yuzu-android/src/video_core/memory_manager.h

// SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later

#pragma once

#include <atomic>
#include <map>
#include <mutex>
#include <optional>
#include <vector>
#include <boost/container/small_vector.hpp>

#include "common/common_types.h"
#include "common/multi_level_page_table.h"
#include "common/range_map.h"
#include "common/scratch_buffer.h"
#include "common/virtual_buffer.h"
#include "video_core/cache_types.h"
#include "video_core/host1x/gpu_device_memory_manager.h"
#include "video_core/pte_kind.h"

namespace VideoCore {
class RasterizerInterface;
}

namespace VideoCommon {
class InvalidationAccumulator;
}

namespace Core {
class System;
} // namespace Core

namespace Tegra {

class MemoryManager final {
public:
    explicit MemoryManager(Core::System& system_, u64 address_space_bits_ = 40,
                           GPUVAddr split_address = 1ULL << 34, u64 big_page_bits_ = 16,
                           u64 page_bits_ = 12);
    explicit MemoryManager(Core::System& system_, MaxwellDeviceMemoryManager& memory_,
                           u64 address_space_bits_ = 40, GPUVAddr split_address = 1ULL << 34,
                           u64 big_page_bits_ = 16, u64 page_bits_ = 12);
    ~MemoryManager();

    static constexpr bool HAS_FLUSH_INVALIDATION = true;

    size_t GetID() const {
        return unique_identifier;
    }

    /// Binds a renderer to the memory manager.
    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);

    [[nodiscard]] std::optional<DAddr> GpuToCpuAddress(GPUVAddr addr) const;

    [[nodiscard]] std::optional<DAddr> GpuToCpuAddress(GPUVAddr addr, std::size_t size) const;

    template <typename T>
    [[nodiscard]] T Read(GPUVAddr addr) const;

    template <typename T>
    void Write(GPUVAddr addr, T data);

    [[nodiscard]] u8* GetPointer(GPUVAddr addr);
    [[nodiscard]] const u8* GetPointer(GPUVAddr addr) const;

    template <typename T>
    [[nodiscard]] T* GetPointer(GPUVAddr addr) {
        const auto address{GpuToCpuAddress(addr)};
        if (!address) {
            return {};
        }
        return memory.GetPointer<T>(*address);
    }

    template <typename T>
    [[nodiscard]] const T* GetPointer(GPUVAddr addr) const {
        return GetPointer<T*>(addr);
    }

    /**
     * ReadBlock and WriteBlock are full read and write operations over virtual
     * GPU Memory. It's important to use these when GPU memory may not be continuous
     * in the Host Memory counterpart. Note: This functions cause Host GPU Memory
     * Flushes and Invalidations, respectively to each operation.
     */
    void ReadBlock(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size,
                   VideoCommon::CacheType which = VideoCommon::CacheType::All) const;
    void WriteBlock(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size,
                    VideoCommon::CacheType which = VideoCommon::CacheType::All);
    void CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size,
                   VideoCommon::CacheType which = VideoCommon::CacheType::All);

    /**
     * ReadBlockUnsafe and WriteBlockUnsafe are special versions of ReadBlock and
     * WriteBlock respectively. In this versions, no flushing or invalidation is actually
     * done and their performance is similar to a memcpy. This functions can be used
     * on either of this 2 scenarios instead of their safe counterpart:
     * - Memory which is sure to never be represented in the Host GPU.
     * - Memory Managed by a Cache Manager. Example: Texture Flushing should use
     * WriteBlockUnsafe instead of WriteBlock since it shouldn't invalidate the texture
     * being flushed.
     */
    void ReadBlockUnsafe(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size) const;
    void WriteBlockUnsafe(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size);
    void WriteBlockCached(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size);

    /**
     * Checks if a gpu region can be simply read with a pointer.
     */
    [[nodiscard]] bool IsGranularRange(GPUVAddr gpu_addr, std::size_t size) const;

    /**
     * Checks if a gpu region is mapped by a single range of device addresses.
     */
    [[nodiscard]] bool IsContinuousRange(GPUVAddr gpu_addr, std::size_t size) const;

    /**
     * Checks if a gpu region is mapped entirely.
     */
    [[nodiscard]] bool IsFullyMappedRange(GPUVAddr gpu_addr, std::size_t size) const;

    /**
     * Returns a vector with all the subranges of device addresses mapped beneath.
     * if the region is continuous, a single pair will be returned. If it's unmapped, an empty
     * vector will be returned;
     */
    boost::container::small_vector<std::pair<GPUVAddr, std::size_t>, 32> GetSubmappedRange(
        GPUVAddr gpu_addr, std::size_t size) const;

    GPUVAddr Map(GPUVAddr gpu_addr, DAddr dev_addr, std::size_t size,
                 PTEKind kind = PTEKind::INVALID, bool is_big_pages = true);
    GPUVAddr MapSparse(GPUVAddr gpu_addr, std::size_t size, bool is_big_pages = true);
    void Unmap(GPUVAddr gpu_addr, std::size_t size);

    void FlushRegion(GPUVAddr gpu_addr, size_t size,
                     VideoCommon::CacheType which = VideoCommon::CacheType::All) const;

    void InvalidateRegion(GPUVAddr gpu_addr, size_t size,
                          VideoCommon::CacheType which = VideoCommon::CacheType::All) const;

    bool IsMemoryDirty(GPUVAddr gpu_addr, size_t size,
                       VideoCommon::CacheType which = VideoCommon::CacheType::All) const;

    size_t MaxContinuousRange(GPUVAddr gpu_addr, size_t size) const;

    bool IsWithinGPUAddressRange(GPUVAddr gpu_addr) const {
        return gpu_addr < address_space_size;
    }

    PTEKind GetPageKind(GPUVAddr gpu_addr) const;

    size_t GetMemoryLayoutSize(GPUVAddr gpu_addr,
                               size_t max_size = std::numeric_limits<size_t>::max()) const;

    void FlushCaching();

    const u8* GetSpan(const GPUVAddr src_addr, const std::size_t size) const;
    u8* GetSpan(const GPUVAddr src_addr, const std::size_t size);

private:
    template <bool is_big_pages, typename FuncMapped, typename FuncReserved, typename FuncUnmapped>
    inline void MemoryOperation(GPUVAddr gpu_src_addr, std::size_t size, FuncMapped&& func_mapped,
                                FuncReserved&& func_reserved, FuncUnmapped&& func_unmapped) const;

    template <bool is_safe>
    void ReadBlockImpl(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size,
                       VideoCommon::CacheType which) const;

    template <bool is_safe>
    void WriteBlockImpl(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size,
                        VideoCommon::CacheType which);

    template <bool is_big_page>
    [[nodiscard]] std::size_t PageEntryIndex(GPUVAddr gpu_addr) const {
        if constexpr (is_big_page) {
            return (gpu_addr >> big_page_bits) & big_page_table_mask;
        } else {
            return (gpu_addr >> page_bits) & page_table_mask;
        }
    }

    inline bool IsBigPageContinuous(size_t big_page_index) const;
    inline void SetBigPageContinuous(size_t big_page_index, bool value);

    template <bool is_gpu_address>
    void GetSubmappedRangeImpl(
        GPUVAddr gpu_addr, std::size_t size,
        boost::container::small_vector<
            std::pair<std::conditional_t<is_gpu_address, GPUVAddr, DAddr>, std::size_t>, 32>&
            result) const;

    Core::System& system;
    MaxwellDeviceMemoryManager& memory;

    const u64 address_space_bits;
    GPUVAddr split_address;
    const u64 page_bits;
    u64 address_space_size;
    u64 page_size;
    u64 page_mask;
    u64 page_table_mask;
    static constexpr u64 cpu_page_bits{12};

    const u64 big_page_bits;
    u64 big_page_size;
    u64 big_page_mask;
    u64 big_page_table_mask;

    VideoCore::RasterizerInterface* rasterizer = nullptr;

    enum class EntryType : u64 {
        Free = 0,
        Reserved = 1,
        Mapped = 2,
    };

    std::vector<u64> entries;
    std::vector<u64> big_entries;

    template <EntryType entry_type>
    GPUVAddr PageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] DAddr dev_addr, size_t size,
                         PTEKind kind);

    template <EntryType entry_type>
    GPUVAddr BigPageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] DAddr dev_addr, size_t size,
                            PTEKind kind);

    template <bool is_big_page>
    inline EntryType GetEntry(size_t position) const;

    template <bool is_big_page>
    inline void SetEntry(size_t position, EntryType entry);

    Common::MultiLevelPageTable<u32> page_table;
    Common::RangeMap<GPUVAddr, PTEKind> kind_map;
    Common::VirtualBuffer<u32> big_page_table_dev;

    std::vector<u64> big_page_continuous;
    boost::container::small_vector<std::pair<DAddr, std::size_t>, 32> page_stash{};
    boost::container::small_vector<std::pair<DAddr, std::size_t>, 32> page_stash2{};

    mutable std::mutex guard;

    static constexpr size_t continuous_bits = 64;

    const size_t unique_identifier;
    std::unique_ptr<VideoCommon::InvalidationAccumulator> accumulator;

    static std::atomic<size_t> unique_identifier_generator;

    Common::ScratchBuffer<u8> tmp_buffer;
};

} // namespace Tegra