early-access version 3534

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3531.
+This is the source code for early-access 3534.
 
 ## Legal Notice
 

src/tests/video_core/memory_tracker.cpp

@@ -427,7 +427,7 @@ TEST_CASE("MemoryTracker: Single page in large region", "[video_core]") {
 
     memory_track->MarkRegionAsCpuModified(c + WORD * 12 + PAGE * 8, PAGE);
     REQUIRE(memory_track->IsRegionCpuModified(c, WORD * 16));
-    REQUIRE(memory_track->IsRegionCpuModified(c + WORD * 10, WORD * 2));
+    REQUIRE(!memory_track->IsRegionCpuModified(c + WORD * 10, WORD * 2));
     REQUIRE(memory_track->IsRegionCpuModified(c + WORD * 11, WORD * 2));
     REQUIRE(memory_track->IsRegionCpuModified(c + WORD * 12, WORD * 2));
     REQUIRE(memory_track->IsRegionCpuModified(c + WORD * 12 + PAGE * 4, PAGE * 8));
@@ -535,6 +535,8 @@ TEST_CASE("MemoryTracker: Cached write downloads") {
     memory_track->MarkRegionAsGpuModified(c + PAGE, PAGE);
     int num = 0;
     memory_track->ForEachDownloadRangeAndClear(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    REQUIRE(num == 1);
+    num = 0;
     memory_track->ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
     REQUIRE(num == 0);
     REQUIRE(!memory_track->IsRegionCpuModified(c + PAGE, PAGE));

src/video_core/buffer_cache/buffer_base.h

@@ -38,10 +38,8 @@ public:
     static constexpr u64 BASE_PAGE_BITS = 16;
     static constexpr u64 BASE_PAGE_SIZE = 1ULL << BASE_PAGE_BITS;
 
-    explicit BufferBase(RasterizerInterface& rasterizer_, VAddr cpu_addr_, u64 size_bytes)
+    explicit BufferBase(RasterizerInterface& rasterizer_, VAddr cpu_addr_, u64 size_bytes_)
-        : cpu_addr{Common::AlignDown(cpu_addr_, BASE_PAGE_SIZE)},
+        : cpu_addr{cpu_addr_}, size_bytes{size_bytes_} {}
-          word_manager(cpu_addr, rasterizer_,
-                       Common::AlignUp(size_bytes + (cpu_addr_ - cpu_addr), BASE_PAGE_SIZE)) {}
 
     explicit BufferBase(NullBufferParams) {}
 
@@ -51,88 +49,6 @@ public:
     BufferBase& operator=(BufferBase&&) = default;
     BufferBase(BufferBase&&) = default;
 
-    /// Returns the inclusive CPU modified range in a begin end pair
-    [[nodiscard]] std::pair<u64, u64> ModifiedCpuRegion(VAddr query_cpu_addr,
-                                                        u64 query_size) const noexcept {
-        const u64 offset = query_cpu_addr - cpu_addr;
-        return word_manager.template ModifiedRegion<Type::CPU>(offset, query_size);
-    }
-
-    /// Returns the inclusive GPU modified range in a begin end pair
-    [[nodiscard]] std::pair<u64, u64> ModifiedGpuRegion(VAddr query_cpu_addr,
-                                                        u64 query_size) const noexcept {
-        const u64 offset = query_cpu_addr - cpu_addr;
-        return word_manager.template ModifiedRegion<Type::GPU>(offset, query_size);
-    }
-
-    /// Returns true if a region has been modified from the CPU
-    [[nodiscard]] bool IsRegionCpuModified(VAddr query_cpu_addr, u64 query_size) const noexcept {
-        const u64 offset = query_cpu_addr - cpu_addr;
-        return word_manager.template IsRegionModified<Type::CPU>(offset, query_size);
-    }
-
-    /// Returns true if a region has been modified from the GPU
-    [[nodiscard]] bool IsRegionGpuModified(VAddr query_cpu_addr, u64 query_size) const noexcept {
-        const u64 offset = query_cpu_addr - cpu_addr;
-        return word_manager.template IsRegionModified<Type::GPU>(offset, query_size);
-    }
-
-    /// Mark region as CPU modified, notifying the rasterizer about this change
-    void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 size) {
-        word_manager.template ChangeRegionState<Type::CPU, true>(dirty_cpu_addr, size);
-    }
-
-    /// Unmark region as CPU modified, notifying the rasterizer about this change
-    void UnmarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 size) {
-        word_manager.template ChangeRegionState<Type::CPU, false>(dirty_cpu_addr, size);
-    }
-
-    /// Mark region as modified from the host GPU
-    void MarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 size) noexcept {
-        word_manager.template ChangeRegionState<Type::GPU, true>(dirty_cpu_addr, size);
-    }
-
-    /// Unmark region as modified from the host GPU
-    void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 size) noexcept {
-        word_manager.template ChangeRegionState<Type::GPU, false>(dirty_cpu_addr, size);
-    }
-
-    /// Mark region as modified from the CPU
-    /// but don't mark it as modified until FlusHCachedWrites is called.
-    void CachedCpuWrite(VAddr dirty_cpu_addr, u64 size) {
-        flags |= BufferFlagBits::CachedWrites;
-        word_manager.template ChangeRegionState<Type::CachedCPU, true>(dirty_cpu_addr, size);
-    }
-
-    /// Flushes cached CPU writes, and notify the rasterizer about the deltas
-    void FlushCachedWrites() noexcept {
-        flags &= ~BufferFlagBits::CachedWrites;
-        word_manager.FlushCachedWrites();
-    }
-
-    /// 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) {
-        word_manager.template 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, bool clear, Func&& func) {
-        word_manager.template ForEachModifiedRange<Type::GPU>(query_cpu_range, size, clear, func);
-    }
-
-    template <typename Func>
-    void ForEachDownloadRangeAndClear(VAddr query_cpu_range, u64 size, Func&& func) {
-        word_manager.template ForEachModifiedRange<Type::GPU>(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(Func&& func) {
-        word_manager.template ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), true, func);
-    }
-
     /// Mark buffer as picked
     void Pick() noexcept {
         flags |= BufferFlagBits::Picked;
@@ -179,11 +95,6 @@ public:
         return static_cast<u32>(other_cpu_addr - cpu_addr);
     }
 
-    /// Returns the size in bytes of the buffer
-    [[nodiscard]] u64 SizeBytes() const noexcept {
-        return word_manager.SizeBytes();
-    }
-
     size_t getLRUID() const noexcept {
         return lru_id;
     }
@@ -192,12 +103,16 @@ public:
         lru_id = lru_id_;
     }
 
+    size_t SizeBytes() const {
+        return size_bytes;
+    }
+
 private:
     VAddr cpu_addr = 0;
-    WordManager<RasterizerInterface> word_manager;
     BufferFlagBits flags{};
     int stream_score = 0;
     size_t lru_id = SIZE_MAX;
+    size_t size_bytes = 0;
 };
 
 } // namespace VideoCommon

src/video_core/buffer_cache/buffer_cache.h

@@ -21,6 +21,7 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
     // Ensure the first slot is used for the null buffer
     void(slot_buffers.insert(runtime, NullBufferParams{}));
     common_ranges.clear();
+    inline_buffer_id = NULL_BUFFER_ID;
 
     active_async_buffers = !Settings::IsGPULevelHigh();
 
@@ -442,9 +443,6 @@ template <class P>
 void BufferCache<P>::FlushCachedWrites() {
     cached_write_buffer_ids.clear();
     memory_tracker.FlushCachedWrites();
-    for (auto& interval : cached_ranges) {
-        ClearDownload(interval);
-    }
     cached_ranges.clear();
 }
 
@@ -689,7 +687,7 @@ void BufferCache<P>::BindHostIndexBuffer() {
     const u32 offset = buffer.Offset(index_buffer.cpu_addr);
     const u32 size = index_buffer.size;
     const auto& draw_state = maxwell3d->draw_manager->GetDrawState();
-    if (!draw_state.inline_index_draw_indexes.empty()) {
+    if (!draw_state.inline_index_draw_indexes.empty()) [[unlikely]] {
         if constexpr (USE_MEMORY_MAPS) {
             auto upload_staging = runtime.UploadStagingBuffer(size);
             std::array<BufferCopy, 1> copies{
@@ -1001,12 +999,20 @@ void BufferCache<P>::UpdateIndexBuffer() {
         return;
     }
     flags[Dirty::IndexBuffer] = false;
-    if (!draw_state.inline_index_draw_indexes.empty()) {
+    if (!draw_state.inline_index_draw_indexes.empty()) [[unlikely]] {
         auto inline_index_size = static_cast<u32>(draw_state.inline_index_draw_indexes.size());
+        u32 buffer_size = Common::AlignUp(inline_index_size, PAGE_SIZE);
+        if (inline_buffer_id == NULL_BUFFER_ID) [[unlikely]] {
+            inline_buffer_id = CreateBuffer(0, buffer_size);
+        }
+        if (slot_buffers[inline_buffer_id].SizeBytes() < buffer_size) [[unlikely]] {
+            slot_buffers.erase(inline_buffer_id);
+            inline_buffer_id = CreateBuffer(0, buffer_size);
+        }
         index_buffer = Binding{
             .cpu_addr = 0,
             .size = inline_index_size,
-            .buffer_id = FindBuffer(0, inline_index_size),
+            .buffer_id = inline_buffer_id,
         };
         return;
     }
@@ -1317,6 +1323,9 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
 
 template <class P>
 BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
+    VAddr cpu_addr_end = Common::AlignUp(cpu_addr + wanted_size, PAGE_SIZE);
+    cpu_addr = Common::AlignDown(cpu_addr, PAGE_SIZE);
+    wanted_size = static_cast<u32>(cpu_addr_end - cpu_addr);
     const OverlapResult overlap = ResolveOverlaps(cpu_addr, wanted_size);
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
     const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);

src/video_core/buffer_cache/buffer_cache_base.h

@@ -112,6 +112,10 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelI
     static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB;
     static constexpr s64 TARGET_THRESHOLD = 4_GiB;
 
+    // Debug Flags.
+
+    static constexpr bool DISABLE_DOWNLOADS = true;
+
     using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
     using Runtime = typename P::Runtime;
@@ -568,6 +572,7 @@ private:
     u64 total_used_memory = 0;
     u64 minimum_memory = 0;
     u64 critical_memory = 0;
+    BufferId inline_buffer_id;
 
     bool active_async_buffers = false;
 

src/video_core/buffer_cache/memory_tracker_base.h

@@ -132,8 +132,8 @@ public:
     void ForEachUploadRange(VAddr query_cpu_range, u64 query_size, Func&& func) {
         IteratePages<true>(query_cpu_range, query_size,
                            [&func](Manager* manager, u64 offset, size_t size) {
-                               manager->template ForEachModifiedRange<Type::CPU>(
+                               manager->template ForEachModifiedRange<Type::CPU, true>(
-                                   manager->GetCpuAddr() + offset, size, true, func);
+                                   manager->GetCpuAddr() + offset, size, func);
                            });
     }
 
@@ -142,8 +142,13 @@ public:
     void ForEachDownloadRange(VAddr query_cpu_range, u64 query_size, bool clear, Func&& func) {
         IteratePages<false>(query_cpu_range, query_size,
                             [&func, clear](Manager* manager, u64 offset, size_t size) {
-                                manager->template ForEachModifiedRange<Type::GPU>(
+                                if (clear) {
-                                    manager->GetCpuAddr() + offset, size, clear, func);
+                                    manager->template ForEachModifiedRange<Type::GPU, true>(
+                                        manager->GetCpuAddr() + offset, size, func);
+                                } else {
+                                    manager->template ForEachModifiedRange<Type::GPU, false>(
+                                        manager->GetCpuAddr() + offset, size, func);
+                                }
                             });
     }
 
@@ -151,8 +156,8 @@ public:
     void ForEachDownloadRangeAndClear(VAddr query_cpu_range, u64 query_size, Func&& func) {
         IteratePages<false>(query_cpu_range, query_size,
                             [&func](Manager* manager, u64 offset, size_t size) {
-                                manager->template ForEachModifiedRange<Type::GPU>(
+                                manager->template ForEachModifiedRange<Type::GPU, true>(
-                                    manager->GetCpuAddr() + offset, size, true, func);
+                                    manager->GetCpuAddr() + offset, size, func);
                             });
     }
 

src/video_core/buffer_cache/word_manager.h

@@ -6,6 +6,7 @@
 #include <algorithm>
 #include <bit>
 #include <limits>
+#include <span>
 #include <utility>
 
 #include "common/alignment.h"
@@ -20,9 +21,16 @@ constexpr u64 PAGES_PER_WORD = 64;
 constexpr u64 BYTES_PER_PAGE = Core::Memory::YUZU_PAGESIZE;
 constexpr u64 BYTES_PER_WORD = PAGES_PER_WORD * BYTES_PER_PAGE;
 
+enum class Type {
+    CPU,
+    GPU,
+    CachedCPU,
+    Untracked,
+};
+
 /// Vector tracking modified pages tightly packed with small vector optimization
 template <size_t stack_words = 1>
-union WordsArray {
+struct WordsArray {
     /// Returns the pointer to the words state
     [[nodiscard]] const u64* Pointer(bool is_short) const noexcept {
         return is_short ? stack.data() : heap;
@@ -41,13 +49,13 @@ template <size_t stack_words = 1>
 struct Words {
     explicit Words() = default;
     explicit Words(u64 size_bytes_) : size_bytes{size_bytes_} {
+        num_words = Common::DivCeil(size_bytes, BYTES_PER_WORD);
         if (IsShort()) {
             cpu.stack.fill(~u64{0});
             gpu.stack.fill(0);
             cached_cpu.stack.fill(0);
             untracked.stack.fill(~u64{0});
         } else {
-            const size_t num_words = NumWords();
             // Share allocation between CPU and GPU pages and set their default values
             u64* const alloc = new u64[num_words * 4];
             cpu.heap = alloc;
@@ -75,6 +83,7 @@ struct Words {
     Words& operator=(Words&& rhs) noexcept {
         Release();
         size_bytes = rhs.size_bytes;
+        num_words = rhs.num_words;
         cpu = rhs.cpu;
         gpu = rhs.gpu;
         cached_cpu = rhs.cached_cpu;
@@ -84,7 +93,7 @@ struct Words {
     }
 
     Words(Words&& rhs) noexcept
-        : size_bytes{rhs.size_bytes}, cpu{rhs.cpu}, gpu{rhs.gpu},
+        : size_bytes{rhs.size_bytes}, num_words{rhs.num_words}, cpu{rhs.cpu}, gpu{rhs.gpu},
           cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked} {
         rhs.cpu.heap = nullptr;
     }
@@ -94,12 +103,12 @@ struct Words {
 
     /// Returns true when the buffer fits in the small vector optimization
     [[nodiscard]] bool IsShort() const noexcept {
-        return size_bytes <= stack_words * BYTES_PER_WORD;
+        return num_words <= stack_words;
     }
 
     /// Returns the number of words of the buffer
     [[nodiscard]] size_t NumWords() const noexcept {
-        return Common::DivCeil(size_bytes, BYTES_PER_WORD);
+        return num_words;
     }
 
     /// Release buffer resources
@@ -110,20 +119,40 @@ struct Words {
         }
     }
 
+    template <Type type>
+    std::span<u64> Span() noexcept {
+        if constexpr (type == Type::CPU) {
+            return std::span<u64>(cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::GPU) {
+            return std::span<u64>(gpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::CachedCPU) {
+            return std::span<u64>(cached_cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Untracked) {
+            return std::span<u64>(untracked.Pointer(IsShort()), num_words);
+        }
+    }
+
+    template <Type type>
+    std::span<const u64> Span() const noexcept {
+        if constexpr (type == Type::CPU) {
+            return std::span<const u64>(cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::GPU) {
+            return std::span<const u64>(gpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::CachedCPU) {
+            return std::span<const u64>(cached_cpu.Pointer(IsShort()), num_words);
+        } else if constexpr (type == Type::Untracked) {
+            return std::span<const u64>(untracked.Pointer(IsShort()), num_words);
+        }
+    }
+
     u64 size_bytes = 0;
+    size_t num_words = 0;
     WordsArray<stack_words> cpu;
     WordsArray<stack_words> gpu;
     WordsArray<stack_words> cached_cpu;
     WordsArray<stack_words> untracked;
 };
 
-enum class Type {
-    CPU,
-    GPU,
-    CachedCPU,
-    Untracked,
-};
-
 template <class RasterizerInterface, size_t stack_words = 1>
 class WordManager {
 public:
@@ -140,6 +169,69 @@ public:
         return cpu_addr;
     }
 
+    static u64 ExtractBits(u64 word, size_t page_start, size_t page_end) {
+        constexpr size_t number_bits = sizeof(u64) * 8;
+        const size_t limit_page_end = number_bits - std::min(page_end, number_bits);
+        u64 bits = (word >> page_start) << page_start;
+        bits = (bits << limit_page_end) >> limit_page_end;
+        return bits;
+    }
+
+    static std::pair<size_t, size_t> GetWordPage(VAddr address) {
+        const size_t converted_address = static_cast<size_t>(address);
+        const size_t word_number = converted_address / BYTES_PER_WORD;
+        const size_t amount_pages = converted_address % BYTES_PER_WORD;
+        return std::make_pair(word_number, amount_pages / BYTES_PER_PAGE);
+    }
+
+    template <typename Func>
+    void IterateWords(size_t offset, size_t size, Func&& func) const {
+        using FuncReturn = std::invoke_result_t<Func, std::size_t, u64>;
+        static constexpr bool BOOL_BREAK = std::is_same_v<FuncReturn, bool>;
+        const size_t start = static_cast<size_t>(std::max<s64>(static_cast<s64>(offset), 0LL));
+        const size_t end = static_cast<size_t>(std::max<s64>(static_cast<s64>(offset + size), 0LL));
+        if (start >= SizeBytes() || end <= start) {
+            return;
+        }
+        auto [start_word, start_page] = GetWordPage(start);
+        auto [end_word, end_page] = GetWordPage(end + BYTES_PER_PAGE - 1ULL);
+        const size_t num_words = NumWords();
+        start_word = std::min(start_word, num_words);
+        end_word = std::min(end_word, num_words);
+        const size_t diff = end_word - start_word;
+        end_word += (end_page + PAGES_PER_WORD - 1ULL) / PAGES_PER_WORD;
+        end_word = std::min(end_word, num_words);
+        end_page += diff * PAGES_PER_WORD;
+        constexpr u64 base_mask{~0ULL};
+        for (size_t word_index = start_word; word_index < end_word; word_index++) {
+            const u64 mask = ExtractBits(base_mask, start_page, end_page);
+            start_page = 0;
+            end_page -= PAGES_PER_WORD;
+            if constexpr (BOOL_BREAK) {
+                if (func(word_index, mask)) {
+                    return;
+                }
+            } else {
+                func(word_index, mask);
+            }
+        }
+    }
+
+    template <typename Func>
+    void IteratePages(u64 mask, Func&& func) const {
+        size_t offset = 0;
+        while (mask != 0) {
+            const size_t empty_bits = std::countr_zero(mask);
+            offset += empty_bits;
+            mask = mask >> empty_bits;
+
+            const size_t continuous_bits = std::countr_one(mask);
+            func(offset, continuous_bits);
+            mask = continuous_bits < PAGES_PER_WORD ? (mask >> continuous_bits) : 0;
+            offset += continuous_bits;
+        }
+    }
+
     /**
      * Change the state of a range of pages
      *
@@ -147,47 +239,33 @@ public:
      * @param size          Size in bytes to mark or unmark as modified
      */
     template <Type type, bool enable>
-    void ChangeRegionState(u64 dirty_addr, s64 size) noexcept(type == Type::GPU) {
+    void ChangeRegionState(u64 dirty_addr, u64 size) noexcept(type == Type::GPU) {
-        const s64 difference = dirty_addr - cpu_addr;
+        std::span<u64> state_words = words.template Span<type>();
-        const u64 offset = std::max<s64>(difference, 0);
+        [[maybe_unused]] std::span<u64> untracked_words = words.template Span<Type::Untracked>();
-        size += std::min<s64>(difference, 0);
+        [[maybe_unused]] std::span<u64> cached_words = words.template Span<Type::CachedCPU>();
-        if (offset >= SizeBytes() || size < 0) {
+        IterateWords(dirty_addr - cpu_addr, size, [&](size_t index, u64 mask) {
-            return;
-        }
-        u64* const untracked_words = Array<Type::Untracked>();
-        u64* const state_words = Array<type>();
-        const u64 offset_end = std::min(offset + size, SizeBytes());
-        const u64 begin_page_index = offset / BYTES_PER_PAGE;
-        const u64 begin_word_index = begin_page_index / PAGES_PER_WORD;
-        const u64 end_page_index = Common::DivCeil(offset_end, BYTES_PER_PAGE);
-        const u64 end_word_index = Common::DivCeil(end_page_index, PAGES_PER_WORD);
-        u64 page_index = begin_page_index % PAGES_PER_WORD;
-        u64 word_index = begin_word_index;
-        while (word_index < end_word_index) {
-            const u64 next_word_first_page = (word_index + 1) * PAGES_PER_WORD;
-            const u64 left_offset =
-                std::min(next_word_first_page - end_page_index, PAGES_PER_WORD) % PAGES_PER_WORD;
-            const u64 right_offset = page_index;
-            u64 bits = ~u64{0};
-            bits = (bits >> right_offset) << right_offset;
-            bits = (bits << left_offset) >> left_offset;
             if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-                NotifyRasterizer<!enable>(word_index, untracked_words[word_index], bits);
+                NotifyRasterizer<!enable>(index, untracked_words[index], mask);
             }
             if constexpr (enable) {
-                state_words[word_index] |= bits;
+                state_words[index] |= mask;
                 if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-                    untracked_words[word_index] |= bits;
+                    untracked_words[index] |= mask;
+                }
+                if constexpr (type == Type::CPU) {
+                    cached_words[index] &= ~mask;
                 }
             } else {
-                state_words[word_index] &= ~bits;
+                if constexpr (type == Type::CPU) {
+                    const u64 word = state_words[index] & mask;
+                    cached_words[index] &= ~word;
+                }
+                state_words[index] &= ~mask;
                 if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-                    untracked_words[word_index] &= ~bits;
+                    untracked_words[index] &= ~mask;
                 }
             }
-            page_index = 0;
+        });
-            ++word_index;
-        }
     }
 
     /**
@@ -198,117 +276,57 @@ public:
      * @param size            Size in bytes of the CPU range to loop over
      * @param func            Function to call for each turned off region
      */
-    template <Type type, typename Func>
+    template <Type type, bool clear, typename Func>
-    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, bool clear, Func&& func) {
+    void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) {
         static_assert(type != Type::Untracked);
 
-        const s64 difference = query_cpu_range - cpu_addr;
+        std::span<u64> state_words = words.template Span<type>();
-        const u64 query_begin = std::max<s64>(difference, 0);
+        [[maybe_unused]] std::span<u64> untracked_words = words.template Span<Type::Untracked>();
-        size += std::min<s64>(difference, 0);
+        [[maybe_unused]] std::span<u64> cached_words = words.template Span<Type::CachedCPU>();
-        if (query_begin >= SizeBytes() || size < 0) {
+        const size_t offset = query_cpu_range - cpu_addr;
-            return;
+        bool pending = false;
-        }
+        size_t pending_offset{};
-        [[maybe_unused]] u64* const untracked_words = Array<Type::Untracked>();
+        size_t pending_pointer{};
-        [[maybe_unused]] u64* const cpu_words = Array<Type::CPU>();
+        const auto release = [&]() {
-        u64* const state_words = Array<type>();
+            func(cpu_addr + pending_offset * BYTES_PER_PAGE,
-        const u64 query_end = query_begin + std::min(static_cast<u64>(size), SizeBytes());
+                 (pending_pointer - pending_offset) * BYTES_PER_PAGE);
-        u64* const words_begin = state_words + query_begin / BYTES_PER_WORD;
+        };
-        u64* const words_end = state_words + Common::DivCeil(query_end, BYTES_PER_WORD);
+        IterateWords(offset, size, [&](size_t index, u64 mask) {
-        u64 first_page = (query_begin / BYTES_PER_PAGE) % PAGES_PER_WORD;
+            const u64 word = state_words[index] & mask;
-
+            if constexpr (clear) {
-        const auto modified = [](u64 word) { return word != 0; };
+                if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-        const auto first_modified_word = std::find_if(words_begin, words_end, modified);
+                    NotifyRasterizer<true>(index, untracked_words[index], mask);
-        if (first_modified_word == words_end) {
-            // Exit early when the buffer is not modified
-            return;
-        }
-        if (first_modified_word != words_begin) {
-            first_page = 0;
-        }
-        std::reverse_iterator<u64*> first_word_reverse(first_modified_word);
-        std::reverse_iterator<u64*> last_word_iterator(words_end);
-        auto last_word_result = std::find_if(last_word_iterator, first_word_reverse, modified);
-        u64* const last_modified_word = &(*last_word_result) + 1;
-
-        const u64 word_index_begin = std::distance(state_words, first_modified_word);
-        const u64 word_index_end = std::distance(state_words, last_modified_word);
-        const unsigned local_page_begin = std::countr_zero(*first_modified_word);
-        const unsigned local_page_end =
-            static_cast<unsigned>(PAGES_PER_WORD) - std::countl_zero(last_modified_word[-1]);
-        const u64 word_page_begin = word_index_begin * PAGES_PER_WORD;
-        const u64 word_page_end = (word_index_end - 1) * PAGES_PER_WORD;
-        const u64 query_page_begin = query_begin / BYTES_PER_PAGE;
-        const u64 query_page_end = Common::DivCeil(query_end, BYTES_PER_PAGE);
-        const u64 page_index_begin = std::max(word_page_begin + local_page_begin, query_page_begin);
-        const u64 page_index_end = std::min(word_page_end + local_page_end, query_page_end);
-        const u64 first_word_page_begin = page_index_begin % PAGES_PER_WORD;
-        const u64 last_word_page_end = (page_index_end - 1) % PAGES_PER_WORD + 1;
-
-        u64 page_begin = std::max(first_word_page_begin, first_page);
-        u64 current_base = 0;
-        u64 current_size = 0;
-        bool on_going = false;
-        for (u64 word_index = word_index_begin; word_index < word_index_end; ++word_index) {
-            const bool is_last_word = word_index + 1 == word_index_end;
-            const u64 page_end = is_last_word ? last_word_page_end : PAGES_PER_WORD;
-            const u64 right_offset = page_begin;
-            const u64 left_offset = PAGES_PER_WORD - page_end;
-            u64 bits = ~u64{0};
-            bits = (bits >> right_offset) << right_offset;
-            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::CachedCPU) {
-                NotifyRasterizer<false>(word_index, untracked_words[word_index], current_word);
-                untracked_words[word_index] |= current_word;
-                cpu_words[word_index] |= current_word;
-            }
-
-            if constexpr (type == Type::CPU) {
-                const u64 current_bits = untracked_words[word_index] & bits;
-                untracked_words[word_index] &= ~bits;
-                NotifyRasterizer<true>(word_index, current_bits, ~u64{0});
-            }
-            const u64 word = current_word & ~(type == Type::GPU ? untracked_words[word_index] : 0);
-            u64 page = page_begin;
-            page_begin = 0;
-
-            while (page < page_end) {
-                const int empty_bits = std::countr_zero(word >> page);
-                if (on_going && empty_bits != 0) {
-                    InvokeModifiedRange(func, current_size, current_base);
-                    current_size = 0;
-                    on_going = false;
                 }
-                if (empty_bits == PAGES_PER_WORD) {
+                state_words[index] &= ~mask;
-                    break;
+                if constexpr (type == Type::CPU || type == Type::CachedCPU) {
+                    untracked_words[index] &= ~mask;
                 }
-                page += empty_bits;
+                if constexpr (type == Type::CPU) {
-
+                    cached_words[index] &= ~word;
-                const int continuous_bits = std::countr_one(word >> page);
-                if (!on_going && continuous_bits != 0) {
-                    current_base = word_index * PAGES_PER_WORD + page;
-                    on_going = true;
                 }
-                current_size += continuous_bits;
-                page += continuous_bits;
             }
+            const size_t base_offset = index * PAGES_PER_WORD;
+            IteratePages(word, [&](size_t pages_offset, size_t pages_size) {
+                const auto reset = [&]() {
+                    pending_offset = base_offset + pages_offset;
+                    pending_pointer = base_offset + pages_offset + pages_size;
+                };
+                if (!pending) {
+                    reset();
+                    pending = true;
+                    return;
+                }
+                if (pending_pointer == base_offset + pages_offset) {
+                    pending_pointer += pages_size;
+                    return;
+                }
+                release();
+                reset();
+            });
+        });
+        if (pending) {
+            release();
         }
-        if (on_going && current_size > 0) {
-            InvokeModifiedRange(func, current_size, current_base);
-        }
-    }
-
-    template <typename Func>
-    void InvokeModifiedRange(Func&& func, u64 current_size, u64 current_base) {
-        const u64 current_size_bytes = current_size * BYTES_PER_PAGE;
-        const u64 offset_begin = current_base * BYTES_PER_PAGE;
-        const u64 offset_end = std::min(offset_begin + current_size_bytes, SizeBytes());
-        func(cpu_addr + offset_begin, offset_end - offset_begin);
     }
 
     /**
@@ -321,27 +339,17 @@ public:
     [[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept {
         static_assert(type != Type::Untracked);
 
-        const u64* const untracked_words = Array<Type::Untracked>();
+        const std::span<const u64> state_words = words.template Span<type>();
-        const u64* const state_words = Array<type>();
+        bool result = false;
-        const u64 num_query_words = size / BYTES_PER_WORD + 1;
+        IterateWords(offset, size, [&](size_t index, u64 mask) {
-        const u64 word_begin = offset / BYTES_PER_WORD;
+            const u64 word = state_words[index] & mask;
-        const u64 word_end = std::min(word_begin + num_query_words, NumWords());
+            if (word != 0) {
-        const u64 page_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE);
+                result = true;
-        u64 page_index = (offset / BYTES_PER_PAGE) % PAGES_PER_WORD;
-        for (u64 word_index = word_begin; word_index < word_end; ++word_index, page_index = 0) {
-            const u64 off_word = type == Type::GPU ? untracked_words[word_index] : 0;
-            const u64 word = state_words[word_index] & ~off_word;
-            if (word == 0) {
-                continue;
-            }
-            const u64 page_end = std::min((word_index + 1) * PAGES_PER_WORD, page_limit);
-            const u64 local_page_end = page_end % PAGES_PER_WORD;
-            const u64 page_end_shift = (PAGES_PER_WORD - local_page_end) % PAGES_PER_WORD;
-            if (((word >> page_index) << page_index) << page_end_shift != 0) {
                 return true;
             }
-        }
+            return false;
-        return false;
+        });
+        return result;
     }
 
     /**
@@ -353,34 +361,20 @@ public:
     template <Type type>
     [[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept {
         static_assert(type != Type::Untracked);
-        const u64* const state_words = Array<type>();
+        const std::span<const u64> state_words = words.template Span<type>();
-        const u64 num_query_words = size / BYTES_PER_WORD + 1;
-        const u64 word_begin = offset / BYTES_PER_WORD;
-        const u64 word_end = std::min(word_begin + num_query_words, NumWords());
-        const u64 page_base = offset / BYTES_PER_PAGE;
-        u64 page_begin = page_base & (PAGES_PER_WORD - 1);
-        u64 page_end =
-            Common::DivCeil(offset + size, BYTES_PER_PAGE) - (page_base & ~(PAGES_PER_WORD - 1));
         u64 begin = std::numeric_limits<u64>::max();
         u64 end = 0;
-        for (u64 word_index = word_begin; word_index < word_end; ++word_index) {
+        IterateWords(offset, size, [&](size_t index, u64 mask) {
-            const u64 base_mask = (1ULL << page_begin) - 1ULL;
+            const u64 word = state_words[index] & mask;
-            const u64 end_mask = page_end >= PAGES_PER_WORD ? 0ULL : ~((1ULL << page_end) - 1ULL);
-            const u64 off_word = end_mask | base_mask;
-            const u64 word = state_words[word_index] & ~off_word;
             if (word == 0) {
-                page_begin = 0;
+                return;
-                page_end -= PAGES_PER_WORD;
-                continue;
             }
             const u64 local_page_begin = std::countr_zero(word);
             const u64 local_page_end = PAGES_PER_WORD - std::countl_zero(word);
-            const u64 page_index = word_index * PAGES_PER_WORD;
+            const u64 page_index = index * PAGES_PER_WORD;
             begin = std::min(begin, page_index + local_page_begin);
             end = page_index + local_page_end;
-            page_begin = 0;
+        });
-            page_end -= PAGES_PER_WORD;
-        }
         static constexpr std::pair<u64, u64> EMPTY{0, 0};
         return begin < end ? std::make_pair(begin * BYTES_PER_PAGE, end * BYTES_PER_PAGE) : EMPTY;
     }
@@ -454,18 +448,10 @@ private:
     void NotifyRasterizer(u64 word_index, u64 current_bits, u64 new_bits) const {
         u64 changed_bits = (add_to_rasterizer ? current_bits : ~current_bits) & new_bits;
         VAddr addr = cpu_addr + word_index * BYTES_PER_WORD;
-        while (changed_bits != 0) {
+        IteratePages(changed_bits, [&](size_t offset, size_t size) {
-            const int empty_bits = std::countr_zero(changed_bits);
+            rasterizer->UpdatePagesCachedCount(addr + offset * BYTES_PER_PAGE,
-            addr += empty_bits * BYTES_PER_PAGE;
+                                               size * BYTES_PER_PAGE, add_to_rasterizer ? 1 : -1);
-            changed_bits >>= empty_bits;
+        });
-
-            const u32 continuous_bits = std::countr_one(changed_bits);
-            const u64 size = continuous_bits * BYTES_PER_PAGE;
-            const VAddr begin_addr = addr;
-            addr += size;
-            changed_bits = continuous_bits < PAGES_PER_WORD ? (changed_bits >> continuous_bits) : 0;
-            rasterizer->UpdatePagesCachedCount(begin_addr, size, add_to_rasterizer ? 1 : -1);
-        }
     }
 
     VAddr cpu_addr = 0;

src/video_core/fence_manager.h

@@ -64,6 +64,7 @@ public:
     }
 
     void SignalFence(std::function<void()>&& func) {
+        rasterizer.InvalidateGPUCache();
         bool delay_fence = Settings::IsGPULevelHigh();
         if constexpr (!can_async_check) {
             TryReleasePendingFences<false>();

src/video_core/memory_manager.cpp

@@ -170,6 +170,7 @@ void MemoryManager::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_)
 
 GPUVAddr MemoryManager::Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size, PTEKind kind,
                             bool is_big_pages) {
+    std::unique_lock<std::mutex> lock(guard);
     if (is_big_pages) [[likely]] {
         return BigPageTableOp<EntryType::Mapped>(gpu_addr, cpu_addr, size, kind);
     }
@@ -177,6 +178,7 @@ GPUVAddr MemoryManager::Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size,
 }
 
 GPUVAddr MemoryManager::MapSparse(GPUVAddr gpu_addr, std::size_t size, bool is_big_pages) {
+    std::unique_lock<std::mutex> lock(guard);
     if (is_big_pages) [[likely]] {
         return BigPageTableOp<EntryType::Reserved>(gpu_addr, 0, size, PTEKind::INVALID);
     }
@@ -187,6 +189,7 @@ void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {
     if (size == 0) {
         return;
     }
+    std::unique_lock<std::mutex> lock(guard);
     GetSubmappedRangeImpl<false>(gpu_addr, size, page_stash);
 
     for (const auto& [map_addr, map_size] : page_stash) {
@@ -553,6 +556,7 @@ size_t MemoryManager::MaxContinuousRange(GPUVAddr gpu_addr, size_t size) const {
 }
 
 size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr, size_t max_size) const {
+    std::unique_lock<std::mutex> lock(guard);
     return kind_map.GetContinuousSizeFrom(gpu_addr);
 }
 
@@ -745,10 +749,10 @@ void MemoryManager::FlushCaching() {
         return;
     }
     accumulator->Callback([this](GPUVAddr addr, size_t size) {
-        GetSubmappedRangeImpl<false>(addr, size, page_stash);
+        GetSubmappedRangeImpl<false>(addr, size, page_stash2);
     });
-    rasterizer->InnerInvalidation(page_stash);
+    rasterizer->InnerInvalidation(page_stash2);
-    page_stash.clear();
+    page_stash2.clear();
     accumulator->Clear();
 }
 

src/video_core/memory_manager.h

@@ -5,6 +5,7 @@
 
 #include <atomic>
 #include <map>
+#include <mutex>
 #include <optional>
 #include <vector>
 
@@ -215,6 +216,9 @@ private:
 
     std::vector<u64> big_page_continuous;
     std::vector<std::pair<VAddr, std::size_t>> page_stash{};
+    std::vector<std::pair<VAddr, std::size_t>> page_stash2{};
+
+    mutable std::mutex guard;
 
     static constexpr size_t continuous_bits = 64;
 

src/video_core/query_cache.h

@@ -310,16 +310,22 @@ private:
         std::function<void()> operation([this, new_async_job_id, timestamp] {
             std::unique_lock local_lock{mutex};
             AsyncJob& async_job = slot_async_jobs[new_async_job_id];
+            u64 value = async_job.value;
+            VAddr address = async_job.query_location;
+            slot_async_jobs.erase(new_async_job_id);
+            local_lock.unlock();
             if (timestamp) {
                 u64 timestamp_value = *timestamp;
-                cpu_memory.WriteBlockUnsafe(async_job.query_location + sizeof(u64),
+                cpu_memory.WriteBlockUnsafe(address + sizeof(u64), &timestamp_value, sizeof(u64));
-                                            &timestamp_value, sizeof(8));
+                cpu_memory.WriteBlockUnsafe(address, &value, sizeof(u64));
-                cpu_memory.WriteBlockUnsafe(async_job.query_location, &async_job.value, sizeof(8));
+                rasterizer.InvalidateRegion(address, sizeof(u64) * 2,
+                                            VideoCommon::CacheType::NoQueryCache);
             } else {
-                u32 small_value = static_cast<u32>(async_job.value);
+                u32 small_value = static_cast<u32>(value);
-                cpu_memory.WriteBlockUnsafe(async_job.query_location, &small_value, sizeof(u32));
+                cpu_memory.WriteBlockUnsafe(address, &small_value, sizeof(u32));
+                rasterizer.InvalidateRegion(address, sizeof(u32),
+                                            VideoCommon::CacheType::NoQueryCache);
             }
-            slot_async_jobs.erase(new_async_job_id);
         });
         rasterizer.SyncOperation(std::move(operation));
     }

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -801,14 +801,22 @@ void Image::UploadMemory(const ImageBufferMap& map,
     UploadMemory(map.buffer, map.offset, copies);
 }
 
-void Image::DownloadMemory(std::span<GLuint> buffer_handles, size_t buffer_offset,
+void Image::DownloadMemory(GLuint buffer_handle, size_t buffer_offset,
+                           std::span<const VideoCommon::BufferImageCopy> copies) {
+    std::array buffer_handles{buffer_handle};
+    std::array buffer_offsets{buffer_offset};
+    DownloadMemory(buffer_handles, buffer_offsets, copies);
+}
+
+void Image::DownloadMemory(std::span<GLuint> buffer_handles, std::span<size_t> buffer_offsets,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
     if (is_rescaled) {
         ScaleDown();
     }
     glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT); // TODO: Move this to its own API
-    for (auto buffer_handle : buffer_handles) {
+    for (size_t i = 0; buffer_handles.size(); i++) {
+        auto& buffer_handle = buffer_handles[i];
         glBindBuffer(GL_PIXEL_PACK_BUFFER, buffer_handle);
         glPixelStorei(GL_PACK_ALIGNMENT, 1);
 
@@ -827,7 +835,7 @@ void Image::DownloadMemory(std::span<GLuint> buffer_handles, size_t buffer_offse
                 current_image_height = copy.buffer_image_height;
                 glPixelStorei(GL_PACK_IMAGE_HEIGHT, current_image_height);
             }
-            CopyImageToBuffer(copy, buffer_offset);
+            CopyImageToBuffer(copy, buffer_offsets[i]);
         }
     }
     if (is_rescaled) {
@@ -837,10 +845,7 @@ void Image::DownloadMemory(std::span<GLuint> buffer_handles, size_t buffer_offse
 
 void Image::DownloadMemory(ImageBufferMap& map,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
-    std::array buffers{
+    DownloadMemory(map.buffer, map.offset, copies);
-        map.buffer,
-    };
-    DownloadMemory(buffers, map.offset, copies);
 }
 
 GLuint Image::StorageHandle() noexcept {

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -212,7 +212,10 @@ public:
     void UploadMemory(const ImageBufferMap& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
-    void DownloadMemory(std::span<GLuint> buffer_handle, size_t buffer_offset,
+    void DownloadMemory(GLuint buffer_handle, size_t buffer_offset,
+                        std::span<const VideoCommon::BufferImageCopy> copies);
+
+    void DownloadMemory(std::span<GLuint> buffer_handle, std::span<size_t> buffer_offset,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 
     void DownloadMemory(ImageBufferMap& map, std::span<const VideoCommon::BufferImageCopy> copies);

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -696,6 +696,13 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
 std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
     ShaderPools& pools, const ComputePipelineCacheKey& key, Shader::Environment& env,
     PipelineStatistics* statistics, bool build_in_parallel) try {
+    // TODO: Remove this when Intel fixes their shader compiler.
+    //       https://github.com/IGCIT/Intel-GPU-Community-Issue-Tracker-IGCIT/issues/159
+    if (device.GetDriverID() == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
+        LOG_ERROR(Render_Vulkan, "Skipping 0x{:016x}", key.Hash());
+        return nullptr;
+    }
+
     LOG_INFO(Render_Vulkan, "0x{:016x}", key.Hash());
 
     Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()};

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -1342,6 +1342,17 @@ void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImag
     UploadMemory(map.buffer, map.offset, copies);
 }
 
+void Image::DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
+                           std::span<const VideoCommon::BufferImageCopy> copies) {
+    std::array buffer_handles{
+        buffer,
+    };
+    std::array buffer_offsets{
+        offset,
+    };
+    DownloadMemory(buffer_handles, buffer_offsets, copies);
+}
+
 void Image::DownloadMemory(std::span<VkBuffer> buffers_span, std::span<VkDeviceSize> offsets_span,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -138,6 +138,9 @@ public:
     void UploadMemory(const StagingBufferRef& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
+    void DownloadMemory(VkBuffer buffer, VkDeviceSize offset,
+                        std::span<const VideoCommon::BufferImageCopy> copies);
+
     void DownloadMemory(std::span<VkBuffer> buffers, std::span<VkDeviceSize> offsets,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 

src/video_core/texture_cache/texture_cache.h

@@ -755,7 +755,7 @@ void TextureCache<P>::PopAsyncFlushes() {
         auto download_map = runtime.DownloadStagingBuffer(total_size_bytes);
         const size_t original_offset = download_map.offset;
         for (const PendingDownload& download_info : download_ids) {
-            if (download_info.is_swizzle) {
+            if (!download_info.is_swizzle) {
                 continue;
             }
             Image& image = slot_images[download_info.object_id];
@@ -768,7 +768,7 @@ void TextureCache<P>::PopAsyncFlushes() {
         download_map.offset = original_offset;
         std::span<u8> download_span = download_map.mapped_span;
         for (const PendingDownload& download_info : download_ids) {
-            if (download_info.is_swizzle) {
+            if (!download_info.is_swizzle) {
                 continue;
             }
             const ImageBase& image = slot_images[download_info.object_id];
@@ -894,10 +894,7 @@ void TextureCache<P>::DownloadImageIntoBuffer(typename TextureCache<P>::Image* i
         };
         image->DownloadMemory(buffers, buffer_offsets, copies);
     } else {
-        std::array buffers{
+        image->DownloadMemory(buffer, buffer_offset, copies);
-            buffer,
-        };
-        image->DownloadMemory(buffers, buffer_offset, copies);
     }
 }