>{std::bit_cast<u8*>(alloc_info.pMappedData), size_bytes};
+    }
+    is_coherent = property_flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
+}
+
+vk::BufferView Buffer::View(u32 offset, u32 size, AmdGpu::DataFormat dfmt,
+                            AmdGpu::NumberFormat nfmt) {
+    const auto it{std::ranges::find_if(views, [offset, size, dfmt, nfmt](const BufferView& view) {
+        return offset == view.offset && size == view.size && dfmt == view.dfmt && nfmt == view.nfmt;
+    })};
+    if (it != views.end()) {
+        return it->handle;
+    }
+    views.push_back({
+        .offset = offset,
+        .size = size,
+        .dfmt = dfmt,
+        .nfmt = nfmt,
+        .handle = instance->GetDevice().createBufferView({
+            .buffer = buffer.buffer,
+            .format = Vulkan::LiverpoolToVK::SurfaceFormat(dfmt, nfmt),
+            .offset = offset,
+            .range = size,
+        }),
+    });
+    return views.back().handle;
+}
+
+constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000;
+constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000;
+
+StreamBuffer::StreamBuffer(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler_,
+                           MemoryUsage usage, u64 size_bytes)
+    : Buffer{instance, usage, 0, size_bytes}, scheduler{scheduler_} {
+    ReserveWatches(current_watches, WATCHES_INITIAL_RESERVE);
+    ReserveWatches(previous_watches, WATCHES_INITIAL_RESERVE);
+    const auto device = instance.GetDevice();
+    if (instance.HasDebuggingToolAttached()) {
+        Vulkan::SetObjectName(device, Handle(), "StreamBuffer({}): {} KiB", BufferTypeName(usage),
+                              size_bytes / 1024);
+    }
+}
+
+std::pair<u8*, u64> StreamBuffer::Map(u64 size, u64 alignment) {
+    if (!is_coherent && usage == MemoryUsage::Stream) {
+        size = Common::AlignUp(size, instance->NonCoherentAtomSize());
+    }
+
+    ASSERT(size <= this->size_bytes);
+    mapped_size = size;
+
+    if (alignment > 0) {
+        offset = Common::AlignUp(offset, alignment);
+    }
+
+    if (offset + size > this->size_bytes) {
+        // The buffer would overflow, save the amount of used watches and reset the state.
+        invalidation_mark = current_watch_cursor;
+        current_watch_cursor = 0;
+        offset = 0;
+
+        // Swap watches and reset waiting cursors.
+        std::swap(previous_watches, current_watches);
+        wait_cursor = 0;
+        wait_bound = 0;
+    }
+
+    const u64 mapped_upper_bound = offset + size;
+    WaitPendingOperations(mapped_upper_bound);
+    return std::make_pair(mapped_data.data() + offset, offset);
+}
+
+void StreamBuffer::Commit() {
+    if (!is_coherent) {
+        if (usage == MemoryUsage::Download) {
+            vmaInvalidateAllocation(instance->GetAllocator(), buffer.allocation, offset,
+                                    mapped_size);
+        } else {
+            vmaFlushAllocation(instance->GetAllocator(), buffer.allocation, offset, mapped_size);
+        }
+    }
+
+    offset += mapped_size;
+    if (current_watch_cursor + 1 >= current_watches.size()) {
+        // Ensure that there are enough watches.
+        ReserveWatches(current_watches, WATCHES_RESERVE_CHUNK);
+    }
+
+    auto& watch = current_watches[current_watch_cursor++];
+    watch.upper_bound = offset;
+    watch.tick = scheduler.CurrentTick();
+}
+
+void StreamBuffer::ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size) {
+    watches.resize(watches.size() + grow_size);
+}
+
+void StreamBuffer::WaitPendingOperations(u64 requested_upper_bound) {
+    if (!invalidation_mark) {
+        return;
+    }
+    while (requested_upper_bound > wait_bound && wait_cursor < *invalidation_mark) {
+        auto& watch = previous_watches[wait_cursor];
+        wait_bound = watch.upper_bound;
+        scheduler.Wait(watch.tick);
+        ++wait_cursor;
+    }
+}
+
+} // namespace VideoCore

src/video_core/buffer_cache/buffer.h

@@ -0,0 +1,173 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <cstddef>
+#include <utility>
+#include <vector>
+#include "common/types.h"
+#include "video_core/amdgpu/resource.h"
+#include "video_core/renderer_vulkan/vk_common.h"
+
+namespace Vulkan {
+class Instance;
+class Scheduler;
+} // namespace Vulkan
+
+VK_DEFINE_HANDLE(VmaAllocation)
+VK_DEFINE_HANDLE(VmaAllocator)
+
+struct VmaAllocationInfo;
+
+namespace VideoCore {
+
+/// Hints and requirements for the backing memory type of a commit
+enum class MemoryUsage {
+    DeviceLocal, ///< Requests device local buffer.
+    Upload,      ///< Requires a host visible memory type optimized for CPU to GPU uploads
+    Download,    ///< Requires a host visible memory type optimized for GPU to CPU readbacks
+    Stream,      ///< Requests device local host visible buffer, falling back host memory.
+};
+
+struct UniqueBuffer {
+    explicit UniqueBuffer(vk::Device device, VmaAllocator allocator);
+    ~UniqueBuffer();
+
+    UniqueBuffer(const UniqueBuffer&) = delete;
+    UniqueBuffer& operator=(const UniqueBuffer&) = delete;
+
+    UniqueBuffer(UniqueBuffer&& other)
+        : buffer{std::exchange(other.buffer, VK_NULL_HANDLE)},
+          allocator{std::exchange(other.allocator, VK_NULL_HANDLE)},
+          allocation{std::exchange(other.allocation, VK_NULL_HANDLE)} {}
+    UniqueBuffer& operator=(UniqueBuffer&& other) {
+        buffer = std::exchange(other.buffer, VK_NULL_HANDLE);
+        allocator = std::exchange(other.allocator, VK_NULL_HANDLE);
+        allocation = std::exchange(other.allocation, VK_NULL_HANDLE);
+        return *this;
+    }
+
+    void Create(const vk::BufferCreateInfo& image_ci, MemoryUsage usage,
+                VmaAllocationInfo* out_alloc_info);
+
+    operator vk::Buffer() const {
+        return buffer;
+    }
+
+    vk::Device device;
+    VmaAllocator allocator;
+    VmaAllocation allocation;
+    vk::Buffer buffer{};
+};
+
+class Buffer {
+public:
+    explicit Buffer(const Vulkan::Instance& instance, MemoryUsage usage, VAddr cpu_addr_,
+                    u64 size_bytes_);
+
+    Buffer& operator=(const Buffer&) = delete;
+    Buffer(const Buffer&) = delete;
+
+    Buffer& operator=(Buffer&&) = default;
+    Buffer(Buffer&&) = default;
+
+    vk::BufferView View(u32 offset, u32 size, AmdGpu::DataFormat dfmt, AmdGpu::NumberFormat nfmt);
+
+    /// Increases the likeliness of this being a stream buffer
+    void IncreaseStreamScore(int score) noexcept {
+        stream_score += score;
+    }
+
+    /// Returns the likeliness of this being a stream buffer
+    [[nodiscard]] int StreamScore() const noexcept {
+        return stream_score;
+    }
+
+    /// Returns true when vaddr -> vaddr+size is fully contained in the buffer
+    [[nodiscard]] bool IsInBounds(VAddr addr, u64 size) const noexcept {
+        return addr >= cpu_addr && addr + size <= cpu_addr + SizeBytes();
+    }
+
+    /// Returns the base CPU address of the buffer
+    [[nodiscard]] VAddr CpuAddr() const noexcept {
+        return cpu_addr;
+    }
+
+    /// Returns the offset relative to the given CPU address
+    [[nodiscard]] u32 Offset(VAddr other_cpu_addr) const noexcept {
+        return static_cast<u32>(other_cpu_addr - cpu_addr);
+    }
+
+    size_t SizeBytes() const {
+        return size_bytes;
+    }
+
+    vk::Buffer Handle() const noexcept {
+        return buffer;
+    }
+
+public:
+    VAddr cpu_addr = 0;
+    bool is_picked{};
+    bool is_coherent{};
+    int stream_score = 0;
+    size_t size_bytes = 0;
+    std::span<u8> mapped_data;
+    const Vulkan::Instance* instance{};
+    MemoryUsage usage;
+    UniqueBuffer buffer;
+    struct BufferView {
+        u32 offset;
+        u32 size;
+        AmdGpu::DataFormat dfmt;
+        AmdGpu::NumberFormat nfmt;
+        vk::BufferView handle;
+    };
+    std::vector<BufferView> views;
+};
+
+class StreamBuffer : public Buffer {
+public:
+    explicit StreamBuffer(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
+                          MemoryUsage usage, u64 size_bytes_);
+
+    /// Reserves a region of memory from the stream buffer.
+    std::pair<u8*, u64> Map(u64 size, u64 alignment = 0);
+
+    /// Ensures that reserved bytes of memory are available to the GPU.
+    void Commit();
+
+    /// Maps and commits a memory region with user provided data
+    u64 Copy(VAddr src, size_t size, size_t alignment = 0) {
+        const auto [data, offset] = Map(size, alignment);
+        std::memcpy(data, reinterpret_cast<const void*>(src), size);
+        Commit();
+        return offset;
+    }
+
+private:
+    struct Watch {
+        u64 tick{};
+        u64 upper_bound{};
+    };
+
+    /// Increases the amount of watches available.
+    void ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size);
+
+    /// Waits pending watches until requested upper bound.
+    void WaitPendingOperations(u64 requested_upper_bound);
+
+private:
+    Vulkan::Scheduler& scheduler;
+    u64 offset{};
+    u64 mapped_size{};
+    std::vector<Watch> current_watches;
+    std::size_t current_watch_cursor{};
+    std::optional<size_t> invalidation_mark;
+    std::vector<Watch> previous_watches;
+    std::size_t wait_cursor{};
+    u64 wait_bound{};
+};
+
+} // namespace VideoCore

src/video_core/buffer_cache/buffer_cache.cpp

@@ -0,0 +1,497 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+#pragma clang optimize off
+#include <algorithm>
+#include "common/alignment.h"
+#include "common/scope_exit.h"
+#include "shader_recompiler/runtime_info.h"
+#include "video_core/amdgpu/liverpool.h"
+#include "video_core/buffer_cache/buffer_cache.h"
+#include "video_core/renderer_vulkan/liverpool_to_vk.h"
+#include "video_core/renderer_vulkan/vk_instance.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+
+namespace VideoCore {
+
+static constexpr size_t StagingBufferSize = 256_MB;
+static constexpr size_t UboStreamBufferSize = 64_MB;
+
+BufferCache::BufferCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
+                         const AmdGpu::Liverpool* liverpool_, PageManager& tracker_)
+    : instance{instance_}, scheduler{scheduler_}, liverpool{liverpool_}, tracker{tracker_},
+      staging_buffer{instance, scheduler, MemoryUsage::Upload, StagingBufferSize},
+      stream_buffer{instance, scheduler, MemoryUsage::Stream, UboStreamBufferSize},
+      memory_tracker{&tracker} {
+    // Ensure the first slot is used for the null buffer
+    void(slot_buffers.insert(instance, MemoryUsage::DeviceLocal, 0, 1));
+}
+
+BufferCache::~BufferCache() = default;
+
+void BufferCache::InvalidateMemory(VAddr device_addr, u64 size) {
+    std::scoped_lock lk{mutex};
+    const bool is_tracked = IsRegionRegistered(device_addr, size);
+    if (!is_tracked) {
+        return;
+    }
+    // Mark the page as CPU modified to stop tracking writes.
+    SCOPE_EXIT {
+        memory_tracker.MarkRegionAsCpuModified(device_addr, size);
+    };
+    if (!memory_tracker.IsRegionGpuModified(device_addr, size)) {
+        // Page has not been modified by the GPU, nothing to do.
+        return;
+    }
+}
+
+void BufferCache::DownloadBufferMemory(Buffer& buffer, VAddr device_addr, u64 size) {
+    boost::container::small_vector<vk::BufferCopy, 1> copies;
+    u64 total_size_bytes = 0;
+    u64 largest_copy = 0;
+    memory_tracker.ForEachDownloadRange<true>(
+        device_addr, size, [&](u64 device_addr_out, u64 range_size) {
+            const VAddr buffer_addr = buffer.CpuAddr();
+            const auto add_download = [&](VAddr start, VAddr end, u64) {
+                const u64 new_offset = start - buffer_addr;
+                const u64 new_size = end - start;
+                copies.push_back(vk::BufferCopy{
+                    .srcOffset = new_offset,
+                    .dstOffset = total_size_bytes,
+                    .size = new_size,
+                });
+                // Align up to avoid cache conflicts
+                constexpr u64 align = 64ULL;
+                constexpr u64 mask = ~(align - 1ULL);
+                total_size_bytes += (new_size + align - 1) & mask;
+                largest_copy = std::max(largest_copy, new_size);
+            };
+        });
+    if (total_size_bytes == 0) {
+        return;
+    }
+    const auto [staging, offset] = staging_buffer.Map(total_size_bytes);
+    for (auto& copy : copies) {
+        // Modify copies to have the staging offset in mind
+        copy.dstOffset += offset;
+    }
+    staging_buffer.Commit();
+    scheduler.EndRendering();
+    const auto cmdbuf = scheduler.CommandBuffer();
+    cmdbuf.copyBuffer(buffer.buffer, staging_buffer.Handle(), copies);
+    scheduler.Finish();
+    for (const auto& copy : copies) {
+        const VAddr copy_device_addr = buffer.CpuAddr() + copy.srcOffset;
+        const u64 dst_offset = copy.dstOffset - offset;
+        std::memcpy(std::bit_cast<u8*>(copy_device_addr), staging + dst_offset, copy.size);
+    }
+}
+
+bool BufferCache::BindVertexBuffers(const Shader::Info& vs_info) {
+    if (vs_info.vs_inputs.empty()) {
+        return false;
+    }
+
+    std::array<vk::Buffer, NUM_VERTEX_BUFFERS> host_buffers;
+    std::array<vk::DeviceSize, NUM_VERTEX_BUFFERS> host_offsets;
+    boost::container::static_vector<AmdGpu::Buffer, NUM_VERTEX_BUFFERS> guest_buffers;
+
+    struct BufferRange {
+        VAddr base_address;
+        VAddr end_address;
+        vk::Buffer vk_buffer;
+        u64 offset;
+
+        size_t GetSize() const {
+            return end_address - base_address;
+        }
+    };
+
+    // Calculate buffers memory overlaps
+    bool has_step_rate = false;
+    boost::container::static_vector<BufferRange, NUM_VERTEX_BUFFERS> ranges{};
+    for (const auto& input : vs_info.vs_inputs) {
+        if (input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate0 ||
+            input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate1) {
+            has_step_rate = true;
+            continue;
+        }
+
+        const auto& buffer = vs_info.ReadUd<AmdGpu::Buffer>(input.sgpr_base, input.dword_offset);
+        if (buffer.GetSize() == 0) {
+            continue;
+        }
+        guest_buffers.emplace_back(buffer);
+        ranges.emplace_back(buffer.base_address, buffer.base_address + buffer.GetSize());
+    }
+
+    std::ranges::sort(ranges, [](const BufferRange& lhv, const BufferRange& rhv) {
+        return lhv.base_address < rhv.base_address;
+    });
+
+    boost::container::static_vector<BufferRange, NUM_VERTEX_BUFFERS> ranges_merged{ranges[0]};
+    for (auto range : ranges) {
+        auto& prev_range = ranges_merged.back();
+        if (prev_range.end_address < range.base_address) {
+            ranges_merged.emplace_back(range);
+        } else {
+            prev_range.end_address = std::max(prev_range.end_address, range.end_address);
+        }
+    }
+
+    // Map buffers
+    for (auto& range : ranges_merged) {
+        const auto [buffer, offset] = ObtainBuffer(range.base_address, range.GetSize(), false);
+        range.vk_buffer = buffer->buffer;
+        range.offset = offset;
+    }
+
+    // Bind vertex buffers
+    const size_t num_buffers = guest_buffers.size();
+    for (u32 i = 0; i < num_buffers; ++i) {
+        const auto& buffer = guest_buffers[i];
+        const auto host_buffer = std::ranges::find_if(ranges_merged, [&](const BufferRange& range) {
+            return (buffer.base_address >= range.base_address &&
+                    buffer.base_address < range.end_address);
+        });
+        ASSERT(host_buffer != ranges_merged.cend());
+
+        host_buffers[i] = host_buffer->vk_buffer;
+        host_offsets[i] = host_buffer->offset + buffer.base_address - host_buffer->base_address;
+    }
+
+    if (num_buffers > 0) {
+        const auto cmdbuf = scheduler.CommandBuffer();
+        cmdbuf.bindVertexBuffers(0, num_buffers, host_buffers.data(), host_offsets.data());
+    }
+
+    return has_step_rate;
+}
+
+u32 BufferCache::BindIndexBuffer(bool& is_indexed, u32 index_offset) {
+    // Emulate QuadList primitive type with CPU made index buffer.
+    const auto& regs = liverpool->regs;
+    if (regs.primitive_type == AmdGpu::Liverpool::PrimitiveType::QuadList) {
+        is_indexed = true;
+
+        // Emit indices.
+        const u32 index_size = 3 * regs.num_indices;
+        const auto [data, offset] = stream_buffer.Map(index_size);
+        Vulkan::LiverpoolToVK::EmitQuadToTriangleListIndices(data, regs.num_indices);
+        stream_buffer.Commit();
+
+        // Bind index buffer.
+        const auto cmdbuf = scheduler.CommandBuffer();
+        cmdbuf.bindIndexBuffer(stream_buffer.Handle(), offset, vk::IndexType::eUint16);
+        return index_size / sizeof(u16);
+    }
+    if (!is_indexed) {
+        return regs.num_indices;
+    }
+
+    // Figure out index type and size.
+    const bool is_index16 =
+        regs.index_buffer_type.index_type == AmdGpu::Liverpool::IndexType::Index16;
+    const vk::IndexType index_type = is_index16 ? vk::IndexType::eUint16 : vk::IndexType::eUint32;
+    const u32 index_size = is_index16 ? sizeof(u16) : sizeof(u32);
+    VAddr index_address = regs.index_base_address.Address<VAddr>();
+    index_address += index_offset * index_size;
+
+    // Bind index buffer.
+    const u32 index_buffer_size = regs.num_indices * index_size;
+    const auto [vk_buffer, offset] = ObtainBuffer(index_address, index_buffer_size, false);
+    const auto cmdbuf = scheduler.CommandBuffer();
+    cmdbuf.bindIndexBuffer(vk_buffer->Handle(), offset, index_type);
+    return regs.num_indices;
+}
+
+std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, bool is_written) {
+    std::scoped_lock lk{mutex};
+    static constexpr u64 StreamThreshold = CACHING_PAGESIZE;
+    const bool is_gpu_dirty = memory_tracker.IsRegionGpuModified(device_addr, size);
+    if (!is_written && size < StreamThreshold && !is_gpu_dirty) {
+        // For small uniform buffers that have not been modified by gpu
+        // use device local stream buffer to reduce renderpass breaks.
+        const u64 offset = stream_buffer.Copy(device_addr, size, instance.UniformMinAlignment());
+        return {&stream_buffer, offset};
+    }
+
+    const BufferId buffer_id = FindBuffer(device_addr, size);
+    Buffer& buffer = slot_buffers[buffer_id];
+    SynchronizeBuffer(buffer, device_addr, size);
+    if (is_written) {
+        memory_tracker.MarkRegionAsGpuModified(device_addr, size);
+    }
+    return {&buffer, buffer.Offset(device_addr)};
+}
+
+bool BufferCache::IsRegionRegistered(VAddr addr, size_t size) {
+    const VAddr end_addr = addr + size;
+    const u64 page_end = Common::DivCeil(end_addr, CACHING_PAGESIZE);
+    for (u64 page = addr >> CACHING_PAGEBITS; page < page_end;) {
+        const BufferId buffer_id = page_table[page];
+        if (!buffer_id) {
+            ++page;
+            continue;
+        }
+        Buffer& buffer = slot_buffers[buffer_id];
+        const VAddr buf_start_addr = buffer.CpuAddr();
+        const VAddr buf_end_addr = buf_start_addr + buffer.SizeBytes();
+        if (buf_start_addr < end_addr && addr < buf_end_addr) {
+            return true;
+        }
+        page = Common::DivCeil(end_addr, CACHING_PAGESIZE);
+    }
+    return false;
+}
+
+bool BufferCache::IsRegionCpuModified(VAddr addr, size_t size) {
+    return memory_tracker.IsRegionCpuModified(addr, size);
+}
+
+BufferId BufferCache::FindBuffer(VAddr device_addr, u32 size) {
+    if (device_addr == 0) {
+        return NULL_BUFFER_ID;
+    }
+    const u64 page = device_addr >> CACHING_PAGEBITS;
+    const BufferId buffer_id = page_table[page];
+    if (!buffer_id) {
+        return CreateBuffer(device_addr, size);
+    }
+    const Buffer& buffer = slot_buffers[buffer_id];
+    if (buffer.IsInBounds(device_addr, size)) {
+        return buffer_id;
+    }
+    return CreateBuffer(device_addr, size);
+}
+
+BufferCache::OverlapResult BufferCache::ResolveOverlaps(VAddr device_addr, u32 wanted_size) {
+    static constexpr int STREAM_LEAP_THRESHOLD = 16;
+    boost::container::small_vector<BufferId, 16> overlap_ids;
+    VAddr begin = device_addr;
+    VAddr end = device_addr + wanted_size;
+    int stream_score = 0;
+    bool has_stream_leap = false;
+    const auto expand_begin = [&](VAddr add_value) {
+        static constexpr VAddr min_page = CACHING_PAGESIZE + DEVICE_PAGESIZE;
+        if (add_value > begin - min_page) {
+            begin = min_page;
+            device_addr = DEVICE_PAGESIZE;
+            return;
+        }
+        begin -= add_value;
+        device_addr = begin - CACHING_PAGESIZE;
+    };
+    const auto expand_end = [&](VAddr add_value) {
+        static constexpr VAddr max_page = 1ULL << MemoryTracker::MAX_CPU_PAGE_BITS;
+        if (add_value > max_page - end) {
+            end = max_page;
+            return;
+        }
+        end += add_value;
+    };
+    if (begin == 0) {
+        return OverlapResult{
+            .ids = std::move(overlap_ids),
+            .begin = begin,
+            .end = end,
+            .has_stream_leap = has_stream_leap,
+        };
+    }
+    for (; device_addr >> CACHING_PAGEBITS < Common::DivCeil(end, CACHING_PAGESIZE);
+         device_addr += CACHING_PAGESIZE) {
+        const BufferId overlap_id = page_table[device_addr >> CACHING_PAGEBITS];
+        if (!overlap_id) {
+            continue;
+        }
+        Buffer& overlap = slot_buffers[overlap_id];
+        if (overlap.is_picked) {
+            continue;
+        }
+        overlap_ids.push_back(overlap_id);
+        overlap.is_picked = true;
+        const VAddr overlap_device_addr = overlap.CpuAddr();
+        const bool expands_left = overlap_device_addr < begin;
+        if (expands_left) {
+            begin = overlap_device_addr;
+        }
+        const VAddr overlap_end = overlap_device_addr + overlap.SizeBytes();
+        const bool expands_right = overlap_end > end;
+        if (overlap_end > end) {
+            end = overlap_end;
+        }
+        stream_score += overlap.StreamScore();
+        if (stream_score > STREAM_LEAP_THRESHOLD && !has_stream_leap) {
+            // When this memory region has been joined a bunch of times, we assume it's being used
+            // as a stream buffer. Increase the size to skip constantly recreating buffers.
+            has_stream_leap = true;
+            if (expands_right) {
+                expand_begin(CACHING_PAGESIZE * 128);
+            }
+            if (expands_left) {
+                expand_end(CACHING_PAGESIZE * 128);
+            }
+        }
+    }
+    return OverlapResult{
+        .ids = std::move(overlap_ids),
+        .begin = begin,
+        .end = end,
+        .has_stream_leap = has_stream_leap,
+    };
+}
+
+void BufferCache::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
+                              bool accumulate_stream_score) {
+    Buffer& new_buffer = slot_buffers[new_buffer_id];
+    Buffer& overlap = slot_buffers[overlap_id];
+    if (accumulate_stream_score) {
+        new_buffer.IncreaseStreamScore(overlap.StreamScore() + 1);
+    }
+    const size_t dst_base_offset = overlap.CpuAddr() - new_buffer.CpuAddr();
+    const vk::BufferCopy copy = {
+        .srcOffset = 0,
+        .dstOffset = dst_base_offset,
+        .size = overlap.SizeBytes(),
+    };
+    scheduler.EndRendering();
+    const auto cmdbuf = scheduler.CommandBuffer();
+    static constexpr vk::MemoryBarrier READ_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eMemoryWrite,
+        .dstAccessMask = vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite,
+    };
+    static constexpr vk::MemoryBarrier WRITE_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
+        .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,
+    };
+    cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
+                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlagBits::eByRegion,
+                           READ_BARRIER, {}, {});
+    cmdbuf.copyBuffer(overlap.buffer, new_buffer.buffer, copy);
+    cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
+                           vk::PipelineStageFlagBits::eAllCommands,
+                           vk::DependencyFlagBits::eByRegion, WRITE_BARRIER, {}, {});
+    DeleteBuffer(overlap_id, true);
+}
+
+BufferId BufferCache::CreateBuffer(VAddr device_addr, u32 wanted_size) {
+    const VAddr device_addr_end = Common::AlignUp(device_addr + wanted_size, CACHING_PAGESIZE);
+    device_addr = Common::AlignDown(device_addr, CACHING_PAGESIZE);
+    wanted_size = static_cast<u32>(device_addr_end - device_addr);
+    const OverlapResult overlap = ResolveOverlaps(device_addr, wanted_size);
+    const u32 size = static_cast<u32>(overlap.end - overlap.begin);
+    const BufferId new_buffer_id =
+        slot_buffers.insert(instance, MemoryUsage::DeviceLocal, overlap.begin, size);
+    auto& new_buffer = slot_buffers[new_buffer_id];
+    const size_t size_bytes = new_buffer.SizeBytes();
+    const auto cmdbuf = scheduler.CommandBuffer();
+    scheduler.EndRendering();
+    cmdbuf.fillBuffer(new_buffer.buffer, 0, size_bytes, 0);
+    for (const BufferId overlap_id : overlap.ids) {
+        JoinOverlap(new_buffer_id, overlap_id, !overlap.has_stream_leap);
+    }
+    Register(new_buffer_id);
+    return new_buffer_id;
+}
+
+void BufferCache::Register(BufferId buffer_id) {
+    ChangeRegister<true>(buffer_id);
+}
+
+void BufferCache::Unregister(BufferId buffer_id) {
+    ChangeRegister<false>(buffer_id);
+}
+
+template <bool insert>
+void BufferCache::ChangeRegister(BufferId buffer_id) {
+    Buffer& buffer = slot_buffers[buffer_id];
+    const auto size = buffer.SizeBytes();
+    const VAddr device_addr_begin = buffer.CpuAddr();
+    const VAddr device_addr_end = device_addr_begin + size;
+    const u64 page_begin = device_addr_begin / CACHING_PAGESIZE;
+    const u64 page_end = Common::DivCeil(device_addr_end, CACHING_PAGESIZE);
+    for (u64 page = page_begin; page != page_end; ++page) {
+        if constexpr (insert) {
+            page_table[page] = buffer_id;
+        } else {
+            page_table[page] = BufferId{};
+        }
+    }
+}
+
+bool BufferCache::SynchronizeBuffer(Buffer& buffer, VAddr device_addr, u32 size) {
+    boost::container::small_vector<vk::BufferCopy, 4> copies;
+    u64 total_size_bytes = 0;
+    u64 largest_copy = 0;
+    VAddr buffer_start = buffer.CpuAddr();
+    const auto add_copy = [&](VAddr device_addr_out, u64 range_size) {
+        copies.push_back(vk::BufferCopy{
+            .srcOffset = total_size_bytes,
+            .dstOffset = device_addr_out - buffer_start,
+            .size = range_size,
+        });
+        total_size_bytes += range_size;
+        largest_copy = std::max(largest_copy, range_size);
+    };
+    memory_tracker.ForEachUploadRange(device_addr, size, [&](u64 device_addr_out, u64 range_size) {
+        add_copy(device_addr_out, range_size);
+        // Prevent uploading to gpu modified regions.
+        // gpu_modified_ranges.ForEachNotInRange(device_addr_out, range_size, add_copy);
+    });
+    if (total_size_bytes == 0) {
+        return true;
+    }
+    vk::Buffer src_buffer = staging_buffer.Handle();
+    if (total_size_bytes < StagingBufferSize) {
+        const auto [staging, offset] = staging_buffer.Map(total_size_bytes);
+        for (auto& copy : copies) {
+            u8* const src_pointer = staging + copy.srcOffset;
+            const VAddr device_addr = buffer.CpuAddr() + copy.dstOffset;
+            std::memcpy(src_pointer, std::bit_cast<const u8*>(device_addr), copy.size);
+            // Apply the staging offset
+            copy.srcOffset += offset;
+        }
+        staging_buffer.Commit();
+    } else {
+        // For large one time transfers use a temporary host buffer.
+        // RenderDoc can lag quite a bit if the stream buffer is too large.
+        Buffer temp_buffer{instance, MemoryUsage::Upload, 0, total_size_bytes};
+        src_buffer = temp_buffer.Handle();
+        u8* const staging = temp_buffer.mapped_data.data();
+        for (auto& copy : copies) {
+            u8* const src_pointer = staging + copy.srcOffset;
+            const VAddr device_addr = buffer.CpuAddr() + copy.dstOffset;
+            std::memcpy(src_pointer, std::bit_cast<const u8*>(device_addr), copy.size);
+        }
+        scheduler.DeferOperation([buffer = std::move(temp_buffer)]() mutable {});
+    }
+    scheduler.EndRendering();
+    const auto cmdbuf = scheduler.CommandBuffer();
+    static constexpr vk::MemoryBarrier READ_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eMemoryWrite,
+        .dstAccessMask = vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite,
+    };
+    static constexpr vk::MemoryBarrier WRITE_BARRIER{
+        .srcAccessMask = vk::AccessFlagBits::eTransferWrite,
+        .dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,
+    };
+    cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
+                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlagBits::eByRegion,
+                           READ_BARRIER, {}, {});
+    cmdbuf.copyBuffer(src_buffer, buffer.buffer, copies);
+    cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
+                           vk::PipelineStageFlagBits::eAllCommands,
+                           vk::DependencyFlagBits::eByRegion, WRITE_BARRIER, {}, {});
+    return false;
+}
+
+void BufferCache::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
+    // Mark the whole buffer as CPU written to stop tracking CPU writes
+    if (!do_not_mark) {
+        Buffer& buffer = slot_buffers[buffer_id];
+        memory_tracker.MarkRegionAsCpuModified(buffer.CpuAddr(), buffer.SizeBytes());
+    }
+    Unregister(buffer_id);
+    scheduler.DeferOperation([this, buffer_id] { slot_buffers.erase(buffer_id); });
+}
+
+} // namespace VideoCore

src/video_core/buffer_cache/buffer_cache.h

@@ -0,0 +1,121 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <array>
+#include <mutex>
+#include <boost/container/small_vector.hpp>
+#include <boost/icl/interval_map.hpp>
+#include <tsl/robin_map.h>
+#include "common/div_ceil.h"
+#include "common/slot_vector.h"
+#include "common/types.h"
+#include "video_core/buffer_cache/buffer.h"
+#include "video_core/buffer_cache/memory_tracker_base.h"
+#include "video_core/buffer_cache/range_set.h"
+
+namespace AmdGpu {
+struct Liverpool;
+}
+
+namespace Shader {
+struct Info;
+}
+
+namespace VideoCore {
+
+using BufferId = Common::SlotId;
+
+static constexpr BufferId NULL_BUFFER_ID{0};
+
+static constexpr u32 NUM_VERTEX_BUFFERS = 32;
+
+class BufferCache {
+public:
+    static constexpr u32 CACHING_PAGEBITS = 12;
+    static constexpr u64 CACHING_PAGESIZE = u64{1} << CACHING_PAGEBITS;
+    static constexpr u64 DEVICE_PAGESIZE = 4_KB;
+
+    struct OverlapResult {
+        boost::container::small_vector<BufferId, 16> ids;
+        VAddr begin;
+        VAddr end;
+        bool has_stream_leap = false;
+    };
+
+public:
+    explicit BufferCache(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
+                         const AmdGpu::Liverpool* liverpool, PageManager& tracker);
+    ~BufferCache();
+
+    /// Invalidates any buffer in the logical page range.
+    void InvalidateMemory(VAddr device_addr, u64 size);
+
+    /// Binds host vertex buffers for the current draw.
+    bool BindVertexBuffers(const Shader::Info& vs_info);
+
+    /// Bind host index buffer for the current draw.
+    u32 BindIndexBuffer(bool& is_indexed, u32 index_offset);
+
+    /// Obtains a buffer for the specified region.
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(VAddr gpu_addr, u32 size, bool is_written);
+
+    /// Return true when a region is registered on the cache
+    [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size);
+
+    /// Return true when a CPU region is modified from the CPU
+    [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
+
+private:
+    template <typename Func>
+    void ForEachBufferInRange(VAddr device_addr, u64 size, Func&& func) {
+        const u64 page_end = Common::DivCeil(device_addr + size, CACHING_PAGESIZE);
+        for (u64 page = device_addr >> CACHING_PAGEBITS; page < page_end;) {
+            const BufferId buffer_id = page_table[page];
+            if (!buffer_id) {
+                ++page;
+                continue;
+            }
+            Buffer& buffer = slot_buffers[buffer_id];
+            func(buffer_id, buffer);
+
+            const VAddr end_addr = buffer.CpuAddr() + buffer.SizeBytes();
+            page = Common::DivCeil(end_addr, CACHING_PAGESIZE);
+        }
+    }
+
+    void DownloadBufferMemory(Buffer& buffer, VAddr device_addr, u64 size);
+
+    [[nodiscard]] BufferId FindBuffer(VAddr device_addr, u32 size);
+
+    [[nodiscard]] OverlapResult ResolveOverlaps(VAddr device_addr, u32 wanted_size);
+
+    void JoinOverlap(BufferId new_buffer_id, BufferId overlap_id, bool accumulate_stream_score);
+
+    [[nodiscard]] BufferId CreateBuffer(VAddr device_addr, u32 wanted_size);
+
+    void Register(BufferId buffer_id);
+
+    void Unregister(BufferId buffer_id);
+
+    template <bool insert>
+    void ChangeRegister(BufferId buffer_id);
+
+    bool SynchronizeBuffer(Buffer& buffer, VAddr device_addr, u32 size);
+
+    void DeleteBuffer(BufferId buffer_id, bool do_not_mark = false);
+
+    const Vulkan::Instance& instance;
+    Vulkan::Scheduler& scheduler;
+    const AmdGpu::Liverpool* liverpool;
+    PageManager& tracker;
+    StreamBuffer staging_buffer;
+    StreamBuffer stream_buffer;
+    std::recursive_mutex mutex;
+    Common::SlotVector<Buffer> slot_buffers;
+    MemoryTracker memory_tracker;
+    std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
+};
+
+} // namespace VideoCore

src/video_core/buffer_cache/memory_tracker_base.h

@@ -0,0 +1,175 @@
+// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <algorithm>
+#include <deque>
+#include <type_traits>
+#include <vector>
+#include "common/types.h"
+#include "video_core/buffer_cache/word_manager.h"
+
+namespace VideoCore {
+
+class MemoryTracker {
+public:
+    static constexpr size_t MAX_CPU_PAGE_BITS = 39;
+    static constexpr size_t HIGHER_PAGE_BITS = 22;
+    static constexpr size_t HIGHER_PAGE_SIZE = 1ULL << HIGHER_PAGE_BITS;
+    static constexpr size_t HIGHER_PAGE_MASK = HIGHER_PAGE_SIZE - 1ULL;
+    static constexpr size_t NUM_HIGH_PAGES = 1ULL << (MAX_CPU_PAGE_BITS - HIGHER_PAGE_BITS);
+    static constexpr size_t MANAGER_POOL_SIZE = 32;
+    static constexpr size_t WORDS_STACK_NEEDED = HIGHER_PAGE_SIZE / BYTES_PER_WORD;
+    using Manager = WordManager<WORDS_STACK_NEEDED>;
+
+public:
+    explicit MemoryTracker(PageManager* tracker_) : tracker{tracker_} {}
+    ~MemoryTracker() = default;
+
+    /// Returns true if a region has been modified from the CPU
+    [[nodiscard]] bool IsRegionCpuModified(VAddr query_cpu_addr, u64 query_size) noexcept {
+        return IteratePages<true>(
+            query_cpu_addr, query_size, [](Manager* manager, u64 offset, size_t size) {
+                return manager->template IsRegionModified<Type::CPU>(offset, size);
+            });
+    }
+
+    /// Returns true if a region has been modified from the GPU
+    [[nodiscard]] bool IsRegionGpuModified(VAddr query_cpu_addr, u64 query_size) noexcept {
+        return IteratePages<false>(
+            query_cpu_addr, query_size, [](Manager* manager, u64 offset, size_t size) {
+                return manager->template IsRegionModified<Type::GPU>(offset, size);
+            });
+    }
+
+    /// Mark region as CPU modified, notifying the device_tracker about this change
+    void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 query_size) {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::CPU, true>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
+    /// Unmark region as CPU modified, notifying the device_tracker about this change
+    void UnmarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 query_size) {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::CPU, false>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
+    /// Mark region as modified from the host GPU
+    void MarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 query_size) noexcept {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::GPU, true>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
+    /// Unmark region as modified from the host GPU
+    void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 query_size) noexcept {
+        IteratePages<true>(dirty_cpu_addr, query_size,
+                           [](Manager* manager, u64 offset, size_t size) {
+                               manager->template ChangeRegionState<Type::GPU, false>(
+                                   manager->GetCpuAddr() + offset, size);
+                           });
+    }
+
+    /// Call 'func' for each CPU modified range and unmark those pages as CPU modified
+    template <typename Func>
+    void ForEachUploadRange(VAddr query_cpu_range, u64 query_size, Func&& func) {
+        IteratePages<true>(query_cpu_range, query_size,
+                           [&func](Manager* manager, u64 offset, size_t size) {
+                               manager->template ForEachModifiedRange<Type::CPU, true>(
+                                   manager->GetCpuAddr() + offset, size, func);
+                           });
+    }
+
+    /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
+    template <bool clear, typename Func>
+    void ForEachDownloadRange(VAddr query_cpu_range, u64 query_size, Func&& func) {
+        IteratePages<false>(query_cpu_range, query_size,
+                            [&func](Manager* manager, u64 offset, size_t size) {
+                                if constexpr (clear) {
+                                    manager->template ForEachModifiedRange<Type::GPU, true>(
+                                        manager->GetCpuAddr() + offset, size, func);
+                                } else {
+                                    manager->template ForEachModifiedRange<Type::GPU, false>(
+                                        manager->GetCpuAddr() + offset, size, func);
+                                }
+                            });
+    }
+
+private:
+    /**
+     * @brief IteratePages Iterates L2 word manager page table.
+     * @param cpu_address Start byte cpu address
+     * @param size Size in bytes of the region of iterate.
+     * @param func Callback for each word manager.
+     * @return
+     */
+    template <bool create_region_on_fail, typename Func>
+    bool IteratePages(VAddr cpu_address, size_t size, Func&& func) {
+        using FuncReturn = typename std::invoke_result<Func, Manager*, u64, size_t>::type;
+        static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
+        std::size_t remaining_size{size};
+        std::size_t page_index{cpu_address >> HIGHER_PAGE_BITS};
+        u64 page_offset{cpu_address & HIGHER_PAGE_MASK};
+        while (remaining_size > 0) {
+            const std::size_t copy_amount{
+                std::min<std::size_t>(HIGHER_PAGE_SIZE - page_offset, remaining_size)};
+            auto* manager{top_tier[page_index]};
+            if (manager) {
+                if constexpr (BOOL_BREAK) {
+                    if (func(manager, page_offset, copy_amount)) {
+                        return true;
+                    }
+                } else {
+                    func(manager, page_offset, copy_amount);
+                }
+            } else if constexpr (create_region_on_fail) {
+                CreateRegion(page_index);
+                manager = top_tier[page_index];
+                if constexpr (BOOL_BREAK) {
+                    if (func(manager, page_offset, copy_amount)) {
+                        return true;
+                    }
+                } else {
+                    func(manager, page_offset, copy_amount);
+                }
+            }
+            page_index++;
+            page_offset = 0;
+            remaining_size -= copy_amount;
+        }
+        return false;
+    }
+
+    void CreateRegion(std::size_t page_index) {
+        const VAddr base_cpu_addr = page_index << HIGHER_PAGE_BITS;
+        if (free_managers.empty()) {
+            manager_pool.emplace_back();
+            auto& last_pool = manager_pool.back();
+            for (size_t i = 0; i < MANAGER_POOL_SIZE; i++) {
+                std::construct_at(&last_pool[i], tracker, 0, HIGHER_PAGE_SIZE);
+                free_managers.push_back(&last_pool[i]);
+            }
+        }
+        // Each manager tracks a 4_MB virtual address space.
+        auto* new_manager = free_managers.back();
+        new_manager->SetCpuAddress(base_cpu_addr);
+        free_managers.pop_back();
+        top_tier[page_index] = new_manager;
+    }
+
+    PageManager* tracker;
+    std::deque<std::array<Manager, MANAGER_POOL_SIZE>> manager_pool;
+    std::vector<Manager*> free_managers;
+    std::array<Manager*, NUM_HIGH_PAGES> top_tier{};
+};
+
+} // namespace VideoCore

src/video_core/buffer_cache/range_set.h

@@ -0,0 +1,159 @@
+// SPDX-FileCopyrightText: 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <boost/icl/interval_map.hpp>
+#include <boost/pool/pool.hpp>
+#include <boost/pool/pool_alloc.hpp>
+#include <boost/pool/poolfwd.hpp>
+#include "common/types.h"
+
+namespace VideoCore {
+
+template <class T>
+using RangeSetsAllocator =
+    boost::fast_pool_allocator<T, boost::default_user_allocator_new_delete,
+                               boost::details::pool::default_mutex, 1024, 2048>;
+
+struct RangeSet {
+    using IntervalSet =
+        boost::icl::interval_set<VAddr, std::less,
+                                 ICL_INTERVAL_INSTANCE(ICL_INTERVAL_DEFAULT, VAddr, std::less),
+                                 RangeSetsAllocator>;
+    using IntervalType = typename IntervalSet::interval_type;
+
+    explicit RangeSet() = default;
+    ~RangeSet() = default;
+
+    void Add(VAddr base_address, size_t size) {
+        const VAddr end_address = base_address + size;
+        IntervalType interval{base_address, end_address};
+        m_ranges_set.add(interval);
+    }
+
+    void Subtract(VAddr base_address, size_t size) {
+        const VAddr end_address = base_address + size;
+        IntervalType interval{base_address, end_address};
+        m_ranges_set.subtract(interval);
+    }
+
+    template <typename Func>
+    void ForEach(Func&& func) const {
+        if (m_ranges_set.empty()) {
+            return;
+        }
+        auto it = m_ranges_set.begin();
+        auto end_it = m_ranges_set.end();
+        for (; it != end_it; it++) {
+            const VAddr inter_addr_end = it->upper();
+            const VAddr inter_addr = it->lower();
+            func(inter_addr, inter_addr_end);
+        }
+    }
+
+    template <typename Func>
+    void ForEachInRange(VAddr base_addr, size_t size, Func&& func) const {
+        if (m_ranges_set.empty()) {
+            return;
+        }
+        const VAddr start_address = base_addr;
+        const VAddr end_address = start_address + size;
+        const IntervalType search_interval{start_address, end_address};
+        auto it = m_ranges_set.lower_bound(search_interval);
+        if (it == m_ranges_set.end()) {
+            return;
+        }
+        auto end_it = m_ranges_set.upper_bound(search_interval);
+        for (; it != end_it; it++) {
+            VAddr inter_addr_end = it->upper();
+            VAddr inter_addr = it->lower();
+            if (inter_addr_end > end_address) {
+                inter_addr_end = end_address;
+            }
+            if (inter_addr < start_address) {
+                inter_addr = start_address;
+            }
+            func(inter_addr, inter_addr_end);
+        }
+    }
+
+    IntervalSet m_ranges_set;
+};
+
+class RangeMap {
+public:
+    using IntervalMap =
+        boost::icl::interval_map<VAddr, u64, boost::icl::partial_absorber, std::less,
+                                 boost::icl::inplace_plus, boost::icl::inter_section,
+                                 ICL_INTERVAL_INSTANCE(ICL_INTERVAL_DEFAULT, VAddr, std::less),
+                                 RangeSetsAllocator>;
+    using IntervalType = typename IntervalMap::interval_type;
+
+public:
+    RangeMap() = default;
+    ~RangeMap() = default;
+
+    RangeMap(RangeMap const&) = delete;
+    RangeMap& operator=(RangeMap const&) = delete;
+
+    RangeMap(RangeMap&& other);
+    RangeMap& operator=(RangeMap&& other);
+
+    void Add(VAddr base_address, size_t size, u64 value) {
+        const VAddr end_address = base_address + size;
+        IntervalType interval{base_address, end_address};
+        m_ranges_map.add({interval, value});
+    }
+
+    void Subtract(VAddr base_address, size_t size) {
+        const VAddr end_address = base_address + size;
+        IntervalType interval{base_address, end_address};
+        m_ranges_map -= interval;
+    }
+
+    template <typename Func>
+    void ForEachInRange(VAddr base_addr, size_t size, Func&& func) const {
+        if (m_ranges_map.empty()) {
+            return;
+        }
+        const VAddr start_address = base_addr;
+        const VAddr end_address = start_address + size;
+        const IntervalType search_interval{start_address, end_address};
+        auto it = m_ranges_map.lower_bound(search_interval);
+        if (it == m_ranges_map.end()) {
+            return;
+        }
+        auto end_it = m_ranges_map.upper_bound(search_interval);
+        for (; it != end_it; it++) {
+            VAddr inter_addr_end = it->first.upper();
+            VAddr inter_addr = it->first.lower();
+            if (inter_addr_end > end_address) {
+                inter_addr_end = end_address;
+            }
+            if (inter_addr < start_address) {
+                inter_addr = start_address;
+            }
+            func(inter_addr, inter_addr_end, it->second);
+        }
+    }
+
+    template <typename Func>
+    void ForEachNotInRange(VAddr base_addr, size_t size, Func&& func) const {
+        const VAddr end_addr = base_addr + size;
+        ForEachInRange(base_addr, size, [&](VAddr range_addr, VAddr range_end, u64) {
+            if (size_t gap_size = range_addr - base_addr; gap_size != 0) {
+                func(base_addr, gap_size);
+            }
+            base_addr = range_end;
+        });
+        if (base_addr != end_addr) {
+            func(base_addr, end_addr - base_addr);
+        }
+    }
+
+private:
+    IntervalMap m_ranges_map;
+};
+
+} // namespace VideoCore

src/video_core/buffer_cache/word_manager.h

@@ -0,0 +1,398 @@
+// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <algorithm>
+#include <span>
+#include <utility>
+#include "common/div_ceil.h"
+#include "common/types.h"
+#include "video_core/page_manager.h"
+
+namespace VideoCore {
+
+constexpr u64 PAGES_PER_WORD = 64;
+constexpr u64 BYTES_PER_PAGE = 4_KB;
+constexpr u64 BYTES_PER_WORD = PAGES_PER_WORD * BYTES_PER_PAGE;
+
+enum class Type {
+    CPU,
+    GPU,
+    Untracked,
+};
+
+/// Vector tracking modified pages tightly packed with small vector optimization
+template <size_t stack_words = 1>
+struct WordsArray {
+    /// Returns the pointer to the words state
+    [[nodiscard]] const u64* Pointer(bool is_short) const noexcept {
+        return is_short ? stack.data() : heap;
+    }
+
+    /// Returns the pointer to the words state
+    [[nodiscard]] u64* Pointer(bool is_short) noexcept {
+        return is_short ? stack.data() : heap;
+    }
+
+    std::array<u64, stack_words> stack{}; ///< Small buffers storage
+    u64* heap;                            ///< Not-small buffers pointer to the storage
+};
+
+template <size_t stack_words = 1>
+struct Words {
+    explicit Words() = default;
+    explicit Words(u64 size_bytes_) : size_bytes{size_bytes_} {
+        num_words = Common::DivCeil(size_bytes, BYTES_PER_WORD);
+        if (IsShort()) {
+            cpu.stack.fill(~u64{0});
+            gpu.stack.fill(0);
+            untracked.stack.fill(~u64{0});
+        } else {
+            // Share allocation between CPU and GPU pages and set their default values
+            u64* const alloc = new u64[num_words * 3];
+            cpu.heap = alloc;
+            gpu.heap = alloc + num_words;
+            untracked.heap = alloc + num_words * 2;
+            std::fill_n(cpu.heap, num_words, ~u64{0});
+            std::fill_n(gpu.heap, num_words, 0);
+            std::fill_n(untracked.heap, num_words, ~u64{0});
+        }
+        // Clean up tailing bits
+        const u64 last_word_size = size_bytes % BYTES_PER_WORD;
+        const u64 last_local_page = Common::DivCeil(last_word_size, BYTES_PER_PAGE);
+        const u64 shift = (PAGES_PER_WORD - last_local_page) % PAGES_PER_WORD;
+        const u64 last_word = (~u64{0} << shift) >> shift;
+        cpu.Pointer(IsShort())[NumWords() - 1] = last_word;
+        untracked.Pointer(IsShort())[NumWords() - 1] = last_word;
+    }
+
+    ~Words() {
+        Release();
+    }
+
+    Words& operator=(Words&& rhs) noexcept {
+        Release();
+        size_bytes = rhs.size_bytes;
+        num_words = rhs.num_words;
+        cpu = rhs.cpu;
+        gpu = rhs.gpu;
+        untracked = rhs.untracked;
+        rhs.cpu.heap = nullptr;
+        return *this;
+    }
+
+    Words(Words&& rhs) noexcept
+        : size_bytes{rhs.size_bytes}, num_words{rhs.num_words}, cpu{rhs.cpu}, gpu{rhs.gpu},
+          untracked{rhs.untracked} {
+        rhs.cpu.heap = nullptr;
+    }
+
+    Words& operator=(const Words&) = delete;
+    Words(const Words&) = delete;
+
+    /// Returns true when the buffer fits in the small vector optimization
+    [[nodiscard]] bool IsShort() const noexcept {
+        return num_words <= stack_words;
+    }
+
+    /// Returns the number of words of the buffer
+    [[nodiscard]] size_t NumWords() const noexcept {
+        return num_words;
+    }
+
+    /// Release buffer resources
+    void Release() {
+        if (!IsShort()) {
+            // CPU written words is the base for the heap allocation
+            delete[] cpu.heap;
+        }
+    }
+
+    template <Type type>
+    std::span<u64> Span() noexcept {
+        if constexpr (type == Type::CPU) {
+            return std::span<u64>(cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::GPU) {
+            return std::span<u64>(gpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Untracked) {
+            return std::span<u64>(untracked.Pointer(IsShort()), num_words);
+        }
+    }
+
+    template <Type type>
+    std::span<const u64> Span() const noexcept {
+        if constexpr (type == Type::CPU) {
+            return std::span<const u64>(cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::GPU) {
+            return std::span<const u64>(gpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Untracked) {
+            return std::span<const u64>(untracked.Pointer(IsShort()), num_words);
+        }
+    }
+
+    u64 size_bytes = 0;
+    size_t num_words = 0;
+    WordsArray<stack_words> cpu;
+    WordsArray<stack_words> gpu;
+    WordsArray<stack_words> untracked;
+};
+
+template <size_t stack_words = 1>
+class WordManager {
+public:
+    explicit WordManager(PageManager* tracker_, VAddr cpu_addr_, u64 size_bytes)
+        : tracker{tracker_}, cpu_addr{cpu_addr_}, words{size_bytes} {}
+
+    explicit WordManager() = default;
+
+    void SetCpuAddress(VAddr new_cpu_addr) {
+        cpu_addr = new_cpu_addr;
+    }
+
+    VAddr GetCpuAddr() const {
+        return cpu_addr;
+    }
+
+    static u64 ExtractBits(u64 word, size_t page_start, size_t page_end) {
+        constexpr size_t number_bits = sizeof(u64) * 8;
+        const size_t limit_page_end = number_bits - std::min(page_end, number_bits);
+        u64 bits = (word >> page_start) << page_start;
+        bits = (bits << limit_page_end) >> limit_page_end;
+        return bits;
+    }
+
+    static std::pair<size_t, size_t> GetWordPage(VAddr address) {
+        const size_t converted_address = static_cast<size_t>(address);
+        const size_t word_number = converted_address / BYTES_PER_WORD;
+        const size_t amount_pages = converted_address % BYTES_PER_WORD;
+        return std::make_pair(word_number, amount_pages / BYTES_PER_PAGE);
+    }
+
+    template <typename Func>
+    void IterateWords(size_t offset, size_t size, Func&& func) const {
+        using FuncReturn = std::invoke_result_t<Func, std::size_t, u64>;
+        static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
+        const size_t start = static_cast<size_t>(std::max<s64>(static_cast<s64>(offset), 0LL));
+        const size_t end = static_cast<size_t>(std::max<s64>(static_cast<s64>(offset + size), 0LL));
+        if (start >= SizeBytes() || end <= start) {
+            return;
+        }
+        auto [start_word, start_page] = GetWordPage(start);
+        auto [end_word, end_page] = GetWordPage(end + BYTES_PER_PAGE - 1ULL);
+        const size_t num_words = NumWords();
+        start_word = std::min(start_word, num_words);
+        end_word = std::min(end_word, num_words);
+        const size_t diff = end_word - start_word;
+        end_word += (end_page + PAGES_PER_WORD - 1ULL) / PAGES_PER_WORD;
+        end_word = std::min(end_word, num_words);
+        end_page += diff * PAGES_PER_WORD;
+        constexpr u64 base_mask{~0ULL};
+        for (size_t word_index = start_word; word_index < end_word; word_index++) {
+            const u64 mask = ExtractBits(base_mask, start_page, end_page);
+            start_page = 0;
+            end_page -= PAGES_PER_WORD;
+            if constexpr (BOOL_BREAK) {
+                if (func(word_index, mask)) {
+                    return;
+                }
+            } else {
+                func(word_index, mask);
+            }
+        }
+    }
+
+    template <typename Func>
+    void IteratePages(u64 mask, Func&& func) const {
+        size_t offset = 0;
+        while (mask != 0) {
+            const size_t empty_bits = std::countr_zero(mask);
+            offset += empty_bits;
+            mask = mask >> empty_bits;
+
+            const size_t continuous_bits = std::countr_one(mask);
+            func(offset, continuous_bits);
+            mask = continuous_bits < PAGES_PER_WORD ? (mask >> continuous_bits) : 0;
+            offset += continuous_bits;
+        }
+    }
+
+    /**
+     * Change the state of a range of pages
+     *
+     * @param dirty_addr    Base address to mark or unmark as modified
+     * @param size          Size in bytes to mark or unmark as modified
+     */
+    template <Type type, bool enable>
+    void ChangeRegionState(u64 dirty_addr, u64 size) noexcept(type == Type::GPU) {
+        std::span<u64> state_words = words.template Span<type>();
+        [[maybe_unused]] std::span<u64> untracked_words = words.template Span<Type::Untracked>();
+        IterateWords(dirty_addr - cpu_addr, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::CPU) {
+                NotifyPageTracker<!enable>(index, untracked_words[index], mask);
+            }
+            if constexpr (enable) {
+                state_words[index] |= mask;
+                if constexpr (type == Type::CPU) {
+                    untracked_words[index] |= mask;
+                }
+            } else {
+                state_words[index] &= ~mask;
+                if constexpr (type == Type::CPU) {
+                    untracked_words[index] &= ~mask;
+                }
+            }
+        });
+    }
+
+    /**
+     * Loop over each page in the given range, turn off those bits and notify the tracker if
+     * needed. Call the given function on each turned off range.
+     *
+     * @param query_cpu_range Base CPU address to loop over
+     * @param size            Size in bytes of the CPU range to loop over
+     * @param func            Function to call for each turned off region
+     */
+    template <Type type, bool clear, typename Func>
+    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) {
+        static_assert(type != Type::Untracked);
+
+        std::span<u64> state_words = words.template Span<type>();
+        [[maybe_unused]] std::span<u64> untracked_words = words.template Span<Type::Untracked>();
+        const size_t offset = query_cpu_range - cpu_addr;
+        bool pending = false;
+        size_t pending_offset{};
+        size_t pending_pointer{};
+        const auto release = [&]() {
+            func(cpu_addr + pending_offset * BYTES_PER_PAGE,
+                 (pending_pointer - pending_offset) * BYTES_PER_PAGE);
+        };
+        IterateWords(offset, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::GPU) {
+                mask &= ~untracked_words[index];
+            }
+            const u64 word = state_words[index] & mask;
+            if constexpr (clear) {
+                if constexpr (type == Type::CPU) {
+                    NotifyPageTracker<true>(index, untracked_words[index], mask);
+                }
+                state_words[index] &= ~mask;
+                if constexpr (type == Type::CPU) {
+                    untracked_words[index] &= ~mask;
+                }
+            }
+            const size_t base_offset = index * PAGES_PER_WORD;
+            IteratePages(word, [&](size_t pages_offset, size_t pages_size) {
+                const auto reset = [&]() {
+                    pending_offset = base_offset + pages_offset;
+                    pending_pointer = base_offset + pages_offset + pages_size;
+                };
+                if (!pending) {
+                    reset();
+                    pending = true;
+                    return;
+                }
+                if (pending_pointer == base_offset + pages_offset) {
+                    pending_pointer += pages_size;
+                    return;
+                }
+                release();
+                reset();
+            });
+        });
+        if (pending) {
+            release();
+        }
+    }
+
+    /**
+     * Returns true when a region has been modified
+     *
+     * @param offset Offset in bytes from the start of the buffer
+     * @param size   Size in bytes of the region to query for modifications
+     */
+    template <Type type>
+    [[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept {
+        static_assert(type != Type::Untracked);
+
+        const std::span<const u64> state_words = words.template Span<type>();
+        [[maybe_unused]] const std::span<const u64> untracked_words =
+            words.template Span<Type::Untracked>();
+        bool result = false;
+        IterateWords(offset, size, [&](size_t index, u64 mask) {
+            if constexpr (type == Type::GPU) {
+                mask &= ~untracked_words[index];
+            }
+            const u64 word = state_words[index] & mask;
+            if (word != 0) {
+                result = true;
+                return true;
+            }
+            return false;
+        });
+        return result;
+    }
+
+    /// Returns the number of words of the manager
+    [[nodiscard]] size_t NumWords() const noexcept {
+        return words.NumWords();
+    }
+
+    /// Returns the size in bytes of the manager
+    [[nodiscard]] u64 SizeBytes() const noexcept {
+        return words.size_bytes;
+    }
+
+    /// Returns true when the buffer fits in the small vector optimization
+    [[nodiscard]] bool IsShort() const noexcept {
+        return words.IsShort();
+    }
+
+private:
+    template <Type type>
+    u64* Array() noexcept {
+        if constexpr (type == Type::CPU) {
+            return words.cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::GPU) {
+            return words.gpu.Pointer(IsShort());
+        } else if constexpr (type == Type::Untracked) {
+            return words.untracked.Pointer(IsShort());
+        }
+    }
+
+    template <Type type>
+    const u64* Array() const noexcept {
+        if constexpr (type == Type::CPU) {
+            return words.cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::GPU) {
+            return words.gpu.Pointer(IsShort());
+        } else if constexpr (type == Type::Untracked) {
+            return words.untracked.Pointer(IsShort());
+        }
+    }
+
+    /**
+     * Notify tracker about changes in the CPU tracking state of a word in the buffer
+     *
+     * @param word_index   Index to the word to notify to the tracker
+     * @param current_bits Current state of the word
+     * @param new_bits     New state of the word
+     *
+     * @tparam add_to_tracker True when the tracker should start tracking the new pages
+     */
+    template <bool add_to_tracker>
+    void NotifyPageTracker(u64 word_index, u64 current_bits, u64 new_bits) const {
+        u64 changed_bits = (add_to_tracker ? current_bits : ~current_bits) & new_bits;
+        VAddr addr = cpu_addr + word_index * BYTES_PER_WORD;
+        IteratePages(changed_bits, [&](size_t offset, size_t size) {
+            tracker->UpdatePagesCachedCount(addr + offset * BYTES_PER_PAGE, size * BYTES_PER_PAGE,
+                                            add_to_tracker ? 1 : -1);
+        });
+    }
+
+    PageManager* tracker;
+    VAddr cpu_addr = 0;
+    Words<stack_words> words;
+};
+
+} // namespace VideoCore

src/video_core/page_manager.cpp

@@ -0,0 +1,209 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <thread>
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "common/error.h"
+#include "video_core/page_manager.h"
+#include "video_core/renderer_vulkan/vk_rasterizer.h"
+
+#ifndef _WIN64
+#include <fcntl.h>
+#include <linux/userfaultfd.h>
+#include <poll.h>
+#include <signal.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#endif
+
+namespace VideoCore {
+
+constexpr size_t PAGESIZE = 4_KB;
+constexpr size_t PAGEBITS = 12;
+
+#ifdef SHADPS4_USERFAULTFD
+struct PageManager::Impl {
+    Impl(Vulkan::Rasterizer* rasterizer_) : rasterizer{rasterizer_} {
+        uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
+        ASSERT_MSG(uffd != -1, "{}", Common::GetLastErrorMsg());
+
+        // Request uffdio features from kernel.
+        uffdio_api api;
+        api.api = UFFD_API;
+        api.features = UFFD_FEATURE_THREAD_ID;
+        const int ret = ioctl(uffd, UFFDIO_API, &api);
+        ASSERT(ret == 0 && api.api == UFFD_API);
+
+        // Create uffd handler thread
+        ufd_thread = std::jthread([&](std::stop_token token) { UffdHandler(token); });
+    }
+
+    void OnMap(VAddr address, size_t size) {
+        uffdio_register reg;
+        reg.range.start = address;
+        reg.range.len = size;
+        reg.mode = UFFDIO_REGISTER_MODE_WP;
+        const int ret = ioctl(uffd, UFFDIO_REGISTER, &reg);
+        ASSERT_MSG(ret != -1, "Uffdio register failed");
+    }
+
+    void OnUnmap(VAddr address, size_t size) {
+        uffdio_range range;
+        range.start = address;
+        range.len = size;
+        const int ret = ioctl(uffd, UFFDIO_UNREGISTER, &range);
+        ASSERT_MSG(ret != -1, "Uffdio unregister failed");
+    }
+
+    void Protect(VAddr address, size_t size, bool allow_write) {
+        uffdio_writeprotect wp;
+        wp.range.start = address;
+        wp.range.len = size;
+        wp.mode = allow_write ? 0 : UFFDIO_WRITEPROTECT_MODE_WP;
+        const int ret = ioctl(uffd, UFFDIO_WRITEPROTECT, &wp);
+        ASSERT_MSG(ret != -1, "Uffdio writeprotect failed with error: {}",
+                   Common::GetLastErrorMsg());
+    }
+
+    void UffdHandler(std::stop_token token) {
+        while (!token.stop_requested()) {
+            pollfd pollfd;
+            pollfd.fd = uffd;
+            pollfd.events = POLLIN;
+
+            // Block until the descriptor is ready for data reads.
+            const int pollres = poll(&pollfd, 1, -1);
+            switch (pollres) {
+            case -1:
+                perror("Poll userfaultfd");
+                continue;
+                break;
+            case 0:
+                continue;
+            case 1:
+                break;
+            default:
+                UNREACHABLE_MSG("Unexpected number of descriptors {} out of poll", pollres);
+            }
+
+            // We don't want an error condition to have occured.
+            ASSERT_MSG(!(pollfd.revents & POLLERR), "POLLERR on userfaultfd");
+
+            // We waited until there is data to read, we don't care about anything else.
+            if (!(pollfd.revents & POLLIN)) {
+                continue;
+            }
+
+            // Read message from kernel.
+            uffd_msg msg;
+            const int readret = read(uffd, &msg, sizeof(msg));
+            ASSERT_MSG(readret != -1 || errno == EAGAIN, "Unexpected result of uffd read");
+            if (errno == EAGAIN) {
+                continue;
+            }
+            ASSERT_MSG(readret == sizeof(msg), "Unexpected short read, exiting");
+            ASSERT(msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP);
+
+            // Notify rasterizer about the fault.
+            const VAddr addr = msg.arg.pagefault.address;
+            const VAddr addr_page = Common::AlignDown(addr, PAGESIZE);
+            rasterizer->InvalidateMemory(addr_page, PAGESIZE);
+        }
+    }
+
+    Vulkan::Rasterizer* rasterizer;
+    std::jthread ufd_thread;
+    int uffd;
+};
+#else
+struct PageManager::Impl {
+    Impl(Vulkan::Rasterizer* rasterizer_) {
+        rasterizer = rasterizer_;
+
+        sigset_t signal_mask;
+        sigemptyset(&signal_mask);
+        sigaddset(&signal_mask, SIGSEGV);
+
+        using HandlerType = decltype(sigaction::sa_sigaction);
+
+        struct sigaction guest_access_fault {};
+        guest_access_fault.sa_flags = SA_SIGINFO | SA_ONSTACK;
+        guest_access_fault.sa_sigaction = &GuestFaultSignalHandler;
+        guest_access_fault.sa_mask = signal_mask;
+        sigaction(SIGSEGV, &guest_access_fault, nullptr);
+    }
+
+    void OnMap(VAddr address, size_t size) {}
+
+    void OnUnmap(VAddr address, size_t size) {}
+
+    void Protect(VAddr address, size_t size, bool allow_write) {
+        mprotect(reinterpret_cast<void*>(address), size,
+                 PROT_READ | (allow_write ? PROT_WRITE : 0));
+    }
+
+    static void GuestFaultSignalHandler(int sig, siginfo_t* info, void* raw_context) {
+        ucontext_t* ctx = reinterpret_cast<ucontext_t*>(raw_context);
+        const VAddr address = reinterpret_cast<VAddr>(info->si_addr);
+        const greg_t err = ctx->uc_mcontext.gregs[REG_ERR];
+        if (err & 0x2) {
+            rasterizer->InvalidateMemory(address, PAGESIZE);
+        } else {
+            // Read not supported!
+            UNREACHABLE();
+        }
+    }
+
+    inline static Vulkan::Rasterizer* rasterizer;
+};
+#endif
+
+PageManager::PageManager(Vulkan::Rasterizer* rasterizer_)
+    : impl{std::make_unique<Impl>(rasterizer_)}, rasterizer{rasterizer_} {}
+
+PageManager::~PageManager() = default;
+
+void PageManager::OnGpuMap(VAddr address, size_t size) {
+    impl->OnMap(address, size);
+}
+
+void PageManager::OnGpuUnmap(VAddr address, size_t size) {
+    impl->OnUnmap(address, size);
+}
+
+void PageManager::UpdatePagesCachedCount(VAddr addr, u64 size, s32 delta) {
+    static constexpr u64 PageShift = 12;
+
+    std::scoped_lock lk{mutex};
+    const u64 num_pages = ((addr + size - 1) >> PageShift) - (addr >> PageShift) + 1;
+    const u64 page_start = addr >> PageShift;
+    const u64 page_end = page_start + num_pages;
+
+    const auto pages_interval =
+        decltype(cached_pages)::interval_type::right_open(page_start, page_end);
+    if (delta > 0) {
+        cached_pages.add({pages_interval, delta});
+    }
+
+    const auto& range = cached_pages.equal_range(pages_interval);
+    for (const auto& [range, count] : boost::make_iterator_range(range)) {
+        const auto interval = range & pages_interval;
+        const VAddr interval_start_addr = boost::icl::first(interval) << PageShift;
+        const VAddr interval_end_addr = boost::icl::last_next(interval) << PageShift;
+        const u32 interval_size = interval_end_addr - interval_start_addr;
+        if (delta > 0 && count == delta) {
+            impl->Protect(interval_start_addr, interval_size, false);
+        } else if (delta < 0 && count == -delta) {
+            impl->Protect(interval_start_addr, interval_size, true);
+        } else {
+            ASSERT(count >= 0);
+        }
+    }
+
+    if (delta < 0) {
+        cached_pages.add({pages_interval, delta});
+    }
+}
+
+} // namespace VideoCore

src/video_core/page_manager.h

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <memory>
+#include <mutex>
+#include <boost/icl/interval_map.hpp>
+#include "common/types.h"
+
+namespace Vulkan {
+class Rasterizer;
+}
+
+namespace VideoCore {
+
+class PageManager {
+public:
+    explicit PageManager(Vulkan::Rasterizer* rasterizer);
+    ~PageManager();
+
+    /// Register a range of mapped gpu memory.
+    void OnGpuMap(VAddr address, size_t size);
+
+    /// Unregister a range of gpu memory that was unmapped.
+    void OnGpuUnmap(VAddr address, size_t size);
+
+    /// Increase/decrease the number of surface in pages touching the specified region
+    void UpdatePagesCachedCount(VAddr addr, u64 size, s32 delta);
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl;
+    Vulkan::Rasterizer* rasterizer;
+    std::mutex mutex;
+    boost::icl::interval_map<VAddr, s32> cached_pages;
+};
+
+} // namespace VideoCore

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -67,8 +67,8 @@ RendererVulkan::RendererVulkan(Frontend::WindowSDL& window_, AmdGpu::Liverpool*
     : window{window_}, liverpool{liverpool_},
       instance{window, Config::getGpuId(), Config::vkValidationEnabled()}, draw_scheduler{instance},
       present_scheduler{instance}, flip_scheduler{instance}, swapchain{instance, window},
-      texture_cache{instance, draw_scheduler} {
+      rasterizer{std::make_unique<Rasterizer>(instance, draw_scheduler, liverpool)},
-    rasterizer = std::make_unique<Rasterizer>(instance, draw_scheduler, texture_cache, liverpool);
+      texture_cache{rasterizer->GetTextureCache()} {
     const u32 num_images = swapchain.GetImageCount();
     const vk::Device device = instance.GetDevice();
 

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -47,7 +47,7 @@ public:
     Frame* PrepareFrame(const Libraries::VideoOut::BufferAttributeGroup& attribute,
                         VAddr cpu_address, bool is_eop) {
         const auto info = VideoCore::ImageInfo{attribute, cpu_address};
-        const auto image_id = texture_cache.FindImage(info, cpu_address);
+        const auto image_id = texture_cache.FindImage(info, false);
         auto& image = texture_cache.GetImage(image_id);
         return PrepareFrameInternal(image, is_eop);
     }
@@ -61,7 +61,7 @@ public:
         const Libraries::VideoOut::BufferAttributeGroup& attribute, VAddr cpu_address) {
         vo_buffers_addr.emplace_back(cpu_address);
         const auto info = VideoCore::ImageInfo{attribute, cpu_address};
-        const auto image_id = texture_cache.FindImage(info, cpu_address);
+        const auto image_id = texture_cache.FindImage(info, false);
         return texture_cache.GetImage(image_id);
     }
 
@@ -88,7 +88,7 @@ private:
     Scheduler flip_scheduler;
     Swapchain swapchain;
     std::unique_ptr<Rasterizer> rasterizer;
-    VideoCore::TextureCache texture_cache;
+    VideoCore::TextureCache& texture_cache;
     vk::UniqueCommandPool command_pool;
     std::vector<Frame> present_frames;
     std::queue<Frame*> free_queue;

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -3,7 +3,7 @@
 
 #include <boost/container/small_vector.hpp>
 #include "common/alignment.h"
-#include "core/memory.h"
+#include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_compute_pipeline.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
@@ -51,6 +51,12 @@ ComputePipeline::ComputePipeline(const Instance& instance_, Scheduler& scheduler
         });
     }
 
+    const vk::PushConstantRange push_constants = {
+        .stageFlags = vk::ShaderStageFlagBits::eCompute,
+        .offset = 0,
+        .size = sizeof(Shader::PushData),
+    };
+
     const vk::DescriptorSetLayoutCreateInfo desc_layout_ci = {
         .flags = vk::DescriptorSetLayoutCreateFlagBits::ePushDescriptorKHR,
         .bindingCount = static_cast<u32>(bindings.size()),
@@ -62,8 +68,8 @@ ComputePipeline::ComputePipeline(const Instance& instance_, Scheduler& scheduler
     const vk::PipelineLayoutCreateInfo layout_info = {
         .setLayoutCount = 1U,
         .pSetLayouts = &set_layout,
-        .pushConstantRangeCount = 0,
+        .pushConstantRangeCount = 1U,
-        .pPushConstantRanges = nullptr,
+        .pPushConstantRanges = &push_constants,
     };
     pipeline_layout = instance.GetDevice().createPipelineLayoutUnique(layout_info);
 
@@ -82,35 +88,18 @@ ComputePipeline::ComputePipeline(const Instance& instance_, Scheduler& scheduler
 
 ComputePipeline::~ComputePipeline() = default;
 
-bool ComputePipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& staging,
+bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
                                     VideoCore::TextureCache& texture_cache) const {
     // Bind resource buffers and textures.
     boost::container::static_vector<vk::DescriptorBufferInfo, 16> buffer_infos;
     boost::container::static_vector<vk::DescriptorImageInfo, 16> image_infos;
     boost::container::small_vector<vk::WriteDescriptorSet, 16> set_writes;
+    Shader::PushData push_data{};
     u32 binding{};
 
     for (const auto& buffer : info.buffers) {
         const auto vsharp = buffer.GetVsharp(info);
-        const u32 size = vsharp.GetSize();
         const VAddr address = vsharp.base_address;
-        if (buffer.is_storage) {
-            texture_cache.OnCpuWrite(address);
-        }
-        const u32 offset = staging.Copy(address, size,
-                                        buffer.is_storage ? instance.StorageMinAlignment()
-                                                          : instance.UniformMinAlignment());
-        buffer_infos.emplace_back(staging.Handle(), offset, size);
-        set_writes.push_back({
-            .dstSet = VK_NULL_HANDLE,
-            .dstBinding = binding++,
-            .dstArrayElement = 0,
-            .descriptorCount = 1,
-            .descriptorType = buffer.is_storage ? vk::DescriptorType::eStorageBuffer
-                                                : vk::DescriptorType::eUniformBuffer,
-            .pBufferInfo = &buffer_infos.back(),
-        });
-
         // Most of the time when a metadata is updated with a shader it gets cleared. It means we
         // can skip the whole dispatch and update the tracked state instead. Also, it is not
         // intended to be consumed and in such rare cases (e.g. HTile introspection, CRAA) we will
@@ -125,6 +114,30 @@ bool ComputePipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& s
                 LOG_WARNING(Render_Vulkan, "Unexpected metadata read by a CS shader (buffer)");
             }
         }
+        const u32 size = vsharp.GetSize();
+        if (buffer.is_written) {
+            texture_cache.InvalidateMemory(address, size);
+        }
+        const u32 alignment =
+            buffer.is_storage ? instance.StorageMinAlignment() : instance.UniformMinAlignment();
+        const auto [vk_buffer, offset] =
+            buffer_cache.ObtainBuffer(address, size, buffer.is_written);
+        const u32 offset_aligned = Common::AlignDown(offset, alignment);
+        const u32 adjust = offset - offset_aligned;
+        if (adjust != 0) {
+            ASSERT(adjust % 4 == 0);
+            push_data.AddOffset(binding, adjust);
+        }
+        buffer_infos.emplace_back(vk_buffer->Handle(), offset_aligned, size + adjust);
+        set_writes.push_back({
+            .dstSet = VK_NULL_HANDLE,
+            .dstBinding = binding++,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = buffer.is_storage ? vk::DescriptorType::eStorageBuffer
+                                                : vk::DescriptorType::eUniformBuffer,
+            .pBufferInfo = &buffer_infos.back(),
+        });
     }
 
     for (const auto& image_desc : info.images) {
@@ -168,6 +181,8 @@ bool ComputePipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& s
     }
 
     const auto cmdbuf = scheduler.CommandBuffer();
+    cmdbuf.pushConstants(*pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0u, sizeof(push_data),
+                         &push_data);
     cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, *pipeline_layout, 0, set_writes);
     return true;
 }

src/video_core/renderer_vulkan/vk_compute_pipeline.h

@@ -6,13 +6,10 @@
 #include "shader_recompiler/runtime_info.h"
 #include "video_core/renderer_vulkan/vk_common.h"
 
-namespace Core {
-class MemoryManager;
-}
-
 namespace VideoCore {
+class BufferCache;
 class TextureCache;
-}
+} // namespace VideoCore
 
 namespace Vulkan {
 
@@ -31,7 +28,7 @@ public:
         return *pipeline;
     }
 
-    bool BindResources(Core::MemoryManager* memory, StreamBuffer& staging,
+    bool BindResources(VideoCore::BufferCache& buffer_cache,
                        VideoCore::TextureCache& texture_cache) const;
 
 private:

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -1,13 +1,14 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
-
+#pragma clang optimize off
 #include <algorithm>
 #include <boost/container/small_vector.hpp>
 #include <boost/container/static_vector.hpp>
 
+#include "common/alignment.h"
 #include "common/assert.h"
-#include "core/memory.h"
 #include "video_core/amdgpu/resource.h"
+#include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_graphics_pipeline.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
@@ -32,9 +33,9 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
     BuildDescSetLayout();
 
     const vk::PushConstantRange push_constants = {
-        .stageFlags = vk::ShaderStageFlagBits::eVertex,
+        .stageFlags = vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eFragment,
         .offset = 0,
-        .size = 2 * sizeof(u32),
+        .size = sizeof(Shader::PushData),
     };
 
     const vk::DescriptorSetLayout set_layout = *desc_layout;
@@ -328,25 +329,36 @@ void GraphicsPipeline::BuildDescSetLayout() {
     desc_layout = instance.GetDevice().createDescriptorSetLayoutUnique(desc_layout_ci);
 }
 
-void GraphicsPipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& staging,
+void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
+                                     VideoCore::BufferCache& buffer_cache,
                                      VideoCore::TextureCache& texture_cache) const {
-    BindVertexBuffers(staging);
-
     // Bind resource buffers and textures.
     boost::container::static_vector<vk::DescriptorBufferInfo, 16> buffer_infos;
     boost::container::static_vector<vk::DescriptorImageInfo, 32> image_infos;
     boost::container::small_vector<vk::WriteDescriptorSet, 16> set_writes;
+    Shader::PushData push_data{};
     u32 binding{};
 
     for (const auto& stage : stages) {
+        if (stage.uses_step_rates) {
+            push_data.step0 = regs.vgt_instance_step_rate_0;
+            push_data.step1 = regs.vgt_instance_step_rate_1;
+        }
         for (const auto& buffer : stage.buffers) {
             const auto vsharp = buffer.GetVsharp(stage);
             const VAddr address = vsharp.base_address;
             const u32 size = vsharp.GetSize();
-            const u32 offset = staging.Copy(address, size,
+            const u32 alignment =
-                                            buffer.is_storage ? instance.StorageMinAlignment()
+                buffer.is_storage ? instance.StorageMinAlignment() : instance.UniformMinAlignment();
-                                                              : instance.UniformMinAlignment());
+            const auto [vk_buffer, offset] =
-            buffer_infos.emplace_back(staging.Handle(), offset, size);
+                buffer_cache.ObtainBuffer(address, size, buffer.is_written);
+            const u32 offset_aligned = Common::AlignDown(offset, alignment);
+            const u32 adjust = offset - offset_aligned;
+            if (adjust != 0) {
+                ASSERT(adjust % 4 == 0);
+                push_data.AddOffset(binding, adjust);
+            }
+            buffer_infos.emplace_back(vk_buffer->Handle(), offset_aligned, size + adjust);
             set_writes.push_back({
                 .dstSet = VK_NULL_HANDLE,
                 .dstBinding = binding++,
@@ -406,86 +418,15 @@ void GraphicsPipeline::BindResources(Core::MemoryManager* memory, StreamBuffer&
         }
     }
 
-    if (!set_writes.empty()) {
     const auto cmdbuf = scheduler.CommandBuffer();
+    if (!set_writes.empty()) {
         cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eGraphics, *pipeline_layout, 0,
                                     set_writes);
     }
-}
+    cmdbuf.pushConstants(*pipeline_layout,
-
+                         vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eFragment, 0U,
-void GraphicsPipeline::BindVertexBuffers(StreamBuffer& staging) const {
+                         sizeof(push_data), &push_data);
-    const auto& vs_info = stages[u32(Shader::Stage::Vertex)];
+    cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, Handle());
-    if (vs_info.vs_inputs.empty()) {
-        return;
-    }
-
-    std::array<vk::Buffer, MaxVertexBufferCount> host_buffers;
-    std::array<vk::DeviceSize, MaxVertexBufferCount> host_offsets;
-    boost::container::static_vector<AmdGpu::Buffer, MaxVertexBufferCount> guest_buffers;
-
-    struct BufferRange {
-        VAddr base_address;
-        VAddr end_address;
-        u64 offset; // offset in the mapped memory
-
-        size_t GetSize() const {
-            return end_address - base_address;
-        }
-    };
-
-    // Calculate buffers memory overlaps
-    boost::container::static_vector<BufferRange, MaxVertexBufferCount> ranges{};
-    for (const auto& input : vs_info.vs_inputs) {
-        if (input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate0 ||
-            input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate1) {
-            continue;
-        }
-
-        const auto& buffer = vs_info.ReadUd<AmdGpu::Buffer>(input.sgpr_base, input.dword_offset);
-        if (buffer.GetSize() == 0) {
-            continue;
-        }
-        guest_buffers.emplace_back(buffer);
-        ranges.emplace_back(buffer.base_address, buffer.base_address + buffer.GetSize());
-    }
-    std::ranges::sort(ranges, [](const BufferRange& lhv, const BufferRange& rhv) {
-        return lhv.base_address < rhv.base_address;
-    });
-
-    boost::container::static_vector<BufferRange, MaxVertexBufferCount> ranges_merged{ranges[0]};
-    for (auto range : ranges) {
-        auto& prev_range = ranges_merged.back();
-        if (prev_range.end_address < range.base_address) {
-            ranges_merged.emplace_back(range);
-        } else {
-            prev_range.end_address = std::max(prev_range.end_address, range.end_address);
-        }
-    }
-
-    // Map buffers
-    for (auto& range : ranges_merged) {
-        range.offset = staging.Copy(range.base_address, range.GetSize(), 4);
-    }
-
-    // Bind vertex buffers
-    const size_t num_buffers = guest_buffers.size();
-    for (u32 i = 0; i < num_buffers; ++i) {
-        const auto& buffer = guest_buffers[i];
-        const auto& host_buffer = std::ranges::find_if(
-            ranges_merged.cbegin(), ranges_merged.cend(), [&](const BufferRange& range) {
-                return (buffer.base_address >= range.base_address &&
-                        buffer.base_address < range.end_address);
-            });
-        assert(host_buffer != ranges_merged.cend());
-
-        host_buffers[i] = staging.Handle();
-        host_offsets[i] = host_buffer->offset + buffer.base_address - host_buffer->base_address;
-    }
-
-    if (num_buffers > 0) {
-        const auto cmdbuf = scheduler.CommandBuffer();
-        cmdbuf.bindVertexBuffers(0, num_buffers, host_buffers.data(), host_offsets.data());
-    }
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -7,13 +7,10 @@
 #include "video_core/renderer_vulkan/liverpool_to_vk.h"
 #include "video_core/renderer_vulkan/vk_common.h"
 
-namespace Core {
-class MemoryManager;
-}
-
 namespace VideoCore {
+class BufferCache;
 class TextureCache;
-}
+} // namespace VideoCore
 
 namespace Vulkan {
 
@@ -64,7 +61,7 @@ public:
                               std::array<vk::ShaderModule, MaxShaderStages> modules);
     ~GraphicsPipeline();
 
-    void BindResources(Core::MemoryManager* memory, StreamBuffer& staging,
+    void BindResources(const Liverpool::Regs& regs, VideoCore::BufferCache& buffer_cache,
                        VideoCore::TextureCache& texture_cache) const;
 
     vk::Pipeline Handle() const noexcept {
@@ -75,6 +72,10 @@ public:
         return *pipeline_layout;
     }
 
+    const Shader::Info& GetStage(Shader::Stage stage) const noexcept {
+        return stages[u32(stage)];
+    }
+
     bool IsEmbeddedVs() const noexcept {
         static constexpr size_t EmbeddedVsHash = 0x9b2da5cf47f8c29f;
         return key.stage_hashes[u32(Shader::Stage::Vertex)] == EmbeddedVsHash;
@@ -90,7 +91,6 @@ public:
 
 private:
     void BuildDescSetLayout();
-    void BindVertexBuffers(StreamBuffer& staging) const;
 
 private:
     const Instance& instance;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -13,22 +13,16 @@
 
 namespace Vulkan {
 
-static constexpr vk::BufferUsageFlags VertexIndexFlags =
-    vk::BufferUsageFlagBits::eVertexBuffer | vk::BufferUsageFlagBits::eIndexBuffer |
-    vk::BufferUsageFlagBits::eTransferDst | vk::BufferUsageFlagBits::eUniformBuffer |
-    vk::BufferUsageFlagBits::eStorageBuffer;
-
 Rasterizer::Rasterizer(const Instance& instance_, Scheduler& scheduler_,
-                       VideoCore::TextureCache& texture_cache_, AmdGpu::Liverpool* liverpool_)
+                       AmdGpu::Liverpool* liverpool_)
-    : instance{instance_}, scheduler{scheduler_}, texture_cache{texture_cache_},
+    : instance{instance_}, scheduler{scheduler_}, page_manager{this},
-      liverpool{liverpool_}, memory{Core::Memory::Instance()},
+      buffer_cache{instance, scheduler, liverpool_, page_manager},
-      pipeline_cache{instance, scheduler, liverpool},
+      texture_cache{instance, scheduler, buffer_cache, page_manager}, liverpool{liverpool_},
-      vertex_index_buffer{instance, scheduler, VertexIndexFlags, 2_GB, BufferType::Upload} {
+      memory{Core::Memory::Instance()}, pipeline_cache{instance, scheduler, liverpool} {
     if (!Config::nullGpu()) {
         liverpool->BindRasterizer(this);
     }
-
+    memory->SetRasterizer(this);
-    memory->SetInstance(&instance);
 }
 
 Rasterizer::~Rasterizer() = default;
@@ -38,29 +32,24 @@ void Rasterizer::Draw(bool is_indexed, u32 index_offset) {
 
     const auto cmdbuf = scheduler.CommandBuffer();
     const auto& regs = liverpool->regs;
-    const u32 num_indices = SetupIndexBuffer(is_indexed, index_offset);
     const GraphicsPipeline* pipeline = pipeline_cache.GetGraphicsPipeline();
     if (!pipeline) {
         return;
     }
 
     try {
-        pipeline->BindResources(memory, vertex_index_buffer, texture_cache);
+        pipeline->BindResources(regs, buffer_cache, texture_cache);
     } catch (...) {
         UNREACHABLE();
     }
 
+    const auto& vs_info = pipeline->GetStage(Shader::Stage::Vertex);
+    buffer_cache.BindVertexBuffers(vs_info);
+    const u32 num_indices = buffer_cache.BindIndexBuffer(is_indexed, index_offset);
+
     BeginRendering();
     UpdateDynamicState(*pipeline);
 
-    cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline->Handle());
-
-    const u32 step_rates[] = {
-        regs.vgt_instance_step_rate_0,
-        regs.vgt_instance_step_rate_1,
-    };
-    cmdbuf.pushConstants(pipeline->GetLayout(), vk::ShaderStageFlagBits::eVertex, 0u,
-                         sizeof(step_rates), &step_rates);
     if (is_indexed) {
         cmdbuf.drawIndexed(num_indices, regs.num_instances.NumInstances(), 0, 0, 0);
     } else {
@@ -82,8 +71,7 @@ void Rasterizer::DispatchDirect() {
     }
 
     try {
-        const auto has_resources =
+        const auto has_resources = pipeline->BindResources(buffer_cache, texture_cache);
-            pipeline->BindResources(memory, vertex_index_buffer, texture_cache);
         if (!has_resources) {
             return;
         }
@@ -131,7 +119,7 @@ void Rasterizer::BeginRendering() {
         state.color_images[state.num_color_attachments] = image.image;
         state.color_attachments[state.num_color_attachments++] = {
             .imageView = *image_view.image_view,
-            .imageLayout = vk::ImageLayout::eGeneral,
+            .imageLayout = vk::ImageLayout::eColorAttachmentOptimal,
             .loadOp = is_clear ? vk::AttachmentLoadOp::eClear : vk::AttachmentLoadOp::eLoad,
             .storeOp = vk::AttachmentStoreOp::eStore,
             .clearValue =
@@ -168,45 +156,19 @@ void Rasterizer::BeginRendering() {
     scheduler.BeginRendering(state);
 }
 
-u32 Rasterizer::SetupIndexBuffer(bool& is_indexed, u32 index_offset) {
+void Rasterizer::InvalidateMemory(VAddr addr, u64 size) {
-    // Emulate QuadList primitive type with CPU made index buffer.
+    buffer_cache.InvalidateMemory(addr, size);
-    const auto& regs = liverpool->regs;
+    texture_cache.InvalidateMemory(addr, size);
-    if (liverpool->regs.primitive_type == Liverpool::PrimitiveType::QuadList) {
-        // ASSERT_MSG(!is_indexed, "Using QuadList primitive with indexed draw");
-        is_indexed = true;
-
-        // Emit indices.
-        const u32 index_size = 3 * regs.num_indices;
-        const auto [data, offset, _] = vertex_index_buffer.Map(index_size);
-        LiverpoolToVK::EmitQuadToTriangleListIndices(data, regs.num_indices);
-        vertex_index_buffer.Commit(index_size);
-
-        // Bind index buffer.
-        const auto cmdbuf = scheduler.CommandBuffer();
-        cmdbuf.bindIndexBuffer(vertex_index_buffer.Handle(), offset, vk::IndexType::eUint16);
-        return index_size / sizeof(u16);
-    }
-    if (!is_indexed) {
-        return regs.num_indices;
 }
 
-    // Figure out index type and size.
+void Rasterizer::MapMemory(VAddr addr, u64 size) {
-    const bool is_index16 = regs.index_buffer_type.index_type == Liverpool::IndexType::Index16;
+    page_manager.OnGpuMap(addr, size);
-    const vk::IndexType index_type = is_index16 ? vk::IndexType::eUint16 : vk::IndexType::eUint32;
+}
-    const u32 index_size = is_index16 ? sizeof(u16) : sizeof(u32);
 
-    // Upload index data to stream buffer.
+void Rasterizer::UnmapMemory(VAddr addr, u64 size) {
-    const auto index_address = regs.index_base_address.Address<const void*>();
+    buffer_cache.InvalidateMemory(addr, size);
-    const u32 index_buffer_size = (index_offset + regs.num_indices) * index_size;
+    texture_cache.UnmapMemory(addr, size);
-    const auto [data, offset, _] = vertex_index_buffer.Map(index_buffer_size);
+    page_manager.OnGpuUnmap(addr, size);
-    std::memcpy(data, index_address, index_buffer_size);
-    vertex_index_buffer.Commit(index_buffer_size);
-
-    // Bind index buffer.
-    const auto cmdbuf = scheduler.CommandBuffer();
-    cmdbuf.bindIndexBuffer(vertex_index_buffer.Handle(), offset + index_offset * index_size,
-                           index_type);
-    return regs.num_indices;
 }
 
 void Rasterizer::UpdateDynamicState(const GraphicsPipeline& pipeline) {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -3,8 +3,10 @@
 
 #pragma once
 
+#include "video_core/buffer_cache/buffer_cache.h"
+#include "video_core/page_manager.h"
 #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
-#include "video_core/renderer_vulkan/vk_stream_buffer.h"
+#include "video_core/texture_cache/texture_cache.h"
 
 namespace AmdGpu {
 struct Liverpool;
@@ -14,10 +16,6 @@ namespace Core {
 class MemoryManager;
 }
 
-namespace VideoCore {
-class TextureCache;
-}
-
 namespace Vulkan {
 
 class Scheduler;
@@ -26,9 +24,13 @@ class GraphicsPipeline;
 class Rasterizer {
 public:
     explicit Rasterizer(const Instance& instance, Scheduler& scheduler,
-                        VideoCore::TextureCache& texture_cache, AmdGpu::Liverpool* liverpool);
+                        AmdGpu::Liverpool* liverpool);
     ~Rasterizer();
 
+    [[nodiscard]] VideoCore::TextureCache& GetTextureCache() noexcept {
+        return texture_cache;
+    }
+
     void Draw(bool is_indexed, u32 index_offset = 0);
 
     void DispatchDirect();
@@ -36,12 +38,13 @@ public:
     void ScopeMarkerBegin(const std::string& str);
     void ScopeMarkerEnd();
 
+    void InvalidateMemory(VAddr addr, u64 size);
+    void MapMemory(VAddr addr, u64 size);
+    void UnmapMemory(VAddr addr, u64 size);
+
     u64 Flush();
 
 private:
-    u32 SetupIndexBuffer(bool& is_indexed, u32 index_offset);
-    void MapMemory(VAddr addr, size_t size);
-
     void BeginRendering();
 
     void UpdateDynamicState(const GraphicsPipeline& pipeline);
@@ -51,11 +54,12 @@ private:
 private:
     const Instance& instance;
     Scheduler& scheduler;
-    VideoCore::TextureCache& texture_cache;
+    VideoCore::PageManager page_manager;
+    VideoCore::BufferCache buffer_cache;
+    VideoCore::TextureCache texture_cache;
     AmdGpu::Liverpool* liverpool;
     Core::MemoryManager* memory;
     PipelineCache pipeline_cache;
-    StreamBuffer vertex_index_buffer;
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -6,6 +6,7 @@
 #include <condition_variable>
 #include <boost/container/static_vector.hpp>
 #include "common/types.h"
+#include "common/unique_function.h"
 #include "video_core/renderer_vulkan/vk_master_semaphore.h"
 #include "video_core/renderer_vulkan/vk_resource_pool.h"
 
@@ -97,8 +98,8 @@ public:
     }
 
     /// Defers an operation until the gpu has reached the current cpu tick.
-    void DeferOperation(auto&& func) {
+    void DeferOperation(Common::UniqueFunction<void>&& func) {
-        pending_ops.emplace(func, CurrentTick());
+        pending_ops.emplace(std::move(func), CurrentTick());
     }
 
     static std::mutex submit_mutex;
@@ -115,7 +116,7 @@ private:
     vk::CommandBuffer current_cmdbuf;
     std::condition_variable_any event_cv;
     struct PendingOp {
-        std::function<void()> callback;
+        Common::UniqueFunction<void> callback;
         u64 gpu_tick;
     };
     std::queue<PendingOp> pending_ops;

src/video_core/renderer_vulkan/vk_stream_buffer.h

@@ -22,8 +22,6 @@ class Instance;
 class Scheduler;
 
 class StreamBuffer final {
-    static constexpr std::size_t MAX_BUFFER_VIEWS = 3;
-
 public:
     explicit StreamBuffer(const Instance& instance, Scheduler& scheduler,
                           vk::BufferUsageFlags usage, u64 size,

src/video_core/texture_cache/image_info.cpp

@@ -260,7 +260,7 @@ ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
         case AmdGpu::TilingMode::Display_MacroTiled:
         case AmdGpu::TilingMode::Texture_MacroTiled:
         case AmdGpu::TilingMode::Depth_MacroTiled: {
-            ASSERT(!props.is_cube && !props.is_block);
+            // ASSERT(!props.is_cube && !props.is_block);
             ASSERT(num_samples == 1);
             std::tie(mip_info.pitch, mip_info.size) =
                 ImageSizeMacroTiled(mip_w, mip_h, bpp, num_samples, image.tiling_index);

src/video_core/texture_cache/image_view.cpp

@@ -61,23 +61,24 @@ vk::Format TrySwizzleFormat(vk::Format format, u32 dst_sel) {
     return format;
 }
 
-ImageViewInfo::ImageViewInfo(const AmdGpu::Image& image, bool is_storage) noexcept
+ImageViewInfo::ImageViewInfo(const AmdGpu::Image& image, bool is_storage_) noexcept
-    : is_storage{is_storage} {
+    : is_storage{is_storage_} {
     type = ConvertImageViewType(image.GetType());
     format = Vulkan::LiverpoolToVK::SurfaceFormat(image.GetDataFmt(), image.GetNumberFmt());
     range.base.level = image.base_level;
     range.base.layer = image.base_array;
     range.extent.levels = image.last_level + 1;
     range.extent.layers = image.last_array + 1;
+    if (!is_storage) {
         mapping.r = ConvertComponentSwizzle(image.dst_sel_x);
         mapping.g = ConvertComponentSwizzle(image.dst_sel_y);
         mapping.b = ConvertComponentSwizzle(image.dst_sel_z);
         mapping.a = ConvertComponentSwizzle(image.dst_sel_w);
+    }
     // Check for unfortunate case of storage images being swizzled
     const u32 num_comps = AmdGpu::NumComponents(image.GetDataFmt());
     const u32 dst_sel = image.DstSelect();
     if (is_storage && !IsIdentityMapping(dst_sel, num_comps)) {
-        mapping = vk::ComponentMapping{};
         if (auto new_format = TrySwizzleFormat(format, dst_sel); new_format != format) {
             format = new_format;
             return;

src/video_core/texture_cache/texture_cache.cpp

@@ -3,103 +3,23 @@
 
 #include <xxhash.h>
 #include "common/assert.h"
-#include "common/config.h"
+#include "video_core/page_manager.h"
-#include "core/virtual_memory.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/texture_cache/texture_cache.h"
 #include "video_core/texture_cache/tile_manager.h"
 
-#ifndef _WIN64
-#include <signal.h>
-#include <sys/mman.h>
-
-#define PAGE_NOACCESS PROT_NONE
-#define PAGE_READWRITE (PROT_READ | PROT_WRITE)
-#define PAGE_READONLY PROT_READ
-#else
-#include <windows.h>
-
-void mprotect(void* addr, size_t len, int prot) {
-    DWORD old_prot{};
-    BOOL result = VirtualProtect(addr, len, prot, &old_prot);
-    ASSERT_MSG(result != 0, "Region protection failed");
-}
-
-#endif
-
 namespace VideoCore {
 
-static TextureCache* g_texture_cache = nullptr;
-
-#ifndef _WIN64
-void GuestFaultSignalHandler(int sig, siginfo_t* info, void* raw_context) {
-    ucontext_t* ctx = reinterpret_cast<ucontext_t*>(raw_context);
-    const VAddr address = reinterpret_cast<VAddr>(info->si_addr);
-
-#ifdef __APPLE__
-    const u32 err = ctx->uc_mcontext->__es.__err;
-#else
-    const greg_t err = ctx->uc_mcontext.gregs[REG_ERR];
-#endif
-
-    if (err & 0x2) {
-        g_texture_cache->OnCpuWrite(address);
-    } else {
-        // Read not supported!
-        UNREACHABLE();
-    }
-}
-#else
-LONG WINAPI GuestFaultSignalHandler(EXCEPTION_POINTERS* pExp) noexcept {
-    const u32 ec = pExp->ExceptionRecord->ExceptionCode;
-    if (ec == EXCEPTION_ACCESS_VIOLATION) {
-        const auto info = pExp->ExceptionRecord->ExceptionInformation;
-        if (info[0] == 1) { // Write violation
-            g_texture_cache->OnCpuWrite(info[1]);
-            return EXCEPTION_CONTINUE_EXECUTION;
-        } /* else {
-            UNREACHABLE();
-        }*/
-    }
-    return EXCEPTION_CONTINUE_SEARCH; // pass further
-}
-#endif
-
 static constexpr u64 StreamBufferSize = 512_MB;
 static constexpr u64 PageShift = 12;
 
-TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_)
+TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
-    : instance{instance_}, scheduler{scheduler_},
+                           BufferCache& buffer_cache_, PageManager& tracker_)
+    : instance{instance_}, scheduler{scheduler_}, buffer_cache{buffer_cache_}, tracker{tracker_},
       staging{instance, scheduler, vk::BufferUsageFlagBits::eTransferSrc, StreamBufferSize,
               Vulkan::BufferType::Upload},
       tile_manager{instance, scheduler} {
-
-#ifndef _WIN64
-#ifdef __APPLE__
-    // Read-only memory write results in SIGBUS on Apple.
-    static constexpr int SignalType = SIGBUS;
-#else
-    static constexpr int SignalType = SIGSEGV;
-#endif
-
-    sigset_t signal_mask;
-    sigemptyset(&signal_mask);
-    sigaddset(&signal_mask, SignalType);
-
-    using HandlerType = decltype(sigaction::sa_sigaction);
-
-    struct sigaction guest_access_fault {};
-    guest_access_fault.sa_flags = SA_SIGINFO | SA_ONSTACK;
-    guest_access_fault.sa_sigaction = &GuestFaultSignalHandler;
-    guest_access_fault.sa_mask = signal_mask;
-    sigaction(SignalType, &guest_access_fault, nullptr);
-#else
-    veh_handle = AddVectoredExceptionHandler(0, GuestFaultSignalHandler);
-    ASSERT_MSG(veh_handle, "Failed to register an exception handler");
-#endif
-    g_texture_cache = this;
-
     ImageInfo info;
     info.pixel_format = vk::Format::eR8G8B8A8Unorm;
     info.type = vk::ImageType::e2D;
@@ -110,15 +30,11 @@ TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler&
     void(slot_image_views.insert(instance, view_info, slot_images[null_id], null_id));
 }
 
-TextureCache::~TextureCache() {
+TextureCache::~TextureCache() = default;
-#if _WIN64
-    RemoveVectoredExceptionHandler(veh_handle);
-#endif
-}
 
-void TextureCache::OnCpuWrite(VAddr address) {
+void TextureCache::InvalidateMemory(VAddr address, size_t size) {
-    std::unique_lock lock{m_page_table};
+    std::unique_lock lock{mutex};
-    ForEachImageInRegion(address, 1 << PageShift, [&](ImageId image_id, Image& image) {
+    ForEachImageInRegion(address, size, [&](ImageId image_id, Image& image) {
         // Ensure image is reuploaded when accessed again.
         image.flags |= ImageFlagBits::CpuModified;
         // Untrack image, so the range is unprotected and the guest can write freely.
@@ -126,8 +42,28 @@ void TextureCache::OnCpuWrite(VAddr address) {
     });
 }
 
+void TextureCache::UnmapMemory(VAddr cpu_addr, size_t size) {
+    std::scoped_lock lk{mutex};
+
+    boost::container::small_vector<ImageId, 16> deleted_images;
+    ForEachImageInRegion(cpu_addr, size, [&](ImageId id, Image&) { deleted_images.push_back(id); });
+    for (const ImageId id : deleted_images) {
+        Image& image = slot_images[id];
+        if (True(image.flags & ImageFlagBits::Tracked)) {
+            UntrackImage(image, id);
+        }
+        // TODO: Download image data back to host.
+        UnregisterImage(id);
+        DeleteImage(id);
+    }
+}
+
 ImageId TextureCache::FindImage(const ImageInfo& info, bool refresh_on_create) {
-    std::unique_lock lock{m_page_table};
+    if (info.guest_address == 0) [[unlikely]] {
+        return NULL_IMAGE_VIEW_ID;
+    }
+
+    std::unique_lock lock{mutex};
     boost::container::small_vector<ImageId, 2> image_ids;
     ForEachImageInRegion(
         info.guest_address, info.guest_size_bytes, [&](ImageId image_id, Image& image) {
@@ -183,10 +119,6 @@ ImageView& TextureCache::RegisterImageView(ImageId image_id, const ImageViewInfo
 }
 
 ImageView& TextureCache::FindTexture(const ImageInfo& info, const ImageViewInfo& view_info) {
-    if (info.guest_address == 0) [[unlikely]] {
-        return slot_image_views[NULL_IMAGE_VIEW_ID];
-    }
-
     const ImageId image_id = FindImage(info);
     Image& image = slot_images[image_id];
     auto& usage = image.info.usage;
@@ -344,9 +276,6 @@ void TextureCache::RefreshImage(Image& image) {
     }
 
     cmdbuf.copyBufferToImage(buffer, image.image, vk::ImageLayout::eTransferDstOptimal, image_copy);
-
-    image.Transit(vk::ImageLayout::eGeneral,
-                  vk::AccessFlagBits::eMemoryWrite | vk::AccessFlagBits::eMemoryRead);
 }
 
 vk::Sampler TextureCache::GetSampler(const AmdGpu::Sampler& sampler) {
@@ -362,8 +291,6 @@ void TextureCache::RegisterImage(ImageId image_id) {
     image.flags |= ImageFlagBits::Registered;
     ForEachPage(image.cpu_addr, image.info.guest_size_bytes,
                 [this, image_id](u64 page) { page_table[page].push_back(image_id); });
-
-    image.Transit(vk::ImageLayout::eGeneral, vk::AccessFlagBits::eNone);
 }
 
 void TextureCache::UnregisterImage(ImageId image_id) {
@@ -393,7 +320,7 @@ void TextureCache::TrackImage(Image& image, ImageId image_id) {
         return;
     }
     image.flags |= ImageFlagBits::Tracked;
-    UpdatePagesCachedCount(image.cpu_addr, image.info.guest_size_bytes, 1);
+    tracker.UpdatePagesCachedCount(image.cpu_addr, image.info.guest_size_bytes, 1);
 }
 
 void TextureCache::UntrackImage(Image& image, ImageId image_id) {
@@ -401,40 +328,34 @@ void TextureCache::UntrackImage(Image& image, ImageId image_id) {
         return;
     }
     image.flags &= ~ImageFlagBits::Tracked;
-    UpdatePagesCachedCount(image.cpu_addr, image.info.guest_size_bytes, -1);
+    tracker.UpdatePagesCachedCount(image.cpu_addr, image.info.guest_size_bytes, -1);
 }
 
-void TextureCache::UpdatePagesCachedCount(VAddr addr, u64 size, s32 delta) {
+void TextureCache::DeleteImage(ImageId image_id) {
-    std::scoped_lock lk{mutex};
+    Image& image = slot_images[image_id];
-    const u64 num_pages = ((addr + size - 1) >> PageShift) - (addr >> PageShift) + 1;
+    ASSERT_MSG(False(image.flags & ImageFlagBits::Tracked), "Image was not untracked");
-    const u64 page_start = addr >> PageShift;
+    ASSERT_MSG(False(image.flags & ImageFlagBits::Registered), "Image was not unregistered");
-    const u64 page_end = page_start + num_pages;
 
-    const auto pages_interval =
+    // Remove any registered meta areas.
-        decltype(cached_pages)::interval_type::right_open(page_start, page_end);
+    const auto& meta_info = image.info.meta_info;
-    if (delta > 0) {
+    if (meta_info.cmask_addr) {
-        cached_pages.add({pages_interval, delta});
+        surface_metas.erase(meta_info.cmask_addr);
+    }
+    if (meta_info.fmask_addr) {
+        surface_metas.erase(meta_info.fmask_addr);
+    }
+    if (meta_info.htile_addr) {
+        surface_metas.erase(meta_info.htile_addr);
     }
 
-    const auto& range = cached_pages.equal_range(pages_interval);
+    // Reclaim image and any image views it references.
-    for (const auto& [range, count] : boost::make_iterator_range(range)) {
+    scheduler.DeferOperation([this, image_id] {
-        const auto interval = range & pages_interval;
+        Image& image = slot_images[image_id];
-        const VAddr interval_start_addr = boost::icl::first(interval) << PageShift;
+        for (const ImageViewId image_view_id : image.image_view_ids) {
-        const VAddr interval_end_addr = boost::icl::last_next(interval) << PageShift;
+            slot_image_views.erase(image_view_id);
-        const u32 interval_size = interval_end_addr - interval_start_addr;
-        void* addr = reinterpret_cast<void*>(interval_start_addr);
-        if (delta > 0 && count == delta) {
-            mprotect(addr, interval_size, PAGE_READONLY);
-        } else if (delta < 0 && count == -delta) {
-            mprotect(addr, interval_size, PAGE_READWRITE);
-        } else {
-            ASSERT(count >= 0);
-        }
-    }
-
-    if (delta < 0) {
-        cached_pages.add({pages_interval, delta});
         }
+        slot_images.erase(image_id);
+    });
 }
 
 } // namespace VideoCore

src/video_core/texture_cache/texture_cache.h

@@ -21,31 +21,27 @@ struct BufferAttributeGroup;
 
 namespace VideoCore {
 
+class BufferCache;
+class PageManager;
+
 class TextureCache {
     // This is the page shift for adding images into the hash map. It isn't related to
     // the page size of the guest or the host and is chosen for convenience. A number too
     // small will increase the number of hash map lookups per image, while too large will
     // increase the number of images per page.
-    static constexpr u64 PageBits = 20;
+    static constexpr u64 PageBits = 22;
     static constexpr u64 PageMask = (1ULL << PageBits) - 1;
 
-    struct MetaDataInfo {
-        enum class Type {
-            CMask,
-            FMask,
-            HTile,
-        };
-
-        Type type;
-        bool is_cleared;
-    };
-
 public:
-    explicit TextureCache(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler);
+    explicit TextureCache(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
+                          BufferCache& buffer_cache, PageManager& tracker);
     ~TextureCache();
 
     /// Invalidates any image in the logical page range.
-    void OnCpuWrite(VAddr address);
+    void InvalidateMemory(VAddr address, size_t size);
+
+    /// Evicts any images that overlap the unmapped range.
+    void UnmapMemory(VAddr cpu_addr, size_t size);
 
     /// Retrieves the image handle of the image with the provided attributes.
     [[nodiscard]] ImageId FindImage(const ImageInfo& info, bool refresh_on_create = true);
@@ -166,12 +162,14 @@ private:
     /// Stop tracking CPU reads and writes for image
     void UntrackImage(Image& image, ImageId image_id);
 
-    /// Increase/decrease the number of surface in pages touching the specified region
+    /// Removes the image and any views/surface metas that reference it.
-    void UpdatePagesCachedCount(VAddr addr, u64 size, s32 delta);
+    void DeleteImage(ImageId image_id);
 
 private:
     const Vulkan::Instance& instance;
     Vulkan::Scheduler& scheduler;
+    BufferCache& buffer_cache;
+    PageManager& tracker;
     Vulkan::StreamBuffer staging;
     TileManager tile_manager;
     Common::SlotVector<Image> slot_images;
@@ -179,12 +177,18 @@ private:
     tsl::robin_map<u64, Sampler> samplers;
     tsl::robin_pg_map<u64, std::vector<ImageId>> page_table;
     boost::icl::interval_map<VAddr, s32> cached_pages;
-    tsl::robin_map<VAddr, MetaDataInfo> surface_metas;
     std::mutex mutex;
-#ifdef _WIN64
+
-    void* veh_handle{};
+    struct MetaDataInfo {
-#endif
+        enum class Type {
-    std::mutex m_page_table;
+            CMask,
+            FMask,
+            HTile,
+        };
+        Type type;
+        bool is_cleared;
+    };
+    tsl::robin_map<VAddr, MetaDataInfo> surface_metas;
 };
 
 } // namespace VideoCore

src/video_core/texture_cache/tile_manager.cpp

@@ -183,10 +183,12 @@ vk::Format DemoteImageFormatForDetiling(vk::Format format) {
     case vk::Format::eB8G8R8A8Srgb:
     case vk::Format::eB8G8R8A8Unorm:
     case vk::Format::eR8G8B8A8Unorm:
+    case vk::Format::eR8G8B8A8Uint:
     case vk::Format::eR32Sfloat:
     case vk::Format::eR32Uint:
     case vk::Format::eR16G16Sfloat:
         return vk::Format::eR32Uint;
+    case vk::Format::eBc1RgbaSrgbBlock:
     case vk::Format::eBc1RgbaUnormBlock:
     case vk::Format::eBc4UnormBlock:
     case vk::Format::eR32G32Sfloat:
@@ -200,6 +202,7 @@ vk::Format DemoteImageFormatForDetiling(vk::Format format) {
     case vk::Format::eBc5UnormBlock:
     case vk::Format::eBc7SrgbBlock:
     case vk::Format::eBc7UnormBlock:
+    case vk::Format::eR32G32B32A32Sfloat:
         return vk::Format::eR32G32B32A32Uint;
     default:
         break;
@@ -236,8 +239,11 @@ struct DetilerParams {
     u32 sizes[14];
 };
 
+static constexpr size_t StreamBufferSize = 128_MB;
+
 TileManager::TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler)
-    : instance{instance}, scheduler{scheduler} {
+    : instance{instance}, scheduler{scheduler},
+      stream_buffer{instance, scheduler, MemoryUsage::Stream, StreamBufferSize} {
     static const std::array detiler_shaders{
         HostShaders::DETILE_M8X1_COMP,  HostShaders::DETILE_M8X2_COMP,
         HostShaders::DETILE_M32X1_COMP, HostShaders::DETILE_M32X2_COMP,
@@ -336,8 +342,7 @@ TileManager::ScratchBuffer TileManager::AllocBuffer(u32 size, bool is_storage /*
         .flags = !is_storage ? VMA_ALLOCATION_CREATE_HOST_ACCESS_ALLOW_TRANSFER_INSTEAD_BIT |
                                    VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT
                              : static_cast<VmaAllocationCreateFlags>(0),
-        .usage = is_large_buffer ? VMA_MEMORY_USAGE_AUTO_PREFER_HOST
+        .usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE,
-                                 : VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE,
         .requiredFlags = !is_storage ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
                                      : static_cast<VkMemoryPropertyFlags>(0),
     };
@@ -373,37 +378,46 @@ std::optional<vk::Buffer> TileManager::TryDetile(Image& image) {
 
     const auto* detiler = GetDetiler(image);
     if (!detiler) {
+        if (image.info.tiling_mode != AmdGpu::TilingMode::Texture_MacroTiled) {
             LOG_ERROR(Render_Vulkan, "Unsupported tiled image: {} ({})",
                       vk::to_string(image.info.pixel_format), NameOf(image.info.tiling_mode));
+        }
         return std::nullopt;
     }
 
     // Prepare input buffer
-    auto in_buffer = AllocBuffer(image.info.guest_size_bytes);
+    const u32 image_size = image.info.guest_size_bytes;
-    Upload(in_buffer, reinterpret_cast<const void*>(image.info.guest_address),
+    const auto [in_buffer, in_offset] = [&] -> std::pair<vk::Buffer, u32> {
-           image.info.guest_size_bytes);
+        // Use stream buffer for smaller textures.
+        if (image_size <= StreamBufferSize) {
+            u32 offset = stream_buffer.Copy(image.info.guest_address, image_size);
+            return {stream_buffer.Handle(), offset};
+        }
+        // Request temporary host buffer for larger sizes.
+        auto in_buffer = AllocBuffer(image_size);
+        const auto addr = reinterpret_cast<const void*>(image.info.guest_address);
+        Upload(in_buffer, addr, image_size);
+        scheduler.DeferOperation([=, this]() { FreeBuffer(in_buffer); });
+        return {in_buffer.first, 0};
+    }();
 
     // Prepare output buffer
-    auto out_buffer = AllocBuffer(image.info.guest_size_bytes, true);
+    auto out_buffer = AllocBuffer(image_size, true);
-
+    scheduler.DeferOperation([=, this]() { FreeBuffer(out_buffer); });
-    scheduler.DeferOperation([=, this]() {
-        FreeBuffer(in_buffer);
-        FreeBuffer(out_buffer);
-    });
 
     auto cmdbuf = scheduler.CommandBuffer();
     cmdbuf.bindPipeline(vk::PipelineBindPoint::eCompute, *detiler->pl);
 
     const vk::DescriptorBufferInfo input_buffer_info{
-        .buffer = in_buffer.first,
+        .buffer = in_buffer,
-        .offset = 0,
+        .offset = in_offset,
-        .range = image.info.guest_size_bytes,
+        .range = image_size,
     };
 
     const vk::DescriptorBufferInfo output_buffer_info{
         .buffer = out_buffer.first,
         .offset = 0,
-        .range = image.info.guest_size_bytes,
+        .range = image_size,
     };
 
     std::vector<vk::WriteDescriptorSet> set_writes{
@@ -442,16 +456,16 @@ std::optional<vk::Buffer> TileManager::TryDetile(Image& image) {
     cmdbuf.pushConstants(*detiler->pl_layout, vk::ShaderStageFlagBits::eCompute, 0u, sizeof(params),
                          &params);
 
-    ASSERT((image.info.guest_size_bytes % 64) == 0);
+    ASSERT((image_size % 64) == 0);
     const auto bpp = image.info.num_bits * (image.info.props.is_block ? 16u : 1u);
-    const auto num_tiles = image.info.guest_size_bytes / (64 * (bpp / 8));
+    const auto num_tiles = image_size / (64 * (bpp / 8));
     cmdbuf.dispatch(num_tiles, 1, 1);
 
     const vk::BufferMemoryBarrier post_barrier{
         .srcAccessMask = vk::AccessFlagBits::eShaderWrite,
         .dstAccessMask = vk::AccessFlagBits::eTransferRead,
         .buffer = out_buffer.first,
-        .size = image.info.guest_size_bytes,
+        .size = image_size,
     };
     cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eComputeShader,
                            vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlagBits::eByRegion,

src/video_core/texture_cache/tile_manager.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "common/types.h"
+#include "video_core/buffer_cache/buffer.h"
 #include "video_core/renderer_vulkan/vk_stream_buffer.h"
 #include "video_core/texture_cache/image.h"
 
@@ -34,7 +35,7 @@ struct DetilerContext {
 
 class TileManager {
 public:
-    using ScratchBuffer = std::pair<VkBuffer, VmaAllocation>;
+    using ScratchBuffer = std::pair<vk::Buffer, VmaAllocation>;
 
     TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler);
     ~TileManager();
@@ -51,6 +52,7 @@ private:
 private:
     const Vulkan::Instance& instance;
     Vulkan::Scheduler& scheduler;
+    StreamBuffer stream_buffer;
     std::array<DetilerContext, DetilerType::Max> detilers;
 };