early-access version 1843

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 1842.
+This is the source code for early-access 1843.
 
 ## Legal Notice
 

src/core/hle/service/nvdrv/nvdrv.cpp

@@ -41,7 +41,6 @@ void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger
 
 Module::Module(Core::System& system)
     : syncpoint_manager{system.GPU()}, service_context{system, "nvdrv"} {
-    auto& kernel = system.Kernel();
     for (u32 i = 0; i < MaxNvEvents; i++) {
         events_interface.events[i].event =
             service_context.CreateEvent(fmt::format("NVDRV::NvEvent_{}", i));

src/tests/video_core/buffer_base.cpp

@@ -536,7 +536,7 @@ TEST_CASE("BufferBase: Cached write downloads") {
     REQUIRE(rasterizer.Count() == 63);
     buffer.MarkRegionAsGpuModified(c + PAGE, PAGE);
     int num = 0;
-    buffer.ForEachDownloadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    buffer.ForEachDownloadRange(c, WORD, true, [&](u64 offset, u64 size) { ++num; });
     buffer.ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
     REQUIRE(num == 0);
     REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));

src/video_core/buffer_cache/buffer_base.h

@@ -226,19 +226,19 @@ public:
     /// Call 'func' for each CPU modified range and unmark those pages as CPU modified
     template <typename Func>
     void ForEachUploadRange(VAddr query_cpu_range, u64 size, Func&& func) {
-        ForEachModifiedRange<Type::CPU>(query_cpu_range, size, func);
+        ForEachModifiedRange<Type::CPU>(query_cpu_range, size, true, func);
     }
 
     /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
     template <typename Func>
-    void ForEachDownloadRange(VAddr query_cpu_range, u64 size, Func&& func) {
+    void ForEachDownloadRange(VAddr query_cpu_range, u64 size, bool clear, Func&& func) {
-        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, func);
+        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, clear, func);
     }
 
     /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
     template <typename Func>
     void ForEachDownloadRange(Func&& func) {
-        ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), func);
+        ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), true, func);
     }
 
     /// Mark buffer as picked
@@ -415,7 +415,7 @@ private:
      * @param func            Function to call for each turned off region
      */
     template <Type type, typename Func>
-    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) {
+    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, bool clear, Func&& func) {
         static_assert(type != Type::Untracked);
 
         const s64 difference = query_cpu_range - cpu_addr;
@@ -467,7 +467,9 @@ private:
             bits = (bits << left_offset) >> left_offset;
 
             const u64 current_word = state_words[word_index] & bits;
+            if (clear) {
                 state_words[word_index] &= ~bits;
+            }
 
             if constexpr (type == Type::CPU) {
                 const u64 current_bits = untracked_words[word_index] & bits;

src/video_core/buffer_cache/buffer_cache.h

@@ -15,6 +15,7 @@
 #include <vector>
 
 #include <boost/container/small_vector.hpp>
+#include <boost/icl/interval_set.hpp>
 
 #include "common/common_types.h"
 #include "common/div_ceil.h"
@@ -77,6 +78,9 @@ class BufferCache {
     using Runtime = typename P::Runtime;
     using Buffer = typename P::Buffer;
 
+    using IntervalSet = boost::icl::interval_set<VAddr>;
+    using IntervalType = typename IntervalSet::interval_type;
+
     struct Empty {};
 
     struct OverlapResult {
@@ -153,6 +157,7 @@ public:
 
     /// Commit asynchronous downloads
     void CommitAsyncFlushes();
+    void CommitAsyncFlushesHigh();
 
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
@@ -160,6 +165,9 @@ public:
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
+    /// Return true when a CPU region is modified from the GPU
+    [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
+
     std::mutex mutex;
 
 private:
@@ -272,8 +280,6 @@ private:
 
     void DeleteBuffer(BufferId buffer_id);
 
-    void ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id);
-
     void NotifyBufferDeletion();
 
     [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr) const;
@@ -327,9 +333,7 @@ private:
 
     std::vector<BufferId> cached_write_buffer_ids;
 
-    // TODO: This data structure is not optimal and it should be reworked
+    IntervalSet uncommitted_ranges;
-    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;
@@ -547,29 +551,18 @@ void BufferCache<P>::FlushCachedWrites() {
 
 template <class P>
 bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
-    return !uncommitted_downloads.empty();
+    return !uncommitted_ranges.empty();
 }
 
 template <class P>
 bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept {
-    return !committed_downloads.empty() && !committed_downloads.front().empty();
+    return false;
 }
 
 template <class P>
-void BufferCache<P>::CommitAsyncFlushes() {
+void BufferCache<P>::CommitAsyncFlushesHigh() {
-    // This is intentionally passing the value by copy
+    const IntervalSet& intervals = uncommitted_ranges;
-    committed_downloads.push_front(uncommitted_downloads);
+    if (intervals.empty()) {
-    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);
@@ -577,18 +570,35 @@ void BufferCache<P>::PopAsyncFlushes() {
     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) {
+    for (auto& interval : intervals) {
-        slot_buffers[buffer_id].ForEachDownloadRange([&](u64 range_offset, u64 range_size) {
+        const std::size_t size = interval.upper() - interval.lower();
+        const VAddr cpu_addr = interval.lower();
+        const VAddr cpu_addr_end = interval.upper();
+        ForEachBufferInRange(cpu_addr, size, [&](BufferId buffer_id, Buffer& buffer) {
+            boost::container::small_vector<BufferCopy, 1> copies;
+            buffer.ForEachDownloadRange(
+                cpu_addr, size, false, [&](u64 range_offset, u64 range_size) {
+                    VAddr cpu_addr_base = buffer.CpuAddr() + range_offset;
+                    VAddr cpu_addr_end2 = cpu_addr_base + range_size;
+                    const s64 difference = s64(cpu_addr_end2 - cpu_addr_end);
+                    cpu_addr_end2 -= u64(std::max<s64>(difference, 0));
+                    const s64 difference2 = s64(cpu_addr - cpu_addr_base);
+                    cpu_addr_base += u64(std::max<s64>(difference2, 0));
+                    const u64 new_size = cpu_addr_end2 - cpu_addr_base;
+                    const u64 new_offset = cpu_addr_base - buffer.CpuAddr();
+                    ASSERT(!IsRegionCpuModified(cpu_addr_base, new_size));
                     downloads.push_back({
                         BufferCopy{
-                    .src_offset = range_offset,
+                            .src_offset = new_offset,
                             .dst_offset = total_size_bytes,
-                    .size = range_size,
+                            .size = new_size,
                         },
                         buffer_id,
                     });
-            total_size_bytes += range_size;
+                    total_size_bytes += new_size;
-            largest_copy = std::max(largest_copy, range_size);
+                    buffer.UnmarkRegionAsGpuModified(cpu_addr_base, new_size);
+                    largest_copy = std::max(largest_copy, new_size);
+                });
         });
     }
     if (downloads.empty()) {
@@ -622,6 +632,18 @@ void BufferCache<P>::PopAsyncFlushes() {
     }
 }
 
+template <class P>
+void BufferCache<P>::CommitAsyncFlushes() {
+    if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) {
+        CommitAsyncFlushesHigh();
+    } else {
+        uncommitted_ranges.clear();
+    }
+}
+
+template <class P>
+void BufferCache<P>::PopAsyncFlushes() {}
+
 template <class P>
 bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
     const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
@@ -641,6 +663,25 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
     return false;
 }
 
+template <class P>
+bool BufferCache<P>::IsRegionCpuModified(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.IsRegionCpuModified(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];
@@ -1010,16 +1051,14 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
     Buffer& buffer = slot_buffers[buffer_id];
     buffer.MarkRegionAsGpuModified(cpu_addr, size);
 
-    const bool is_accuracy_high = Settings::IsGPULevelHigh();
+    const bool is_accuracy_high =
+        Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High;
     const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
-    if (!is_accuracy_high || !is_async) {
+    if (!is_async && !is_accuracy_high) {
         return;
     }
-    if (std::ranges::find(uncommitted_downloads, buffer_id) != uncommitted_downloads.end()) {
+    const IntervalType base_interval{cpu_addr, cpu_addr + size};
-        // Already inserted
+    uncommitted_ranges.add(base_interval);
-        return;
-    }
-    uncommitted_downloads.push_back(buffer_id);
 }
 
 template <class P>
@@ -1103,7 +1142,6 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
     if (!copies.empty()) {
         runtime.CopyBuffer(slot_buffers[new_buffer_id], overlap, copies);
     }
-    ReplaceBufferDownloads(overlap_id, new_buffer_id);
     DeleteBuffer(overlap_id);
 }
 
@@ -1244,7 +1282,7 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
     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) {
+    buffer.ForEachDownloadRange(cpu_addr, size, true, [&](u64 range_offset, u64 range_size) {
         copies.push_back(BufferCopy{
             .src_offset = range_offset,
             .dst_offset = total_size_bytes,
@@ -1315,18 +1353,6 @@ void BufferCache<P>::DeleteBuffer(BufferId 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) {

src/video_core/engines/fermi_2d.cpp

@@ -65,19 +65,20 @@ void Fermi2D::Blit() {
         .src_x1 = static_cast<s32>((args.du_dx * args.dst_width + args.src_x0) >> 32),
         .src_y1 = static_cast<s32>((args.dv_dy * args.dst_height + args.src_y0) >> 32),
     };
-
     Surface src = regs.src;
-    s32 src_address_offset = 0;
     const auto bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(src.format));
-    if (src.linear == Tegra::Engines::Fermi2D::MemoryLayout::Pitch && src.width == config.src_x1 &&
+    const auto is_copy_out_of_bound =
-        config.src_x1 > static_cast<s32>(src.pitch / bytes_per_pixel) && config.src_x0 > 0) {
+        src.linear == Tegra::Engines::Fermi2D::MemoryLayout::Pitch && src.width == config.src_x1 &&
-        src_address_offset = config.src_x0 * bytes_per_pixel;
+        config.src_x1 > static_cast<s32>(src.pitch / bytes_per_pixel) && config.src_x0 > 0;
+    if (is_copy_out_of_bound) {
+        auto address = src.Address() + config.src_x0 * bytes_per_pixel;
+        src.addr_upper = static_cast<u32>(address >> 32);
+        src.addr_lower = static_cast<u32>(address);
         src.width -= config.src_x0;
         config.src_x1 -= config.src_x0;
         config.src_x0 = 0;
     }
-
+    if (!rasterizer->AccelerateSurfaceCopy(src, regs.dst, config)) {
-    if (!rasterizer->AccelerateSurfaceCopy(src, src_address_offset, regs.dst, config)) {
         UNIMPLEMENTED();
     }
 }

src/video_core/fence_manager.h

@@ -96,6 +96,23 @@ public:
         }
     }
 
+    void TryReleasePendingFences() {
+        while (!fences.empty()) {
+            TFence& current_fence = fences.front();
+            if (ShouldWait() && !IsFenceSignaled(current_fence)) {
+                return;
+            }
+            PopAsyncFlushes();
+            if (current_fence->IsSemaphore()) {
+                gpu_memory.template Write<u32>(current_fence->GetAddress(),
+                                               current_fence->GetPayload());
+            } else {
+                gpu.IncrementSyncPoint(current_fence->GetPayload());
+            }
+            PopFence();
+        }
+    }
+
 protected:
     explicit FenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
                           TTextureCache& texture_cache_, TTBufferCache& buffer_cache_,
@@ -125,23 +142,6 @@ protected:
     TQueryCache& query_cache;
 
 private:
-    void TryReleasePendingFences() {
-        while (!fences.empty()) {
-            TFence& current_fence = fences.front();
-            if (ShouldWait() && !IsFenceSignaled(current_fence)) {
-                return;
-            }
-            PopAsyncFlushes();
-            if (current_fence->IsSemaphore()) {
-                gpu_memory.template Write<u32>(current_fence->GetAddress(),
-                                               current_fence->GetPayload());
-            } else {
-                gpu.IncrementSyncPoint(current_fence->GetPayload());
-            }
-            PopFence();
-        }
-    }
-
     bool ShouldWait() const {
         std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         return texture_cache.ShouldWaitAsyncFlushes() || buffer_cache.ShouldWaitAsyncFlushes() ||

src/video_core/gpu_thread.cpp

@@ -8,6 +8,7 @@
 #include "common/settings.h"
 #include "common/thread.h"
 #include "core/core.h"
+#include "core/core_timing.h"
 #include "core/frontend/emu_window.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/gpu.h"
@@ -83,6 +84,17 @@ void ThreadManager::StartThread(VideoCore::RendererBase& renderer,
     rasterizer = renderer.ReadRasterizer();
     thread = std::thread(RunThread, std::ref(system), std::ref(renderer), std::ref(context),
                          std::ref(dma_pusher), std::ref(state));
+    gpu_sync_event = Core::Timing::CreateEvent(
+        "GPUHostSyncCallback", [this](std::uintptr_t, std::chrono::nanoseconds) {
+            if (!state.is_running) {
+                return;
+            }
+
+            OnCommandListEnd();
+            const auto time_interval = std::chrono::nanoseconds{500 * 1000};
+            system.CoreTiming().ScheduleEvent(time_interval, gpu_sync_event);
+        });
+    system.CoreTiming().ScheduleEvent(std::chrono::nanoseconds{500 * 1000}, gpu_sync_event);
 }
 
 void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
@@ -128,6 +140,9 @@ void ThreadManager::ShutDown() {
         state.cv.notify_all();
     }
 
+    system.CoreTiming().UnscheduleEvent(gpu_sync_event, 0);
+    system.CoreTiming().RemoveEvent(gpu_sync_event);
+
     if (!thread.joinable()) {
         return;
     }

src/video_core/gpu_thread.h

@@ -20,6 +20,10 @@ class DmaPusher;
 } // namespace Tegra
 
 namespace Core {
+namespace Timing {
+class CoreTiming;
+struct EventType;
+} // namespace Timing
 namespace Frontend {
 class GraphicsContext;
 }
@@ -150,6 +154,7 @@ private:
 
     SynchState state;
     std::thread thread;
+    std::shared_ptr<Core::Timing::EventType> gpu_sync_event;
 };
 
 } // namespace VideoCommon::GPUThread

src/video_core/memory_manager.cpp

@@ -69,11 +69,16 @@ void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {
     } else {
         UNREACHABLE_MSG("Unmapping non-existent GPU address=0x{:x}", gpu_addr);
     }
+
+    const auto submapped_ranges = GetSubmappedRange(gpu_addr, size);
+
+    for (const auto& map : submapped_ranges) {
         // Flush and invalidate through the GPU interface, to be asynchronous if possible.
-    const std::optional<VAddr> cpu_addr = GpuToCpuAddress(gpu_addr);
+        const std::optional<VAddr> cpu_addr = GpuToCpuAddress(map.first);
         ASSERT(cpu_addr);
 
-    rasterizer->UnmapMemory(*cpu_addr, size);
+        rasterizer->UnmapMemory(*cpu_addr, map.second);
+    }
 
     UpdateRange(gpu_addr, PageEntry::State::Unmapped, size);
 }
@@ -146,8 +151,14 @@ void MemoryManager::SetPageEntry(GPUVAddr gpu_addr, PageEntry page_entry, std::s
 
     //// Lock the new page
     // TryLockPage(page_entry, size);
+    auto& current_page = page_table[PageEntryIndex(gpu_addr)];
 
-    page_table[PageEntryIndex(gpu_addr)] = page_entry;
+    if ((!current_page.IsValid() && page_entry.IsValid()) ||
+        current_page.ToAddress() != page_entry.ToAddress()) {
+        rasterizer->ModifyGPUMemory(gpu_addr, size);
+    }
+
+    current_page = page_entry;
 }
 
 std::optional<GPUVAddr> MemoryManager::FindFreeRange(std::size_t size, std::size_t align,
@@ -193,6 +204,19 @@ std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) const {
     return page_entry.ToAddress() + (gpu_addr & page_mask);
 }
 
+std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr addr, std::size_t size) const {
+    size_t page_index{addr >> page_bits};
+    const size_t page_last{(addr + size + page_size - 1) >> page_bits};
+    while (page_index < page_last) {
+        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
+        if (page_addr && *page_addr != 0) {
+            return page_addr;
+        }
+        ++page_index;
+    }
+    return std::nullopt;
+}
+
 template <typename T>
 T MemoryManager::Read(GPUVAddr addr) const {
     if (auto page_pointer{GetPointer(addr)}; page_pointer) {
@@ -389,4 +413,79 @@ bool MemoryManager::IsGranularRange(GPUVAddr gpu_addr, std::size_t size) const {
     return page <= Core::Memory::PAGE_SIZE;
 }
 
+bool MemoryManager::IsContinousRange(GPUVAddr gpu_addr, std::size_t size) const {
+    size_t page_index{gpu_addr >> page_bits};
+    const size_t page_last{(gpu_addr + size + page_size - 1) >> page_bits};
+    std::optional<VAddr> old_page_addr{};
+    while (page_index != page_last) {
+        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
+        if (!page_addr || *page_addr == 0) {
+            return false;
+        }
+        if (old_page_addr) {
+            if (*old_page_addr + page_size != *page_addr) {
+                return false;
+            }
+        }
+        old_page_addr = page_addr;
+        ++page_index;
+    }
+    return true;
+}
+
+bool MemoryManager::IsFullyMappedRange(GPUVAddr gpu_addr, std::size_t size) const {
+    size_t page_index{gpu_addr >> page_bits};
+    const size_t page_last{(gpu_addr + size + page_size - 1) >> page_bits};
+    while (page_index < page_last) {
+        if (!page_table[page_index].IsValid() || page_table[page_index].ToAddress() == 0) {
+            return false;
+        }
+        ++page_index;
+    }
+    return true;
+}
+
+std::vector<std::pair<GPUVAddr, std::size_t>> MemoryManager::GetSubmappedRange(
+    GPUVAddr gpu_addr, std::size_t size) const {
+    std::vector<std::pair<GPUVAddr, std::size_t>> result{};
+    size_t page_index{gpu_addr >> page_bits};
+    size_t remaining_size{size};
+    size_t page_offset{gpu_addr & page_mask};
+    std::optional<std::pair<GPUVAddr, std::size_t>> last_segment{};
+    std::optional<VAddr> old_page_addr{};
+    const auto extend_size = [this, &last_segment, &page_index](std::size_t bytes) {
+        if (!last_segment) {
+            GPUVAddr new_base_addr = page_index << page_bits;
+            last_segment = {new_base_addr, bytes};
+        } else {
+            last_segment->second += bytes;
+        }
+    };
+    const auto split = [this, &last_segment, &result] {
+        if (last_segment) {
+            result.push_back(*last_segment);
+            last_segment = std::nullopt;
+        }
+    };
+    while (remaining_size > 0) {
+        const size_t num_bytes{std::min(page_size - page_offset, remaining_size)};
+        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
+        if (!page_addr) {
+            split();
+        } else if (old_page_addr) {
+            if (*old_page_addr + page_size != *page_addr) {
+                split();
+            }
+            extend_size(num_bytes);
+        } else {
+            extend_size(num_bytes);
+        }
+        ++page_index;
+        page_offset = 0;
+        remaining_size -= num_bytes;
+    }
+    split();
+    return result;
+}
+
 } // namespace Tegra

src/video_core/memory_manager.h

@@ -76,6 +76,8 @@ public:
 
     [[nodiscard]] std::optional<VAddr> GpuToCpuAddress(GPUVAddr addr) const;
 
+    [[nodiscard]] std::optional<VAddr> GpuToCpuAddress(GPUVAddr addr, std::size_t size) const;
+
     template <typename T>
     [[nodiscard]] T Read(GPUVAddr addr) const;
 
@@ -112,10 +114,28 @@ public:
     void WriteBlockUnsafe(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size);
 
     /**
-     * IsGranularRange checks if a gpu region can be simply read with a pointer.
+     * 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 cpu addresses.
+     */
+    [[nodiscard]] bool IsContinousRange(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 cpu addresses mapped beneath.
+     * if the region is continous, a single pair will be returned. If it's unmapped, an empty vector
+     * will be returned;
+     */
+    std::vector<std::pair<GPUVAddr, std::size_t>> GetSubmappedRange(GPUVAddr gpu_addr,
+                                                                    std::size_t size) const;
+
     [[nodiscard]] GPUVAddr Map(VAddr cpu_addr, GPUVAddr gpu_addr, std::size_t size);
     [[nodiscard]] GPUVAddr MapAllocate(VAddr cpu_addr, std::size_t size, std::size_t align);
     [[nodiscard]] GPUVAddr MapAllocate32(VAddr cpu_addr, std::size_t size);

src/video_core/rasterizer_interface.h

@@ -93,6 +93,9 @@ public:
     /// Unmap memory range
     virtual void UnmapMemory(VAddr addr, u64 size) = 0;
 
+    /// Remap GPU memory range. This means underneath backing memory changed
+    virtual void ModifyGPUMemory(GPUVAddr addr, u64 size) = 0;
+
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     /// and invalidated
     virtual void FlushAndInvalidateRegion(VAddr addr, u64 size) = 0;
@@ -114,8 +117,7 @@ public:
 
     /// Attempt to use a faster method to perform a surface copy
     [[nodiscard]] virtual bool AccelerateSurfaceCopy(
-        const Tegra::Engines::Fermi2D::Surface& src, s32 src_address_offset,
+        const Tegra::Engines::Fermi2D::Surface& src, const Tegra::Engines::Fermi2D::Surface& dst,
-        const Tegra::Engines::Fermi2D::Surface& dst,
         const Tegra::Engines::Fermi2D::Config& copy_config) {
         return false;
     }

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -631,6 +631,13 @@ void RasterizerOpenGL::UnmapMemory(VAddr addr, u64 size) {
     shader_cache.OnCPUWrite(addr, size);
 }
 
+void RasterizerOpenGL::ModifyGPUMemory(GPUVAddr addr, u64 size) {
+    {
+        std::scoped_lock lock{texture_cache.mutex};
+        texture_cache.UnmapGPUMemory(addr, size);
+    }
+}
+
 void RasterizerOpenGL::SignalSemaphore(GPUVAddr addr, u32 value) {
     if (!gpu.IsAsync()) {
         gpu_memory.Write<u32>(addr, value);
@@ -698,12 +705,11 @@ void RasterizerOpenGL::TickFrame() {
 }
 
 bool RasterizerOpenGL::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
-                                             s32 src_address_offset,
                                              const Tegra::Engines::Fermi2D::Surface& dst,
                                              const Tegra::Engines::Fermi2D::Config& copy_config) {
     MICROPROFILE_SCOPE(OpenGL_Blits);
     std::scoped_lock lock{texture_cache.mutex};
-    texture_cache.BlitImage(dst, src, src_address_offset, copy_config);
+    texture_cache.BlitImage(dst, src, copy_config);
     return true;
 }
 

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -82,6 +82,7 @@ public:
     void OnCPUWrite(VAddr addr, u64 size) override;
     void SyncGuestHost() override;
     void UnmapMemory(VAddr addr, u64 size) override;
+    void ModifyGPUMemory(GPUVAddr addr, u64 size) override;
     void SignalSemaphore(GPUVAddr addr, u32 value) override;
     void SignalSyncPoint(u32 value) override;
     void ReleaseFences() override;
@@ -91,7 +92,7 @@ public:
     void TiledCacheBarrier() override;
     void FlushCommands() override;
     void TickFrame() override;
-    bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src, s32 src_address_offset,
+    bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                const Tegra::Engines::Fermi2D::Surface& dst,
                                const Tegra::Engines::Fermi2D::Config& copy_config) override;
     bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -34,6 +34,10 @@ bool InnerFence::IsSignaled() const {
     if (is_stubbed) {
         return true;
     }
+    if (scheduler.IsFree(wait_tick)) {
+        return true;
+    }
+    scheduler.Refresh();
     return scheduler.IsFree(wait_tick);
 }
 

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -577,6 +577,13 @@ void RasterizerVulkan::UnmapMemory(VAddr addr, u64 size) {
     pipeline_cache.OnCPUWrite(addr, size);
 }
 
+void RasterizerVulkan::ModifyGPUMemory(GPUVAddr addr, u64 size) {
+    {
+        std::scoped_lock lock{texture_cache.mutex};
+        texture_cache.UnmapGPUMemory(addr, size);
+    }
+}
+
 void RasterizerVulkan::SignalSemaphore(GPUVAddr addr, u32 value) {
     if (!gpu.IsAsync()) {
         gpu_memory.Write<u32>(addr, value);
@@ -597,7 +604,7 @@ void RasterizerVulkan::ReleaseFences() {
     if (!gpu.IsAsync()) {
         return;
     }
-    fence_manager.WaitPendingFences();
+    fence_manager.TryReleasePendingFences();
 }
 
 void RasterizerVulkan::FlushAndInvalidateRegion(VAddr addr, u64 size) {
@@ -658,11 +665,10 @@ void RasterizerVulkan::TickFrame() {
 }
 
 bool RasterizerVulkan::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
-                                             s32 src_address_offset,
                                              const Tegra::Engines::Fermi2D::Surface& dst,
                                              const Tegra::Engines::Fermi2D::Config& copy_config) {
     std::scoped_lock lock{texture_cache.mutex};
-    texture_cache.BlitImage(dst, src, src_address_offset, copy_config);
+    texture_cache.BlitImage(dst, src, copy_config);
     return true;
 }
 

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -74,6 +74,7 @@ public:
     void OnCPUWrite(VAddr addr, u64 size) override;
     void SyncGuestHost() override;
     void UnmapMemory(VAddr addr, u64 size) override;
+    void ModifyGPUMemory(GPUVAddr addr, u64 size) override;
     void SignalSemaphore(GPUVAddr addr, u32 value) override;
     void SignalSyncPoint(u32 value) override;
     void ReleaseFences() override;
@@ -83,7 +84,7 @@ public:
     void TiledCacheBarrier() override;
     void FlushCommands() override;
     void TickFrame() override;
-    bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src, s32 src_address_offset,
+    bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                const Tegra::Engines::Fermi2D::Surface& dst,
                                const Tegra::Engines::Fermi2D::Config& copy_config) override;
     bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -83,6 +83,10 @@ public:
         return master_semaphore->IsFree(tick);
     }
 
+    void Refresh() const noexcept {
+        return master_semaphore->Refresh();
+    }
+
     /// Waits for the given tick to trigger on the GPU.
     void Wait(u64 tick) {
         master_semaphore->Wait(tick);

src/video_core/texture_cache/image_base.cpp

@@ -69,6 +69,9 @@ ImageBase::ImageBase(const ImageInfo& info_, GPUVAddr gpu_addr_, VAddr cpu_addr_
     }
 }
 
+ImageMapView::ImageMapView(GPUVAddr gpu_addr_, VAddr cpu_addr_, size_t size_, ImageId image_id_)
+    : gpu_addr{gpu_addr_}, cpu_addr{cpu_addr_}, size{size_}, image_id{image_id_} {}
+
 std::optional<SubresourceBase> ImageBase::TryFindBase(GPUVAddr other_addr) const noexcept {
     if (other_addr < gpu_addr) {
         // Subresource address can't be lower than the base

src/video_core/texture_cache/image_base.h

@@ -25,11 +25,13 @@ enum class ImageFlagBits : u32 {
     Strong = 1 << 5,      ///< Exists in the image table, the dimensions are can be trusted
     Registered = 1 << 6,  ///< True when the image is registered
     Picked = 1 << 7,      ///< Temporary flag to mark the image as picked
+    Remapped = 1 << 8,    ///< Image has been remapped.
+    Sparse = 1 << 9,      ///< Image has non continous submemory.
 
     // Garbage Collection Flags
-    BadOverlap = 1 << 8, ///< This image overlaps other but doesn't fit, has higher
+    BadOverlap = 1 << 10, ///< This image overlaps other but doesn't fit, has higher
                           ///< garbage collection priority
-    Alias = 1 << 9,      ///< This image has aliases and has priority on garbage
+    Alias = 1 << 11,      ///< This image has aliases and has priority on garbage
                           ///< collection
 };
 DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)
@@ -57,6 +59,12 @@ struct ImageBase {
         return cpu_addr < overlap_end && overlap_cpu_addr < cpu_addr_end;
     }
 
+    [[nodiscard]] bool OverlapsGPU(GPUVAddr overlap_gpu_addr, size_t overlap_size) const noexcept {
+        const VAddr overlap_end = overlap_gpu_addr + overlap_size;
+        const GPUVAddr gpu_addr_end = gpu_addr + guest_size_bytes;
+        return gpu_addr < overlap_end && overlap_gpu_addr < gpu_addr_end;
+    }
+
     void CheckBadOverlapState();
     void CheckAliasState();
 
@@ -84,6 +92,29 @@ struct ImageBase {
 
     std::vector<AliasedImage> aliased_images;
     std::vector<ImageId> overlapping_images;
+    ImageMapId map_view_id{};
+};
+
+struct ImageMapView {
+    explicit ImageMapView(GPUVAddr gpu_addr, VAddr cpu_addr, size_t size, ImageId image_id);
+
+    [[nodiscard]] bool Overlaps(VAddr overlap_cpu_addr, size_t overlap_size) const noexcept {
+        const VAddr overlap_end = overlap_cpu_addr + overlap_size;
+        const VAddr cpu_addr_end = cpu_addr + size;
+        return cpu_addr < overlap_end && overlap_cpu_addr < cpu_addr_end;
+    }
+
+    [[nodiscard]] bool OverlapsGPU(GPUVAddr overlap_gpu_addr, size_t overlap_size) const noexcept {
+        const GPUVAddr overlap_end = overlap_gpu_addr + overlap_size;
+        const GPUVAddr gpu_addr_end = gpu_addr + size;
+        return gpu_addr < overlap_end && overlap_gpu_addr < gpu_addr_end;
+    }
+
+    GPUVAddr gpu_addr;
+    VAddr cpu_addr;
+    size_t size;
+    ImageId image_id;
+    bool picked{};
 };
 
 struct ImageAllocBase {

src/video_core/texture_cache/texture_cache.h

@@ -13,6 +13,7 @@
 #include <span>
 #include <type_traits>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 
@@ -152,9 +153,12 @@ public:
     /// Remove images in a region
     void UnmapMemory(VAddr cpu_addr, size_t size);
 
+    /// Remove images in a region
+    void UnmapGPUMemory(GPUVAddr gpu_addr, size_t size);
+
     /// Blit an image with the given parameters
     void BlitImage(const Tegra::Engines::Fermi2D::Surface& dst,
-                   const Tegra::Engines::Fermi2D::Surface& src, s32 src_address_offset,
+                   const Tegra::Engines::Fermi2D::Surface& src,
                    const Tegra::Engines::Fermi2D::Config& copy);
 
     /// Invalidate the contents of the color buffer index
@@ -188,7 +192,22 @@ public:
 private:
     /// Iterate over all page indices in a range
     template <typename Func>
-    static void ForEachPage(VAddr addr, size_t size, Func&& func) {
+    static void ForEachCPUPage(VAddr addr, size_t size, Func&& func) {
+        static constexpr bool RETURNS_BOOL = std::is_same_v<std::invoke_result<Func, u64>, bool>;
+        const u64 page_end = (addr + size - 1) >> PAGE_BITS;
+        for (u64 page = addr >> PAGE_BITS; page <= page_end; ++page) {
+            if constexpr (RETURNS_BOOL) {
+                if (func(page)) {
+                    break;
+                }
+            } else {
+                func(page);
+            }
+        }
+    }
+
+    template <typename Func>
+    static void ForEachGPUPage(GPUVAddr addr, size_t size, Func&& func) {
         static constexpr bool RETURNS_BOOL = std::is_same_v<std::invoke_result<Func, u64>, bool>;
         const u64 page_end = (addr + size - 1) >> PAGE_BITS;
         for (u64 page = addr >> PAGE_BITS; page <= page_end; ++page) {
@@ -218,7 +237,7 @@ private:
     FramebufferId GetFramebufferId(const RenderTargets& key);
 
     /// Refresh the contents (pixel data) of an image
-    void RefreshContents(Image& image);
+    void RefreshContents(Image& image, ImageId image_id);
 
     /// Upload data from guest to an image
     template <typename StagingBuffer>
@@ -248,8 +267,7 @@ private:
 
     /// Return a blit image pair from the given guest blit parameters
     [[nodiscard]] BlitImages GetBlitImages(const Tegra::Engines::Fermi2D::Surface& dst,
-                                           const Tegra::Engines::Fermi2D::Surface& src,
+                                           const Tegra::Engines::Fermi2D::Surface& src);
-                                           s32 src_address_offset);
 
     /// Find or create a sampler from a guest descriptor sampler
     [[nodiscard]] SamplerId FindSampler(const TSCEntry& config);
@@ -268,6 +286,16 @@ private:
     template <typename Func>
     void ForEachImageInRegion(VAddr cpu_addr, size_t size, Func&& func);
 
+    template <typename Func>
+    void ForEachImageInRegionGPU(GPUVAddr gpu_addr, size_t size, Func&& func);
+
+    template <typename Func>
+    void ForEachSparseImageInRegion(GPUVAddr gpu_addr, size_t size, Func&& func);
+
+    /// Iterates over all the images in a region calling func
+    template <typename Func>
+    void ForEachSparseSegment(ImageBase& image, Func&& func);
+
     /// Find or create an image view in the given image with the passed parameters
     [[nodiscard]] ImageViewId FindOrEmplaceImageView(ImageId image_id, const ImageViewInfo& info);
 
@@ -278,10 +306,10 @@ private:
     void UnregisterImage(ImageId image);
 
     /// Track CPU reads and writes for image
-    void TrackImage(ImageBase& image);
+    void TrackImage(ImageBase& image, ImageId image_id);
 
     /// Stop tracking CPU reads and writes for image
-    void UntrackImage(ImageBase& image);
+    void UntrackImage(ImageBase& image, ImageId image_id);
 
     /// Delete image from the cache
     void DeleteImage(ImageId image);
@@ -339,7 +367,13 @@ private:
     std::unordered_map<TSCEntry, SamplerId> samplers;
     std::unordered_map<RenderTargets, FramebufferId> framebuffers;
 
-    std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> page_table;
+    std::unordered_map<u64, std::vector<ImageMapId>, IdentityHash<u64>> page_table;
+    std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> gpu_page_table;
+    std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> sparse_page_table;
+
+    std::unordered_map<ImageId, std::vector<ImageViewId>> sparse_views;
+
+    VAddr virtual_invalid_space{};
 
     bool has_deleted_images = false;
     u64 total_used_memory = 0;
@@ -348,6 +382,7 @@ private:
     u64 critical_memory;
 
     SlotVector<Image> slot_images;
+    SlotVector<ImageMapView> slot_map_views;
     SlotVector<ImageView> slot_image_views;
     SlotVector<ImageAlloc> slot_image_allocs;
     SlotVector<Sampler> slot_samplers;
@@ -458,7 +493,7 @@ void TextureCache<P>::RunGarbageCollector() {
                 }
             }
             if (True(image->flags & ImageFlagBits::Tracked)) {
-                UntrackImage(*image);
+                UntrackImage(*image, image_id);
             }
             UnregisterImage(image_id);
             DeleteImage(image_id);
@@ -657,7 +692,9 @@ void TextureCache<P>::WriteMemory(VAddr cpu_addr, size_t size) {
             return;
         }
         image.flags |= ImageFlagBits::CpuModified;
-        UntrackImage(image);
+        if (True(image.flags & ImageFlagBits::Tracked)) {
+            UntrackImage(image, image_id);
+        }
     });
 }
 
@@ -694,18 +731,35 @@ void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
     for (const ImageId id : deleted_images) {
         Image& image = slot_images[id];
         if (True(image.flags & ImageFlagBits::Tracked)) {
-            UntrackImage(image);
+            UntrackImage(image, id);
         }
         UnregisterImage(id);
         DeleteImage(id);
     }
 }
 
+template <class P>
+void TextureCache<P>::UnmapGPUMemory(GPUVAddr gpu_addr, size_t size) {
+    std::vector<ImageId> deleted_images;
+    ForEachImageInRegionGPU(gpu_addr, size,
+                            [&](ImageId id, Image&) { deleted_images.push_back(id); });
+    for (const ImageId id : deleted_images) {
+        Image& image = slot_images[id];
+        if (True(image.flags & ImageFlagBits::Remapped)) {
+            continue;
+        }
+        image.flags |= ImageFlagBits::Remapped;
+        if (True(image.flags & ImageFlagBits::Tracked)) {
+            UntrackImage(image, id);
+        }
+    }
+}
+
 template <class P>
 void TextureCache<P>::BlitImage(const Tegra::Engines::Fermi2D::Surface& dst,
-                                const Tegra::Engines::Fermi2D::Surface& src, s32 src_address_offset,
+                                const Tegra::Engines::Fermi2D::Surface& src,
                                 const Tegra::Engines::Fermi2D::Config& copy) {
-    const BlitImages images = GetBlitImages(dst, src, src_address_offset);
+    const BlitImages images = GetBlitImages(dst, src);
     const ImageId dst_id = images.dst_id;
     const ImageId src_id = images.src_id;
     PrepareImage(src_id, false, false);
@@ -725,7 +779,7 @@ void TextureCache<P>::BlitImage(const Tegra::Engines::Fermi2D::Surface& dst,
         Offset2D{.x = copy.src_x1 >> src_samples_x, .y = copy.src_y1 >> src_samples_y},
     };
 
-    const std::optional src_base = src_image.TryFindBase(src.Address() + src_address_offset);
+    const std::optional src_base = src_image.TryFindBase(src.Address());
     const SubresourceRange src_range{.base = src_base.value(), .extent = {1, 1}};
     const ImageViewInfo src_view_info(ImageViewType::e2D, images.src_format, src_range);
     const auto [src_framebuffer_id, src_view_id] = RenderTargetFromImage(src_id, src_view_info);
@@ -793,9 +847,10 @@ typename P::ImageView* TextureCache<P>::TryFindFramebufferImageView(VAddr cpu_ad
     if (it == page_table.end()) {
         return nullptr;
     }
-    const auto& image_ids = it->second;
+    const auto& image_map_ids = it->second;
-    for (const ImageId image_id : image_ids) {
+    for (const ImageMapId map_id : image_map_ids) {
-        const ImageBase& image = slot_images[image_id];
+        const ImageMapView& map = slot_map_views[map_id];
+        const ImageBase& image = slot_images[map.image_id];
         if (image.cpu_addr != cpu_addr) {
             continue;
         }
@@ -875,13 +930,13 @@ bool TextureCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
 }
 
 template <class P>
-void TextureCache<P>::RefreshContents(Image& image) {
+void TextureCache<P>::RefreshContents(Image& image, ImageId image_id) {
     if (False(image.flags & ImageFlagBits::CpuModified)) {
         // Only upload modified images
         return;
     }
     image.flags &= ~ImageFlagBits::CpuModified;
-    TrackImage(image);
+    TrackImage(image, image_id);
 
     if (image.info.num_samples > 1) {
         LOG_WARNING(HW_GPU, "MSAA image uploads are not implemented");
@@ -918,7 +973,7 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
 
 template <class P>
 ImageViewId TextureCache<P>::FindImageView(const TICEntry& config) {
-    if (!IsValidAddress(gpu_memory, config)) {
+    if (!IsValidEntry(gpu_memory, config)) {
         return NULL_IMAGE_VIEW_ID;
     }
     const auto [pair, is_new] = image_views.try_emplace(config);
@@ -960,14 +1015,20 @@ ImageId TextureCache<P>::FindOrInsertImage(const ImageInfo& info, GPUVAddr gpu_a
 template <class P>
 ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
                                    RelaxedOptions options) {
-    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
+    std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
+    if (!cpu_addr) {
+        cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr, CalculateGuestSizeInBytes(info));
         if (!cpu_addr) {
             return ImageId{};
         }
+    }
     const bool broken_views = runtime.HasBrokenTextureViewFormats();
     const bool native_bgr = runtime.HasNativeBgr();
     ImageId image_id;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
+        if (True(existing_image.flags & ImageFlagBits::Remapped)) {
+            return false;
+        }
         if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear) {
             const bool strict_size = False(options & RelaxedOptions::Size) &&
                                      True(existing_image.flags & ImageFlagBits::Strong);
@@ -993,7 +1054,16 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
 template <class P>
 ImageId TextureCache<P>::InsertImage(const ImageInfo& info, GPUVAddr gpu_addr,
                                      RelaxedOptions options) {
-    const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
+    std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
+    if (!cpu_addr) {
+        const auto size = CalculateGuestSizeInBytes(info);
+        cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr, size);
+        if (!cpu_addr) {
+            const VAddr fake_addr = ~(1ULL << 40ULL) + virtual_invalid_space;
+            virtual_invalid_space += Common::AlignUp(size, 32);
+            cpu_addr = std::optional<VAddr>(fake_addr);
+        }
+    }
     ASSERT_MSG(cpu_addr, "Tried to insert an image to an invalid gpu_addr=0x{:x}", gpu_addr);
     const ImageId image_id = JoinImages(info, gpu_addr, *cpu_addr);
     const Image& image = slot_images[image_id];
@@ -1013,10 +1083,16 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
     const bool broken_views = runtime.HasBrokenTextureViewFormats();
     const bool native_bgr = runtime.HasNativeBgr();
     std::vector<ImageId> overlap_ids;
+    std::unordered_set<ImageId> overlaps_found;
     std::vector<ImageId> left_aliased_ids;
     std::vector<ImageId> right_aliased_ids;
+    std::unordered_set<ImageId> ignore_textures;
     std::vector<ImageId> bad_overlap_ids;
-    ForEachImageInRegion(cpu_addr, size_bytes, [&](ImageId overlap_id, ImageBase& overlap) {
+    const auto region_check = [&](ImageId overlap_id, ImageBase& overlap) {
+        if (True(overlap.flags & ImageFlagBits::Remapped)) {
+            ignore_textures.insert(overlap_id);
+            return;
+        }
         if (info.type == ImageType::Linear) {
             if (info.pitch == overlap.info.pitch && gpu_addr == overlap.gpu_addr) {
                 // Alias linear images with the same pitch
@@ -1024,6 +1100,7 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
             }
             return;
         }
+        overlaps_found.insert(overlap_id);
         static constexpr bool strict_size = true;
         const std::optional<OverlapResult> solution = ResolveOverlap(
             new_info, gpu_addr, cpu_addr, overlap, strict_size, broken_views, native_bgr);
@@ -1047,12 +1124,40 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
             bad_overlap_ids.push_back(overlap_id);
             overlap.flags |= ImageFlagBits::BadOverlap;
         }
-    });
+    };
+    ForEachImageInRegion(cpu_addr, size_bytes, region_check);
+    const auto region_check_gpu = [&](ImageId overlap_id, ImageBase& overlap) {
+        if (!overlaps_found.contains(overlap_id)) {
+            if (True(overlap.flags & ImageFlagBits::Remapped)) {
+                ignore_textures.insert(overlap_id);
+            }
+            if (overlap.gpu_addr == gpu_addr && overlap.guest_size_bytes == size_bytes) {
+                ignore_textures.insert(overlap_id);
+            }
+        }
+    };
+    ForEachSparseImageInRegion(gpu_addr, size_bytes, region_check_gpu);
     const ImageId new_image_id = slot_images.insert(runtime, new_info, gpu_addr, cpu_addr);
     Image& new_image = slot_images[new_image_id];
 
+    if (!gpu_memory.IsContinousRange(new_image.gpu_addr, new_image.guest_size_bytes)) {
+        new_image.flags |= ImageFlagBits::Sparse;
+    }
+
+    for (const ImageId overlap_id : ignore_textures) {
+        Image& overlap = slot_images[overlap_id];
+        if (True(overlap.flags & ImageFlagBits::GpuModified)) {
+            UNIMPLEMENTED();
+        }
+        if (True(overlap.flags & ImageFlagBits::Tracked)) {
+            UntrackImage(overlap, overlap_id);
+        }
+        UnregisterImage(overlap_id);
+        DeleteImage(overlap_id);
+    }
+
     // TODO: Only upload what we need
-    RefreshContents(new_image);
+    RefreshContents(new_image, new_image_id);
 
     for (const ImageId overlap_id : overlap_ids) {
         Image& overlap = slot_images[overlap_id];
@@ -1064,7 +1169,7 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
             runtime.CopyImage(new_image, overlap, copies);
         }
         if (True(overlap.flags & ImageFlagBits::Tracked)) {
-            UntrackImage(overlap);
+            UntrackImage(overlap, overlap_id);
         }
         UnregisterImage(overlap_id);
         DeleteImage(overlap_id);
@@ -1092,11 +1197,10 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
 
 template <class P>
 typename TextureCache<P>::BlitImages TextureCache<P>::GetBlitImages(
-    const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src,
+    const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src) {
-    s32 src_address_offset) {
     static constexpr auto FIND_OPTIONS = RelaxedOptions::Format | RelaxedOptions::Samples;
     const GPUVAddr dst_addr = dst.Address();
-    const GPUVAddr src_addr = src.Address() + src_address_offset;
+    const GPUVAddr src_addr = src.Address();
     ImageInfo dst_info(dst);
     ImageInfo src_info(src);
     ImageId dst_id;
@@ -1200,7 +1304,8 @@ void TextureCache<P>::ForEachImageInRegion(VAddr cpu_addr, size_t size, Func&& f
     using FuncReturn = typename std::invoke_result<Func, ImageId, Image&>::type;
     static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
     boost::container::small_vector<ImageId, 32> images;
-    ForEachPage(cpu_addr, size, [this, &images, cpu_addr, size, func](u64 page) {
+    boost::container::small_vector<ImageMapId, 32> maps;
+    ForEachCPUPage(cpu_addr, size, [this, &images, &maps, cpu_addr, size, func](u64 page) {
         const auto it = page_table.find(page);
         if (it == page_table.end()) {
             if constexpr (BOOL_BREAK) {
@@ -1209,12 +1314,63 @@ void TextureCache<P>::ForEachImageInRegion(VAddr cpu_addr, size_t size, Func&& f
                 return;
             }
         }
+        for (const ImageMapId map_id : it->second) {
+            ImageMapView& map = slot_map_views[map_id];
+            if (map.picked) {
+                continue;
+            }
+            if (!map.Overlaps(cpu_addr, size)) {
+                continue;
+            }
+            map.picked = true;
+            maps.push_back(map_id);
+            Image& image = slot_images[map.image_id];
+            if (True(image.flags & ImageFlagBits::Picked)) {
+                continue;
+            }
+            image.flags |= ImageFlagBits::Picked;
+            images.push_back(map.image_id);
+            if constexpr (BOOL_BREAK) {
+                if (func(map.image_id, image)) {
+                    return true;
+                }
+            } else {
+                func(map.image_id, image);
+            }
+        }
+        if constexpr (BOOL_BREAK) {
+            return false;
+        }
+    });
+    for (const ImageId image_id : images) {
+        slot_images[image_id].flags &= ~ImageFlagBits::Picked;
+    }
+    for (const ImageMapId map_id : maps) {
+        slot_map_views[map_id].picked = false;
+    }
+}
+
+template <class P>
+template <typename Func>
+void TextureCache<P>::ForEachImageInRegionGPU(GPUVAddr gpu_addr, size_t size, Func&& func) {
+    using FuncReturn = typename std::invoke_result<Func, ImageId, Image&>::type;
+    static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
+    boost::container::small_vector<ImageId, 8> images;
+    ForEachGPUPage(gpu_addr, size, [this, &images, gpu_addr, size, func](u64 page) {
+        const auto it = gpu_page_table.find(page);
+        if (it == gpu_page_table.end()) {
+            if constexpr (BOOL_BREAK) {
+                return false;
+            } else {
+                return;
+            }
+        }
         for (const ImageId image_id : it->second) {
             Image& image = slot_images[image_id];
             if (True(image.flags & ImageFlagBits::Picked)) {
                 continue;
             }
-            if (!image.Overlaps(cpu_addr, size)) {
+            if (!image.OverlapsGPU(gpu_addr, size)) {
                 continue;
             }
             image.flags |= ImageFlagBits::Picked;
@@ -1236,6 +1392,69 @@ void TextureCache<P>::ForEachImageInRegion(VAddr cpu_addr, size_t size, Func&& f
     }
 }
 
+template <class P>
+template <typename Func>
+void TextureCache<P>::ForEachSparseImageInRegion(GPUVAddr gpu_addr, size_t size, Func&& func) {
+    using FuncReturn = typename std::invoke_result<Func, ImageId, Image&>::type;
+    static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
+    boost::container::small_vector<ImageId, 8> images;
+    ForEachGPUPage(gpu_addr, size, [this, &images, gpu_addr, size, func](u64 page) {
+        const auto it = sparse_page_table.find(page);
+        if (it == sparse_page_table.end()) {
+            if constexpr (BOOL_BREAK) {
+                return false;
+            } else {
+                return;
+            }
+        }
+        for (const ImageId image_id : it->second) {
+            Image& image = slot_images[image_id];
+            if (True(image.flags & ImageFlagBits::Picked)) {
+                continue;
+            }
+            if (!image.OverlapsGPU(gpu_addr, size)) {
+                continue;
+            }
+            image.flags |= ImageFlagBits::Picked;
+            images.push_back(image_id);
+            if constexpr (BOOL_BREAK) {
+                if (func(image_id, image)) {
+                    return true;
+                }
+            } else {
+                func(image_id, image);
+            }
+        }
+        if constexpr (BOOL_BREAK) {
+            return false;
+        }
+    });
+    for (const ImageId image_id : images) {
+        slot_images[image_id].flags &= ~ImageFlagBits::Picked;
+    }
+}
+
+template <class P>
+template <typename Func>
+void TextureCache<P>::ForEachSparseSegment(ImageBase& image, Func&& func) {
+    using FuncReturn = typename std::invoke_result<Func, GPUVAddr, VAddr, size_t>::type;
+    static constexpr bool RETURNS_BOOL = std::is_same_v<FuncReturn, bool>;
+    const auto segments = gpu_memory.GetSubmappedRange(image.gpu_addr, image.guest_size_bytes);
+    for (auto& segment : segments) {
+        const auto gpu_addr = segment.first;
+        const auto size = segment.second;
+        std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr);
+        ASSERT(cpu_addr);
+        if constexpr (RETURNS_BOOL) {
+            if (func(gpu_addr, *cpu_addr, size)) {
+                break;
+            }
+        } else {
+            func(gpu_addr, *cpu_addr, size);
+        }
+    }
+}
+
 template <class P>
 ImageViewId TextureCache<P>::FindOrEmplaceImageView(ImageId image_id, const ImageViewInfo& info) {
     Image& image = slot_images[image_id];
@@ -1253,8 +1472,6 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
     ASSERT_MSG(False(image.flags & ImageFlagBits::Registered),
                "Trying to register an already registered image");
     image.flags |= ImageFlagBits::Registered;
-    ForEachPage(image.cpu_addr, image.guest_size_bytes,
-                [this, image_id](u64 page) { page_table[page].push_back(image_id); });
     u64 tentative_size = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
     if ((IsPixelFormatASTC(image.info.format) &&
          True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
@@ -1262,6 +1479,27 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
         tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
     }
     total_used_memory += Common::AlignUp(tentative_size, 1024);
+    ForEachGPUPage(image.gpu_addr, image.guest_size_bytes,
+                   [this, image_id](u64 page) { gpu_page_table[page].push_back(image_id); });
+    if (False(image.flags & ImageFlagBits::Sparse)) {
+        auto map_id =
+            slot_map_views.insert(image.gpu_addr, image.cpu_addr, image.guest_size_bytes, image_id);
+        ForEachCPUPage(image.cpu_addr, image.guest_size_bytes,
+                       [this, map_id](u64 page) { page_table[page].push_back(map_id); });
+        image.map_view_id = map_id;
+        return;
+    }
+    std::vector<ImageViewId> sparse_maps{};
+    ForEachSparseSegment(
+        image, [this, image_id, &sparse_maps](GPUVAddr gpu_addr, VAddr cpu_addr, size_t size) {
+            auto map_id = slot_map_views.insert(gpu_addr, cpu_addr, size, image_id);
+            ForEachCPUPage(cpu_addr, size,
+                           [this, map_id](u64 page) { page_table[page].push_back(map_id); });
+            sparse_maps.push_back(map_id);
+        });
+    sparse_views.emplace(image_id, std::move(sparse_maps));
+    ForEachGPUPage(image.gpu_addr, image.guest_size_bytes,
+                   [this, image_id](u64 page) { sparse_page_table[page].push_back(image_id); });
 }
 
 template <class P>
@@ -1278,34 +1516,125 @@ void TextureCache<P>::UnregisterImage(ImageId image_id) {
         tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
     }
     total_used_memory -= Common::AlignUp(tentative_size, 1024);
-    ForEachPage(image.cpu_addr, image.guest_size_bytes, [this, image_id](u64 page) {
+    const auto& clear_page_table =
-        const auto page_it = page_table.find(page);
+        [this, image_id](
-        if (page_it == page_table.end()) {
+            u64 page,
+            std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>>& selected_page_table) {
+            const auto page_it = selected_page_table.find(page);
+            if (page_it == selected_page_table.end()) {
                 UNREACHABLE_MSG("Unregistering unregistered page=0x{:x}", page << PAGE_BITS);
                 return;
             }
             std::vector<ImageId>& image_ids = page_it->second;
             const auto vector_it = std::ranges::find(image_ids, image_id);
             if (vector_it == image_ids.end()) {
-            UNREACHABLE_MSG("Unregistering unregistered image in page=0x{:x}", page << PAGE_BITS);
+                UNREACHABLE_MSG("Unregistering unregistered image in page=0x{:x}",
+                                page << PAGE_BITS);
                 return;
             }
             image_ids.erase(vector_it);
+        };
+    ForEachGPUPage(image.gpu_addr, image.guest_size_bytes,
+                   [this, &clear_page_table](u64 page) { clear_page_table(page, gpu_page_table); });
+    if (False(image.flags & ImageFlagBits::Sparse)) {
+        const auto map_id = image.map_view_id;
+        ForEachCPUPage(image.cpu_addr, image.guest_size_bytes, [this, map_id](u64 page) {
+            const auto page_it = page_table.find(page);
+            if (page_it == page_table.end()) {
+                UNREACHABLE_MSG("Unregistering unregistered page=0x{:x}", page << PAGE_BITS);
+                return;
+            }
+            std::vector<ImageMapId>& image_map_ids = page_it->second;
+            const auto vector_it = std::ranges::find(image_map_ids, map_id);
+            if (vector_it == image_map_ids.end()) {
+                UNREACHABLE_MSG("Unregistering unregistered image in page=0x{:x}",
+                                page << PAGE_BITS);
+                return;
+            }
+            image_map_ids.erase(vector_it);
+        });
+        slot_map_views.erase(map_id);
+        return;
+    }
+    ForEachGPUPage(image.gpu_addr, image.guest_size_bytes, [this, &clear_page_table](u64 page) {
+        clear_page_table(page, sparse_page_table);
+    });
+    auto it = sparse_views.find(image_id);
+    ASSERT(it != sparse_views.end());
+    auto& sparse_maps = it->second;
+    for (auto& map_view_id : sparse_maps) {
+        const auto& map_range = slot_map_views[map_view_id];
+        const VAddr cpu_addr = map_range.cpu_addr;
+        const std::size_t size = map_range.size;
+        ForEachCPUPage(cpu_addr, size, [this, image_id](u64 page) {
+            const auto page_it = page_table.find(page);
+            if (page_it == page_table.end()) {
+                UNREACHABLE_MSG("Unregistering unregistered page=0x{:x}", page << PAGE_BITS);
+                return;
+            }
+            std::vector<ImageMapId>& image_map_ids = page_it->second;
+            auto vector_it = image_map_ids.begin();
+            while (vector_it != image_map_ids.end()) {
+                ImageMapView& map = slot_map_views[*vector_it];
+                if (map.image_id != image_id) {
+                    vector_it++;
+                    continue;
+                }
+                if (!map.picked) {
+                    map.picked = true;
+                }
+                vector_it = image_map_ids.erase(vector_it);
+            }
+        });
+        slot_map_views.erase(map_view_id);
+    }
+    sparse_views.erase(it);
+}
+
+template <class P>
+void TextureCache<P>::TrackImage(ImageBase& image, ImageId image_id) {
+    ASSERT(False(image.flags & ImageFlagBits::Tracked));
+    image.flags |= ImageFlagBits::Tracked;
+    if (False(image.flags & ImageFlagBits::Sparse)) {
+        rasterizer.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, 1);
+        return;
+    }
+    if (True(image.flags & ImageFlagBits::Registered)) {
+        auto it = sparse_views.find(image_id);
+        ASSERT(it != sparse_views.end());
+        auto& sparse_maps = it->second;
+        for (auto& map_view_id : sparse_maps) {
+            const auto& map = slot_map_views[map_view_id];
+            const VAddr cpu_addr = map.cpu_addr;
+            const std::size_t size = map.size;
+            rasterizer.UpdatePagesCachedCount(cpu_addr, size, 1);
+        }
+        return;
+    }
+    ForEachSparseSegment(image,
+                         [this]([[maybe_unused]] GPUVAddr gpu_addr, VAddr cpu_addr, size_t size) {
+                             rasterizer.UpdatePagesCachedCount(cpu_addr, size, 1);
                          });
 }
 
 template <class P>
-void TextureCache<P>::TrackImage(ImageBase& image) {
+void TextureCache<P>::UntrackImage(ImageBase& image, ImageId image_id) {
-    ASSERT(False(image.flags & ImageFlagBits::Tracked));
-    image.flags |= ImageFlagBits::Tracked;
-    rasterizer.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, 1);
-}
-
-template <class P>
-void TextureCache<P>::UntrackImage(ImageBase& image) {
     ASSERT(True(image.flags & ImageFlagBits::Tracked));
     image.flags &= ~ImageFlagBits::Tracked;
+    if (False(image.flags & ImageFlagBits::Sparse)) {
         rasterizer.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, -1);
+        return;
+    }
+    ASSERT(True(image.flags & ImageFlagBits::Registered));
+    auto it = sparse_views.find(image_id);
+    ASSERT(it != sparse_views.end());
+    auto& sparse_maps = it->second;
+    for (auto& map_view_id : sparse_maps) {
+        const auto& map = slot_map_views[map_view_id];
+        const VAddr cpu_addr = map.cpu_addr;
+        const std::size_t size = map.size;
+        rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1);
+    }
 }
 
 template <class P>
@@ -1447,10 +1776,10 @@ void TextureCache<P>::PrepareImage(ImageId image_id, bool is_modification, bool
     if (invalidate) {
         image.flags &= ~(ImageFlagBits::CpuModified | ImageFlagBits::GpuModified);
         if (False(image.flags & ImageFlagBits::Tracked)) {
-            TrackImage(image);
+            TrackImage(image, image_id);
         }
     } else {
-        RefreshContents(image);
+        RefreshContents(image, image_id);
         SynchronizeAliases(image_id);
     }
     if (is_modification) {

src/video_core/texture_cache/types.h

@@ -16,6 +16,7 @@ constexpr size_t MAX_MIP_LEVELS = 14;
 constexpr SlotId CORRUPT_ID{0xfffffffe};
 
 using ImageId = SlotId;
+using ImageMapId = SlotId;
 using ImageViewId = SlotId;
 using ImageAllocId = SlotId;
 using SamplerId = SlotId;
@@ -132,8 +133,8 @@ struct BufferImageCopy {
 };
 
 struct BufferCopy {
-    size_t src_offset;
+    u64 src_offset;
-    size_t dst_offset;
+    u64 dst_offset;
     size_t size;
 };
 

src/video_core/texture_cache/util.cpp

@@ -664,6 +664,16 @@ LevelArray CalculateMipLevelOffsets(const ImageInfo& info) noexcept {
     return offsets;
 }
 
+LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept {
+    const u32 num_levels = info.resources.levels;
+    const LevelInfo level_info = MakeLevelInfo(info);
+    LevelArray sizes{};
+    for (u32 level = 0; level < num_levels; ++level) {
+        sizes[level] = CalculateLevelSize(level_info, level);
+    }
+    return sizes;
+}
+
 std::vector<u32> CalculateSliceOffsets(const ImageInfo& info) {
     ASSERT(info.type == ImageType::e3D);
     std::vector<u32> offsets;
@@ -776,14 +786,20 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
     return copies;
 }
 
-bool IsValidAddress(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) {
+bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) {
-    if (config.Address() == 0) {
+    const GPUVAddr address = config.Address();
+    if (address == 0) {
         return false;
     }
-    if (config.Address() > (u64(1) << 48)) {
+    if (address > (1ULL << 48)) {
         return false;
     }
-    return gpu_memory.GpuToCpuAddress(config.Address()).has_value();
+    if (gpu_memory.GpuToCpuAddress(address).has_value()) {
+        return true;
+    }
+    const ImageInfo info{config};
+    const size_t guest_size_bytes = CalculateGuestSizeInBytes(info);
+    return gpu_memory.GpuToCpuAddress(address, guest_size_bytes).has_value();
 }
 
 std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,

src/video_core/texture_cache/util.h

@@ -40,6 +40,8 @@ struct OverlapResult {
 
 [[nodiscard]] LevelArray CalculateMipLevelOffsets(const ImageInfo& info) noexcept;
 
+[[nodiscard]] LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept;
+
 [[nodiscard]] std::vector<u32> CalculateSliceOffsets(const ImageInfo& info);
 
 [[nodiscard]] std::vector<SubresourceBase> CalculateSliceSubresources(const ImageInfo& info);
@@ -55,7 +57,7 @@ struct OverlapResult {
                                                            const ImageInfo& src,
                                                            SubresourceBase base);
 
-[[nodiscard]] bool IsValidAddress(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
+[[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
 
 [[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
                                                           GPUVAddr gpu_addr, const ImageInfo& info,