Merge pull request #10996 from Kelebek1/readblock_optimisation

Use spans over guest memory where possible instead of copying data

src/audio_core/device/device_session.cpp

@@ -92,9 +92,9 @@ void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) {
         if (type == Sink::StreamType::In) {
             stream->AppendBuffer(new_buffer, tmp_samples);
         } else {
-            system.ApplicationMemory().ReadBlockUnsafe(buffer.samples, tmp_samples.data(),
+            Core::Memory::CpuGuestMemory<s16, Core::Memory::GuestMemoryFlags::UnsafeRead> samples(
-                                                       buffer.size);
+                system.ApplicationMemory(), buffer.samples, buffer.size / sizeof(s16));
-            stream->AppendBuffer(new_buffer, tmp_samples);
+            stream->AppendBuffer(new_buffer, samples);
         }
     }
 }

src/audio_core/renderer/command/data_source/decode.cpp

@@ -28,7 +28,6 @@ constexpr std::array<u8, 3> PitchBySrcQuality = {4, 8, 4};
 template <typename T>
 static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
                      const DecodeArg& req) {
-    std::array<T, TempBufferSize> tmp_samples{};
     constexpr s32 min{std::numeric_limits<s16>::min()};
     constexpr s32 max{std::numeric_limits<s16>::max()};
 
@@ -49,19 +48,18 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
         const VAddr source{req.buffer +
                            (((req.start_offset + req.offset) * channel_count) * sizeof(T))};
         const u64 size{channel_count * samples_to_decode};
-        const u64 size_bytes{size * sizeof(T)};
-
-        memory.ReadBlockUnsafe(source, tmp_samples.data(), size_bytes);
 
+        Core::Memory::CpuGuestMemory<T, Core::Memory::GuestMemoryFlags::UnsafeRead> samples(
+            memory, source, size);
         if constexpr (std::is_floating_point_v<T>) {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                auto sample{static_cast<s32>(tmp_samples[i * channel_count + req.target_channel] *
+                auto sample{static_cast<s32>(samples[i * channel_count + req.target_channel] *
                                              std::numeric_limits<s16>::max())};
                 out_buffer[i] = static_cast<s16>(std::clamp(sample, min, max));
             }
         } else {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                out_buffer[i] = tmp_samples[i * channel_count + req.target_channel];
+                out_buffer[i] = samples[i * channel_count + req.target_channel];
             }
         }
     } break;
@@ -74,16 +72,17 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
         }
 
         const VAddr source{req.buffer + ((req.start_offset + req.offset) * sizeof(T))};
-        memory.ReadBlockUnsafe(source, tmp_samples.data(), samples_to_decode * sizeof(T));
+        Core::Memory::CpuGuestMemory<T, Core::Memory::GuestMemoryFlags::UnsafeRead> samples(
+            memory, source, samples_to_decode);
 
         if constexpr (std::is_floating_point_v<T>) {
             for (u32 i = 0; i < samples_to_decode; i++) {
-                auto sample{static_cast<s32>(tmp_samples[i * channel_count + req.target_channel] *
+                auto sample{static_cast<s32>(samples[i * channel_count + req.target_channel] *
                                              std::numeric_limits<s16>::max())};
                 out_buffer[i] = static_cast<s16>(std::clamp(sample, min, max));
             }
         } else {
-            std::memcpy(out_buffer.data(), tmp_samples.data(), samples_to_decode * sizeof(s16));
+            std::memcpy(out_buffer.data(), samples.data(), samples_to_decode * sizeof(s16));
         }
         break;
     }
@@ -101,7 +100,6 @@ static u32 DecodePcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
  */
 static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
                        const DecodeArg& req) {
-    std::array<u8, TempBufferSize> wavebuffer{};
     constexpr u32 SamplesPerFrame{14};
     constexpr u32 NibblesPerFrame{16};
 
@@ -139,7 +137,8 @@ static u32 DecodeAdpcm(Core::Memory::Memory& memory, std::span<s16> out_buffer,
     }
 
     const auto size{std::max((samples_to_process / 8U) * SamplesPerFrame, 8U)};
-    memory.ReadBlockUnsafe(req.buffer + position_in_frame / 2, wavebuffer.data(), size);
+    Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> wavebuffer(
+        memory, req.buffer + position_in_frame / 2, size);
 
     auto context{req.adpcm_context};
     auto header{context->header};

src/audio_core/renderer/command/effect/aux_.cpp

@@ -21,23 +21,13 @@ static void ResetAuxBufferDsp(Core::Memory::Memory& memory, const CpuAddr aux_in
     }
 
     AuxInfo::AuxInfoDsp info{};
-    auto info_ptr{&info};
+    memory.ReadBlockUnsafe(aux_info, &info, sizeof(AuxInfo::AuxInfoDsp));
-    bool host_safe{(aux_info & Core::Memory::YUZU_PAGEMASK) <=
-                   (Core::Memory::YUZU_PAGESIZE - sizeof(AuxInfo::AuxInfoDsp))};
 
-    if (host_safe) [[likely]] {
+    info.read_offset = 0;
-        info_ptr = memory.GetPointer<AuxInfo::AuxInfoDsp>(aux_info);
+    info.write_offset = 0;
-    } else {
+    info.total_sample_count = 0;
-        memory.ReadBlockUnsafe(aux_info, info_ptr, sizeof(AuxInfo::AuxInfoDsp));
-    }
 
-    info_ptr->read_offset = 0;
+    memory.WriteBlockUnsafe(aux_info, &info, sizeof(AuxInfo::AuxInfoDsp));
-    info_ptr->write_offset = 0;
-    info_ptr->total_sample_count = 0;
-
-    if (!host_safe) [[unlikely]] {
-        memory.WriteBlockUnsafe(aux_info, info_ptr, sizeof(AuxInfo::AuxInfoDsp));
-    }
 }
 
 /**
@@ -86,17 +76,9 @@ static u32 WriteAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr send_info_,
     }
 
     AuxInfo::AuxInfoDsp send_info{};
-    auto send_ptr = &send_info;
+    memory.ReadBlockUnsafe(send_info_, &send_info, sizeof(AuxInfo::AuxInfoDsp));
-    bool host_safe = (send_info_ & Core::Memory::YUZU_PAGEMASK) <=
-                     (Core::Memory::YUZU_PAGESIZE - sizeof(AuxInfo::AuxInfoDsp));
 
-    if (host_safe) [[likely]] {
+    u32 target_write_offset{send_info.write_offset + write_offset};
-        send_ptr = memory.GetPointer<AuxInfo::AuxInfoDsp>(send_info_);
-    } else {
-        memory.ReadBlockUnsafe(send_info_, send_ptr, sizeof(AuxInfo::AuxInfoDsp));
-    }
-
-    u32 target_write_offset{send_ptr->write_offset + write_offset};
     if (target_write_offset > count_max) {
         return 0;
     }
@@ -105,15 +87,9 @@ static u32 WriteAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr send_info_,
     u32 read_pos{0};
     while (write_count > 0) {
         u32 to_write{std::min(count_max - target_write_offset, write_count)};
-        const auto write_addr = send_buffer + target_write_offset * sizeof(s32);
+        if (to_write > 0) {
-        bool write_safe{(write_addr & Core::Memory::YUZU_PAGEMASK) <=
+            const auto write_addr = send_buffer + target_write_offset * sizeof(s32);
-                        (Core::Memory::YUZU_PAGESIZE - (write_addr + to_write * sizeof(s32)))};
+            memory.WriteBlockUnsafe(write_addr, &input[read_pos], to_write * sizeof(s32));
-        if (write_safe) [[likely]] {
-            auto ptr = memory.GetPointer(write_addr);
-            std::memcpy(ptr, &input[read_pos], to_write * sizeof(s32));
-        } else {
-            memory.WriteBlockUnsafe(send_buffer + target_write_offset * sizeof(s32),
-                                    &input[read_pos], to_write * sizeof(s32));
         }
         target_write_offset = (target_write_offset + to_write) % count_max;
         write_count -= to_write;
@@ -121,13 +97,10 @@ static u32 WriteAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr send_info_,
     }
 
     if (update_count) {
-        send_ptr->write_offset = (send_ptr->write_offset + update_count) % count_max;
+        send_info.write_offset = (send_info.write_offset + update_count) % count_max;
-    }
-
-    if (!host_safe) [[unlikely]] {
-        memory.WriteBlockUnsafe(send_info_, send_ptr, sizeof(AuxInfo::AuxInfoDsp));
     }
 
+    memory.WriteBlockUnsafe(send_info_, &send_info, sizeof(AuxInfo::AuxInfoDsp));
     return write_count_;
 }
 
@@ -174,17 +147,9 @@ static u32 ReadAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr return_info_,
     }
 
     AuxInfo::AuxInfoDsp return_info{};
-    auto return_ptr = &return_info;
+    memory.ReadBlockUnsafe(return_info_, &return_info, sizeof(AuxInfo::AuxInfoDsp));
-    bool host_safe = (return_info_ & Core::Memory::YUZU_PAGEMASK) <=
-                     (Core::Memory::YUZU_PAGESIZE - sizeof(AuxInfo::AuxInfoDsp));
 
-    if (host_safe) [[likely]] {
+    u32 target_read_offset{return_info.read_offset + read_offset};
-        return_ptr = memory.GetPointer<AuxInfo::AuxInfoDsp>(return_info_);
-    } else {
-        memory.ReadBlockUnsafe(return_info_, return_ptr, sizeof(AuxInfo::AuxInfoDsp));
-    }
-
-    u32 target_read_offset{return_ptr->read_offset + read_offset};
     if (target_read_offset > count_max) {
         return 0;
     }
@@ -193,15 +158,9 @@ static u32 ReadAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr return_info_,
     u32 write_pos{0};
     while (read_count > 0) {
         u32 to_read{std::min(count_max - target_read_offset, read_count)};
-        const auto read_addr = return_buffer + target_read_offset * sizeof(s32);
+        if (to_read > 0) {
-        bool read_safe{(read_addr & Core::Memory::YUZU_PAGEMASK) <=
+            const auto read_addr = return_buffer + target_read_offset * sizeof(s32);
-                       (Core::Memory::YUZU_PAGESIZE - (read_addr + to_read * sizeof(s32)))};
+            memory.ReadBlockUnsafe(read_addr, &output[write_pos], to_read * sizeof(s32));
-        if (read_safe) [[likely]] {
-            auto ptr = memory.GetPointer(read_addr);
-            std::memcpy(&output[write_pos], ptr, to_read * sizeof(s32));
-        } else {
-            memory.ReadBlockUnsafe(return_buffer + target_read_offset * sizeof(s32),
-                                   &output[write_pos], to_read * sizeof(s32));
         }
         target_read_offset = (target_read_offset + to_read) % count_max;
         read_count -= to_read;
@@ -209,13 +168,10 @@ static u32 ReadAuxBufferDsp(Core::Memory::Memory& memory, CpuAddr return_info_,
     }
 
     if (update_count) {
-        return_ptr->read_offset = (return_ptr->read_offset + update_count) % count_max;
+        return_info.read_offset = (return_info.read_offset + update_count) % count_max;
-    }
-
-    if (!host_safe) [[unlikely]] {
-        memory.WriteBlockUnsafe(return_info_, return_ptr, sizeof(AuxInfo::AuxInfoDsp));
     }
 
+    memory.WriteBlockUnsafe(return_info_, &return_info, sizeof(AuxInfo::AuxInfoDsp));
     return read_count_;
 }
 

src/common/page_table.cpp

@@ -66,6 +66,7 @@ void PageTable::Resize(std::size_t address_space_width_in_bits, std::size_t page
                                              << (address_space_width_in_bits - page_size_in_bits)};
     pointers.resize(num_page_table_entries);
     backing_addr.resize(num_page_table_entries);
+    blocks.resize(num_page_table_entries);
     current_address_space_width_in_bits = address_space_width_in_bits;
     page_size = 1ULL << page_size_in_bits;
 }

src/common/page_table.h

@@ -122,6 +122,7 @@ struct PageTable {
      * corresponding attribute element is of type `Memory`.
      */
     VirtualBuffer<PageInfo> pointers;
+    VirtualBuffer<u64> blocks;
 
     VirtualBuffer<u64> backing_addr;
 

src/common/scratch_buffer.h

@@ -40,8 +40,21 @@ public:
     ~ScratchBuffer() = default;
     ScratchBuffer(const ScratchBuffer&) = delete;
     ScratchBuffer& operator=(const ScratchBuffer&) = delete;
-    ScratchBuffer(ScratchBuffer&&) = default;
+
-    ScratchBuffer& operator=(ScratchBuffer&&) = default;
+    ScratchBuffer(ScratchBuffer&& other) noexcept {
+        swap(other);
+        other.last_requested_size = 0;
+        other.buffer_capacity = 0;
+        other.buffer.reset();
+    }
+
+    ScratchBuffer& operator=(ScratchBuffer&& other) noexcept {
+        swap(other);
+        other.last_requested_size = 0;
+        other.buffer_capacity = 0;
+        other.buffer.reset();
+        return *this;
+    }
 
     /// This will only grow the buffer's capacity if size is greater than the current capacity.
     /// The previously held data will remain intact.

src/core/core_timing.cpp

@@ -70,7 +70,7 @@ void CoreTiming::Initialize(std::function<void()>&& on_thread_init_) {
         -> std::optional<std::chrono::nanoseconds> { return std::nullopt; };
     ev_lost = CreateEvent("_lost_event", empty_timed_callback);
     if (is_multicore) {
-        timer_thread = std::make_unique<std::thread>(ThreadEntry, std::ref(*this));
+        timer_thread = std::make_unique<std::jthread>(ThreadEntry, std::ref(*this));
     }
 }
 
@@ -255,7 +255,6 @@ void CoreTiming::ThreadLoop() {
 #ifdef _WIN32
                     while (!paused && !event.IsSet() && wait_time > 0) {
                         wait_time = *next_time - GetGlobalTimeNs().count();
-
                         if (wait_time >= timer_resolution_ns) {
                             Common::Windows::SleepForOneTick();
                         } else {

src/core/core_timing.h

@@ -163,7 +163,7 @@ private:
     Common::Event pause_event{};
     std::mutex basic_lock;
     std::mutex advance_lock;
-    std::unique_ptr<std::thread> timer_thread;
+    std::unique_ptr<std::jthread> timer_thread;
     std::atomic<bool> paused{};
     std::atomic<bool> paused_set{};
     std::atomic<bool> wait_set{};

src/core/hle/service/hle_ipc.cpp

@@ -329,8 +329,22 @@ std::vector<u8> HLERequestContext::ReadBufferCopy(std::size_t buffer_index) cons
 }
 
 std::span<const u8> HLERequestContext::ReadBuffer(std::size_t buffer_index) const {
-    static thread_local std::array<Common::ScratchBuffer<u8>, 2> read_buffer_a;
+    static thread_local std::array read_buffer_a{
-    static thread_local std::array<Common::ScratchBuffer<u8>, 2> read_buffer_x;
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+    };
+    static thread_local std::array read_buffer_data_a{
+        Common::ScratchBuffer<u8>(),
+        Common::ScratchBuffer<u8>(),
+    };
+    static thread_local std::array read_buffer_x{
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+    };
+    static thread_local std::array read_buffer_data_x{
+        Common::ScratchBuffer<u8>(),
+        Common::ScratchBuffer<u8>(),
+    };
 
     const bool is_buffer_a{BufferDescriptorA().size() > buffer_index &&
                            BufferDescriptorA()[buffer_index].Size()};
@@ -339,19 +353,17 @@ std::span<const u8> HLERequestContext::ReadBuffer(std::size_t buffer_index) cons
             BufferDescriptorA().size() > buffer_index, { return {}; },
             "BufferDescriptorA invalid buffer_index {}", buffer_index);
         auto& read_buffer = read_buffer_a[buffer_index];
-        read_buffer.resize_destructive(BufferDescriptorA()[buffer_index].Size());
+        return read_buffer.Read(BufferDescriptorA()[buffer_index].Address(),
-        memory.ReadBlock(BufferDescriptorA()[buffer_index].Address(), read_buffer.data(),
+                                BufferDescriptorA()[buffer_index].Size(),
-                         read_buffer.size());
+                                &read_buffer_data_a[buffer_index]);
-        return read_buffer;
     } else {
         ASSERT_OR_EXECUTE_MSG(
             BufferDescriptorX().size() > buffer_index, { return {}; },
             "BufferDescriptorX invalid buffer_index {}", buffer_index);
         auto& read_buffer = read_buffer_x[buffer_index];
-        read_buffer.resize_destructive(BufferDescriptorX()[buffer_index].Size());
+        return read_buffer.Read(BufferDescriptorX()[buffer_index].Address(),
-        memory.ReadBlock(BufferDescriptorX()[buffer_index].Address(), read_buffer.data(),
+                                BufferDescriptorX()[buffer_index].Size(),
-                         read_buffer.size());
+                                &read_buffer_data_x[buffer_index]);
-        return read_buffer;
     }
 }
 

src/core/memory.cpp

@@ -266,6 +266,22 @@ struct Memory::Impl {
         ReadBlockImpl<true>(*system.ApplicationProcess(), src_addr, dest_buffer, size);
     }
 
+    const u8* GetSpan(const VAddr src_addr, const std::size_t size) const {
+        if (current_page_table->blocks[src_addr >> YUZU_PAGEBITS] ==
+            current_page_table->blocks[(src_addr + size) >> YUZU_PAGEBITS]) {
+            return GetPointerSilent(src_addr);
+        }
+        return nullptr;
+    }
+
+    u8* GetSpan(const VAddr src_addr, const std::size_t size) {
+        if (current_page_table->blocks[src_addr >> YUZU_PAGEBITS] ==
+            current_page_table->blocks[(src_addr + size) >> YUZU_PAGEBITS]) {
+            return GetPointerSilent(src_addr);
+        }
+        return nullptr;
+    }
+
     template <bool UNSAFE>
     void WriteBlockImpl(const Kernel::KProcess& process, const Common::ProcessAddress dest_addr,
                         const void* src_buffer, const std::size_t size) {
@@ -559,7 +575,7 @@ struct Memory::Impl {
             }
         }
 
-        const Common::ProcessAddress end = base + size;
+        const auto end = base + size;
         ASSERT_MSG(end <= page_table.pointers.size(), "out of range mapping at {:016X}",
                    base + page_table.pointers.size());
 
@@ -570,14 +586,18 @@ struct Memory::Impl {
             while (base != end) {
                 page_table.pointers[base].Store(nullptr, type);
                 page_table.backing_addr[base] = 0;
-
+                page_table.blocks[base] = 0;
                 base += 1;
             }
         } else {
+            auto orig_base = base;
             while (base != end) {
-                page_table.pointers[base].Store(
+                auto host_ptr =
-                    system.DeviceMemory().GetPointer<u8>(target) - (base << YUZU_PAGEBITS), type);
+                    system.DeviceMemory().GetPointer<u8>(target) - (base << YUZU_PAGEBITS);
-                page_table.backing_addr[base] = GetInteger(target) - (base << YUZU_PAGEBITS);
+                auto backing = GetInteger(target) - (base << YUZU_PAGEBITS);
+                page_table.pointers[base].Store(host_ptr, type);
+                page_table.backing_addr[base] = backing;
+                page_table.blocks[base] = orig_base << YUZU_PAGEBITS;
 
                 ASSERT_MSG(page_table.pointers[base].Pointer(),
                            "memory mapping base yield a nullptr within the table");
@@ -747,6 +767,14 @@ struct Memory::Impl {
         VAddr last_address;
     };
 
+    void InvalidateRegion(Common::ProcessAddress dest_addr, size_t size) {
+        system.GPU().InvalidateRegion(GetInteger(dest_addr), size);
+    }
+
+    void FlushRegion(Common::ProcessAddress dest_addr, size_t size) {
+        system.GPU().FlushRegion(GetInteger(dest_addr), size);
+    }
+
     Core::System& system;
     Common::PageTable* current_page_table = nullptr;
     std::array<VideoCore::RasterizerDownloadArea, Core::Hardware::NUM_CPU_CORES>
@@ -881,6 +909,14 @@ void Memory::ReadBlockUnsafe(const Common::ProcessAddress src_addr, void* dest_b
     impl->ReadBlockUnsafe(src_addr, dest_buffer, size);
 }
 
+const u8* Memory::GetSpan(const VAddr src_addr, const std::size_t size) const {
+    return impl->GetSpan(src_addr, size);
+}
+
+u8* Memory::GetSpan(const VAddr src_addr, const std::size_t size) {
+    return impl->GetSpan(src_addr, size);
+}
+
 void Memory::WriteBlock(const Common::ProcessAddress dest_addr, const void* src_buffer,
                         const std::size_t size) {
     impl->WriteBlock(dest_addr, src_buffer, size);
@@ -924,4 +960,12 @@ void Memory::MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug)
     impl->MarkRegionDebug(GetInteger(vaddr), size, debug);
 }
 
+void Memory::InvalidateRegion(Common::ProcessAddress dest_addr, size_t size) {
+    impl->InvalidateRegion(dest_addr, size);
+}
+
+void Memory::FlushRegion(Common::ProcessAddress dest_addr, size_t size) {
+    impl->FlushRegion(dest_addr, size);
+}
+
 } // namespace Core::Memory

src/core/memory.h

@@ -5,8 +5,12 @@
 
 #include <cstddef>
 #include <memory>
+#include <optional>
 #include <span>
 #include <string>
+#include <vector>
+
+#include "common/scratch_buffer.h"
 #include "common/typed_address.h"
 #include "core/hle/result.h"
 
@@ -24,6 +28,10 @@ class PhysicalMemory;
 class KProcess;
 } // namespace Kernel
 
+namespace Tegra {
+class MemoryManager;
+}
+
 namespace Core::Memory {
 
 /**
@@ -343,6 +351,9 @@ public:
      */
     void ReadBlockUnsafe(Common::ProcessAddress src_addr, void* dest_buffer, std::size_t size);
 
+    const u8* GetSpan(const VAddr src_addr, const std::size_t size) const;
+    u8* GetSpan(const VAddr src_addr, const std::size_t size);
+
     /**
      * Writes a range of bytes into the current process' address space at the specified
      * virtual address.
@@ -461,6 +472,8 @@ public:
     void MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug);
 
     void SetGPUDirtyManagers(std::span<Core::GPUDirtyMemoryManager> managers);
+    void InvalidateRegion(Common::ProcessAddress dest_addr, size_t size);
+    void FlushRegion(Common::ProcessAddress dest_addr, size_t size);
 
 private:
     Core::System& system;
@@ -469,4 +482,203 @@ private:
     std::unique_ptr<Impl> impl;
 };
 
+enum GuestMemoryFlags : u32 {
+    Read = 1 << 0,
+    Write = 1 << 1,
+    Safe = 1 << 2,
+    Cached = 1 << 3,
+
+    SafeRead = Read | Safe,
+    SafeWrite = Write | Safe,
+    SafeReadWrite = SafeRead | SafeWrite,
+    SafeReadCachedWrite = SafeReadWrite | Cached,
+
+    UnsafeRead = Read,
+    UnsafeWrite = Write,
+    UnsafeReadWrite = UnsafeRead | UnsafeWrite,
+    UnsafeReadCachedWrite = UnsafeReadWrite | Cached,
+};
+
+namespace {
+template <typename M, typename T, GuestMemoryFlags FLAGS>
+class GuestMemory {
+    using iterator = T*;
+    using const_iterator = const T*;
+    using value_type = T;
+    using element_type = T;
+    using iterator_category = std::contiguous_iterator_tag;
+
+public:
+    GuestMemory() = delete;
+    explicit GuestMemory(M& memory_, u64 addr_, std::size_t size_,
+                         Common::ScratchBuffer<T>* backup = nullptr)
+        : memory{memory_}, addr{addr_}, size{size_} {
+        static_assert(FLAGS & GuestMemoryFlags::Read || FLAGS & GuestMemoryFlags::Write);
+        if constexpr (FLAGS & GuestMemoryFlags::Read) {
+            Read(addr, size, backup);
+        }
+    }
+
+    ~GuestMemory() = default;
+
+    T* data() noexcept {
+        return data_span.data();
+    }
+
+    const T* data() const noexcept {
+        return data_span.data();
+    }
+
+    [[nodiscard]] T* begin() noexcept {
+        return data();
+    }
+
+    [[nodiscard]] const T* begin() const noexcept {
+        return data();
+    }
+
+    [[nodiscard]] T* end() noexcept {
+        return data() + size;
+    }
+
+    [[nodiscard]] const T* end() const noexcept {
+        return data() + size;
+    }
+
+    T& operator[](size_t index) noexcept {
+        return data_span[index];
+    }
+
+    const T& operator[](size_t index) const noexcept {
+        return data_span[index];
+    }
+
+    void SetAddressAndSize(u64 addr_, std::size_t size_) noexcept {
+        addr = addr_;
+        size = size_;
+        addr_changed = true;
+    }
+
+    std::span<T> Read(u64 addr_, std::size_t size_,
+                      Common::ScratchBuffer<T>* backup = nullptr) noexcept {
+        addr = addr_;
+        size = size_;
+        if (size == 0) {
+            is_data_copy = true;
+            return {};
+        }
+
+        if (TrySetSpan()) {
+            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                memory.FlushRegion(addr, size * sizeof(T));
+            }
+        } else {
+            if (backup) {
+                backup->resize_destructive(size);
+                data_span = *backup;
+            } else {
+                data_copy.resize(size);
+                data_span = std::span(data_copy);
+            }
+            is_data_copy = true;
+            span_valid = true;
+            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                memory.ReadBlock(addr, data_span.data(), size * sizeof(T));
+            } else {
+                memory.ReadBlockUnsafe(addr, data_span.data(), size * sizeof(T));
+            }
+        }
+        return data_span;
+    }
+
+    void Write(std::span<T> write_data) noexcept {
+        if constexpr (FLAGS & GuestMemoryFlags::Cached) {
+            memory.WriteBlockCached(addr, write_data.data(), size * sizeof(T));
+        } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+            memory.WriteBlock(addr, write_data.data(), size * sizeof(T));
+        } else {
+            memory.WriteBlockUnsafe(addr, write_data.data(), size * sizeof(T));
+        }
+    }
+
+    bool TrySetSpan() noexcept {
+        if (u8* ptr = memory.GetSpan(addr, size * sizeof(T)); ptr) {
+            data_span = {reinterpret_cast<T*>(ptr), size};
+            span_valid = true;
+            return true;
+        }
+        return false;
+    }
+
+protected:
+    bool IsDataCopy() const noexcept {
+        return is_data_copy;
+    }
+
+    bool AddressChanged() const noexcept {
+        return addr_changed;
+    }
+
+    M& memory;
+    u64 addr;
+    size_t size;
+    std::span<T> data_span{};
+    std::vector<T> data_copy;
+    bool span_valid{false};
+    bool is_data_copy{false};
+    bool addr_changed{false};
+};
+
+template <typename M, typename T, GuestMemoryFlags FLAGS>
+class GuestMemoryScoped : public GuestMemory<M, T, FLAGS> {
+public:
+    GuestMemoryScoped() = delete;
+    explicit GuestMemoryScoped(M& memory_, u64 addr_, std::size_t size_,
+                               Common::ScratchBuffer<T>* backup = nullptr)
+        : GuestMemory<M, T, FLAGS>(memory_, addr_, size_, backup) {
+        if constexpr (!(FLAGS & GuestMemoryFlags::Read)) {
+            if (!this->TrySetSpan()) {
+                if (backup) {
+                    this->data_span = *backup;
+                    this->span_valid = true;
+                    this->is_data_copy = true;
+                }
+            }
+        }
+    }
+
+    ~GuestMemoryScoped() {
+        if constexpr (FLAGS & GuestMemoryFlags::Write) {
+            if (this->size == 0) [[unlikely]] {
+                return;
+            }
+
+            if (this->AddressChanged() || this->IsDataCopy()) {
+                ASSERT(this->span_valid);
+                if constexpr (FLAGS & GuestMemoryFlags::Cached) {
+                    this->memory.WriteBlockCached(this->addr, this->data_span.data(),
+                                                  this->size * sizeof(T));
+                } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                    this->memory.WriteBlock(this->addr, this->data_span.data(),
+                                            this->size * sizeof(T));
+                } else {
+                    this->memory.WriteBlockUnsafe(this->addr, this->data_span.data(),
+                                                  this->size * sizeof(T));
+                }
+            } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                this->memory.InvalidateRegion(this->addr, this->size * sizeof(T));
+            }
+        }
+    }
+};
+} // namespace
+
+template <typename T, GuestMemoryFlags FLAGS>
+using CpuGuestMemory = GuestMemory<Memory, T, FLAGS>;
+template <typename T, GuestMemoryFlags FLAGS>
+using CpuGuestMemoryScoped = GuestMemoryScoped<Memory, T, FLAGS>;
+template <typename T, GuestMemoryFlags FLAGS>
+using GpuGuestMemory = GuestMemory<Tegra::MemoryManager, T, FLAGS>;
+template <typename T, GuestMemoryFlags FLAGS>
+using GpuGuestMemoryScoped = GuestMemoryScoped<Tegra::MemoryManager, T, FLAGS>;
 } // namespace Core::Memory

src/video_core/buffer_cache/buffer_cache.h

@@ -234,9 +234,10 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
-    tmp_buffer.resize_destructive(amount);
+
-    cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
+    Core::Memory::CpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::UnsafeReadWrite> tmp(
-    cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
+        cpu_memory, *cpu_src_address, amount, &tmp_buffer);
+    tmp.SetAddressAndSize(*cpu_dest_address, amount);
     return true;
 }
 

src/video_core/dma_pusher.cpp

@@ -5,6 +5,7 @@
 #include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
+#include "core/memory.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/gpu.h"
@@ -12,6 +13,8 @@
 
 namespace Tegra {
 
+constexpr u32 MacroRegistersStart = 0xE00;
+
 DmaPusher::DmaPusher(Core::System& system_, GPU& gpu_, MemoryManager& memory_manager_,
                      Control::ChannelState& channel_state_)
     : gpu{gpu_}, system{system_}, memory_manager{memory_manager_}, puller{gpu_, memory_manager_,
@@ -74,25 +77,16 @@ bool DmaPusher::Step() {
         }
 
         // Push buffer non-empty, read a word
-        command_headers.resize_destructive(command_list_header.size);
+        if (dma_state.method >= MacroRegistersStart) {
-        constexpr u32 MacroRegistersStart = 0xE00;
-        if (dma_state.method < MacroRegistersStart) {
-            if (Settings::IsGPULevelHigh()) {
-                memory_manager.ReadBlock(dma_state.dma_get, command_headers.data(),
-                                         command_list_header.size * sizeof(u32));
-            } else {
-                memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(),
-                                               command_list_header.size * sizeof(u32));
-            }
-        } else {
-            const size_t copy_size = command_list_header.size * sizeof(u32);
             if (subchannels[dma_state.subchannel]) {
-                subchannels[dma_state.subchannel]->current_dirty =
+                subchannels[dma_state.subchannel]->current_dirty = memory_manager.IsMemoryDirty(
-                    memory_manager.IsMemoryDirty(dma_state.dma_get, copy_size);
+                    dma_state.dma_get, command_list_header.size * sizeof(u32));
             }
-            memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(), copy_size);
         }
-        ProcessCommands(command_headers);
+        Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+                                     Core::Memory::GuestMemoryFlags::UnsafeRead>
+            headers(memory_manager, dma_state.dma_get, command_list_header.size, &command_headers);
+        ProcessCommands(headers);
     }
 
     return true;

src/video_core/engines/engine_upload.cpp

@@ -5,6 +5,7 @@
 
 #include "common/algorithm.h"
 #include "common/assert.h"
+#include "core/memory.h"
 #include "video_core/engines/engine_upload.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
@@ -46,15 +47,11 @@ void State::ProcessData(const u32* data, size_t num_data) {
 void State::ProcessData(std::span<const u8> read_buffer) {
     const GPUVAddr address{regs.dest.Address()};
     if (is_linear) {
-        if (regs.line_count == 1) {
+        for (size_t line = 0; line < regs.line_count; ++line) {
-            rasterizer->AccelerateInlineToMemory(address, copy_size, read_buffer);
+            const GPUVAddr dest_line = address + line * regs.dest.pitch;
-        } else {
+            std::span<const u8> buffer(read_buffer.data() + line * regs.line_length_in,
-            for (size_t line = 0; line < regs.line_count; ++line) {
+                                       regs.line_length_in);
-                const GPUVAddr dest_line = address + line * regs.dest.pitch;
+            rasterizer->AccelerateInlineToMemory(dest_line, regs.line_length_in, buffer);
-                std::span<const u8> buffer(read_buffer.data() + line * regs.line_length_in,
-                                           regs.line_length_in);
-                rasterizer->AccelerateInlineToMemory(dest_line, regs.line_length_in, buffer);
-            }
         }
     } else {
         u32 width = regs.dest.width;
@@ -70,13 +67,14 @@ void State::ProcessData(std::span<const u8> read_buffer) {
         const std::size_t dst_size = Tegra::Texture::CalculateSize(
             true, bytes_per_pixel, width, regs.dest.height, regs.dest.depth,
             regs.dest.BlockHeight(), regs.dest.BlockDepth());
-        tmp_buffer.resize_destructive(dst_size);
+
-        memory_manager.ReadBlock(address, tmp_buffer.data(), dst_size);
+        Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
-        Tegra::Texture::SwizzleSubrect(tmp_buffer, read_buffer, bytes_per_pixel, width,
+            tmp(memory_manager, address, dst_size, &tmp_buffer);
-                                       regs.dest.height, regs.dest.depth, x_offset, regs.dest.y,
+
-                                       x_elements, regs.line_count, regs.dest.BlockHeight(),
+        Tegra::Texture::SwizzleSubrect(tmp, read_buffer, bytes_per_pixel, width, regs.dest.height,
+                                       regs.dest.depth, x_offset, regs.dest.y, x_elements,
+                                       regs.line_count, regs.dest.BlockHeight(),
                                        regs.dest.BlockDepth(), regs.line_length_in);
-        memory_manager.WriteBlockCached(address, tmp_buffer.data(), dst_size);
     }
 }
 

src/video_core/engines/kepler_compute.cpp

@@ -84,7 +84,6 @@ Texture::TICEntry KeplerCompute::GetTICEntry(u32 tic_index) const {
 
     Texture::TICEntry tic_entry;
     memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
-
     return tic_entry;
 }
 

src/video_core/engines/maxwell_3d.cpp

@@ -9,6 +9,7 @@
 #include "common/settings.h"
 #include "core/core.h"
 #include "core/core_timing.h"
+#include "core/memory.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/draw_manager.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -679,17 +680,14 @@ void Maxwell3D::ProcessCBData(u32 value) {
 Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
     const GPUVAddr tic_address_gpu{regs.tex_header.Address() +
                                    tic_index * sizeof(Texture::TICEntry)};
-
     Texture::TICEntry tic_entry;
     memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
-
     return tic_entry;
 }
 
 Texture::TSCEntry Maxwell3D::GetTSCEntry(u32 tsc_index) const {
     const GPUVAddr tsc_address_gpu{regs.tex_sampler.Address() +
                                    tsc_index * sizeof(Texture::TSCEntry)};
-
     Texture::TSCEntry tsc_entry;
     memory_manager.ReadBlockUnsafe(tsc_address_gpu, &tsc_entry, sizeof(Texture::TSCEntry));
     return tsc_entry;

src/video_core/engines/maxwell_dma.cpp

@@ -7,6 +7,7 @@
 #include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
+#include "core/memory.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/memory_manager.h"
@@ -130,11 +131,12 @@ void MaxwellDMA::Launch() {
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
                 read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
-                    memory_manager.ReadBlock(
+                    Core::Memory::GpuGuestMemoryScoped<
-                        convert_linear_2_blocklinear_addr(regs.offset_in + offset),
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
-                        read_buffer.data(), read_buffer.size());
+                        tmp_write_buffer(memory_manager,
-                    memory_manager.WriteBlockCached(regs.offset_out + offset, read_buffer.data(),
+                                         convert_linear_2_blocklinear_addr(regs.offset_in + offset),
-                                                    read_buffer.size());
+                                         16, &read_buffer);
+                    tmp_write_buffer.SetAddressAndSize(regs.offset_out + offset, 16);
                 }
             } else if (is_src_pitch && !is_dst_pitch) {
                 UNIMPLEMENTED_IF(regs.line_length_in % 16 != 0);
@@ -142,20 +144,19 @@ void MaxwellDMA::Launch() {
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
                 read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
-                    memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(),
+                    Core::Memory::GpuGuestMemoryScoped<
-                                             read_buffer.size());
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
-                    memory_manager.WriteBlockCached(
+                        tmp_write_buffer(memory_manager, regs.offset_in + offset, 16, &read_buffer);
-                        convert_linear_2_blocklinear_addr(regs.offset_out + offset),
+                    tmp_write_buffer.SetAddressAndSize(
-                        read_buffer.data(), read_buffer.size());
+                        convert_linear_2_blocklinear_addr(regs.offset_out + offset), 16);
                 }
             } else {
                 if (!accelerate.BufferCopy(regs.offset_in, regs.offset_out, regs.line_length_in)) {
-                    read_buffer.resize_destructive(regs.line_length_in);
+                    Core::Memory::GpuGuestMemoryScoped<
-                    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(),
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
-                                             regs.line_length_in,
+                        tmp_write_buffer(memory_manager, regs.offset_in, regs.line_length_in,
-                                             VideoCommon::CacheType::NoBufferCache);
+                                         &read_buffer);
-                    memory_manager.WriteBlockCached(regs.offset_out, read_buffer.data(),
+                    tmp_write_buffer.SetAddressAndSize(regs.offset_out, regs.line_length_in);
-                                                    regs.line_length_in);
                 }
             }
         }
@@ -222,17 +223,15 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
 
     const size_t dst_size = dst_operand.pitch * regs.line_count;
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
 
-    memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
-    memory_manager.ReadBlock(dst_operand.address, write_buffer.data(), dst_size);
+        memory_manager, src_operand.address, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, dst_operand.address, dst_size, &write_buffer);
 
-    UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
+    UnswizzleSubrect(tmp_write_buffer, tmp_read_buffer, bytes_per_pixel, width, height, depth,
-                     src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
+                     x_offset, src_params.origin.y, x_elements, regs.line_count, block_height,
-                     dst_operand.pitch);
+                     block_depth, dst_operand.pitch);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }
 
 void MaxwellDMA::CopyPitchToBlockLinear() {
@@ -287,18 +286,17 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
     const size_t src_size = static_cast<size_t>(regs.pitch_in) * regs.line_count;
 
-    read_buffer.resize_destructive(src_size);
+    GPUVAddr src_addr = regs.offset_in;
-    write_buffer.resize_destructive(dst_size);
+    GPUVAddr dst_addr = regs.offset_out;
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
+        memory_manager, src_addr, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, dst_addr, dst_size, &write_buffer);
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+    //  If the input is linear and the output is tiled, swizzle the input and copy it over.
-    memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    SwizzleSubrect(tmp_write_buffer, tmp_read_buffer, bytes_per_pixel, width, height, depth,
-
+                   x_offset, dst_params.origin.y, x_elements, regs.line_count, block_height,
-    // If the input is linear and the output is tiled, swizzle the input and copy it over.
+                   block_depth, regs.pitch_in);
-    SwizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
-                   dst_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
-                   regs.pitch_in);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }
 
 void MaxwellDMA::CopyBlockLinearToBlockLinear() {
@@ -342,23 +340,20 @@ void MaxwellDMA::CopyBlockLinearToBlockLinear() {
     const u32 pitch = x_elements * bytes_per_pixel;
     const size_t mid_buffer_size = pitch * regs.line_count;
 
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
-
     intermediate_buffer.resize_destructive(mid_buffer_size);
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
-    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+        memory_manager, regs.offset_in, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, regs.offset_out, dst_size, &write_buffer);
 
-    UnswizzleSubrect(intermediate_buffer, read_buffer, bytes_per_pixel, src_width, src.height,
+    UnswizzleSubrect(intermediate_buffer, tmp_read_buffer, bytes_per_pixel, src_width, src.height,
                      src.depth, src_x_offset, src.origin.y, x_elements, regs.line_count,
                      src.block_size.height, src.block_size.depth, pitch);
 
-    SwizzleSubrect(write_buffer, intermediate_buffer, bytes_per_pixel, dst_width, dst.height,
+    SwizzleSubrect(tmp_write_buffer, intermediate_buffer, bytes_per_pixel, dst_width, dst.height,
                    dst.depth, dst_x_offset, dst.origin.y, x_elements, regs.line_count,
                    dst.block_size.height, dst.block_size.depth, pitch);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }
 
 void MaxwellDMA::ReleaseSemaphore() {

src/video_core/engines/sw_blitter/blitter.cpp

@@ -159,11 +159,11 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
     const auto src_bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(src.format));
     const auto dst_bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(dst.format));
     const size_t src_size = get_surface_size(src, src_bytes_per_pixel);
-    impl->tmp_buffer.resize_destructive(src_size);
+
-    memory_manager.ReadBlock(src.Address(), impl->tmp_buffer.data(), src_size);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_buffer(
+        memory_manager, src.Address(), src_size, &impl->tmp_buffer);
 
     const size_t src_copy_size = src_extent_x * src_extent_y * src_bytes_per_pixel;
-
     const size_t dst_copy_size = dst_extent_x * dst_extent_y * dst_bytes_per_pixel;
 
     impl->src_buffer.resize_destructive(src_copy_size);
@@ -200,12 +200,11 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
 
     impl->dst_buffer.resize_destructive(dst_copy_size);
     if (src.linear == Fermi2D::MemoryLayout::BlockLinear) {
-        UnswizzleSubrect(impl->src_buffer, impl->tmp_buffer, src_bytes_per_pixel, src.width,
+        UnswizzleSubrect(impl->src_buffer, tmp_buffer, src_bytes_per_pixel, src.width, src.height,
-                         src.height, src.depth, config.src_x0, config.src_y0, src_extent_x,
+                         src.depth, config.src_x0, config.src_y0, src_extent_x, src_extent_y,
-                         src_extent_y, src.block_height, src.block_depth,
+                         src.block_height, src.block_depth, src_extent_x * src_bytes_per_pixel);
-                         src_extent_x * src_bytes_per_pixel);
     } else {
-        process_pitch_linear(false, impl->tmp_buffer, impl->src_buffer, src_extent_x, src_extent_y,
+        process_pitch_linear(false, tmp_buffer, impl->src_buffer, src_extent_x, src_extent_y,
                              src.pitch, config.src_x0, config.src_y0, src_bytes_per_pixel);
     }
 
@@ -221,20 +220,18 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
     }
 
     const size_t dst_size = get_surface_size(dst, dst_bytes_per_pixel);
-    impl->tmp_buffer.resize_destructive(dst_size);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadWrite>
-    memory_manager.ReadBlock(dst.Address(), impl->tmp_buffer.data(), dst_size);
+        tmp_buffer2(memory_manager, dst.Address(), dst_size, &impl->tmp_buffer);
 
     if (dst.linear == Fermi2D::MemoryLayout::BlockLinear) {
-        SwizzleSubrect(impl->tmp_buffer, impl->dst_buffer, dst_bytes_per_pixel, dst.width,
+        SwizzleSubrect(tmp_buffer2, impl->dst_buffer, dst_bytes_per_pixel, dst.width, dst.height,
-                       dst.height, dst.depth, config.dst_x0, config.dst_y0, dst_extent_x,
+                       dst.depth, config.dst_x0, config.dst_y0, dst_extent_x, dst_extent_y,
-                       dst_extent_y, dst.block_height, dst.block_depth,
+                       dst.block_height, dst.block_depth, dst_extent_x * dst_bytes_per_pixel);
-                       dst_extent_x * dst_bytes_per_pixel);
     } else {
-        process_pitch_linear(true, impl->dst_buffer, impl->tmp_buffer, dst_extent_x, dst_extent_y,
+        process_pitch_linear(true, impl->dst_buffer, tmp_buffer2, dst_extent_x, dst_extent_y,
                              dst.pitch, config.dst_x0, config.dst_y0,
                              static_cast<size_t>(dst_bytes_per_pixel));
     }
-    memory_manager.WriteBlock(dst.Address(), impl->tmp_buffer.data(), dst_size);
     return true;
 }
 

src/video_core/memory_manager.cpp

@@ -10,13 +10,13 @@
 #include "core/device_memory.h"
 #include "core/hle/kernel/k_page_table.h"
 #include "core/hle/kernel/k_process.h"
-#include "core/memory.h"
 #include "video_core/invalidation_accumulator.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 
 namespace Tegra {
+using Core::Memory::GuestMemoryFlags;
 
 std::atomic<size_t> MemoryManager::unique_identifier_generator{};
 
@@ -587,13 +587,10 @@ void MemoryManager::InvalidateRegion(GPUVAddr gpu_addr, size_t size,
 
 void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size,
                               VideoCommon::CacheType which) {
-    tmp_buffer.resize_destructive(size);
+    Core::Memory::GpuGuestMemoryScoped<u8, GuestMemoryFlags::SafeReadWrite> data(
-    ReadBlock(gpu_src_addr, tmp_buffer.data(), size, which);
+        *this, gpu_src_addr, size);
-
+    data.SetAddressAndSize(gpu_dest_addr, size);
-    // The output block must be flushed in case it has data modified from the GPU.
-    // Fixes NPC geometry in Zombie Panic in Wonderland DX
     FlushRegion(gpu_dest_addr, size, which);
-    WriteBlock(gpu_dest_addr, tmp_buffer.data(), size, which);
 }
 
 bool MemoryManager::IsGranularRange(GPUVAddr gpu_addr, std::size_t size) const {
@@ -758,4 +755,23 @@ void MemoryManager::FlushCaching() {
     accumulator->Clear();
 }
 
+const u8* MemoryManager::GetSpan(const GPUVAddr src_addr, const std::size_t size) const {
+    auto cpu_addr = GpuToCpuAddress(src_addr);
+    if (cpu_addr) {
+        return memory.GetSpan(*cpu_addr, size);
+    }
+    return nullptr;
+}
+
+u8* MemoryManager::GetSpan(const GPUVAddr src_addr, const std::size_t size) {
+    if (!IsContinuousRange(src_addr, size)) {
+        return nullptr;
+    }
+    auto cpu_addr = GpuToCpuAddress(src_addr);
+    if (cpu_addr) {
+        return memory.GetSpan(*cpu_addr, size);
+    }
+    return nullptr;
+}
+
 } // namespace Tegra

src/video_core/memory_manager.h

@@ -15,6 +15,7 @@
 #include "common/range_map.h"
 #include "common/scratch_buffer.h"
 #include "common/virtual_buffer.h"
+#include "core/memory.h"
 #include "video_core/cache_types.h"
 #include "video_core/pte_kind.h"
 
@@ -62,6 +63,20 @@ public:
     [[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(*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
@@ -139,6 +154,9 @@ public:
 
     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,

src/video_core/texture_cache/texture_cache.h

@@ -8,6 +8,7 @@
 
 #include "common/alignment.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/control/channel_state.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/kepler_compute.h"
@@ -1026,19 +1027,19 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
         runtime.AccelerateImageUpload(image, staging, uploads);
         return;
     }
-    const size_t guest_size_bytes = image.guest_size_bytes;
+
-    swizzle_data_buffer.resize_destructive(guest_size_bytes);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(
-    gpu_memory->ReadBlockUnsafe(gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
+        *gpu_memory, gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
 
     if (True(image.flags & ImageFlagBits::Converted)) {
         unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes);
-        auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer,
+        auto copies =
-                                     unswizzle_data_buffer);
+            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data, unswizzle_data_buffer);
         ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies);
         image.UploadMemory(staging, copies);
     } else {
         const auto copies =
-            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, mapped_span);
+            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data, mapped_span);
         image.UploadMemory(staging, copies);
     }
 }
@@ -1231,11 +1232,12 @@ void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
     decode->image_id = image_id;
     async_decodes.push_back(std::move(decode));
 
-    Common::ScratchBuffer<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes);
+    static Common::ScratchBuffer<u8> local_unswizzle_data_buffer;
-    const size_t guest_size_bytes = image.guest_size_bytes;
+    local_unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes);
-    swizzle_data_buffer.resize_destructive(guest_size_bytes);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(
-    gpu_memory->ReadBlockUnsafe(image.gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
+        *gpu_memory, image.gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
-    auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data_buffer,
+
+    auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data,
                                  local_unswizzle_data_buffer);
     const size_t out_size = MapSizeBytes(image);
 

src/video_core/texture_cache/util.cpp

@@ -20,6 +20,7 @@
 #include "common/div_ceil.h"
 #include "common/scratch_buffer.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
@@ -544,17 +545,15 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
                        tile_size.height, info.tile_width_spacing);
     const size_t subresource_size = sizes[level];
 
-    tmp_buffer.resize_destructive(subresource_size);
-    const std::span<u8> dst(tmp_buffer);
-
     for (s32 layer = 0; layer < info.resources.layers; ++layer) {
         const std::span<const u8> src = input.subspan(host_offset);
-        gpu_memory.ReadBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
+        {
+            Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::UnsafeReadWrite>
+                dst(gpu_memory, gpu_addr + guest_offset, subresource_size, &tmp_buffer);
 
-        SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
+            SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
-                       num_tiles.depth, block.height, block.depth);
+                           num_tiles.depth, block.height, block.depth);
-
+        }
-        gpu_memory.WriteBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
 
         host_offset += host_bytes_per_layer;
         guest_offset += layer_stride;
@@ -837,6 +836,7 @@ boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::Memory
     const Extent3D size = info.size;
 
     if (info.type == ImageType::Linear) {
+        ASSERT(output.size_bytes() >= guest_size_bytes);
         gpu_memory.ReadBlockUnsafe(gpu_addr, output.data(), guest_size_bytes);
 
         ASSERT((info.pitch >> bpp_log2) << bpp_log2 == info.pitch);
@@ -904,16 +904,6 @@ boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::Memory
     return copies;
 }
 
-BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                            const ImageBase& image, std::span<u8> output) {
-    gpu_memory.ReadBlockUnsafe(gpu_addr, output.data(), image.guest_size_bytes);
-    return BufferCopy{
-        .src_offset = 0,
-        .dst_offset = 0,
-        .size = image.guest_size_bytes,
-    };
-}
-
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies) {
     u32 output_offset = 0;

src/video_core/texture_cache/util.h

@@ -66,9 +66,6 @@ struct OverlapResult {
     Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
     std::span<const u8> input, std::span<u8> output);
 
-[[nodiscard]] BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                                          const ImageBase& image, std::span<u8> output);
-
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies);