suyu/src/video_core/buffer_cache/buffer_cache.h

// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.

#pragma once

#include <algorithm>
#include <array>
#include <deque>
#include <memory>
#include <mutex>
#include <numeric>
#include <span>
#include <unordered_map>
#include <vector>

#include <boost/container/small_vector.hpp>

#include "common/common_types.h"
#include "common/div_ceil.h"
#include "common/microprofile.h"
#include "common/scope_exit.h"
#include "common/settings.h"
#include "core/memory.h"
#include "video_core/buffer_cache/buffer_base.h"
#include "video_core/delayed_destruction_ring.h"
#include "video_core/dirty_flags.h"
#include "video_core/engines/kepler_compute.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/memory_manager.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/texture_cache/slot_vector.h"
#include "video_core/texture_cache/types.h"

namespace VideoCommon {

MICROPROFILE_DECLARE(GPU_PrepareBuffers);
MICROPROFILE_DECLARE(GPU_BindUploadBuffers);
MICROPROFILE_DECLARE(GPU_DownloadMemory);

using BufferId = SlotId;

constexpr u32 NUM_VERTEX_BUFFERS = 32;
constexpr u32 NUM_TRANSFORM_FEEDBACK_BUFFERS = 4;
constexpr u32 NUM_GRAPHICS_UNIFORM_BUFFERS = 18;
constexpr u32 NUM_COMPUTE_UNIFORM_BUFFERS = 8;
constexpr u32 NUM_STORAGE_BUFFERS = 16;
constexpr u32 NUM_STAGES = 5;

template <typename P>
class BufferCache {
    // Page size for caching purposes.
    // This is unrelated to the CPU page size and it can be changed as it seems optimal.
    static constexpr u32 PAGE_BITS = 16;
    static constexpr u64 PAGE_SIZE = u64{1} << PAGE_BITS;

    static constexpr bool IS_OPENGL = P::IS_OPENGL;
    static constexpr bool HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS =
        P::HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS;
    static constexpr bool HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT =
        P::HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT;
    static constexpr bool NEEDS_BIND_UNIFORM_INDEX = P::NEEDS_BIND_UNIFORM_INDEX;
    static constexpr bool NEEDS_BIND_STORAGE_INDEX = P::NEEDS_BIND_STORAGE_INDEX;
    static constexpr bool USE_MEMORY_MAPS = P::USE_MEMORY_MAPS;

    static constexpr BufferId NULL_BUFFER_ID{0};

    static constexpr u64 expected_memory = 512ULL * 1024ULL * 1024ULL;
    static constexpr u64 critical_memory = 1024ULL * 1024ULL * 1024ULL;

    using Maxwell = Tegra::Engines::Maxwell3D::Regs;

    using Runtime = typename P::Runtime;
    using Buffer = typename P::Buffer;

    struct Empty {};

    struct OverlapResult {
        std::vector<BufferId> ids;
        VAddr begin;
        VAddr end;
        bool has_stream_leap = false;
    };

    struct Binding {
        VAddr cpu_addr{};
        u32 size{};
        BufferId buffer_id;
    };

    static constexpr Binding NULL_BINDING{
        .cpu_addr = 0,
        .size = 0,
        .buffer_id = NULL_BUFFER_ID,
    };

public:
    static constexpr u32 DEFAULT_SKIP_CACHE_SIZE = 4096;

    explicit BufferCache(VideoCore::RasterizerInterface& rasterizer_,
                         Tegra::Engines::Maxwell3D& maxwell3d_,
                         Tegra::Engines::KeplerCompute& kepler_compute_,
                         Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
                         Runtime& runtime_);

    void TickFrame();

    void WriteMemory(VAddr cpu_addr, u64 size);

    void CachedWriteMemory(VAddr cpu_addr, u64 size);

    void DownloadMemory(VAddr cpu_addr, u64 size);

    void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size);

    void DisableGraphicsUniformBuffer(size_t stage, u32 index);

    void UpdateGraphicsBuffers(bool is_indexed);

    void UpdateComputeBuffers();

    void BindHostGeometryBuffers(bool is_indexed);

    void BindHostStageBuffers(size_t stage);

    void BindHostComputeBuffers();

    void SetEnabledUniformBuffers(size_t stage, u32 enabled);

    void SetEnabledComputeUniformBuffers(u32 enabled);

    void UnbindGraphicsStorageBuffers(size_t stage);

    void BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset,
                                   bool is_written);

    void UnbindComputeStorageBuffers();

    void BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset,
                                  bool is_written);

    void FlushCachedWrites();

    /// Return true when there are uncommitted buffers to be downloaded
    [[nodiscard]] bool HasUncommittedFlushes() const noexcept;

    /// Return true when the caller should wait for async downloads
    [[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept;

    /// Commit asynchronous downloads
    void CommitAsyncFlushes();

    /// Pop asynchronous downloads
    void PopAsyncFlushes();

    /// Return true when a CPU region is modified from the GPU
    [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);

    std::mutex mutex;

private:
    template <typename Func>
    static void ForEachEnabledBit(u32 enabled_mask, Func&& func) {
        for (u32 index = 0; enabled_mask != 0; ++index, enabled_mask >>= 1) {
            const int disabled_bits = std::countr_zero(enabled_mask);
            index += disabled_bits;
            enabled_mask >>= disabled_bits;
            func(index);
        }
    }

    template <typename Func>
    void ForEachBufferInRange(VAddr cpu_addr, u64 size, Func&& func) {
        const u64 page_end = Common::DivCeil(cpu_addr + size, PAGE_SIZE);
        for (u64 page = cpu_addr >> PAGE_BITS; 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, PAGE_SIZE);
        }
    }

    static bool IsRangeGranular(VAddr cpu_addr, size_t size) {
        return (cpu_addr & ~Core::Memory::PAGE_MASK) ==
               ((cpu_addr + size) & ~Core::Memory::PAGE_MASK);
    }

    void BindHostIndexBuffer();

    void BindHostVertexBuffers();

    void BindHostGraphicsUniformBuffers(size_t stage);

    void BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 binding_index, bool needs_bind);

    void BindHostGraphicsStorageBuffers(size_t stage);

    void BindHostTransformFeedbackBuffers();

    void BindHostComputeUniformBuffers();

    void BindHostComputeStorageBuffers();

    void DoUpdateGraphicsBuffers(bool is_indexed);

    void DoUpdateComputeBuffers();

    void UpdateIndexBuffer();

    void UpdateVertexBuffers();

    void UpdateVertexBuffer(u32 index);

    void UpdateUniformBuffers(size_t stage);

    void UpdateStorageBuffers(size_t stage);

    void UpdateTransformFeedbackBuffers();

    void UpdateTransformFeedbackBuffer(u32 index);

    void UpdateComputeUniformBuffers();

    void UpdateComputeStorageBuffers();

    void MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 size);

    [[nodiscard]] BufferId FindBuffer(VAddr cpu_addr, u32 size);

    [[nodiscard]] OverlapResult ResolveOverlaps(VAddr cpu_addr, u32 wanted_size);

    void JoinOverlap(BufferId new_buffer_id, BufferId overlap_id, bool accumulate_stream_score);

    [[nodiscard]] BufferId CreateBuffer(VAddr cpu_addr, u32 wanted_size);

    void Register(BufferId buffer_id);

    void Unregister(BufferId buffer_id);

    template <bool insert>
    void ChangeRegister(BufferId buffer_id);

    void TouchBuffer(Buffer& buffer) const noexcept;

    bool SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size);

    bool SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size);

    void UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                      std::span<BufferCopy> copies);

    void ImmediateUploadMemory(Buffer& buffer, u64 largest_copy,
                               std::span<const BufferCopy> copies);

    void MappedUploadMemory(Buffer& buffer, u64 total_size_bytes, std::span<BufferCopy> copies);

    void DownloadBufferMemory(Buffer& buffer_id);

    void DownloadBufferMemory(Buffer& buffer_id, VAddr cpu_addr, u64 size);

    void DeleteBuffer(BufferId buffer_id);

    void ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id);

    void NotifyBufferDeletion();

    [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr) const;

    [[nodiscard]] std::span<const u8> ImmediateBufferWithData(VAddr cpu_addr, size_t size);

    [[nodiscard]] std::span<u8> ImmediateBuffer(size_t wanted_capacity);

    [[nodiscard]] bool HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept;

    VideoCore::RasterizerInterface& rasterizer;
    Tegra::Engines::Maxwell3D& maxwell3d;
    Tegra::Engines::KeplerCompute& kepler_compute;
    Tegra::MemoryManager& gpu_memory;
    Core::Memory::Memory& cpu_memory;
    Runtime& runtime;

    SlotVector<Buffer> slot_buffers;
    DelayedDestructionRing<Buffer, 8> delayed_destruction_ring;

    u32 last_index_count = 0;

    Binding index_buffer;
    std::array<Binding, NUM_VERTEX_BUFFERS> vertex_buffers;
    std::array<std::array<Binding, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES> uniform_buffers;
    std::array<std::array<Binding, NUM_STORAGE_BUFFERS>, NUM_STAGES> storage_buffers;
    std::array<Binding, NUM_TRANSFORM_FEEDBACK_BUFFERS> transform_feedback_buffers;

    std::array<Binding, NUM_COMPUTE_UNIFORM_BUFFERS> compute_uniform_buffers;
    std::array<Binding, NUM_STORAGE_BUFFERS> compute_storage_buffers;

    std::array<u32, NUM_STAGES> enabled_uniform_buffers{};
    u32 enabled_compute_uniform_buffers = 0;

    std::array<u32, NUM_STAGES> enabled_storage_buffers{};
    std::array<u32, NUM_STAGES> written_storage_buffers{};
    u32 enabled_compute_storage_buffers = 0;
    u32 written_compute_storage_buffers = 0;

    std::array<u32, NUM_STAGES> fast_bound_uniform_buffers{};

    std::array<u32, 16> uniform_cache_hits{};
    std::array<u32, 16> uniform_cache_shots{};

    u32 uniform_buffer_skip_cache_size = DEFAULT_SKIP_CACHE_SIZE;

    bool has_deleted_buffers = false;

    std::conditional_t<HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS, std::array<u32, NUM_STAGES>, Empty>
        dirty_uniform_buffers{};

    std::vector<BufferId> cached_write_buffer_ids;

    // TODO: This data structure is not optimal and it should be reworked
    std::vector<BufferId> uncommitted_downloads;
    std::deque<std::vector<BufferId>> committed_downloads;

    size_t immediate_buffer_capacity = 0;
    std::unique_ptr<u8[]> immediate_buffer_alloc;

    typename SlotVector<Buffer>::Iterator deletion_iterator;
    u64 frame_tick = 0;
    u64 total_used_memory = 0;

    std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
};

template <class P>
BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
                            Tegra::Engines::Maxwell3D& maxwell3d_,
                            Tegra::Engines::KeplerCompute& kepler_compute_,
                            Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
                            Runtime& runtime_)
    : rasterizer{rasterizer_}, maxwell3d{maxwell3d_}, kepler_compute{kepler_compute_},
      gpu_memory{gpu_memory_}, cpu_memory{cpu_memory_}, runtime{runtime_} {
    // Ensure the first slot is used for the null buffer
    void(slot_buffers.insert(runtime, NullBufferParams{}));
    deletion_iterator = slot_buffers.end();
}

template <class P>
void BufferCache<P>::TickFrame() {
    SCOPE_EXIT({
      ++frame_tick;
      delayed_destruction_ring.Tick();
    });
    // Calculate hits and shots and move hit bits to the right
    const u32 hits = std::reduce(uniform_cache_hits.begin(), uniform_cache_hits.end());
    const u32 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end());
    std::copy_n(uniform_cache_hits.begin(), uniform_cache_hits.size() - 1,
                uniform_cache_hits.begin() + 1);
    std::copy_n(uniform_cache_shots.begin(), uniform_cache_shots.size() - 1,
                uniform_cache_shots.begin() + 1);
    uniform_cache_hits[0] = 0;
    uniform_cache_shots[0] = 0;

    const bool skip_preferred = hits * 256 < shots * 251;
    uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;

    const bool activate_gc = total_used_memory >= expected_memory;
    if (!activate_gc) {
        return;
    }
    const bool agressive_gc = total_used_memory >= critical_memory;
    const u64 ticks_to_destroy = agressive_gc ? 60 : 120;
    int num_iterations = agressive_gc ? 64 : 32;
    for (; num_iterations > 0; --num_iterations) {
        if (deletion_iterator == slot_buffers.end()) {
            deletion_iterator = slot_buffers.begin();
        }
        ++deletion_iterator;
        if (deletion_iterator == slot_buffers.end()) {
            break;
        }
        const auto [buffer_id, buffer] = *deletion_iterator;
        if (buffer->FrameTick() + ticks_to_destroy < frame_tick) {
            DownloadBufferMemory(*buffer);
            DeleteBuffer(buffer_id);
        }
    }
}

template <class P>
void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
    ForEachBufferInRange(cpu_addr, size, [&](BufferId, Buffer& buffer) {
        buffer.MarkRegionAsCpuModified(cpu_addr, size);
    });
}

template <class P>
void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
    ForEachBufferInRange(cpu_addr, size, [&](BufferId buffer_id, Buffer& buffer) {
        if (!buffer.HasCachedWrites()) {
            cached_write_buffer_ids.push_back(buffer_id);
        }
        buffer.CachedCpuWrite(cpu_addr, size);
    });
}

template <class P>
void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
    ForEachBufferInRange(cpu_addr, size,
                         [&](BufferId, Buffer& buffer) { DownloadBufferMemory(buffer); });
}

template <class P>
void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
                                               u32 size) {
    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
    const Binding binding{
        .cpu_addr = *cpu_addr,
        .size = size,
        .buffer_id = BufferId{},
    };
    uniform_buffers[stage][index] = binding;
}

template <class P>
void BufferCache<P>::DisableGraphicsUniformBuffer(size_t stage, u32 index) {
    uniform_buffers[stage][index] = NULL_BINDING;
}

template <class P>
void BufferCache<P>::UpdateGraphicsBuffers(bool is_indexed) {
    MICROPROFILE_SCOPE(GPU_PrepareBuffers);
    do {
        has_deleted_buffers = false;
        DoUpdateGraphicsBuffers(is_indexed);
    } while (has_deleted_buffers);
}

template <class P>
void BufferCache<P>::UpdateComputeBuffers() {
    MICROPROFILE_SCOPE(GPU_PrepareBuffers);
    do {
        has_deleted_buffers = false;
        DoUpdateComputeBuffers();
    } while (has_deleted_buffers);
}

template <class P>
void BufferCache<P>::BindHostGeometryBuffers(bool is_indexed) {
    MICROPROFILE_SCOPE(GPU_BindUploadBuffers);
    if (is_indexed) {
        BindHostIndexBuffer();
    } else if constexpr (!HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) {
        const auto& regs = maxwell3d.regs;
        if (regs.draw.topology == Maxwell::PrimitiveTopology::Quads) {
            runtime.BindQuadArrayIndexBuffer(regs.vertex_buffer.first, regs.vertex_buffer.count);
        }
    }
    BindHostVertexBuffers();
    BindHostTransformFeedbackBuffers();
}

template <class P>
void BufferCache<P>::BindHostStageBuffers(size_t stage) {
    MICROPROFILE_SCOPE(GPU_BindUploadBuffers);
    BindHostGraphicsUniformBuffers(stage);
    BindHostGraphicsStorageBuffers(stage);
}

template <class P>
void BufferCache<P>::BindHostComputeBuffers() {
    MICROPROFILE_SCOPE(GPU_BindUploadBuffers);
    BindHostComputeUniformBuffers();
    BindHostComputeStorageBuffers();
}

template <class P>
void BufferCache<P>::SetEnabledUniformBuffers(size_t stage, u32 enabled) {
    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
        if (enabled_uniform_buffers[stage] != enabled) {
            dirty_uniform_buffers[stage] = ~u32{0};
        }
    }
    enabled_uniform_buffers[stage] = enabled;
}

template <class P>
void BufferCache<P>::SetEnabledComputeUniformBuffers(u32 enabled) {
    enabled_compute_uniform_buffers = enabled;
}

template <class P>
void BufferCache<P>::UnbindGraphicsStorageBuffers(size_t stage) {
    enabled_storage_buffers[stage] = 0;
    written_storage_buffers[stage] = 0;
}

template <class P>
void BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index,
                                               u32 cbuf_offset, bool is_written) {
    enabled_storage_buffers[stage] |= 1U << ssbo_index;
    written_storage_buffers[stage] |= (is_written ? 1U : 0U) << ssbo_index;

    const auto& cbufs = maxwell3d.state.shader_stages[stage];
    const GPUVAddr ssbo_addr = cbufs.const_buffers[cbuf_index].address + cbuf_offset;
    storage_buffers[stage][ssbo_index] = StorageBufferBinding(ssbo_addr);
}

template <class P>
void BufferCache<P>::UnbindComputeStorageBuffers() {
    enabled_compute_storage_buffers = 0;
    written_compute_storage_buffers = 0;
}

template <class P>
void BufferCache<P>::BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset,
                                              bool is_written) {
    enabled_compute_storage_buffers |= 1U << ssbo_index;
    written_compute_storage_buffers |= (is_written ? 1U : 0U) << ssbo_index;

    const auto& launch_desc = kepler_compute.launch_description;
    ASSERT(((launch_desc.const_buffer_enable_mask >> cbuf_index) & 1) != 0);

    const auto& cbufs = launch_desc.const_buffer_config;
    const GPUVAddr ssbo_addr = cbufs[cbuf_index].Address() + cbuf_offset;
    compute_storage_buffers[ssbo_index] = StorageBufferBinding(ssbo_addr);
}

template <class P>
void BufferCache<P>::FlushCachedWrites() {
    for (const BufferId buffer_id : cached_write_buffer_ids) {
        slot_buffers[buffer_id].FlushCachedWrites();
    }
    cached_write_buffer_ids.clear();
}

template <class P>
bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
    return !uncommitted_downloads.empty();
}

template <class P>
bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
    return !committed_downloads.empty() && !committed_downloads.front().empty();
}

template <class P>
void BufferCache<P>::CommitAsyncFlushes() {
    // This is intentionally passing the value by copy
    committed_downloads.push_front(uncommitted_downloads);
    uncommitted_downloads.clear();
}

template <class P>
void BufferCache<P>::PopAsyncFlushes() {
    if (committed_downloads.empty()) {
        return;
    }
    auto scope_exit_pop_download = detail::ScopeExit([this] { committed_downloads.pop_back(); });
    const std::span<const BufferId> download_ids = committed_downloads.back();
    if (download_ids.empty()) {
        return;
    }
    MICROPROFILE_SCOPE(GPU_DownloadMemory);

    boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads;
    u64 total_size_bytes = 0;
    u64 largest_copy = 0;
    for (const BufferId buffer_id : download_ids) {
        slot_buffers[buffer_id].ForEachDownloadRange([&](u64 range_offset, u64 range_size) {
            downloads.push_back({
                BufferCopy{
                    .src_offset = range_offset,
                    .dst_offset = total_size_bytes,
                    .size = range_size,
                },
                buffer_id,
            });
            total_size_bytes += range_size;
            largest_copy = std::max(largest_copy, range_size);
        });
    }
    if (downloads.empty()) {
        return;
    }
    if constexpr (USE_MEMORY_MAPS) {
        auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
        for (auto& [copy, buffer_id] : downloads) {
            // Have in mind the staging buffer offset for the copy
            copy.dst_offset += download_staging.offset;
            const std::array copies{copy};
            runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies);
        }
        runtime.Finish();
        for (const auto& [copy, buffer_id] : downloads) {
            const Buffer& buffer = slot_buffers[buffer_id];
            const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
            // Undo the modified offset
            const u64 dst_offset = copy.dst_offset - download_staging.offset;
            const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
            cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
        }
    } else {
        const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
        for (const auto& [copy, buffer_id] : downloads) {
            Buffer& buffer = slot_buffers[buffer_id];
            buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
            const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
            cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
        }
    }
}

template <class P>
bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
    const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
    for (u64 page = addr >> PAGE_BITS; page < page_end;) {
        const BufferId image_id = page_table[page];
        if (!image_id) {
            ++page;
            continue;
        }
        Buffer& buffer = slot_buffers[image_id];
        if (buffer.IsRegionGpuModified(addr, size)) {
            return true;
        }
        const VAddr end_addr = buffer.CpuAddr() + buffer.SizeBytes();
        page = Common::DivCeil(end_addr, PAGE_SIZE);
    }
    return false;
}

template <class P>
void BufferCache<P>::BindHostIndexBuffer() {
    Buffer& buffer = slot_buffers[index_buffer.buffer_id];
    TouchBuffer(buffer);
    const u32 offset = buffer.Offset(index_buffer.cpu_addr);
    const u32 size = index_buffer.size;
    SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
    if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) {
        runtime.BindIndexBuffer(buffer, offset, size);
    } else {
        runtime.BindIndexBuffer(maxwell3d.regs.draw.topology, maxwell3d.regs.index_array.format,
                                maxwell3d.regs.index_array.first, maxwell3d.regs.index_array.count,
                                buffer, offset, size);
    }
}

template <class P>
void BufferCache<P>::BindHostVertexBuffers() {
    auto& flags = maxwell3d.dirty.flags;
    for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
        const Binding& binding = vertex_buffers[index];
        Buffer& buffer = slot_buffers[binding.buffer_id];
        TouchBuffer(buffer);
        SynchronizeBuffer(buffer, binding.cpu_addr, binding.size);
        if (!flags[Dirty::VertexBuffer0 + index]) {
            continue;
        }
        flags[Dirty::VertexBuffer0 + index] = false;

        const u32 stride = maxwell3d.regs.vertex_array[index].stride;
        const u32 offset = buffer.Offset(binding.cpu_addr);
        runtime.BindVertexBuffer(index, buffer, offset, binding.size, stride);
    }
}

template <class P>
void BufferCache<P>::BindHostGraphicsUniformBuffers(size_t stage) {
    u32 dirty = ~0U;
    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
        dirty = std::exchange(dirty_uniform_buffers[stage], 0);
    }
    u32 binding_index = 0;
    ForEachEnabledBit(enabled_uniform_buffers[stage], [&](u32 index) {
        const bool needs_bind = ((dirty >> index) & 1) != 0;
        BindHostGraphicsUniformBuffer(stage, index, binding_index, needs_bind);
        if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
            ++binding_index;
        }
    });
}

template <class P>
void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 binding_index,
                                                   bool needs_bind) {
    const Binding& binding = uniform_buffers[stage][index];
    const VAddr cpu_addr = binding.cpu_addr;
    const u32 size = binding.size;
    Buffer& buffer = slot_buffers[binding.buffer_id];
    TouchBuffer(buffer);
    const bool use_fast_buffer = binding.buffer_id != NULL_BUFFER_ID &&
                                 size <= uniform_buffer_skip_cache_size &&
                                 !buffer.IsRegionGpuModified(cpu_addr, size);
    if (use_fast_buffer) {
        if constexpr (IS_OPENGL) {
            if (runtime.HasFastBufferSubData()) {
                // Fast path for Nvidia
                if (!HasFastUniformBufferBound(stage, binding_index)) {
                    // We only have to bind when the currently bound buffer is not the fast version
                    runtime.BindFastUniformBuffer(stage, binding_index, size);
                }
                const auto span = ImmediateBufferWithData(cpu_addr, size);
                runtime.PushFastUniformBuffer(stage, binding_index, span);
                return;
            }
        }
        fast_bound_uniform_buffers[stage] |= 1U << binding_index;

        // Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
        const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
        cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size);
        return;
    }
    // Classic cached path
    const bool sync_cached = SynchronizeBuffer(buffer, cpu_addr, size);
    if (sync_cached) {
        ++uniform_cache_hits[0];
    }
    ++uniform_cache_shots[0];

    if (!needs_bind && !HasFastUniformBufferBound(stage, binding_index)) {
        // Skip binding if it's not needed and if the bound buffer is not the fast version
        // This exists to avoid instances where the fast buffer is bound and a GPU write happens
        return;
    }
    fast_bound_uniform_buffers[stage] &= ~(1U << binding_index);

    const u32 offset = buffer.Offset(cpu_addr);
    if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
        runtime.BindUniformBuffer(stage, binding_index, buffer, offset, size);
    } else {
        runtime.BindUniformBuffer(buffer, offset, size);
    }
}

template <class P>
void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
    u32 binding_index = 0;
    ForEachEnabledBit(enabled_storage_buffers[stage], [&](u32 index) {
        const Binding& binding = storage_buffers[stage][index];
        Buffer& buffer = slot_buffers[binding.buffer_id];
        TouchBuffer(buffer);
        const u32 size = binding.size;
        SynchronizeBuffer(buffer, binding.cpu_addr, size);

        const u32 offset = buffer.Offset(binding.cpu_addr);
        const bool is_written = ((written_storage_buffers[stage] >> index) & 1) != 0;
        if constexpr (NEEDS_BIND_STORAGE_INDEX) {
            runtime.BindStorageBuffer(stage, binding_index, buffer, offset, size, is_written);
            ++binding_index;
        } else {
            runtime.BindStorageBuffer(buffer, offset, size, is_written);
        }
    });
}

template <class P>
void BufferCache<P>::BindHostTransformFeedbackBuffers() {
    if (maxwell3d.regs.tfb_enabled == 0) {
        return;
    }
    for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
        const Binding& binding = transform_feedback_buffers[index];
        Buffer& buffer = slot_buffers[binding.buffer_id];
        TouchBuffer(buffer);
        const u32 size = binding.size;
        SynchronizeBuffer(buffer, binding.cpu_addr, size);

        const u32 offset = buffer.Offset(binding.cpu_addr);
        runtime.BindTransformFeedbackBuffer(index, buffer, offset, size);
    }
}

template <class P>
void BufferCache<P>::BindHostComputeUniformBuffers() {
    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
        // Mark all uniform buffers as dirty
        dirty_uniform_buffers.fill(~u32{0});
    }
    u32 binding_index = 0;
    ForEachEnabledBit(enabled_compute_uniform_buffers, [&](u32 index) {
        const Binding& binding = compute_uniform_buffers[index];
        Buffer& buffer = slot_buffers[binding.buffer_id];
        TouchBuffer(buffer);
        const u32 size = binding.size;
        SynchronizeBuffer(buffer, binding.cpu_addr, size);

        const u32 offset = buffer.Offset(binding.cpu_addr);
        if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
            runtime.BindComputeUniformBuffer(binding_index, buffer, offset, size);
            ++binding_index;
        } else {
            runtime.BindUniformBuffer(buffer, offset, size);
        }
    });
}

template <class P>
void BufferCache<P>::BindHostComputeStorageBuffers() {
    u32 binding_index = 0;
    ForEachEnabledBit(enabled_compute_storage_buffers, [&](u32 index) {
        const Binding& binding = compute_storage_buffers[index];
        Buffer& buffer = slot_buffers[binding.buffer_id];
        TouchBuffer(buffer);
        const u32 size = binding.size;
        SynchronizeBuffer(buffer, binding.cpu_addr, size);

        const u32 offset = buffer.Offset(binding.cpu_addr);
        const bool is_written = ((written_compute_storage_buffers >> index) & 1) != 0;
        if constexpr (NEEDS_BIND_STORAGE_INDEX) {
            runtime.BindComputeStorageBuffer(binding_index, buffer, offset, size, is_written);
            ++binding_index;
        } else {
            runtime.BindStorageBuffer(buffer, offset, size, is_written);
        }
    });
}

template <class P>
void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) {
    if (is_indexed) {
        UpdateIndexBuffer();
    }
    UpdateVertexBuffers();
    UpdateTransformFeedbackBuffers();
    for (size_t stage = 0; stage < NUM_STAGES; ++stage) {
        UpdateUniformBuffers(stage);
        UpdateStorageBuffers(stage);
    }
}

template <class P>
void BufferCache<P>::DoUpdateComputeBuffers() {
    UpdateComputeUniformBuffers();
    UpdateComputeStorageBuffers();
}

template <class P>
void BufferCache<P>::UpdateIndexBuffer() {
    // We have to check for the dirty flags and index count
    // The index count is currently changed without updating the dirty flags
    const auto& index_array = maxwell3d.regs.index_array;
    auto& flags = maxwell3d.dirty.flags;
    if (!flags[Dirty::IndexBuffer] && last_index_count == index_array.count) {
        return;
    }
    flags[Dirty::IndexBuffer] = false;
    last_index_count = index_array.count;

    const GPUVAddr gpu_addr_begin = index_array.StartAddress();
    const GPUVAddr gpu_addr_end = index_array.EndAddress();
    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr_begin);
    const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
    const u32 draw_size = index_array.count * index_array.FormatSizeInBytes();
    const u32 size = std::min(address_size, draw_size);
    if (size == 0 || !cpu_addr) {
        index_buffer = NULL_BINDING;
        return;
    }
    index_buffer = Binding{
        .cpu_addr = *cpu_addr,
        .size = size,
        .buffer_id = FindBuffer(*cpu_addr, size),
    };
}

template <class P>
void BufferCache<P>::UpdateVertexBuffers() {
    auto& flags = maxwell3d.dirty.flags;
    if (!maxwell3d.dirty.flags[Dirty::VertexBuffers]) {
        return;
    }
    flags[Dirty::VertexBuffers] = false;

    for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
        UpdateVertexBuffer(index);
    }
}

template <class P>
void BufferCache<P>::UpdateVertexBuffer(u32 index) {
    if (!maxwell3d.dirty.flags[Dirty::VertexBuffer0 + index]) {
        return;
    }
    const auto& array = maxwell3d.regs.vertex_array[index];
    const auto& limit = maxwell3d.regs.vertex_array_limit[index];
    const GPUVAddr gpu_addr_begin = array.StartAddress();
    const GPUVAddr gpu_addr_end = limit.LimitAddress() + 1;
    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr_begin);
    const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
    const u32 size = address_size; // TODO: Analyze stride and number of vertices
    if (array.enable == 0 || size == 0 || !cpu_addr) {
        vertex_buffers[index] = NULL_BINDING;
        return;
    }
    vertex_buffers[index] = Binding{
        .cpu_addr = *cpu_addr,
        .size = size,
        .buffer_id = FindBuffer(*cpu_addr, size),
    };
}

template <class P>
void BufferCache<P>::UpdateUniformBuffers(size_t stage) {
    ForEachEnabledBit(enabled_uniform_buffers[stage], [&](u32 index) {
        Binding& binding = uniform_buffers[stage][index];
        if (binding.buffer_id) {
            // Already updated
            return;
        }
        // Mark as dirty
        if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
            dirty_uniform_buffers[stage] |= 1U << index;
        }
        // Resolve buffer
        binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size);
    });
}

template <class P>
void BufferCache<P>::UpdateStorageBuffers(size_t stage) {
    const u32 written_mask = written_storage_buffers[stage];
    ForEachEnabledBit(enabled_storage_buffers[stage], [&](u32 index) {
        // Resolve buffer
        Binding& binding = storage_buffers[stage][index];
        const BufferId buffer_id = FindBuffer(binding.cpu_addr, binding.size);
        binding.buffer_id = buffer_id;
        // Mark buffer as written if needed
        if (((written_mask >> index) & 1) != 0) {
            MarkWrittenBuffer(buffer_id, binding.cpu_addr, binding.size);
        }
    });
}

template <class P>
void BufferCache<P>::UpdateTransformFeedbackBuffers() {
    if (maxwell3d.regs.tfb_enabled == 0) {
        return;
    }
    for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
        UpdateTransformFeedbackBuffer(index);
    }
}

template <class P>
void BufferCache<P>::UpdateTransformFeedbackBuffer(u32 index) {
    const auto& binding = maxwell3d.regs.tfb_bindings[index];
    const GPUVAddr gpu_addr = binding.Address() + binding.buffer_offset;
    const u32 size = binding.buffer_size;
    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
    if (binding.buffer_enable == 0 || size == 0 || !cpu_addr) {
        transform_feedback_buffers[index] = NULL_BINDING;
        return;
    }
    const BufferId buffer_id = FindBuffer(*cpu_addr, size);
    transform_feedback_buffers[index] = Binding{
        .cpu_addr = *cpu_addr,
        .size = size,
        .buffer_id = buffer_id,
    };
    MarkWrittenBuffer(buffer_id, *cpu_addr, size);
}

template <class P>
void BufferCache<P>::UpdateComputeUniformBuffers() {
    ForEachEnabledBit(enabled_compute_uniform_buffers, [&](u32 index) {
        Binding& binding = compute_uniform_buffers[index];
        binding = NULL_BINDING;
        const auto& launch_desc = kepler_compute.launch_description;
        if (((launch_desc.const_buffer_enable_mask >> index) & 1) != 0) {
            const auto& cbuf = launch_desc.const_buffer_config[index];
            const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(cbuf.Address());
            if (cpu_addr) {
                binding.cpu_addr = *cpu_addr;
                binding.size = cbuf.size;
            }
        }
        binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size);
    });
}

template <class P>
void BufferCache<P>::UpdateComputeStorageBuffers() {
    ForEachEnabledBit(enabled_compute_storage_buffers, [&](u32 index) {
        // Resolve buffer
        Binding& binding = compute_storage_buffers[index];
        const BufferId buffer_id = FindBuffer(binding.cpu_addr, binding.size);
        binding.buffer_id = buffer_id;
        // Mark as written if needed
        if (((written_compute_storage_buffers >> index) & 1) != 0) {
            MarkWrittenBuffer(buffer_id, binding.cpu_addr, binding.size);
        }
    });
}

template <class P>
void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 size) {
    Buffer& buffer = slot_buffers[buffer_id];
    buffer.MarkRegionAsGpuModified(cpu_addr, size);

    const bool is_accuracy_high = Settings::IsGPULevelHigh();
    const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
    if (!is_accuracy_high || !is_async) {
        return;
    }
    if (std::ranges::find(uncommitted_downloads, buffer_id) != uncommitted_downloads.end()) {
        // Already inserted
        return;
    }
    uncommitted_downloads.push_back(buffer_id);
}

template <class P>
BufferId BufferCache<P>::FindBuffer(VAddr cpu_addr, u32 size) {
    if (cpu_addr == 0) {
        return NULL_BUFFER_ID;
    }
    const u64 page = cpu_addr >> PAGE_BITS;
    const BufferId buffer_id = page_table[page];
    if (!buffer_id) {
        return CreateBuffer(cpu_addr, size);
    }
    const Buffer& buffer = slot_buffers[buffer_id];
    if (buffer.IsInBounds(cpu_addr, size)) {
        return buffer_id;
    }
    return CreateBuffer(cpu_addr, size);
}

template <class P>
typename BufferCache<P>::OverlapResult BufferCache<P>::ResolveOverlaps(VAddr cpu_addr,
                                                                       u32 wanted_size) {
    static constexpr int STREAM_LEAP_THRESHOLD = 16;
    std::vector<BufferId> overlap_ids;
    VAddr begin = cpu_addr;
    VAddr end = cpu_addr + wanted_size;
    int stream_score = 0;
    bool has_stream_leap = false;
    for (; cpu_addr >> PAGE_BITS < Common::DivCeil(end, PAGE_SIZE); cpu_addr += PAGE_SIZE) {
        const BufferId overlap_id = page_table[cpu_addr >> PAGE_BITS];
        if (!overlap_id) {
            continue;
        }
        Buffer& overlap = slot_buffers[overlap_id];
        if (overlap.IsPicked()) {
            continue;
        }
        overlap_ids.push_back(overlap_id);
        overlap.Pick();
        const VAddr overlap_cpu_addr = overlap.CpuAddr();
        if (overlap_cpu_addr < begin) {
            cpu_addr = begin = overlap_cpu_addr;
        }
        end = std::max(end, overlap_cpu_addr + overlap.SizeBytes());

        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;
            end += PAGE_SIZE * 256;
        }
    }
    return OverlapResult{
        .ids = std::move(overlap_ids),
        .begin = begin,
        .end = end,
        .has_stream_leap = has_stream_leap,
    };
}

template <class P>
void BufferCache<P>::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);
    }
    std::vector<BufferCopy> copies;
    const size_t dst_base_offset = overlap.CpuAddr() - new_buffer.CpuAddr();
    overlap.ForEachDownloadRange([&](u64 begin, u64 range_size) {
        copies.push_back(BufferCopy{
            .src_offset = begin,
            .dst_offset = dst_base_offset + begin,
            .size = range_size,
        });
        new_buffer.UnmarkRegionAsCpuModified(begin, range_size);
        new_buffer.MarkRegionAsGpuModified(begin, range_size);
    });
    if (!copies.empty()) {
        runtime.CopyBuffer(slot_buffers[new_buffer_id], overlap, copies);
    }
    ReplaceBufferDownloads(overlap_id, new_buffer_id);
    DeleteBuffer(overlap_id);
}

template <class P>
BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
    const OverlapResult overlap = ResolveOverlaps(cpu_addr, wanted_size);
    const u32 size = static_cast<u32>(overlap.end - overlap.begin);
    const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);
    TouchBuffer(slot_buffers[new_buffer_id]);
    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;
}

template <class P>
void BufferCache<P>::Register(BufferId buffer_id) {
    ChangeRegister<true>(buffer_id);
}

template <class P>
void BufferCache<P>::Unregister(BufferId buffer_id) {
    ChangeRegister<false>(buffer_id);
}

template <class P>
template <bool insert>
void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
    const Buffer& buffer = slot_buffers[buffer_id];
    const auto size = buffer.SizeBytes();
    if (insert) {
        total_used_memory += Common::AlignUp(size, 1024);
    } else {
        total_used_memory -= Common::AlignUp(size, 1024);
    }
    const VAddr cpu_addr_begin = buffer.CpuAddr();
    const VAddr cpu_addr_end = cpu_addr_begin + size;
    const u64 page_begin = cpu_addr_begin / PAGE_SIZE;
    const u64 page_end = Common::DivCeil(cpu_addr_end, PAGE_SIZE);
    for (u64 page = page_begin; page != page_end; ++page) {
        if constexpr (insert) {
            page_table[page] = buffer_id;
        } else {
            page_table[page] = BufferId{};
        }
    }
}

template <class P>
void BufferCache<P>::TouchBuffer(Buffer& buffer) const noexcept {
    buffer.SetFrameTick(frame_tick);
}

template <class P>
bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size) {
    if (buffer.CpuAddr() == 0) {
        return true;
    }
    return SynchronizeBufferImpl(buffer, cpu_addr, size);
}

template <class P>
bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size) {
    boost::container::small_vector<BufferCopy, 4> copies;
    u64 total_size_bytes = 0;
    u64 largest_copy = 0;
    buffer.ForEachUploadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
        copies.push_back(BufferCopy{
            .src_offset = total_size_bytes,
            .dst_offset = range_offset,
            .size = range_size,
        });
        total_size_bytes += range_size;
        largest_copy = std::max(largest_copy, range_size);
    });
    if (total_size_bytes == 0) {
        return true;
    }
    const std::span<BufferCopy> copies_span(copies.data(), copies.size());
    UploadMemory(buffer, total_size_bytes, largest_copy, copies_span);
    return false;
}

template <class P>
void BufferCache<P>::UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                                  std::span<BufferCopy> copies) {
    if constexpr (USE_MEMORY_MAPS) {
        MappedUploadMemory(buffer, total_size_bytes, copies);
    } else {
        ImmediateUploadMemory(buffer, largest_copy, copies);
    }
}

template <class P>
void BufferCache<P>::ImmediateUploadMemory(Buffer& buffer, u64 largest_copy,
                                           std::span<const BufferCopy> copies) {
    std::span<u8> immediate_buffer;
    for (const BufferCopy& copy : copies) {
        std::span<const u8> upload_span;
        const VAddr cpu_addr = buffer.CpuAddr() + copy.dst_offset;
        if (IsRangeGranular(cpu_addr, copy.size)) {
            upload_span = std::span(cpu_memory.GetPointer(cpu_addr), copy.size);
        } else {
            if (immediate_buffer.empty()) {
                immediate_buffer = ImmediateBuffer(largest_copy);
            }
            cpu_memory.ReadBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
            upload_span = immediate_buffer.subspan(0, copy.size);
        }
        buffer.ImmediateUpload(copy.dst_offset, upload_span);
    }
}

template <class P>
void BufferCache<P>::MappedUploadMemory(Buffer& buffer, u64 total_size_bytes,
                                        std::span<BufferCopy> copies) {
    auto upload_staging = runtime.UploadStagingBuffer(total_size_bytes);
    const std::span<u8> staging_pointer = upload_staging.mapped_span;
    for (BufferCopy& copy : copies) {
        u8* const src_pointer = staging_pointer.data() + copy.src_offset;
        const VAddr cpu_addr = buffer.CpuAddr() + copy.dst_offset;
        cpu_memory.ReadBlockUnsafe(cpu_addr, src_pointer, copy.size);

        // Apply the staging offset
        copy.src_offset += upload_staging.offset;
    }
    runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
}

template <class P>
void BufferCache<P>::DownloadBufferMemory(Buffer& buffer) {
    DownloadBufferMemory(buffer, buffer.CpuAddr(), buffer.SizeBytes());
}

template <class P>
void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 size) {
    boost::container::small_vector<BufferCopy, 1> copies;
    u64 total_size_bytes = 0;
    u64 largest_copy = 0;
    buffer.ForEachDownloadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
        copies.push_back(BufferCopy{
            .src_offset = range_offset,
            .dst_offset = total_size_bytes,
            .size = range_size,
        });
        total_size_bytes += range_size;
        largest_copy = std::max(largest_copy, range_size);
    });
    if (total_size_bytes == 0) {
        return;
    }
    MICROPROFILE_SCOPE(GPU_DownloadMemory);

    if constexpr (USE_MEMORY_MAPS) {
        auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
        const u8* const mapped_memory = download_staging.mapped_span.data();
        const std::span<BufferCopy> copies_span(copies.data(), copies.data() + copies.size());
        for (BufferCopy& copy : copies) {
            // Modify copies to have the staging offset in mind
            copy.dst_offset += download_staging.offset;
        }
        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span);
        runtime.Finish();
        for (const BufferCopy& copy : copies) {
            const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
            // Undo the modified offset
            const u64 dst_offset = copy.dst_offset - download_staging.offset;
            const u8* copy_mapped_memory = mapped_memory + dst_offset;
            cpu_memory.WriteBlockUnsafe(copy_cpu_addr, copy_mapped_memory, copy.size);
        }
    } else {
        const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
        for (const BufferCopy& copy : copies) {
            buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
            const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
            cpu_memory.WriteBlockUnsafe(copy_cpu_addr, immediate_buffer.data(), copy.size);
        }
    }
}

template <class P>
void BufferCache<P>::DeleteBuffer(BufferId buffer_id) {
    const auto scalar_replace = [buffer_id](Binding& binding) {
        if (binding.buffer_id == buffer_id) {
            binding.buffer_id = BufferId{};
        }
    };
    const auto replace = [scalar_replace](std::span<Binding> bindings) {
        std::ranges::for_each(bindings, scalar_replace);
    };
    scalar_replace(index_buffer);
    replace(vertex_buffers);
    std::ranges::for_each(uniform_buffers, replace);
    std::ranges::for_each(storage_buffers, replace);
    replace(transform_feedback_buffers);
    replace(compute_uniform_buffers);
    replace(compute_storage_buffers);
    std::erase(cached_write_buffer_ids, buffer_id);

    // Mark the whole buffer as CPU written to stop tracking CPU writes
    Buffer& buffer = slot_buffers[buffer_id];
    buffer.MarkRegionAsCpuModified(buffer.CpuAddr(), buffer.SizeBytes());

    Unregister(buffer_id);
    delayed_destruction_ring.Push(std::move(slot_buffers[buffer_id]));
    slot_buffers.erase(buffer_id);

    NotifyBufferDeletion();
}

template <class P>
void BufferCache<P>::ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id) {
    const auto replace = [old_buffer_id, new_buffer_id](std::vector<BufferId>& buffers) {
        std::ranges::replace(buffers, old_buffer_id, new_buffer_id);
        if (auto it = std::ranges::find(buffers, new_buffer_id); it != buffers.end()) {
            buffers.erase(std::remove(it + 1, buffers.end(), new_buffer_id), buffers.end());
        }
    };
    replace(uncommitted_downloads);
    std::ranges::for_each(committed_downloads, replace);
}

template <class P>
void BufferCache<P>::NotifyBufferDeletion() {
    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
        dirty_uniform_buffers.fill(~u32{0});
    }
    auto& flags = maxwell3d.dirty.flags;
    flags[Dirty::IndexBuffer] = true;
    flags[Dirty::VertexBuffers] = true;
    for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
        flags[Dirty::VertexBuffer0 + index] = true;
    }
    has_deleted_buffers = true;
}

template <class P>
typename BufferCache<P>::Binding BufferCache<P>::StorageBufferBinding(GPUVAddr ssbo_addr) const {
    const GPUVAddr gpu_addr = gpu_memory.Read<u64>(ssbo_addr);
    const u32 size = gpu_memory.Read<u32>(ssbo_addr + 8);
    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
    if (!cpu_addr || size == 0) {
        return NULL_BINDING;
    }
    // HACK(Rodrigo): This is the number of bytes bound in host beyond the guest API's range.
    // It exists due to some games like Astral Chain operate out of bounds.
    // Binding the whole map range would be technically correct, but games have large maps that make
    // this approach unaffordable for now.
    static constexpr u32 arbitrary_extra_bytes = 0xc000;
    const u32 bytes_to_map_end = static_cast<u32>(gpu_memory.BytesToMapEnd(gpu_addr));
    const Binding binding{
        .cpu_addr = *cpu_addr,
        .size = std::min(size + arbitrary_extra_bytes, bytes_to_map_end),
        .buffer_id = BufferId{},
    };
    return binding;
}

template <class P>
std::span<const u8> BufferCache<P>::ImmediateBufferWithData(VAddr cpu_addr, size_t size) {
    u8* const base_pointer = cpu_memory.GetPointer(cpu_addr);
    if (IsRangeGranular(cpu_addr, size) ||
        base_pointer + size == cpu_memory.GetPointer(cpu_addr + size)) {
        return std::span(base_pointer, size);
    } else {
        const std::span<u8> span = ImmediateBuffer(size);
        cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size);
        return span;
    }
}

template <class P>
std::span<u8> BufferCache<P>::ImmediateBuffer(size_t wanted_capacity) {
    if (wanted_capacity > immediate_buffer_capacity) {
        immediate_buffer_capacity = wanted_capacity;
        immediate_buffer_alloc = std::make_unique<u8[]>(wanted_capacity);
    }
    return std::span<u8>(immediate_buffer_alloc.get(), wanted_capacity);
}

template <class P>
bool BufferCache<P>::HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept {
    if constexpr (IS_OPENGL) {
        return ((fast_bound_uniform_buffers[stage] >> binding_index) & 1) != 0;
    } else {
        // Only OpenGL has fast uniform buffers
        return false;
    }
}

} // namespace VideoCommon