Merge pull request #5741 from ReinUsesLisp/new-bufcache

video_core: Reimplement the buffer cache

src/tests/video_core/buffer_base.cpp

@@ -471,3 +471,79 @@ TEST_CASE("BufferBase: Unaligned page region query") {
     REQUIRE(buffer.IsRegionCpuModified(c + 4000, 1000));
     REQUIRE(buffer.IsRegionCpuModified(c + 4000, 1));
 }
+
+TEST_CASE("BufferBase: Cached write") {
+    RasterizerInterface rasterizer;
+    BufferBase buffer(rasterizer, c, WORD);
+    buffer.UnmarkRegionAsCpuModified(c, WORD);
+    buffer.CachedCpuWrite(c + PAGE, PAGE);
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.FlushCachedWrites();
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.MarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 0);
+}
+
+TEST_CASE("BufferBase: Multiple cached write") {
+    RasterizerInterface rasterizer;
+    BufferBase buffer(rasterizer, c, WORD);
+    buffer.UnmarkRegionAsCpuModified(c, WORD);
+    buffer.CachedCpuWrite(c + PAGE, PAGE);
+    buffer.CachedCpuWrite(c + PAGE * 3, PAGE);
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE * 3, PAGE));
+    buffer.FlushCachedWrites();
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE * 3, PAGE));
+    buffer.MarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 0);
+}
+
+TEST_CASE("BufferBase: Cached write unmarked") {
+    RasterizerInterface rasterizer;
+    BufferBase buffer(rasterizer, c, WORD);
+    buffer.UnmarkRegionAsCpuModified(c, WORD);
+    buffer.CachedCpuWrite(c + PAGE, PAGE);
+    buffer.UnmarkRegionAsCpuModified(c + PAGE, PAGE);
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.FlushCachedWrites();
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.MarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 0);
+}
+
+TEST_CASE("BufferBase: Cached write iterated") {
+    RasterizerInterface rasterizer;
+    BufferBase buffer(rasterizer, c, WORD);
+    buffer.UnmarkRegionAsCpuModified(c, WORD);
+    buffer.CachedCpuWrite(c + PAGE, PAGE);
+    int num = 0;
+    buffer.ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    REQUIRE(num == 0);
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.FlushCachedWrites();
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    buffer.MarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 0);
+}
+
+TEST_CASE("BufferBase: Cached write downloads") {
+    RasterizerInterface rasterizer;
+    BufferBase buffer(rasterizer, c, WORD);
+    buffer.UnmarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 64);
+    buffer.CachedCpuWrite(c + PAGE, PAGE);
+    REQUIRE(rasterizer.Count() == 63);
+    buffer.MarkRegionAsGpuModified(c + PAGE, PAGE);
+    int num = 0;
+    buffer.ForEachDownloadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    buffer.ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; });
+    REQUIRE(num == 0);
+    REQUIRE(!buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    REQUIRE(!buffer.IsRegionGpuModified(c + PAGE, PAGE));
+    buffer.FlushCachedWrites();
+    REQUIRE(buffer.IsRegionCpuModified(c + PAGE, PAGE));
+    REQUIRE(!buffer.IsRegionGpuModified(c + PAGE, PAGE));
+    buffer.MarkRegionAsCpuModified(c, WORD);
+    REQUIRE(rasterizer.Count() == 0);
+}

src/video_core/CMakeLists.txt

@@ -2,10 +2,8 @@ add_subdirectory(host_shaders)
 
 add_library(video_core STATIC
     buffer_cache/buffer_base.h
-    buffer_cache/buffer_block.h
+    buffer_cache/buffer_cache.cpp
     buffer_cache/buffer_cache.h
-    buffer_cache/map_interval.cpp
-    buffer_cache/map_interval.h
     cdma_pusher.cpp
     cdma_pusher.h
     command_classes/codecs/codec.cpp
@@ -152,8 +150,6 @@ add_library(video_core STATIC
     renderer_vulkan/vk_staging_buffer_pool.h
     renderer_vulkan/vk_state_tracker.cpp
     renderer_vulkan/vk_state_tracker.h
-    renderer_vulkan/vk_stream_buffer.cpp
-    renderer_vulkan/vk_stream_buffer.h
     renderer_vulkan/vk_swapchain.cpp
     renderer_vulkan/vk_swapchain.h
     renderer_vulkan/vk_texture_cache.cpp

src/video_core/buffer_cache/buffer_base.h

@@ -19,6 +19,7 @@ namespace VideoCommon {
 
 enum class BufferFlagBits {
     Picked = 1 << 0,
+    CachedWrites = 1 << 1,
 };
 DECLARE_ENUM_FLAG_OPERATORS(BufferFlagBits)
 
@@ -40,7 +41,7 @@ class BufferBase {
     static constexpr u64 BYTES_PER_WORD = PAGES_PER_WORD * BYTES_PER_PAGE;
 
     /// Vector tracking modified pages tightly packed with small vector optimization
-    union WrittenWords {
+    union WordsArray {
         /// Returns the pointer to the words state
         [[nodiscard]] const u64* Pointer(bool is_short) const noexcept {
             return is_short ? &stack : heap;
@@ -55,49 +56,59 @@ class BufferBase {
         u64* heap;     ///< Not-small buffers pointer to the storage
     };
 
-    struct GpuCpuWords {
-        explicit GpuCpuWords() = default;
-        explicit GpuCpuWords(u64 size_bytes_) : size_bytes{size_bytes_} {
+    struct Words {
+        explicit Words() = default;
+        explicit Words(u64 size_bytes_) : size_bytes{size_bytes_} {
             if (IsShort()) {
                 cpu.stack = ~u64{0};
                 gpu.stack = 0;
+                cached_cpu.stack = 0;
+                untracked.stack = ~u64{0};
             } else {
                 // Share allocation between CPU and GPU pages and set their default values
                 const size_t num_words = NumWords();
-                u64* const alloc = new u64[num_words * 2];
+                u64* const alloc = new u64[num_words * 4];
                 cpu.heap = alloc;
                 gpu.heap = alloc + num_words;
+                cached_cpu.heap = alloc + num_words * 2;
+                untracked.heap = alloc + num_words * 3;
                 std::fill_n(cpu.heap, num_words, ~u64{0});
                 std::fill_n(gpu.heap, num_words, 0);
+                std::fill_n(cached_cpu.heap, num_words, 0);
+                std::fill_n(untracked.heap, num_words, ~u64{0});
             }
             // Clean up tailing bits
-            const u64 last_local_page =
-                Common::DivCeil(size_bytes % BYTES_PER_WORD, BYTES_PER_PAGE);
+            const u64 last_word_size = size_bytes % BYTES_PER_WORD;
+            const u64 last_local_page = Common::DivCeil(last_word_size, BYTES_PER_PAGE);
             const u64 shift = (PAGES_PER_WORD - last_local_page) % PAGES_PER_WORD;
-            u64& last_word = cpu.Pointer(IsShort())[NumWords() - 1];
-            last_word = (last_word << shift) >> shift;
+            const u64 last_word = (~u64{0} << shift) >> shift;
+            cpu.Pointer(IsShort())[NumWords() - 1] = last_word;
+            untracked.Pointer(IsShort())[NumWords() - 1] = last_word;
         }
 
-        ~GpuCpuWords() {
+        ~Words() {
             Release();
         }
 
-        GpuCpuWords& operator=(GpuCpuWords&& rhs) noexcept {
+        Words& operator=(Words&& rhs) noexcept {
             Release();
             size_bytes = rhs.size_bytes;
             cpu = rhs.cpu;
             gpu = rhs.gpu;
+            cached_cpu = rhs.cached_cpu;
+            untracked = rhs.untracked;
             rhs.cpu.heap = nullptr;
             return *this;
         }
 
-        GpuCpuWords(GpuCpuWords&& rhs) noexcept
-            : size_bytes{rhs.size_bytes}, cpu{rhs.cpu}, gpu{rhs.gpu} {
+        Words(Words&& rhs) noexcept
+            : size_bytes{rhs.size_bytes}, cpu{rhs.cpu}, gpu{rhs.gpu},
+              cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked} {
             rhs.cpu.heap = nullptr;
         }
 
-        GpuCpuWords& operator=(const GpuCpuWords&) = delete;
-        GpuCpuWords(const GpuCpuWords&) = delete;
+        Words& operator=(const Words&) = delete;
+        Words(const Words&) = delete;
 
         /// Returns true when the buffer fits in the small vector optimization
         [[nodiscard]] bool IsShort() const noexcept {
@@ -118,8 +129,17 @@ class BufferBase {
         }
 
         u64 size_bytes = 0;
-        WrittenWords cpu;
-        WrittenWords gpu;
+        WordsArray cpu;
+        WordsArray gpu;
+        WordsArray cached_cpu;
+        WordsArray untracked;
+    };
+
+    enum class Type {
+        CPU,
+        GPU,
+        CachedCPU,
+        Untracked,
     };
 
 public:
@@ -132,68 +152,93 @@ public:
     BufferBase& operator=(const BufferBase&) = delete;
     BufferBase(const BufferBase&) = delete;
 
+    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 ModifiedRegion<false>(offset, query_size);
+        return 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 ModifiedRegion<true>(offset, query_size);
+        return 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 IsRegionModified<false>(offset, query_size);
+        return 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 IsRegionModified<true>(offset, query_size);
+        return 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) {
-        ChangeRegionState<true, true>(words.cpu, dirty_cpu_addr, size);
+        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) {
-        ChangeRegionState<false, true>(words.cpu, dirty_cpu_addr, size);
+        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 {
-        ChangeRegionState<true, false>(words.gpu, dirty_cpu_addr, size);
+        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 {
-        ChangeRegionState<false, false>(words.gpu, dirty_cpu_addr, size);
+        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;
+        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;
+        const u64 num_words = NumWords();
+        const u64* const cached_words = Array<Type::CachedCPU>();
+        u64* const untracked_words = Array<Type::Untracked>();
+        u64* const cpu_words = Array<Type::CPU>();
+        for (u64 word_index = 0; word_index < num_words; ++word_index) {
+            const u64 cached_bits = cached_words[word_index];
+            NotifyRasterizer<false>(word_index, untracked_words[word_index], cached_bits);
+            untracked_words[word_index] |= cached_bits;
+            cpu_words[word_index] |= cached_bits;
+        }
     }
 
     /// 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<false, true>(query_cpu_range, size, func);
+        ForEachModifiedRange<Type::CPU>(query_cpu_range, size, 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) {
-        ForEachModifiedRange<true, false>(query_cpu_range, size, func);
+        ForEachModifiedRange<Type::GPU>(query_cpu_range, size, func);
     }
 
     /// Call 'func' for each GPU modified range and unmark those pages as GPU modified
     template <typename Func>
     void ForEachDownloadRange(Func&& func) {
-        ForEachModifiedRange<true, false>(cpu_addr, SizeBytes(), func);
+        ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), func);
     }
 
     /// Mark buffer as picked
@@ -206,6 +251,16 @@ public:
         flags &= ~BufferFlagBits::Picked;
     }
 
+    /// Increases the likeliness of this being a stream buffer
+    void IncreaseStreamScore(int score) noexcept {
+        stream_score += score;
+    }
+
+    /// Returns the likeliness of this being a stream buffer
+    [[nodiscard]] int StreamScore() const noexcept {
+        return stream_score;
+    }
+
     /// Returns true when vaddr -> vaddr+size is fully contained in the buffer
     [[nodiscard]] bool IsInBounds(VAddr addr, u64 size) const noexcept {
         return addr >= cpu_addr && addr + size <= cpu_addr + SizeBytes();
@@ -216,6 +271,11 @@ public:
         return True(flags & BufferFlagBits::Picked);
     }
 
+    /// Returns true when the buffer has pending cached writes
+    [[nodiscard]] bool HasCachedWrites() const noexcept {
+        return True(flags & BufferFlagBits::CachedWrites);
+    }
+
     /// Returns the base CPU address of the buffer
     [[nodiscard]] VAddr CpuAddr() const noexcept {
         return cpu_addr;
@@ -233,26 +293,48 @@ public:
     }
 
 private:
+    template <Type type>
+    u64* Array() noexcept {
+        if constexpr (type == Type::CPU) {
+            return words.cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::GPU) {
+            return words.gpu.Pointer(IsShort());
+        } else if constexpr (type == Type::CachedCPU) {
+            return words.cached_cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::Untracked) {
+            return words.untracked.Pointer(IsShort());
+        }
+    }
+
+    template <Type type>
+    const u64* Array() const noexcept {
+        if constexpr (type == Type::CPU) {
+            return words.cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::GPU) {
+            return words.gpu.Pointer(IsShort());
+        } else if constexpr (type == Type::CachedCPU) {
+            return words.cached_cpu.Pointer(IsShort());
+        } else if constexpr (type == Type::Untracked) {
+            return words.untracked.Pointer(IsShort());
+        }
+    }
+
     /**
      * Change the state of a range of pages
      *
-     * @param written_words Pages to be marked or unmarked as modified
      * @param dirty_addr    Base address to mark or unmark as modified
      * @param size          Size in bytes to mark or unmark as modified
-     *
-     * @tparam enable            True when the bits will be set to one, false for zero
-     * @tparam notify_rasterizer True when the rasterizer has to be notified about the changes
      */
-    template <bool enable, bool notify_rasterizer>
-    void ChangeRegionState(WrittenWords& written_words, u64 dirty_addr,
-                           s64 size) noexcept(!notify_rasterizer) {
+    template <Type type, bool enable>
+    void ChangeRegionState(u64 dirty_addr, s64 size) noexcept(type == Type::GPU) {
         const s64 difference = dirty_addr - cpu_addr;
         const u64 offset = std::max<s64>(difference, 0);
         size += std::min<s64>(difference, 0);
         if (offset >= SizeBytes() || size < 0) {
             return;
         }
-        u64* const state_words = written_words.Pointer(IsShort());
+        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;
@@ -268,13 +350,19 @@ private:
             u64 bits = ~u64{0};
             bits = (bits >> right_offset) << right_offset;
             bits = (bits << left_offset) >> left_offset;
-            if constexpr (notify_rasterizer) {
-                NotifyRasterizer<!enable>(word_index, state_words[word_index], bits);
+            if constexpr (type == Type::CPU || type == Type::CachedCPU) {
+                NotifyRasterizer<!enable>(word_index, untracked_words[word_index], bits);
             }
             if constexpr (enable) {
                 state_words[word_index] |= bits;
+                if constexpr (type == Type::CPU || type == Type::CachedCPU) {
+                    untracked_words[word_index] |= bits;
+                }
             } else {
                 state_words[word_index] &= ~bits;
+                if constexpr (type == Type::CPU || type == Type::CachedCPU) {
+                    untracked_words[word_index] &= ~bits;
+                }
             }
             page_index = 0;
             ++word_index;
@@ -291,7 +379,7 @@ private:
      * @tparam add_to_rasterizer True when the rasterizer should start tracking the new pages
      */
     template <bool add_to_rasterizer>
-    void NotifyRasterizer(u64 word_index, u64 current_bits, u64 new_bits) {
+    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) {
@@ -315,21 +403,20 @@ private:
      * @param query_cpu_range Base CPU address to loop over
      * @param size            Size in bytes of the CPU range to loop over
      * @param func            Function to call for each turned off region
-     *
-     * @tparam gpu               True for host GPU pages, false for CPU pages
-     * @tparam notify_rasterizer True when the rasterizer should be notified about state changes
      */
-    template <bool gpu, bool notify_rasterizer, typename Func>
+    template <Type type, typename Func>
     void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) {
+        static_assert(type != Type::Untracked);
+
         const s64 difference = query_cpu_range - cpu_addr;
         const u64 query_begin = std::max<s64>(difference, 0);
         size += std::min<s64>(difference, 0);
         if (query_begin >= SizeBytes() || size < 0) {
             return;
         }
-        const u64* const cpu_words = words.cpu.Pointer(IsShort());
+        u64* const untracked_words = Array<Type::Untracked>();
+        u64* const state_words = Array<type>();
         const u64 query_end = query_begin + std::min(static_cast<u64>(size), SizeBytes());
-        u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort());
         u64* const words_begin = state_words + query_begin / BYTES_PER_WORD;
         u64* const words_end = state_words + Common::DivCeil(query_end, BYTES_PER_WORD);
 
@@ -345,7 +432,8 @@ private:
         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 = PAGES_PER_WORD - std::countl_zero(last_modified_word[-1]);
+        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;
@@ -371,11 +459,13 @@ private:
             const u64 current_word = state_words[word_index] & bits;
             state_words[word_index] &= ~bits;
 
-            // Exclude CPU modified pages when visiting GPU pages
-            const u64 word = current_word & ~(gpu ? cpu_words[word_index] : 0);
-            if constexpr (notify_rasterizer) {
-                NotifyRasterizer<true>(word_index, word, ~u64{0});
+            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});
             }
+            // Exclude CPU modified pages when visiting GPU pages
+            const u64 word = current_word & ~(type == Type::GPU ? untracked_words[word_index] : 0);
             u64 page = page_begin;
             page_begin = 0;
 
@@ -416,17 +506,20 @@ private:
      * @param offset Offset in bytes from the start of the buffer
      * @param size   Size in bytes of the region to query for modifications
      */
-    template <bool gpu>
+    template <Type type>
     [[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept {
-        const u64* const cpu_words = words.cpu.Pointer(IsShort());
-        const u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort());
+        static_assert(type != Type::Untracked);
+
+        const u64* const untracked_words = Array<Type::Untracked>();
+        const u64* const state_words = Array<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_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE);
         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 word = state_words[word_index] & ~(gpu ? cpu_words[word_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;
             }
@@ -445,13 +538,13 @@ private:
      *
      * @param offset Offset in bytes from the start of the buffer
      * @param size   Size in bytes of the region to query for modifications
-     *
-     * @tparam gpu True to query GPU modified pages, false for CPU pages
      */
-    template <bool gpu>
+    template <Type type>
     [[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept {
-        const u64* const cpu_words = words.cpu.Pointer(IsShort());
-        const u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort());
+        static_assert(type != Type::Untracked);
+
+        const u64* const untracked_words = Array<Type::Untracked>();
+        const u64* const state_words = Array<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());
@@ -460,7 +553,8 @@ private:
         u64 begin = std::numeric_limits<u64>::max();
         u64 end = 0;
         for (u64 word_index = word_begin; word_index < word_end; ++word_index) {
-            const u64 word = state_words[word_index] & ~(gpu ? cpu_words[word_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;
             }
@@ -488,8 +582,9 @@ private:
 
     RasterizerInterface* rasterizer = nullptr;
     VAddr cpu_addr = 0;
-    GpuCpuWords words;
+    Words words;
     BufferFlagBits flags{};
+    int stream_score = 0;
 };
 
 } // namespace VideoCommon

src/video_core/buffer_cache/buffer_block.h

@@ -1,62 +0,0 @@
-// Copyright 2019 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include "common/common_types.h"
-
-namespace VideoCommon {
-
-class BufferBlock {
-public:
-    [[nodiscard]] bool Overlaps(VAddr start, VAddr end) const {
-        return (cpu_addr < end) && (cpu_addr_end > start);
-    }
-
-    [[nodiscard]] bool IsInside(VAddr other_start, VAddr other_end) const {
-        return cpu_addr <= other_start && other_end <= cpu_addr_end;
-    }
-
-    [[nodiscard]] std::size_t Offset(VAddr in_addr) const {
-        return static_cast<std::size_t>(in_addr - cpu_addr);
-    }
-
-    [[nodiscard]] VAddr CpuAddr() const {
-        return cpu_addr;
-    }
-
-    [[nodiscard]] VAddr CpuAddrEnd() const {
-        return cpu_addr_end;
-    }
-
-    void SetCpuAddr(VAddr new_addr) {
-        cpu_addr = new_addr;
-        cpu_addr_end = new_addr + size;
-    }
-
-    [[nodiscard]] std::size_t Size() const {
-        return size;
-    }
-
-    [[nodiscard]] u64 Epoch() const {
-        return epoch;
-    }
-
-    void SetEpoch(u64 new_epoch) {
-        epoch = new_epoch;
-    }
-
-protected:
-    explicit BufferBlock(VAddr cpu_addr_, std::size_t size_) : size{size_} {
-        SetCpuAddr(cpu_addr_);
-    }
-
-private:
-    VAddr cpu_addr{};
-    VAddr cpu_addr_end{};
-    std::size_t size{};
-    u64 epoch{};
-};
-
-} // namespace VideoCommon

src/video_core/buffer_cache/buffer_cache.cpp

@@ -0,0 +1,13 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/microprofile.h"
+
+namespace VideoCommon {
+
+MICROPROFILE_DEFINE(GPU_PrepareBuffers, "GPU", "Prepare buffers", MP_RGB(224, 128, 128));
+MICROPROFILE_DEFINE(GPU_BindUploadBuffers, "GPU", "Bind and upload buffers", MP_RGB(224, 128, 128));
+MICROPROFILE_DEFINE(GPU_DownloadMemory, "GPU", "Download buffers", MP_RGB(224, 128, 128));
+
+} // namespace VideoCommon

src/video_core/buffer_cache/map_interval.cpp

@@ -1,33 +0,0 @@
-// Copyright 2020 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#include <algorithm>
-#include <array>
-#include <cstddef>
-#include <memory>
-
-#include "video_core/buffer_cache/map_interval.h"
-
-namespace VideoCommon {
-
-MapIntervalAllocator::MapIntervalAllocator() {
-    FillFreeList(first_chunk);
-}
-
-MapIntervalAllocator::~MapIntervalAllocator() = default;
-
-void MapIntervalAllocator::AllocateNewChunk() {
-    *new_chunk = std::make_unique<Chunk>();
-    FillFreeList(**new_chunk);
-    new_chunk = &(*new_chunk)->next;
-}
-
-void MapIntervalAllocator::FillFreeList(Chunk& chunk) {
-    const std::size_t old_size = free_list.size();
-    free_list.resize(old_size + chunk.data.size());
-    std::transform(chunk.data.rbegin(), chunk.data.rend(), free_list.begin() + old_size,
-                   [](MapInterval& interval) { return &interval; });
-}
-
-} // namespace VideoCommon

src/video_core/buffer_cache/map_interval.h

@@ -1,93 +0,0 @@
-// Copyright 2019 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include <array>
-#include <cstddef>
-#include <memory>
-#include <vector>
-
-#include <boost/intrusive/set_hook.hpp>
-
-#include "common/common_types.h"
-#include "video_core/gpu.h"
-
-namespace VideoCommon {
-
-struct MapInterval : public boost::intrusive::set_base_hook<boost::intrusive::optimize_size<true>> {
-    MapInterval() = default;
-
-    /*implicit*/ MapInterval(VAddr start_) noexcept : start{start_} {}
-
-    explicit MapInterval(VAddr start_, VAddr end_, GPUVAddr gpu_addr_) noexcept
-        : start{start_}, end{end_}, gpu_addr{gpu_addr_} {}
-
-    bool IsInside(VAddr other_start, VAddr other_end) const noexcept {
-        return start <= other_start && other_end <= end;
-    }
-
-    bool Overlaps(VAddr other_start, VAddr other_end) const noexcept {
-        return start < other_end && other_start < end;
-    }
-
-    void MarkAsModified(bool is_modified_, u64 ticks_) noexcept {
-        is_modified = is_modified_;
-        ticks = ticks_;
-    }
-
-    boost::intrusive::set_member_hook<> member_hook_;
-    VAddr start = 0;
-    VAddr end = 0;
-    GPUVAddr gpu_addr = 0;
-    u64 ticks = 0;
-    bool is_written = false;
-    bool is_modified = false;
-    bool is_registered = false;
-    bool is_memory_marked = false;
-    bool is_sync_pending = false;
-};
-
-struct MapIntervalCompare {
-    constexpr bool operator()(const MapInterval& lhs, const MapInterval& rhs) const noexcept {
-        return lhs.start < rhs.start;
-    }
-};
-
-class MapIntervalAllocator {
-public:
-    MapIntervalAllocator();
-    ~MapIntervalAllocator();
-
-    MapInterval* Allocate() {
-        if (free_list.empty()) {
-            AllocateNewChunk();
-        }
-        MapInterval* const interval = free_list.back();
-        free_list.pop_back();
-        return interval;
-    }
-
-    void Release(MapInterval* interval) {
-        free_list.push_back(interval);
-    }
-
-private:
-    struct Chunk {
-        std::unique_ptr<Chunk> next;
-        std::array<MapInterval, 0x8000> data;
-    };
-
-    void AllocateNewChunk();
-
-    void FillFreeList(Chunk& chunk);
-
-    std::vector<MapInterval*> free_list;
-
-    Chunk first_chunk;
-
-    std::unique_ptr<Chunk>* new_chunk = &first_chunk.next;
-};
-
-} // namespace VideoCommon

src/video_core/command_classes/vic.cpp

@@ -110,12 +110,10 @@ void Vic::Execute() {
                                            converted_frame_buffer.get(), block_height, 0, 0);
 
             gpu.MemoryManager().WriteBlock(output_surface_luma_address, swizzled_data.data(), size);
-            gpu.Maxwell3D().OnMemoryWrite();
         } else {
             // send pitch linear frame
             gpu.MemoryManager().WriteBlock(output_surface_luma_address, converted_frame_buf_addr,
                                            linear_size);
-            gpu.Maxwell3D().OnMemoryWrite();
         }
         break;
     }
@@ -163,7 +161,6 @@ void Vic::Execute() {
         }
         gpu.MemoryManager().WriteBlock(output_surface_chroma_u_address, chroma_buffer.data(),
                                        chroma_buffer.size());
-        gpu.Maxwell3D().OnMemoryWrite();
         break;
     }
     default:

src/video_core/dirty_flags.cpp

@@ -12,13 +12,30 @@
 #define NUM(field_name) (sizeof(::Tegra::Engines::Maxwell3D::Regs::field_name) / (sizeof(u32)))
 
 namespace VideoCommon::Dirty {
-
+namespace {
 using Tegra::Engines::Maxwell3D;
 
-void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables) {
+void SetupDirtyVertexBuffers(Maxwell3D::DirtyState::Tables& tables) {
+    static constexpr std::size_t num_array = 3;
+    for (std::size_t i = 0; i < Maxwell3D::Regs::NumVertexArrays; ++i) {
+        const std::size_t array_offset = OFF(vertex_array) + i * NUM(vertex_array[0]);
+        const std::size_t limit_offset = OFF(vertex_array_limit) + i * NUM(vertex_array_limit[0]);
+
+        FillBlock(tables, array_offset, num_array, VertexBuffer0 + i, VertexBuffers);
+        FillBlock(tables, limit_offset, NUM(vertex_array_limit), VertexBuffer0 + i, VertexBuffers);
+    }
+}
+
+void SetupIndexBuffer(Maxwell3D::DirtyState::Tables& tables) {
+    FillBlock(tables[0], OFF(index_array), NUM(index_array), IndexBuffer);
+}
+
+void SetupDirtyDescriptors(Maxwell3D::DirtyState::Tables& tables) {
     FillBlock(tables[0], OFF(tic), NUM(tic), Descriptors);
     FillBlock(tables[0], OFF(tsc), NUM(tsc), Descriptors);
+}
 
+void SetupDirtyRenderTargets(Maxwell3D::DirtyState::Tables& tables) {
     static constexpr std::size_t num_per_rt = NUM(rt[0]);
     static constexpr std::size_t begin = OFF(rt);
     static constexpr std::size_t num = num_per_rt * Maxwell3D::Regs::NumRenderTargets;
@@ -41,5 +58,13 @@ void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tabl
         FillBlock(table, OFF(zeta), NUM(zeta), flag);
     }
 }
+} // Anonymous namespace
+
+void SetupDirtyFlags(Maxwell3D::DirtyState::Tables& tables) {
+    SetupDirtyVertexBuffers(tables);
+    SetupIndexBuffer(tables);
+    SetupDirtyDescriptors(tables);
+    SetupDirtyRenderTargets(tables);
+}
 
 } // namespace VideoCommon::Dirty

src/video_core/dirty_flags.h

@@ -30,6 +30,12 @@ enum : u8 {
     ColorBuffer7,
     ZetaBuffer,
 
+    VertexBuffers,
+    VertexBuffer0,
+    VertexBuffer31 = VertexBuffer0 + 31,
+
+    IndexBuffer,
+
     LastCommonEntry,
 };
 
@@ -47,6 +53,6 @@ void FillBlock(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables, std::size_
     FillBlock(tables[1], begin, num, index_b);
 }
 
-void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables);
+void SetupDirtyFlags(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables);
 
 } // namespace VideoCommon::Dirty

src/video_core/dma_pusher.cpp

@@ -23,8 +23,6 @@ void DmaPusher::DispatchCalls() {
     MICROPROFILE_SCOPE(DispatchCalls);
 
     gpu.SyncGuestHost();
-    // On entering GPU code, assume all memory may be touched by the ARM core.
-    gpu.Maxwell3D().OnMemoryWrite();
 
     dma_pushbuffer_subindex = 0;
 

src/video_core/engines/fermi_2d.cpp

@@ -18,8 +18,8 @@ Fermi2D::Fermi2D() {
 
 Fermi2D::~Fermi2D() = default;
 
-void Fermi2D::BindRasterizer(VideoCore::RasterizerInterface& rasterizer_) {
-    rasterizer = &rasterizer_;
+void Fermi2D::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
 }
 
 void Fermi2D::CallMethod(u32 method, u32 method_argument, bool is_last_call) {

src/video_core/engines/fermi_2d.h

@@ -38,7 +38,7 @@ public:
     ~Fermi2D();
 
     /// Binds a rasterizer to this engine.
-    void BindRasterizer(VideoCore::RasterizerInterface& rasterizer);
+    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
 
     /// Write the value to the register identified by method.
     void CallMethod(u32 method, u32 method_argument, bool is_last_call) override;

src/video_core/engines/kepler_compute.cpp

@@ -21,8 +21,8 @@ KeplerCompute::KeplerCompute(Core::System& system_, MemoryManager& memory_manage
 
 KeplerCompute::~KeplerCompute() = default;
 
-void KeplerCompute::BindRasterizer(VideoCore::RasterizerInterface& rasterizer_) {
-    rasterizer = &rasterizer_;
+void KeplerCompute::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
 }
 
 void KeplerCompute::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
@@ -39,7 +39,6 @@ void KeplerCompute::CallMethod(u32 method, u32 method_argument, bool is_last_cal
     case KEPLER_COMPUTE_REG_INDEX(data_upload): {
         upload_state.ProcessData(method_argument, is_last_call);
         if (is_last_call) {
-            system.GPU().Maxwell3D().OnMemoryWrite();
         }
         break;
     }

src/video_core/engines/kepler_compute.h

@@ -46,7 +46,7 @@ public:
     ~KeplerCompute();
 
     /// Binds a rasterizer to this engine.
-    void BindRasterizer(VideoCore::RasterizerInterface& rasterizer);
+    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
 
     static constexpr std::size_t NumConstBuffers = 8;
 

src/video_core/engines/kepler_memory.cpp

@@ -33,7 +33,6 @@ void KeplerMemory::CallMethod(u32 method, u32 method_argument, bool is_last_call
     case KEPLERMEMORY_REG_INDEX(data): {
         upload_state.ProcessData(method_argument, is_last_call);
         if (is_last_call) {
-            system.GPU().Maxwell3D().OnMemoryWrite();
         }
         break;
     }

src/video_core/engines/maxwell_3d.cpp

@@ -30,8 +30,8 @@ Maxwell3D::Maxwell3D(Core::System& system_, MemoryManager& memory_manager_)
 
 Maxwell3D::~Maxwell3D() = default;
 
-void Maxwell3D::BindRasterizer(VideoCore::RasterizerInterface& rasterizer_) {
-    rasterizer = &rasterizer_;
+void Maxwell3D::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
 }
 
 void Maxwell3D::InitializeRegisterDefaults() {
@@ -223,7 +223,6 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume
     case MAXWELL3D_REG_INDEX(data_upload):
         upload_state.ProcessData(argument, is_last_call);
         if (is_last_call) {
-            OnMemoryWrite();
         }
         return;
     case MAXWELL3D_REG_INDEX(fragment_barrier):
@@ -570,17 +569,18 @@ std::optional<u64> Maxwell3D::GetQueryResult() {
     }
 }
 
-void Maxwell3D::ProcessCBBind(std::size_t stage_index) {
+void Maxwell3D::ProcessCBBind(size_t stage_index) {
     // Bind the buffer currently in CB_ADDRESS to the specified index in the desired shader stage.
-    auto& shader = state.shader_stages[stage_index];
-    auto& bind_data = regs.cb_bind[stage_index];
-
-    ASSERT(bind_data.index < Regs::MaxConstBuffers);
-    auto& buffer = shader.const_buffers[bind_data.index];
-
+    const auto& bind_data = regs.cb_bind[stage_index];
+    auto& buffer = state.shader_stages[stage_index].const_buffers[bind_data.index];
     buffer.enabled = bind_data.valid.Value() != 0;
     buffer.address = regs.const_buffer.BufferAddress();
     buffer.size = regs.const_buffer.cb_size;
+
+    const bool is_enabled = bind_data.valid.Value() != 0;
+    const GPUVAddr gpu_addr = is_enabled ? regs.const_buffer.BufferAddress() : 0;
+    const u32 size = is_enabled ? regs.const_buffer.cb_size : 0;
+    rasterizer->BindGraphicsUniformBuffer(stage_index, bind_data.index, gpu_addr, size);
 }
 
 void Maxwell3D::ProcessCBData(u32 value) {
@@ -635,7 +635,6 @@ void Maxwell3D::FinishCBData() {
 
     const u32 id = cb_data_state.id;
     memory_manager.WriteBlock(address, cb_data_state.buffer[id].data(), size);
-    OnMemoryWrite();
 
     cb_data_state.id = null_cb_data;
     cb_data_state.current = null_cb_data;

src/video_core/engines/maxwell_3d.h

@@ -55,7 +55,7 @@ public:
     ~Maxwell3D();
 
     /// Binds a rasterizer to this engine.
-    void BindRasterizer(VideoCore::RasterizerInterface& rasterizer);
+    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
 
     /// Register structure of the Maxwell3D engine.
     /// TODO(Subv): This structure will need to be made bigger as more registers are discovered.
@@ -1314,8 +1314,7 @@ public:
 
                     GPUVAddr LimitAddress() const {
                         return static_cast<GPUVAddr>((static_cast<GPUVAddr>(limit_high) << 32) |
-                                                     limit_low) +
-                               1;
+                                                     limit_low);
                     }
                 } vertex_array_limit[NumVertexArrays];
 
@@ -1403,6 +1402,7 @@ public:
         };
 
         std::array<ShaderStageInfo, Regs::MaxShaderStage> shader_stages;
+
         u32 current_instance = 0; ///< Current instance to be used to simulate instanced rendering.
     };
 
@@ -1452,11 +1452,6 @@ public:
         return *rasterizer;
     }
 
-    /// Notify a memory write has happened.
-    void OnMemoryWrite() {
-        dirty.flags |= dirty.on_write_stores;
-    }
-
     enum class MMEDrawMode : u32 {
         Undefined,
         Array,
@@ -1478,7 +1473,6 @@ public:
         using Tables = std::array<Table, 2>;
 
         Flags flags;
-        Flags on_write_stores;
         Tables tables{};
     } dirty;
 
@@ -1541,7 +1535,7 @@ private:
     void FinishCBData();
 
     /// Handles a write to the CB_BIND register.
-    void ProcessCBBind(std::size_t stage_index);
+    void ProcessCBBind(size_t stage_index);
 
     /// Handles a write to the VERTEX_END_GL register, triggering a draw.
     void DrawArrays();

src/video_core/engines/maxwell_dma.cpp

@@ -60,9 +60,6 @@ void MaxwellDMA::Launch() {
         return;
     }
 
-    // All copies here update the main memory, so mark all rasterizer states as invalid.
-    system.GPU().Maxwell3D().OnMemoryWrite();
-
     if (is_src_pitch && is_dst_pitch) {
         CopyPitchToPitch();
     } else {

src/video_core/fence_manager.h

@@ -143,22 +143,26 @@ private:
     }
 
     bool ShouldWait() const {
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         return texture_cache.ShouldWaitAsyncFlushes() || buffer_cache.ShouldWaitAsyncFlushes() ||
                query_cache.ShouldWaitAsyncFlushes();
     }
 
     bool ShouldFlush() const {
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         return texture_cache.HasUncommittedFlushes() || buffer_cache.HasUncommittedFlushes() ||
                query_cache.HasUncommittedFlushes();
     }
 
     void PopAsyncFlushes() {
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         texture_cache.PopAsyncFlushes();
         buffer_cache.PopAsyncFlushes();
         query_cache.PopAsyncFlushes();
     }
 
     void CommitAsyncFlushes() {
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         texture_cache.CommitAsyncFlushes();
         buffer_cache.CommitAsyncFlushes();
         query_cache.CommitAsyncFlushes();

src/video_core/gpu.cpp

@@ -44,8 +44,8 @@ GPU::~GPU() = default;
 
 void GPU::BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer_) {
     renderer = std::move(renderer_);
+    rasterizer = renderer->ReadRasterizer();
 
-    VideoCore::RasterizerInterface& rasterizer = renderer->Rasterizer();
     memory_manager->BindRasterizer(rasterizer);
     maxwell_3d->BindRasterizer(rasterizer);
     fermi_2d->BindRasterizer(rasterizer);
@@ -171,7 +171,7 @@ void GPU::TickWork() {
         const std::size_t size = request.size;
         flush_requests.pop_front();
         flush_request_mutex.unlock();
-        renderer->Rasterizer().FlushRegion(addr, size);
+        rasterizer->FlushRegion(addr, size);
         current_flush_fence.store(fence);
         flush_request_mutex.lock();
     }
@@ -193,11 +193,11 @@ u64 GPU::GetTicks() const {
 }
 
 void GPU::FlushCommands() {
-    renderer->Rasterizer().FlushCommands();
+    rasterizer->FlushCommands();
 }
 
 void GPU::SyncGuestHost() {
-    renderer->Rasterizer().SyncGuestHost();
+    rasterizer->SyncGuestHost();
 }
 
 enum class GpuSemaphoreOperation {

src/video_core/gpu.h

@@ -366,6 +366,7 @@ protected:
     std::unique_ptr<Tegra::DmaPusher> dma_pusher;
     std::unique_ptr<Tegra::CDmaPusher> cdma_pusher;
     std::unique_ptr<VideoCore::RendererBase> renderer;
+    VideoCore::RasterizerInterface* rasterizer = nullptr;
     const bool use_nvdec;
 
 private:

src/video_core/gpu_thread.cpp

@@ -38,6 +38,7 @@ static void RunThread(Core::System& system, VideoCore::RendererBase& renderer,
     }
 
     auto current_context = context.Acquire();
+    VideoCore::RasterizerInterface* const rasterizer = renderer.ReadRasterizer();
 
     CommandDataContainer next;
     while (state.is_running) {
@@ -52,13 +53,13 @@ static void RunThread(Core::System& system, VideoCore::RendererBase& renderer,
         } else if (const auto* data = std::get_if<SwapBuffersCommand>(&next.data)) {
             renderer.SwapBuffers(data->framebuffer ? &*data->framebuffer : nullptr);
         } else if (std::holds_alternative<OnCommandListEndCommand>(next.data)) {
-            renderer.Rasterizer().ReleaseFences();
+            rasterizer->ReleaseFences();
         } else if (std::holds_alternative<GPUTickCommand>(next.data)) {
             system.GPU().TickWork();
         } else if (const auto* flush = std::get_if<FlushRegionCommand>(&next.data)) {
-            renderer.Rasterizer().FlushRegion(flush->addr, flush->size);
+            rasterizer->FlushRegion(flush->addr, flush->size);
         } else if (const auto* invalidate = std::get_if<InvalidateRegionCommand>(&next.data)) {
-            renderer.Rasterizer().OnCPUWrite(invalidate->addr, invalidate->size);
+            rasterizer->OnCPUWrite(invalidate->addr, invalidate->size);
         } else if (std::holds_alternative<EndProcessingCommand>(next.data)) {
             return;
         } else {
@@ -84,6 +85,7 @@ ThreadManager::~ThreadManager() {
 void ThreadManager::StartThread(VideoCore::RendererBase& renderer,
                                 Core::Frontend::GraphicsContext& context,
                                 Tegra::DmaPusher& dma_pusher, Tegra::CDmaPusher& cdma_pusher) {
+    rasterizer = renderer.ReadRasterizer();
     thread = std::thread(RunThread, std::ref(system), std::ref(renderer), std::ref(context),
                          std::ref(dma_pusher), std::ref(state), std::ref(cdma_pusher));
 }
@@ -129,12 +131,12 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) {
 }
 
 void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
-    system.Renderer().Rasterizer().OnCPUWrite(addr, size);
+    rasterizer->OnCPUWrite(addr, size);
 }
 
 void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
     // Skip flush on asynch mode, as FlushAndInvalidateRegion is not used for anything too important
-    system.Renderer().Rasterizer().OnCPUWrite(addr, size);
+    rasterizer->OnCPUWrite(addr, size);
 }
 
 void ThreadManager::WaitIdle() const {

src/video_core/gpu_thread.h

@@ -27,6 +27,7 @@ class System;
 } // namespace Core
 
 namespace VideoCore {
+class RasterizerInterface;
 class RendererBase;
 } // namespace VideoCore
 
@@ -151,11 +152,12 @@ private:
     /// Pushes a command to be executed by the GPU thread
     u64 PushCommand(CommandData&& command_data);
 
-    SynchState state;
     Core::System& system;
-    std::thread thread;
-    std::thread::id thread_id;
     const bool is_async;
+    VideoCore::RasterizerInterface* rasterizer = nullptr;
+
+    SynchState state;
+    std::thread thread;
 };
 
 } // namespace VideoCommon::GPUThread

src/video_core/host_shaders/CMakeLists.txt

@@ -12,7 +12,6 @@ set(SHADER_FILES
     vulkan_blit_depth_stencil.frag
     vulkan_present.frag
     vulkan_present.vert
-    vulkan_quad_array.comp
     vulkan_quad_indexed.comp
     vulkan_uint8.comp
 )

src/video_core/host_shaders/vulkan_quad_array.comp

@@ -1,28 +0,0 @@
-// Copyright 2019 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 460 core
-
-layout (local_size_x = 1024) in;
-
-layout (std430, set = 0, binding = 0) buffer OutputBuffer {
-    uint output_indexes[];
-};
-
-layout (push_constant) uniform PushConstants {
-    uint first;
-};
-
-void main() {
-    uint primitive = gl_GlobalInvocationID.x;
-    if (primitive * 6 >= output_indexes.length()) {
-        return;
-    }
-
-    const uint quad_map[6] = uint[](0, 1, 2, 0, 2, 3);
-    for (uint vertex = 0; vertex < 6; ++vertex) {
-        uint index = first + primitive * 4 + quad_map[vertex];
-        output_indexes[primitive * 6 + vertex] = index;
-    }
-}

src/video_core/host_shaders/vulkan_uint8.comp

@@ -16,9 +16,16 @@ layout (std430, set = 0, binding = 1) writeonly buffer OutputBuffer {
     uint16_t output_indexes[];
 };
 
+uint AssembleIndex(uint id) {
+    // Most primitive restart indices are 0xFF
+    // Hardcode this to 0xFF for now
+    uint index = uint(input_indexes[id]);
+    return index == 0xFF ? 0xFFFF : index;
+}
+
 void main() {
     uint id = gl_GlobalInvocationID.x;
     if (id < input_indexes.length()) {
-        output_indexes[id] = uint16_t(input_indexes[id]);
+        output_indexes[id] = uint16_t(AssembleIndex(id));
     }
 }

src/video_core/memory_manager.cpp

@@ -21,8 +21,8 @@ MemoryManager::MemoryManager(Core::System& system_)
 
 MemoryManager::~MemoryManager() = default;
 
-void MemoryManager::BindRasterizer(VideoCore::RasterizerInterface& rasterizer_) {
-    rasterizer = &rasterizer_;
+void MemoryManager::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
+    rasterizer = rasterizer_;
 }
 
 GPUVAddr MemoryManager::UpdateRange(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size) {

src/video_core/memory_manager.h

@@ -72,7 +72,7 @@ public:
     ~MemoryManager();
 
     /// Binds a renderer to the memory manager.
-    void BindRasterizer(VideoCore::RasterizerInterface& rasterizer);
+    void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
 
     [[nodiscard]] std::optional<VAddr> GpuToCpuAddress(GPUVAddr addr) const;
 
@@ -157,6 +157,8 @@ private:
 
     using MapRange = std::pair<GPUVAddr, size_t>;
     std::vector<MapRange> map_ranges;
+
+    std::vector<std::pair<VAddr, std::size_t>> cache_invalidate_queue;
 };
 
 } // namespace Tegra

src/video_core/rasterizer_interface.h

@@ -7,6 +7,7 @@
 #include <atomic>
 #include <functional>
 #include <optional>
+#include <span>
 #include "common/common_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
@@ -49,6 +50,10 @@ public:
     /// Records a GPU query and caches it
     virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
 
+    /// Signal an uniform buffer binding
+    virtual void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
+                                           u32 size) = 0;
+
     /// Signal a GPU based semaphore as a fence
     virtual void SignalSemaphore(GPUVAddr addr, u32 value) = 0;
 

src/video_core/renderer_base.h

@@ -37,15 +37,11 @@ public:
                           std::unique_ptr<Core::Frontend::GraphicsContext> context);
     virtual ~RendererBase();
 
-    /// Initialize the renderer
-    [[nodiscard]] virtual bool Init() = 0;
-
-    /// Shutdown the renderer
-    virtual void ShutDown() = 0;
-
     /// Finalize rendering the guest frame and draw into the presentation texture
     virtual void SwapBuffers(const Tegra::FramebufferConfig* framebuffer) = 0;
 
+    [[nodiscard]] virtual RasterizerInterface* ReadRasterizer() = 0;
+
     // Getter/setter functions:
     // ------------------------
 
@@ -57,14 +53,6 @@ public:
         return m_current_frame;
     }
 
-    [[nodiscard]] RasterizerInterface& Rasterizer() {
-        return *rasterizer;
-    }
-
-    [[nodiscard]] const RasterizerInterface& Rasterizer() const {
-        return *rasterizer;
-    }
-
     [[nodiscard]] Core::Frontend::GraphicsContext& Context() {
         return *context;
     }
@@ -98,7 +86,6 @@ public:
 
 protected:
     Core::Frontend::EmuWindow& render_window; ///< Reference to the render window handle.
-    std::unique_ptr<RasterizerInterface> rasterizer;
     std::unique_ptr<Core::Frontend::GraphicsContext> context;
     f32 m_current_fps = 0.0f; ///< Current framerate, should be set by the renderer
     int m_current_frame = 0;  ///< Current frame, should be set by the renderer

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -2,98 +2,208 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include <memory>
+#include <span>
 
-#include <glad/glad.h>
-
-#include "common/assert.h"
-#include "common/microprofile.h"
 #include "video_core/buffer_cache/buffer_cache.h"
-#include "video_core/engines/maxwell_3d.h"
-#include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_opengl/gl_buffer_cache.h"
 #include "video_core/renderer_opengl/gl_device.h"
-#include "video_core/renderer_opengl/gl_rasterizer.h"
-#include "video_core/renderer_opengl/gl_resource_manager.h"
 
 namespace OpenGL {
+namespace {
+struct BindlessSSBO {
+    GLuint64EXT address;
+    GLsizei length;
+    GLsizei padding;
+};
+static_assert(sizeof(BindlessSSBO) == sizeof(GLuint) * 4);
 
-using Maxwell = Tegra::Engines::Maxwell3D::Regs;
+constexpr std::array PROGRAM_LUT{
+    GL_VERTEX_PROGRAM_NV,   GL_TESS_CONTROL_PROGRAM_NV, GL_TESS_EVALUATION_PROGRAM_NV,
+    GL_GEOMETRY_PROGRAM_NV, GL_FRAGMENT_PROGRAM_NV,
+};
+} // Anonymous namespace
 
-MICROPROFILE_DEFINE(OpenGL_Buffer_Download, "OpenGL", "Buffer Download", MP_RGB(192, 192, 128));
+Buffer::Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params)
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params) {}
 
-Buffer::Buffer(const Device& device_, VAddr cpu_addr_, std::size_t size_)
-    : BufferBlock{cpu_addr_, size_} {
-    gl_buffer.Create();
-    glNamedBufferData(gl_buffer.handle, static_cast<GLsizeiptr>(size_), nullptr, GL_DYNAMIC_DRAW);
-    if (device_.UseAssemblyShaders() || device_.HasVertexBufferUnifiedMemory()) {
-        glMakeNamedBufferResidentNV(gl_buffer.handle, GL_READ_WRITE);
-        glGetNamedBufferParameterui64vNV(gl_buffer.handle, GL_BUFFER_GPU_ADDRESS_NV, &gpu_address);
+Buffer::Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rasterizer_,
+               VAddr cpu_addr_, u64 size_bytes_)
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(rasterizer_, cpu_addr_, size_bytes_) {
+    buffer.Create();
+    const std::string name = fmt::format("Buffer 0x{:x}", CpuAddr());
+    glObjectLabel(GL_BUFFER, buffer.handle, static_cast<GLsizei>(name.size()), name.data());
+    glNamedBufferData(buffer.handle, SizeBytes(), nullptr, GL_DYNAMIC_DRAW);
+
+    if (runtime.has_unified_vertex_buffers) {
+        glGetNamedBufferParameterui64vNV(buffer.handle, GL_BUFFER_GPU_ADDRESS_NV, &address);
     }
 }
 
-Buffer::~Buffer() = default;
-
-void Buffer::Upload(std::size_t offset, std::size_t data_size, const u8* data) {
-    glNamedBufferSubData(Handle(), static_cast<GLintptr>(offset),
-                         static_cast<GLsizeiptr>(data_size), data);
+void Buffer::ImmediateUpload(size_t offset, std::span<const u8> data) noexcept {
+    glNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+                         static_cast<GLsizeiptr>(data.size_bytes()), data.data());
 }
 
-void Buffer::Download(std::size_t offset, std::size_t data_size, u8* data) {
-    MICROPROFILE_SCOPE(OpenGL_Buffer_Download);
-    const GLsizeiptr gl_size = static_cast<GLsizeiptr>(data_size);
-    const GLintptr gl_offset = static_cast<GLintptr>(offset);
-    if (read_buffer.handle == 0) {
-        read_buffer.Create();
-        glNamedBufferData(read_buffer.handle, static_cast<GLsizeiptr>(Size()), nullptr,
-                          GL_STREAM_READ);
-    }
-    glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
-    glCopyNamedBufferSubData(gl_buffer.handle, read_buffer.handle, gl_offset, gl_offset, gl_size);
-    glGetNamedBufferSubData(read_buffer.handle, gl_offset, gl_size, data);
+void Buffer::ImmediateDownload(size_t offset, std::span<u8> data) noexcept {
+    glGetNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+                            static_cast<GLsizeiptr>(data.size_bytes()), data.data());
 }
 
-void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
-                      std::size_t copy_size) {
-    glCopyNamedBufferSubData(src.Handle(), Handle(), static_cast<GLintptr>(src_offset),
-                             static_cast<GLintptr>(dst_offset), static_cast<GLsizeiptr>(copy_size));
-}
-
-OGLBufferCache::OGLBufferCache(VideoCore::RasterizerInterface& rasterizer_,
-                               Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
-                               const Device& device_, OGLStreamBuffer& stream_buffer_,
-                               StateTracker& state_tracker)
-    : GenericBufferCache{rasterizer_, gpu_memory_, cpu_memory_, stream_buffer_}, device{device_} {
-    if (!device.HasFastBufferSubData()) {
+void Buffer::MakeResident(GLenum access) noexcept {
+    // Abuse GLenum's order to exit early
+    // GL_NONE (default) < GL_READ_ONLY < GL_READ_WRITE
+    if (access <= current_residency_access || buffer.handle == 0) {
         return;
     }
+    if (std::exchange(current_residency_access, access) != GL_NONE) {
+        // If the buffer is already resident, remove its residency before promoting it
+        glMakeNamedBufferNonResidentNV(buffer.handle);
+    }
+    glMakeNamedBufferResidentNV(buffer.handle, access);
+}
 
-    static constexpr GLsizeiptr size = static_cast<GLsizeiptr>(Maxwell::MaxConstBufferSize);
-    glCreateBuffers(static_cast<GLsizei>(std::size(cbufs)), std::data(cbufs));
-    for (const GLuint cbuf : cbufs) {
-        glNamedBufferData(cbuf, size, nullptr, GL_STREAM_DRAW);
+BufferCacheRuntime::BufferCacheRuntime(const Device& device_)
+    : device{device_}, has_fast_buffer_sub_data{device.HasFastBufferSubData()},
+      use_assembly_shaders{device.UseAssemblyShaders()},
+      has_unified_vertex_buffers{device.HasVertexBufferUnifiedMemory()},
+      stream_buffer{has_fast_buffer_sub_data ? std::nullopt : std::make_optional<StreamBuffer>()} {
+    GLint gl_max_attributes;
+    glGetIntegerv(GL_MAX_VERTEX_ATTRIBS, &gl_max_attributes);
+    max_attributes = static_cast<u32>(gl_max_attributes);
+    for (auto& stage_uniforms : fast_uniforms) {
+        for (OGLBuffer& buffer : stage_uniforms) {
+            buffer.Create();
+            glNamedBufferData(buffer.handle, BufferCache::SKIP_CACHE_SIZE, nullptr, GL_STREAM_DRAW);
+        }
+    }
+    for (auto& stage_uniforms : copy_uniforms) {
+        for (OGLBuffer& buffer : stage_uniforms) {
+            buffer.Create();
+            glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
+        }
+    }
+    for (OGLBuffer& buffer : copy_compute_uniforms) {
+        buffer.Create();
+        glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
     }
 }
 
-OGLBufferCache::~OGLBufferCache() {
-    glDeleteBuffers(static_cast<GLsizei>(std::size(cbufs)), std::data(cbufs));
+void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
+                                    std::span<const VideoCommon::BufferCopy> copies) {
+    for (const VideoCommon::BufferCopy& copy : copies) {
+        glCopyNamedBufferSubData(
+            src_buffer.Handle(), dst_buffer.Handle(), static_cast<GLintptr>(copy.src_offset),
+            static_cast<GLintptr>(copy.dst_offset), static_cast<GLsizeiptr>(copy.size));
+    }
 }
 
-std::shared_ptr<Buffer> OGLBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
-    return std::make_shared<Buffer>(device, cpu_addr, size);
+void BufferCacheRuntime::BindIndexBuffer(Buffer& buffer, u32 offset, u32 size) {
+    if (has_unified_vertex_buffers) {
+        buffer.MakeResident(GL_READ_ONLY);
+        glBufferAddressRangeNV(GL_ELEMENT_ARRAY_ADDRESS_NV, 0, buffer.HostGpuAddr() + offset,
+                               static_cast<GLsizeiptr>(size));
+    } else {
+        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, buffer.Handle());
+        index_buffer_offset = offset;
+    }
 }
 
-OGLBufferCache::BufferInfo OGLBufferCache::GetEmptyBuffer(std::size_t) {
-    return {0, 0, 0};
+void BufferCacheRuntime::BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size,
+                                          u32 stride) {
+    if (index >= max_attributes) {
+        return;
+    }
+    if (has_unified_vertex_buffers) {
+        buffer.MakeResident(GL_READ_ONLY);
+        glBindVertexBuffer(index, 0, 0, static_cast<GLsizei>(stride));
+        glBufferAddressRangeNV(GL_VERTEX_ATTRIB_ARRAY_ADDRESS_NV, index,
+                               buffer.HostGpuAddr() + offset, static_cast<GLsizeiptr>(size));
+    } else {
+        glBindVertexBuffer(index, buffer.Handle(), static_cast<GLintptr>(offset),
+                           static_cast<GLsizei>(stride));
+    }
 }
 
-OGLBufferCache::BufferInfo OGLBufferCache::ConstBufferUpload(const void* raw_pointer,
-                                                             std::size_t size) {
-    DEBUG_ASSERT(cbuf_cursor < std::size(cbufs));
-    const GLuint cbuf = cbufs[cbuf_cursor++];
+void BufferCacheRuntime::BindUniformBuffer(size_t stage, u32 binding_index, Buffer& buffer,
+                                           u32 offset, u32 size) {
+    if (use_assembly_shaders) {
+        GLuint handle;
+        if (offset != 0) {
+            handle = copy_uniforms[stage][binding_index].handle;
+            glCopyNamedBufferSubData(buffer.Handle(), handle, offset, 0, size);
+        } else {
+            handle = buffer.Handle();
+        }
+        glBindBufferRangeNV(PABO_LUT[stage], binding_index, handle, 0,
+                            static_cast<GLsizeiptr>(size));
+    } else {
+        const GLuint base_binding = device.GetBaseBindings(stage).uniform_buffer;
+        const GLuint binding = base_binding + binding_index;
+        glBindBufferRange(GL_UNIFORM_BUFFER, binding, buffer.Handle(),
+                          static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
+    }
+}
 
-    glNamedBufferSubData(cbuf, 0, static_cast<GLsizeiptr>(size), raw_pointer);
-    return {cbuf, 0, 0};
+void BufferCacheRuntime::BindComputeUniformBuffer(u32 binding_index, Buffer& buffer, u32 offset,
+                                                  u32 size) {
+    if (use_assembly_shaders) {
+        GLuint handle;
+        if (offset != 0) {
+            handle = copy_compute_uniforms[binding_index].handle;
+            glCopyNamedBufferSubData(buffer.Handle(), handle, offset, 0, size);
+        } else {
+            handle = buffer.Handle();
+        }
+        glBindBufferRangeNV(GL_COMPUTE_PROGRAM_PARAMETER_BUFFER_NV, binding_index, handle, 0,
+                            static_cast<GLsizeiptr>(size));
+    } else {
+        glBindBufferRange(GL_UNIFORM_BUFFER, binding_index, buffer.Handle(),
+                          static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
+    }
+}
+
+void BufferCacheRuntime::BindStorageBuffer(size_t stage, u32 binding_index, Buffer& buffer,
+                                           u32 offset, u32 size, bool is_written) {
+    if (use_assembly_shaders) {
+        const BindlessSSBO ssbo{
+            .address = buffer.HostGpuAddr() + offset,
+            .length = static_cast<GLsizei>(size),
+            .padding = 0,
+        };
+        buffer.MakeResident(is_written ? GL_READ_WRITE : GL_READ_ONLY);
+        glProgramLocalParametersI4uivNV(PROGRAM_LUT[stage], binding_index, 1,
+                                        reinterpret_cast<const GLuint*>(&ssbo));
+    } else {
+        const GLuint base_binding = device.GetBaseBindings(stage).shader_storage_buffer;
+        const GLuint binding = base_binding + binding_index;
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, binding, buffer.Handle(),
+                          static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
+    }
+}
+
+void BufferCacheRuntime::BindComputeStorageBuffer(u32 binding_index, Buffer& buffer, u32 offset,
+                                                  u32 size, bool is_written) {
+    if (use_assembly_shaders) {
+        const BindlessSSBO ssbo{
+            .address = buffer.HostGpuAddr() + offset,
+            .length = static_cast<GLsizei>(size),
+            .padding = 0,
+        };
+        buffer.MakeResident(is_written ? GL_READ_WRITE : GL_READ_ONLY);
+        glProgramLocalParametersI4uivNV(GL_COMPUTE_PROGRAM_NV, binding_index, 1,
+                                        reinterpret_cast<const GLuint*>(&ssbo));
+    } else if (size == 0) {
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, binding_index, 0, 0, 0);
+    } else {
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, binding_index, buffer.Handle(),
+                          static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
+    }
+}
+
+void BufferCacheRuntime::BindTransformFeedbackBuffer(u32 index, Buffer& buffer, u32 offset,
+                                                     u32 size) {
+    glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, index, buffer.Handle(),
+                      static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -5,79 +5,157 @@
 #pragma once
 
 #include <array>
-#include <memory>
+#include <span>
 
+#include "common/alignment.h"
 #include "common/common_types.h"
+#include "common/dynamic_library.h"
 #include "video_core/buffer_cache/buffer_cache.h"
-#include "video_core/engines/maxwell_3d.h"
+#include "video_core/rasterizer_interface.h"
+#include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_stream_buffer.h"
 
-namespace Core {
-class System;
-}
-
 namespace OpenGL {
 
-class Device;
-class OGLStreamBuffer;
-class RasterizerOpenGL;
-class StateTracker;
+class BufferCacheRuntime;
 
-class Buffer : public VideoCommon::BufferBlock {
+class Buffer : public VideoCommon::BufferBase<VideoCore::RasterizerInterface> {
 public:
-    explicit Buffer(const Device& device_, VAddr cpu_addr_, std::size_t size_);
-    ~Buffer();
+    explicit Buffer(BufferCacheRuntime&, VideoCore::RasterizerInterface& rasterizer, VAddr cpu_addr,
+                    u64 size_bytes);
+    explicit Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams);
 
-    void Upload(std::size_t offset, std::size_t data_size, const u8* data);
+    void ImmediateUpload(size_t offset, std::span<const u8> data) noexcept;
 
-    void Download(std::size_t offset, std::size_t data_size, u8* data);
+    void ImmediateDownload(size_t offset, std::span<u8> data) noexcept;
 
-    void CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
-                  std::size_t copy_size);
+    void MakeResident(GLenum access) noexcept;
 
-    GLuint Handle() const noexcept {
-        return gl_buffer.handle;
+    [[nodiscard]] GLuint64EXT HostGpuAddr() const noexcept {
+        return address;
     }
 
-    u64 Address() const noexcept {
-        return gpu_address;
+    [[nodiscard]] GLuint Handle() const noexcept {
+        return buffer.handle;
     }
 
 private:
-    OGLBuffer gl_buffer;
-    OGLBuffer read_buffer;
-    u64 gpu_address = 0;
+    GLuint64EXT address = 0;
+    OGLBuffer buffer;
+    GLenum current_residency_access = GL_NONE;
 };
 
-using GenericBufferCache = VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer>;
-class OGLBufferCache final : public GenericBufferCache {
+class BufferCacheRuntime {
+    friend Buffer;
+
 public:
-    explicit OGLBufferCache(VideoCore::RasterizerInterface& rasterizer,
-                            Tegra::MemoryManager& gpu_memory, Core::Memory::Memory& cpu_memory,
-                            const Device& device, OGLStreamBuffer& stream_buffer,
-                            StateTracker& state_tracker);
-    ~OGLBufferCache();
+    static constexpr u8 INVALID_BINDING = std::numeric_limits<u8>::max();
 
-    BufferInfo GetEmptyBuffer(std::size_t) override;
+    explicit BufferCacheRuntime(const Device& device_);
 
-    void Acquire() noexcept {
-        cbuf_cursor = 0;
+    void CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
+                    std::span<const VideoCommon::BufferCopy> copies);
+
+    void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);
+
+    void BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size, u32 stride);
+
+    void BindUniformBuffer(size_t stage, u32 binding_index, Buffer& buffer, u32 offset, u32 size);
+
+    void BindComputeUniformBuffer(u32 binding_index, Buffer& buffer, u32 offset, u32 size);
+
+    void BindStorageBuffer(size_t stage, u32 binding_index, Buffer& buffer, u32 offset, u32 size,
+                           bool is_written);
+
+    void BindComputeStorageBuffer(u32 binding_index, Buffer& buffer, u32 offset, u32 size,
+                                  bool is_written);
+
+    void BindTransformFeedbackBuffer(u32 index, Buffer& buffer, u32 offset, u32 size);
+
+    void BindFastUniformBuffer(size_t stage, u32 binding_index, u32 size) {
+        if (use_assembly_shaders) {
+            const GLuint handle = fast_uniforms[stage][binding_index].handle;
+            const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
+            glBindBufferRangeNV(PABO_LUT[stage], binding_index, handle, 0, gl_size);
+        } else {
+            const GLuint base_binding = device.GetBaseBindings(stage).uniform_buffer;
+            const GLuint binding = base_binding + binding_index;
+            glBindBufferRange(GL_UNIFORM_BUFFER, binding,
+                              fast_uniforms[stage][binding_index].handle, 0,
+                              static_cast<GLsizeiptr>(size));
+        }
     }
 
-protected:
-    std::shared_ptr<Buffer> CreateBlock(VAddr cpu_addr, std::size_t size) override;
+    void PushFastUniformBuffer(size_t stage, u32 binding_index, std::span<const u8> data) {
+        if (use_assembly_shaders) {
+            glProgramBufferParametersIuivNV(
+                PABO_LUT[stage], binding_index, 0,
+                static_cast<GLsizei>(data.size_bytes() / sizeof(GLuint)),
+                reinterpret_cast<const GLuint*>(data.data()));
+        } else {
+            glNamedBufferSubData(fast_uniforms[stage][binding_index].handle, 0,
+                                 static_cast<GLsizeiptr>(data.size_bytes()), data.data());
+        }
+    }
 
-    BufferInfo ConstBufferUpload(const void* raw_pointer, std::size_t size) override;
+    std::span<u8> BindMappedUniformBuffer(size_t stage, u32 binding_index, u32 size) noexcept {
+        const auto [mapped_span, offset] = stream_buffer->Request(static_cast<size_t>(size));
+        const GLuint base_binding = device.GetBaseBindings(stage).uniform_buffer;
+        const GLuint binding = base_binding + binding_index;
+        glBindBufferRange(GL_UNIFORM_BUFFER, binding, stream_buffer->Handle(),
+                          static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));
+        return mapped_span;
+    }
+
+    [[nodiscard]] const GLvoid* IndexOffset() const noexcept {
+        return reinterpret_cast<const GLvoid*>(static_cast<uintptr_t>(index_buffer_offset));
+    }
+
+    [[nodiscard]] bool HasFastBufferSubData() const noexcept {
+        return has_fast_buffer_sub_data;
+    }
 
 private:
-    static constexpr std::size_t NUM_CBUFS = Tegra::Engines::Maxwell3D::Regs::MaxConstBuffers *
-                                             Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram;
+    static constexpr std::array PABO_LUT{
+        GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV,          GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
+        GL_TESS_EVALUATION_PROGRAM_PARAMETER_BUFFER_NV, GL_GEOMETRY_PROGRAM_PARAMETER_BUFFER_NV,
+        GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV,
+    };
 
     const Device& device;
 
-    std::size_t cbuf_cursor = 0;
-    std::array<GLuint, NUM_CBUFS> cbufs{};
+    bool has_fast_buffer_sub_data = false;
+    bool use_assembly_shaders = false;
+    bool has_unified_vertex_buffers = false;
+
+    u32 max_attributes = 0;
+
+    std::optional<StreamBuffer> stream_buffer;
+
+    std::array<std::array<OGLBuffer, VideoCommon::NUM_GRAPHICS_UNIFORM_BUFFERS>,
+               VideoCommon::NUM_STAGES>
+        fast_uniforms;
+    std::array<std::array<OGLBuffer, VideoCommon::NUM_GRAPHICS_UNIFORM_BUFFERS>,
+               VideoCommon::NUM_STAGES>
+        copy_uniforms;
+    std::array<OGLBuffer, VideoCommon::NUM_COMPUTE_UNIFORM_BUFFERS> copy_compute_uniforms;
+
+    u32 index_buffer_offset = 0;
 };
 
+struct BufferCacheParams {
+    using Runtime = OpenGL::BufferCacheRuntime;
+    using Buffer = OpenGL::Buffer;
+
+    static constexpr bool IS_OPENGL = true;
+    static constexpr bool HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS = true;
+    static constexpr bool HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT = true;
+    static constexpr bool NEEDS_BIND_UNIFORM_INDEX = true;
+    static constexpr bool NEEDS_BIND_STORAGE_INDEX = true;
+    static constexpr bool USE_MEMORY_MAPS = false;
+};
+
+using BufferCache = VideoCommon::BufferCache<BufferCacheParams>;
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_device.cpp

@@ -21,9 +21,7 @@
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
 namespace OpenGL {
-
 namespace {
-
 // One uniform block is reserved for emulation purposes
 constexpr u32 ReservedUniformBlocks = 1;
 
@@ -197,11 +195,13 @@ bool IsASTCSupported() {
     const bool nsight = std::getenv("NVTX_INJECTION64_PATH") || std::getenv("NSIGHT_LAUNCHED");
     return nsight || HasExtension(extensions, "GL_EXT_debug_tool");
 }
-
 } // Anonymous namespace
 
-Device::Device()
-    : max_uniform_buffers{BuildMaxUniformBuffers()}, base_bindings{BuildBaseBindings()} {
+Device::Device() {
+    if (!GLAD_GL_VERSION_4_6) {
+        LOG_ERROR(Render_OpenGL, "OpenGL 4.6 is not available");
+        throw std::runtime_error{"Insufficient version"};
+    }
     const std::string_view vendor = reinterpret_cast<const char*>(glGetString(GL_VENDOR));
     const std::string_view version = reinterpret_cast<const char*>(glGetString(GL_VERSION));
     const std::vector extensions = GetExtensions();
@@ -217,6 +217,9 @@ Device::Device()
             "Beta driver 443.24 is known to have issues. There might be performance issues.");
         disable_fast_buffer_sub_data = true;
     }
+
+    max_uniform_buffers = BuildMaxUniformBuffers();
+    base_bindings = BuildBaseBindings();
     uniform_buffer_alignment = GetInteger<size_t>(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT);
     shader_storage_alignment = GetInteger<size_t>(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT);
     max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS);

src/video_core/renderer_opengl/gl_device.h

@@ -10,11 +10,9 @@
 
 namespace OpenGL {
 
-static constexpr u32 EmulationUniformBlockBinding = 0;
-
-class Device final {
+class Device {
 public:
-    struct BaseBindings final {
+    struct BaseBindings {
         u32 uniform_buffer{};
         u32 shader_storage_buffer{};
         u32 sampler{};

src/video_core/renderer_opengl/gl_fence_manager.cpp

@@ -47,7 +47,7 @@ void GLInnerFence::Wait() {
 
 FenceManagerOpenGL::FenceManagerOpenGL(VideoCore::RasterizerInterface& rasterizer_,
                                        Tegra::GPU& gpu_, TextureCache& texture_cache_,
-                                       OGLBufferCache& buffer_cache_, QueryCache& query_cache_)
+                                       BufferCache& buffer_cache_, QueryCache& query_cache_)
     : GenericFenceManager{rasterizer_, gpu_, texture_cache_, buffer_cache_, query_cache_} {}
 
 Fence FenceManagerOpenGL::CreateFence(u32 value, bool is_stubbed) {

src/video_core/renderer_opengl/gl_fence_manager.h

@@ -32,14 +32,13 @@ private:
 };
 
 using Fence = std::shared_ptr<GLInnerFence>;
-using GenericFenceManager =
-    VideoCommon::FenceManager<Fence, TextureCache, OGLBufferCache, QueryCache>;
+using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCache, BufferCache, QueryCache>;
 
 class FenceManagerOpenGL final : public GenericFenceManager {
 public:
-    explicit FenceManagerOpenGL(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
-                                TextureCache& texture_cache_, OGLBufferCache& buffer_cache_,
-                                QueryCache& query_cache_);
+    explicit FenceManagerOpenGL(VideoCore::RasterizerInterface& rasterizer, Tegra::GPU& gpu,
+                                TextureCache& texture_cache, BufferCache& buffer_cache,
+                                QueryCache& query_cache);
 
 protected:
     Fence CreateFence(u32 value, bool is_stubbed) override;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -44,28 +44,14 @@ using VideoCore::Surface::PixelFormat;
 using VideoCore::Surface::SurfaceTarget;
 using VideoCore::Surface::SurfaceType;
 
-MICROPROFILE_DEFINE(OpenGL_VAO, "OpenGL", "Vertex Format Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_VB, "OpenGL", "Vertex Buffer Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_Shader, "OpenGL", "Shader Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_UBO, "OpenGL", "Const Buffer Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_Index, "OpenGL", "Index Buffer Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_Texture, "OpenGL", "Texture Setup", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_Framebuffer, "OpenGL", "Framebuffer Setup", MP_RGB(128, 128, 192));
 MICROPROFILE_DEFINE(OpenGL_Drawing, "OpenGL", "Drawing", MP_RGB(128, 128, 192));
+MICROPROFILE_DEFINE(OpenGL_Clears, "OpenGL", "Clears", MP_RGB(128, 128, 192));
 MICROPROFILE_DEFINE(OpenGL_Blits, "OpenGL", "Blits", MP_RGB(128, 128, 192));
-MICROPROFILE_DEFINE(OpenGL_CacheManagement, "OpenGL", "Cache Mgmt", MP_RGB(100, 255, 100));
-MICROPROFILE_DEFINE(OpenGL_PrimitiveAssembly, "OpenGL", "Prim Asmbl", MP_RGB(255, 100, 100));
+MICROPROFILE_DEFINE(OpenGL_CacheManagement, "OpenGL", "Cache Management", MP_RGB(100, 255, 100));
 
 namespace {
 
-constexpr size_t NUM_CONST_BUFFERS_PER_STAGE = 18;
-constexpr size_t NUM_CONST_BUFFERS_BYTES_PER_STAGE =
-    NUM_CONST_BUFFERS_PER_STAGE * Maxwell::MaxConstBufferSize;
-constexpr size_t TOTAL_CONST_BUFFER_BYTES =
-    NUM_CONST_BUFFERS_BYTES_PER_STAGE * Maxwell::MaxShaderStage;
-
 constexpr size_t NUM_SUPPORTED_VERTEX_ATTRIBUTES = 16;
-constexpr size_t NUM_SUPPORTED_VERTEX_BINDINGS = 16;
 
 struct TextureHandle {
     constexpr TextureHandle(u32 data, bool via_header_index) {
@@ -101,20 +87,6 @@ TextureHandle GetTextureInfo(const Engine& engine, bool via_header_index, const
     return TextureHandle(engine.AccessConstBuffer32(shader_type, buffer, offset), via_header_index);
 }
 
-std::size_t GetConstBufferSize(const Tegra::Engines::ConstBufferInfo& buffer,
-                               const ConstBufferEntry& entry) {
-    if (!entry.IsIndirect()) {
-        return entry.GetSize();
-    }
-    if (buffer.size > Maxwell::MaxConstBufferSize) {
-        LOG_WARNING(Render_OpenGL, "Indirect constbuffer size {} exceeds maximum {}", buffer.size,
-                    Maxwell::MaxConstBufferSize);
-        return Maxwell::MaxConstBufferSize;
-    }
-
-    return buffer.size;
-}
-
 /// Translates hardware transform feedback indices
 /// @param location Hardware location
 /// @return Pair of ARB_transform_feedback3 token stream first and third arguments
@@ -147,14 +119,6 @@ void oglEnable(GLenum cap, bool state) {
     (state ? glEnable : glDisable)(cap);
 }
 
-void UpdateBindlessSSBOs(GLenum target, const BindlessSSBO* ssbos, size_t num_ssbos) {
-    if (num_ssbos == 0) {
-        return;
-    }
-    glProgramLocalParametersI4uivNV(target, 0, static_cast<GLsizei>(num_ssbos),
-                                    reinterpret_cast<const GLuint*>(ssbos));
-}
-
 ImageViewType ImageViewTypeFromEntry(const SamplerEntry& entry) {
     if (entry.is_buffer) {
         return ImageViewType::Buffer;
@@ -201,44 +165,28 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra
     : RasterizerAccelerated(cpu_memory_), gpu(gpu_), maxwell3d(gpu.Maxwell3D()),
       kepler_compute(gpu.KeplerCompute()), gpu_memory(gpu.MemoryManager()), device(device_),
       screen_info(screen_info_), program_manager(program_manager_), state_tracker(state_tracker_),
-      stream_buffer(device, state_tracker),
       texture_cache_runtime(device, program_manager, state_tracker),
       texture_cache(texture_cache_runtime, *this, maxwell3d, kepler_compute, gpu_memory),
+      buffer_cache_runtime(device),
+      buffer_cache(*this, maxwell3d, kepler_compute, gpu_memory, cpu_memory_, buffer_cache_runtime),
       shader_cache(*this, emu_window_, gpu, maxwell3d, kepler_compute, gpu_memory, device),
       query_cache(*this, maxwell3d, gpu_memory),
-      buffer_cache(*this, gpu_memory, cpu_memory_, device, stream_buffer, state_tracker),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache),
       async_shaders(emu_window_) {
-    unified_uniform_buffer.Create();
-    glNamedBufferStorage(unified_uniform_buffer.handle, TOTAL_CONST_BUFFER_BYTES, nullptr, 0);
-
-    if (device.UseAssemblyShaders()) {
-        glCreateBuffers(static_cast<GLsizei>(staging_cbufs.size()), staging_cbufs.data());
-        for (const GLuint cbuf : staging_cbufs) {
-            glNamedBufferStorage(cbuf, static_cast<GLsizeiptr>(Maxwell::MaxConstBufferSize),
-                                 nullptr, 0);
-        }
-    }
     if (device.UseAsynchronousShaders()) {
         async_shaders.AllocateWorkers();
     }
 }
 
-RasterizerOpenGL::~RasterizerOpenGL() {
-    if (device.UseAssemblyShaders()) {
-        glDeleteBuffers(static_cast<GLsizei>(staging_cbufs.size()), staging_cbufs.data());
-    }
-}
+RasterizerOpenGL::~RasterizerOpenGL() = default;
 
-void RasterizerOpenGL::SetupVertexFormat() {
+void RasterizerOpenGL::SyncVertexFormats() {
     auto& flags = maxwell3d.dirty.flags;
     if (!flags[Dirty::VertexFormats]) {
         return;
     }
     flags[Dirty::VertexFormats] = false;
 
-    MICROPROFILE_SCOPE(OpenGL_VAO);
-
     // Use the vertex array as-is, assumes that the data is formatted correctly for OpenGL. Enables
     // the first 16 vertex attributes always, as we don't know which ones are actually used until
     // shader time. Note, Tegra technically supports 32, but we're capping this to 16 for now to
@@ -274,55 +222,7 @@ void RasterizerOpenGL::SetupVertexFormat() {
     }
 }
 
-void RasterizerOpenGL::SetupVertexBuffer() {
-    auto& flags = maxwell3d.dirty.flags;
-    if (!flags[Dirty::VertexBuffers]) {
-        return;
-    }
-    flags[Dirty::VertexBuffers] = false;
-
-    MICROPROFILE_SCOPE(OpenGL_VB);
-
-    const bool use_unified_memory = device.HasVertexBufferUnifiedMemory();
-
-    // Upload all guest vertex arrays sequentially to our buffer
-    const auto& regs = maxwell3d.regs;
-    for (std::size_t index = 0; index < NUM_SUPPORTED_VERTEX_BINDINGS; ++index) {
-        if (!flags[Dirty::VertexBuffer0 + index]) {
-            continue;
-        }
-        flags[Dirty::VertexBuffer0 + index] = false;
-
-        const auto& vertex_array = regs.vertex_array[index];
-        if (!vertex_array.IsEnabled()) {
-            continue;
-        }
-
-        const GPUVAddr start = vertex_array.StartAddress();
-        const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
-        ASSERT(end >= start);
-
-        const GLuint gl_index = static_cast<GLuint>(index);
-        const u64 size = end - start;
-        if (size == 0) {
-            glBindVertexBuffer(gl_index, 0, 0, vertex_array.stride);
-            if (use_unified_memory) {
-                glBufferAddressRangeNV(GL_VERTEX_ATTRIB_ARRAY_ADDRESS_NV, gl_index, 0, 0);
-            }
-            continue;
-        }
-        const auto info = buffer_cache.UploadMemory(start, size);
-        if (use_unified_memory) {
-            glBindVertexBuffer(gl_index, 0, 0, vertex_array.stride);
-            glBufferAddressRangeNV(GL_VERTEX_ATTRIB_ARRAY_ADDRESS_NV, gl_index,
-                                   info.address + info.offset, size);
-        } else {
-            glBindVertexBuffer(gl_index, info.handle, info.offset, vertex_array.stride);
-        }
-    }
-}
-
-void RasterizerOpenGL::SetupVertexInstances() {
+void RasterizerOpenGL::SyncVertexInstances() {
     auto& flags = maxwell3d.dirty.flags;
     if (!flags[Dirty::VertexInstances]) {
         return;
@@ -343,17 +243,7 @@ void RasterizerOpenGL::SetupVertexInstances() {
     }
 }
 
-GLintptr RasterizerOpenGL::SetupIndexBuffer() {
-    MICROPROFILE_SCOPE(OpenGL_Index);
-    const auto& regs = maxwell3d.regs;
-    const std::size_t size = CalculateIndexBufferSize();
-    const auto info = buffer_cache.UploadMemory(regs.index_array.IndexStart(), size);
-    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, info.handle);
-    return info.offset;
-}
-
-void RasterizerOpenGL::SetupShaders() {
-    MICROPROFILE_SCOPE(OpenGL_Shader);
+void RasterizerOpenGL::SetupShaders(bool is_indexed) {
     u32 clip_distances = 0;
 
     std::array<Shader*, Maxwell::MaxShaderStage> shaders{};
@@ -410,11 +300,19 @@ void RasterizerOpenGL::SetupShaders() {
         const size_t stage = index == 0 ? 0 : index - 1;
         shaders[stage] = shader;
 
-        SetupDrawConstBuffers(stage, shader);
-        SetupDrawGlobalMemory(stage, shader);
         SetupDrawTextures(shader, stage);
         SetupDrawImages(shader, stage);
 
+        buffer_cache.SetEnabledUniformBuffers(stage, shader->GetEntries().enabled_uniform_buffers);
+
+        buffer_cache.UnbindGraphicsStorageBuffers(stage);
+        u32 ssbo_index = 0;
+        for (const auto& buffer : shader->GetEntries().global_memory_entries) {
+            buffer_cache.BindGraphicsStorageBuffer(stage, ssbo_index, buffer.cbuf_index,
+                                                   buffer.cbuf_offset, buffer.is_written);
+            ++ssbo_index;
+        }
+
         // Workaround for Intel drivers.
         // When a clip distance is enabled but not set in the shader it crops parts of the screen
         // (sometimes it's half the screen, sometimes three quarters). To avoid this, enable the
@@ -430,43 +328,26 @@ void RasterizerOpenGL::SetupShaders() {
     SyncClipEnabled(clip_distances);
     maxwell3d.dirty.flags[Dirty::Shaders] = false;
 
+    buffer_cache.UpdateGraphicsBuffers(is_indexed);
+
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillGraphicsImageViews(indices_span, image_view_ids);
 
+    buffer_cache.BindHostGeometryBuffers(is_indexed);
+
     size_t image_view_index = 0;
     size_t texture_index = 0;
     size_t image_index = 0;
     for (size_t stage = 0; stage < Maxwell::MaxShaderStage; ++stage) {
         const Shader* const shader = shaders[stage];
-        if (shader) {
-            const auto base = device.GetBaseBindings(stage);
-            BindTextures(shader->GetEntries(), base.sampler, base.image, image_view_index,
-                         texture_index, image_index);
-        }
-    }
-}
-
-std::size_t RasterizerOpenGL::CalculateVertexArraysSize() const {
-    const auto& regs = maxwell3d.regs;
-
-    std::size_t size = 0;
-    for (u32 index = 0; index < Maxwell::NumVertexArrays; ++index) {
-        if (!regs.vertex_array[index].IsEnabled())
+        if (!shader) {
             continue;
-
-        const GPUVAddr start = regs.vertex_array[index].StartAddress();
-        const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
-
-        size += end - start;
-        ASSERT(end >= start);
+        }
+        buffer_cache.BindHostStageBuffers(stage);
+        const auto& base = device.GetBaseBindings(stage);
+        BindTextures(shader->GetEntries(), base.sampler, base.image, image_view_index,
+                     texture_index, image_index);
     }
-
-    return size;
-}
-
-std::size_t RasterizerOpenGL::CalculateIndexBufferSize() const {
-    return static_cast<std::size_t>(maxwell3d.regs.index_array.count) *
-           static_cast<std::size_t>(maxwell3d.regs.index_array.FormatSizeInBytes());
 }
 
 void RasterizerOpenGL::LoadDiskResources(u64 title_id, const std::atomic_bool& stop_loading,
@@ -475,6 +356,7 @@ void RasterizerOpenGL::LoadDiskResources(u64 title_id, const std::atomic_bool& s
 }
 
 void RasterizerOpenGL::Clear() {
+    MICROPROFILE_SCOPE(OpenGL_Clears);
     if (!maxwell3d.ShouldExecute()) {
         return;
     }
@@ -525,11 +407,9 @@ void RasterizerOpenGL::Clear() {
     }
     UNIMPLEMENTED_IF(regs.clear_flags.viewport);
 
-    {
-        auto lock = texture_cache.AcquireLock();
-        texture_cache.UpdateRenderTargets(true);
-        state_tracker.BindFramebuffer(texture_cache.GetFramebuffer()->Handle());
-    }
+    std::scoped_lock lock{texture_cache.mutex};
+    texture_cache.UpdateRenderTargets(true);
+    state_tracker.BindFramebuffer(texture_cache.GetFramebuffer()->Handle());
 
     if (use_color) {
         glClearBufferfv(GL_COLOR, regs.clear_buffers.RT, regs.clear_color);
@@ -541,7 +421,6 @@ void RasterizerOpenGL::Clear() {
     } else if (use_stencil) {
         glClearBufferiv(GL_STENCIL, 0, &regs.clear_stencil);
     }
-
     ++num_queued_commands;
 }
 
@@ -550,75 +429,12 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
 
     query_cache.UpdateCounters();
 
-    SyncViewport();
-    SyncRasterizeEnable();
-    SyncPolygonModes();
-    SyncColorMask();
-    SyncFragmentColorClampState();
-    SyncMultiSampleState();
-    SyncDepthTestState();
-    SyncDepthClamp();
-    SyncStencilTestState();
-    SyncBlendState();
-    SyncLogicOpState();
-    SyncCullMode();
-    SyncPrimitiveRestart();
-    SyncScissorTest();
-    SyncPointState();
-    SyncLineState();
-    SyncPolygonOffset();
-    SyncAlphaTest();
-    SyncFramebufferSRGB();
-
-    buffer_cache.Acquire();
-    current_cbuf = 0;
-
-    std::size_t buffer_size = CalculateVertexArraysSize();
-
-    // Add space for index buffer
-    if (is_indexed) {
-        buffer_size = Common::AlignUp(buffer_size, 4) + CalculateIndexBufferSize();
-    }
-
-    // Uniform space for the 5 shader stages
-    buffer_size =
-        Common::AlignUp<std::size_t>(buffer_size, 4) +
-        (sizeof(MaxwellUniformData) + device.GetUniformBufferAlignment()) * Maxwell::MaxShaderStage;
-
-    // Add space for at least 18 constant buffers
-    buffer_size += Maxwell::MaxConstBuffers *
-                   (Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
-
-    // Prepare the vertex array.
-    buffer_cache.Map(buffer_size);
-
-    // Prepare vertex array format.
-    SetupVertexFormat();
-
-    // Upload vertex and index data.
-    SetupVertexBuffer();
-    SetupVertexInstances();
-    GLintptr index_buffer_offset = 0;
-    if (is_indexed) {
-        index_buffer_offset = SetupIndexBuffer();
-    }
-
-    // Setup emulation uniform buffer.
-    if (!device.UseAssemblyShaders()) {
-        MaxwellUniformData ubo;
-        ubo.SetFromRegs(maxwell3d);
-        const auto info =
-            buffer_cache.UploadHostMemory(&ubo, sizeof(ubo), device.GetUniformBufferAlignment());
-        glBindBufferRange(GL_UNIFORM_BUFFER, EmulationUniformBlockBinding, info.handle, info.offset,
-                          static_cast<GLsizeiptr>(sizeof(ubo)));
-    }
+    SyncState();
 
     // Setup shaders and their used resources.
-    auto lock = texture_cache.AcquireLock();
-    SetupShaders();
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
+    SetupShaders(is_indexed);
 
-    // Signal the buffer cache that we are not going to upload more things.
-    buffer_cache.Unmap();
     texture_cache.UpdateRenderTargets(false);
     state_tracker.BindFramebuffer(texture_cache.GetFramebuffer()->Handle());
     program_manager.BindGraphicsPipeline();
@@ -632,7 +448,7 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     if (is_indexed) {
         const GLint base_vertex = static_cast<GLint>(maxwell3d.regs.vb_element_base);
         const GLsizei num_vertices = static_cast<GLsizei>(maxwell3d.regs.index_array.count);
-        const GLvoid* offset = reinterpret_cast<const GLvoid*>(index_buffer_offset);
+        const GLvoid* const offset = buffer_cache_runtime.IndexOffset();
         const GLenum format = MaxwellToGL::IndexFormat(maxwell3d.regs.index_array.format);
         if (num_instances == 1 && base_instance == 0 && base_vertex == 0) {
             glDrawElements(primitive_mode, num_vertices, format, offset);
@@ -672,22 +488,22 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
 }
 
 void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
-    buffer_cache.Acquire();
-    current_cbuf = 0;
-
     Shader* const kernel = shader_cache.GetComputeKernel(code_addr);
 
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     BindComputeTextures(kernel);
 
-    const size_t buffer_size = Tegra::Engines::KeplerCompute::NumConstBuffers *
-                               (Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
-    buffer_cache.Map(buffer_size);
-
-    SetupComputeConstBuffers(kernel);
-    SetupComputeGlobalMemory(kernel);
-
-    buffer_cache.Unmap();
+    const auto& entries = kernel->GetEntries();
+    buffer_cache.SetEnabledComputeUniformBuffers(entries.enabled_uniform_buffers);
+    buffer_cache.UnbindComputeStorageBuffers();
+    u32 ssbo_index = 0;
+    for (const auto& buffer : entries.global_memory_entries) {
+        buffer_cache.BindComputeStorageBuffer(ssbo_index, buffer.cbuf_index, buffer.cbuf_offset,
+                                              buffer.is_written);
+        ++ssbo_index;
+    }
+    buffer_cache.UpdateComputeBuffers();
+    buffer_cache.BindHostComputeBuffers();
 
     const auto& launch_desc = kepler_compute.launch_description;
     glDispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
@@ -703,6 +519,12 @@ void RasterizerOpenGL::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
     query_cache.Query(gpu_addr, type, timestamp);
 }
 
+void RasterizerOpenGL::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
+                                                 u32 size) {
+    std::scoped_lock lock{buffer_cache.mutex};
+    buffer_cache.BindGraphicsUniformBuffer(stage, index, gpu_addr, size);
+}
+
 void RasterizerOpenGL::FlushAll() {}
 
 void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
@@ -711,19 +533,23 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
         return;
     }
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.DownloadMemory(addr, size);
     }
-    buffer_cache.FlushRegion(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.DownloadMemory(addr, size);
+    }
     query_cache.FlushRegion(addr, size);
 }
 
 bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     if (!Settings::IsGPULevelHigh()) {
-        return buffer_cache.MustFlushRegion(addr, size);
+        return buffer_cache.IsRegionGpuModified(addr, size);
     }
     return texture_cache.IsRegionGpuModified(addr, size) ||
-           buffer_cache.MustFlushRegion(addr, size);
+           buffer_cache.IsRegionGpuModified(addr, size);
 }
 
 void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
@@ -732,11 +558,14 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
         return;
     }
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
     }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
     shader_cache.InvalidateRegion(addr, size);
-    buffer_cache.InvalidateRegion(addr, size);
     query_cache.InvalidateRegion(addr, size);
 }
 
@@ -745,26 +574,35 @@ void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
     if (addr == 0 || size == 0) {
         return;
     }
+    shader_cache.OnCPUWrite(addr, size);
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
     }
-    shader_cache.OnCPUWrite(addr, size);
-    buffer_cache.OnCPUWrite(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.CachedWriteMemory(addr, size);
+    }
 }
 
 void RasterizerOpenGL::SyncGuestHost() {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
-    buffer_cache.SyncGuestHost();
     shader_cache.SyncGuestHost();
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.FlushCachedWrites();
+    }
 }
 
 void RasterizerOpenGL::UnmapMemory(VAddr addr, u64 size) {
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.UnmapMemory(addr, size);
     }
-    buffer_cache.OnCPUWrite(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
     shader_cache.OnCPUWrite(addr, size);
 }
 
@@ -799,14 +637,7 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
 }
 
 void RasterizerOpenGL::WaitForIdle() {
-    // Place a barrier on everything that is not framebuffer related.
-    // This is related to another flag that is not currently implemented.
-    glMemoryBarrier(GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT | GL_ELEMENT_ARRAY_BARRIER_BIT |
-                    GL_UNIFORM_BARRIER_BIT | GL_TEXTURE_FETCH_BARRIER_BIT |
-                    GL_SHADER_IMAGE_ACCESS_BARRIER_BIT | GL_COMMAND_BARRIER_BIT |
-                    GL_PIXEL_BUFFER_BARRIER_BIT | GL_TEXTURE_UPDATE_BARRIER_BIT |
-                    GL_BUFFER_UPDATE_BARRIER_BIT | GL_TRANSFORM_FEEDBACK_BARRIER_BIT |
-                    GL_SHADER_STORAGE_BARRIER_BIT | GL_QUERY_BUFFER_BARRIER_BIT);
+    glMemoryBarrier(GL_ALL_BARRIER_BITS);
 }
 
 void RasterizerOpenGL::FragmentBarrier() {
@@ -831,18 +662,21 @@ void RasterizerOpenGL::TickFrame() {
     num_queued_commands = 0;
 
     fence_manager.TickFrame();
-    buffer_cache.TickFrame();
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.TickFrame();
     }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.TickFrame();
+    }
 }
 
 bool RasterizerOpenGL::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                              const Tegra::Engines::Fermi2D::Surface& dst,
                                              const Tegra::Engines::Fermi2D::Config& copy_config) {
     MICROPROFILE_SCOPE(OpenGL_Blits);
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{texture_cache.mutex};
     texture_cache.BlitImage(dst, src, copy_config);
     return true;
 }
@@ -854,7 +688,7 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
     }
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
 
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{texture_cache.mutex};
     ImageView* const image_view{texture_cache.TryFindFramebufferImageView(framebuffer_addr)};
     if (!image_view) {
         return false;
@@ -921,166 +755,6 @@ void RasterizerOpenGL::BindTextures(const ShaderEntries& entries, GLuint base_te
     }
 }
 
-void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, Shader* shader) {
-    static constexpr std::array PARAMETER_LUT{
-        GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV,          GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
-        GL_TESS_EVALUATION_PROGRAM_PARAMETER_BUFFER_NV, GL_GEOMETRY_PROGRAM_PARAMETER_BUFFER_NV,
-        GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV,
-    };
-    MICROPROFILE_SCOPE(OpenGL_UBO);
-    const auto& stages = maxwell3d.state.shader_stages;
-    const auto& shader_stage = stages[stage_index];
-    const auto& entries = shader->GetEntries();
-    const bool use_unified = entries.use_unified_uniforms;
-    const std::size_t base_unified_offset = stage_index * NUM_CONST_BUFFERS_BYTES_PER_STAGE;
-
-    const auto base_bindings = device.GetBaseBindings(stage_index);
-    u32 binding = device.UseAssemblyShaders() ? 0 : base_bindings.uniform_buffer;
-    for (const auto& entry : entries.const_buffers) {
-        const u32 index = entry.GetIndex();
-        const auto& buffer = shader_stage.const_buffers[index];
-        SetupConstBuffer(PARAMETER_LUT[stage_index], binding, buffer, entry, use_unified,
-                         base_unified_offset + index * Maxwell::MaxConstBufferSize);
-        ++binding;
-    }
-    if (use_unified) {
-        const u32 index = static_cast<u32>(base_bindings.shader_storage_buffer +
-                                           entries.global_memory_entries.size());
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, index, unified_uniform_buffer.handle,
-                          base_unified_offset, NUM_CONST_BUFFERS_BYTES_PER_STAGE);
-    }
-}
-
-void RasterizerOpenGL::SetupComputeConstBuffers(Shader* kernel) {
-    MICROPROFILE_SCOPE(OpenGL_UBO);
-    const auto& launch_desc = kepler_compute.launch_description;
-    const auto& entries = kernel->GetEntries();
-    const bool use_unified = entries.use_unified_uniforms;
-
-    u32 binding = 0;
-    for (const auto& entry : entries.const_buffers) {
-        const auto& config = launch_desc.const_buffer_config[entry.GetIndex()];
-        const std::bitset<8> mask = launch_desc.const_buffer_enable_mask.Value();
-        Tegra::Engines::ConstBufferInfo buffer;
-        buffer.address = config.Address();
-        buffer.size = config.size;
-        buffer.enabled = mask[entry.GetIndex()];
-        SetupConstBuffer(GL_COMPUTE_PROGRAM_PARAMETER_BUFFER_NV, binding, buffer, entry,
-                         use_unified, entry.GetIndex() * Maxwell::MaxConstBufferSize);
-        ++binding;
-    }
-    if (use_unified) {
-        const GLuint index = static_cast<GLuint>(entries.global_memory_entries.size());
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, index, unified_uniform_buffer.handle, 0,
-                          NUM_CONST_BUFFERS_BYTES_PER_STAGE);
-    }
-}
-
-void RasterizerOpenGL::SetupConstBuffer(GLenum stage, u32 binding,
-                                        const Tegra::Engines::ConstBufferInfo& buffer,
-                                        const ConstBufferEntry& entry, bool use_unified,
-                                        std::size_t unified_offset) {
-    if (!buffer.enabled) {
-        // Set values to zero to unbind buffers
-        if (device.UseAssemblyShaders()) {
-            glBindBufferRangeNV(stage, entry.GetIndex(), 0, 0, 0);
-        } else {
-            glBindBufferRange(GL_UNIFORM_BUFFER, binding, 0, 0, sizeof(float));
-        }
-        return;
-    }
-
-    // Align the actual size so it ends up being a multiple of vec4 to meet the OpenGL std140
-    // UBO alignment requirements.
-    const std::size_t size = Common::AlignUp(GetConstBufferSize(buffer, entry), sizeof(GLvec4));
-
-    const bool fast_upload = !use_unified && device.HasFastBufferSubData();
-
-    const std::size_t alignment = use_unified ? 4 : device.GetUniformBufferAlignment();
-    const GPUVAddr gpu_addr = buffer.address;
-    auto info = buffer_cache.UploadMemory(gpu_addr, size, alignment, false, fast_upload);
-
-    if (device.UseAssemblyShaders()) {
-        UNIMPLEMENTED_IF(use_unified);
-        if (info.offset != 0) {
-            const GLuint staging_cbuf = staging_cbufs[current_cbuf++];
-            glCopyNamedBufferSubData(info.handle, staging_cbuf, info.offset, 0, size);
-            info.handle = staging_cbuf;
-            info.offset = 0;
-        }
-        glBindBufferRangeNV(stage, binding, info.handle, info.offset, size);
-        return;
-    }
-
-    if (use_unified) {
-        glCopyNamedBufferSubData(info.handle, unified_uniform_buffer.handle, info.offset,
-                                 unified_offset, size);
-    } else {
-        glBindBufferRange(GL_UNIFORM_BUFFER, binding, info.handle, info.offset, size);
-    }
-}
-
-void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, Shader* shader) {
-    static constexpr std::array TARGET_LUT = {
-        GL_VERTEX_PROGRAM_NV,   GL_TESS_CONTROL_PROGRAM_NV, GL_TESS_EVALUATION_PROGRAM_NV,
-        GL_GEOMETRY_PROGRAM_NV, GL_FRAGMENT_PROGRAM_NV,
-    };
-    const auto& cbufs{maxwell3d.state.shader_stages[stage_index]};
-    const auto& entries{shader->GetEntries().global_memory_entries};
-
-    std::array<BindlessSSBO, 32> ssbos;
-    ASSERT(entries.size() < ssbos.size());
-
-    const bool assembly_shaders = device.UseAssemblyShaders();
-    u32 binding = assembly_shaders ? 0 : device.GetBaseBindings(stage_index).shader_storage_buffer;
-    for (const auto& entry : entries) {
-        const GPUVAddr addr{cbufs.const_buffers[entry.cbuf_index].address + entry.cbuf_offset};
-        const GPUVAddr gpu_addr{gpu_memory.Read<u64>(addr)};
-        const u32 size{gpu_memory.Read<u32>(addr + 8)};
-        SetupGlobalMemory(binding, entry, gpu_addr, size, &ssbos[binding]);
-        ++binding;
-    }
-    if (assembly_shaders) {
-        UpdateBindlessSSBOs(TARGET_LUT[stage_index], ssbos.data(), entries.size());
-    }
-}
-
-void RasterizerOpenGL::SetupComputeGlobalMemory(Shader* kernel) {
-    const auto& cbufs{kepler_compute.launch_description.const_buffer_config};
-    const auto& entries{kernel->GetEntries().global_memory_entries};
-
-    std::array<BindlessSSBO, 32> ssbos;
-    ASSERT(entries.size() < ssbos.size());
-
-    u32 binding = 0;
-    for (const auto& entry : entries) {
-        const GPUVAddr addr{cbufs[entry.cbuf_index].Address() + entry.cbuf_offset};
-        const GPUVAddr gpu_addr{gpu_memory.Read<u64>(addr)};
-        const u32 size{gpu_memory.Read<u32>(addr + 8)};
-        SetupGlobalMemory(binding, entry, gpu_addr, size, &ssbos[binding]);
-        ++binding;
-    }
-    if (device.UseAssemblyShaders()) {
-        UpdateBindlessSSBOs(GL_COMPUTE_PROGRAM_NV, ssbos.data(), ssbos.size());
-    }
-}
-
-void RasterizerOpenGL::SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& entry,
-                                         GPUVAddr gpu_addr, size_t size, BindlessSSBO* ssbo) {
-    const size_t alignment{device.GetShaderStorageBufferAlignment()};
-    const auto info = buffer_cache.UploadMemory(gpu_addr, size, alignment, entry.is_written);
-    if (device.UseAssemblyShaders()) {
-        *ssbo = BindlessSSBO{
-            .address = static_cast<GLuint64EXT>(info.address + info.offset),
-            .length = static_cast<GLsizei>(size),
-            .padding = 0,
-        };
-    } else {
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, binding, info.handle, info.offset,
-                          static_cast<GLsizeiptr>(size));
-    }
-}
-
 void RasterizerOpenGL::SetupDrawTextures(const Shader* shader, size_t stage_index) {
     const bool via_header_index =
         maxwell3d.regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex;
@@ -1128,6 +802,30 @@ void RasterizerOpenGL::SetupComputeImages(const Shader* shader) {
     }
 }
 
+void RasterizerOpenGL::SyncState() {
+    SyncViewport();
+    SyncRasterizeEnable();
+    SyncPolygonModes();
+    SyncColorMask();
+    SyncFragmentColorClampState();
+    SyncMultiSampleState();
+    SyncDepthTestState();
+    SyncDepthClamp();
+    SyncStencilTestState();
+    SyncBlendState();
+    SyncLogicOpState();
+    SyncCullMode();
+    SyncPrimitiveRestart();
+    SyncScissorTest();
+    SyncPointState();
+    SyncLineState();
+    SyncPolygonOffset();
+    SyncAlphaTest();
+    SyncFramebufferSRGB();
+    SyncVertexFormats();
+    SyncVertexInstances();
+}
+
 void RasterizerOpenGL::SyncViewport() {
     auto& flags = maxwell3d.dirty.flags;
     const auto& regs = maxwell3d.regs;
@@ -1163,9 +861,11 @@ void RasterizerOpenGL::SyncViewport() {
         if (regs.screen_y_control.y_negate != 0) {
             flip_y = !flip_y;
         }
-        glClipControl(flip_y ? GL_UPPER_LEFT : GL_LOWER_LEFT,
-                      regs.depth_mode == Maxwell::DepthMode::ZeroToOne ? GL_ZERO_TO_ONE
-                                                                       : GL_NEGATIVE_ONE_TO_ONE);
+        const bool is_zero_to_one = regs.depth_mode == Maxwell::DepthMode::ZeroToOne;
+        const GLenum origin = flip_y ? GL_UPPER_LEFT : GL_LOWER_LEFT;
+        const GLenum depth = is_zero_to_one ? GL_ZERO_TO_ONE : GL_NEGATIVE_ONE_TO_ONE;
+        state_tracker.ClipControl(origin, depth);
+        state_tracker.SetYNegate(regs.screen_y_control.y_negate != 0);
     }
 
     if (dirty_viewport) {
@@ -1649,36 +1349,13 @@ void RasterizerOpenGL::BeginTransformFeedback(GLenum primitive_mode) {
     if (regs.tfb_enabled == 0) {
         return;
     }
-
     if (device.UseAssemblyShaders()) {
         SyncTransformFeedback();
     }
-
     UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationControl) ||
                      regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationEval) ||
                      regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::Geometry));
-
-    for (std::size_t index = 0; index < Maxwell::NumTransformFeedbackBuffers; ++index) {
-        const auto& binding = regs.tfb_bindings[index];
-        if (!binding.buffer_enable) {
-            if (enabled_transform_feedback_buffers[index]) {
-                glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, static_cast<GLuint>(index), 0, 0,
-                                  0);
-            }
-            enabled_transform_feedback_buffers[index] = false;
-            continue;
-        }
-        enabled_transform_feedback_buffers[index] = true;
-
-        auto& tfb_buffer = transform_feedback_buffers[index];
-        tfb_buffer.Create();
-
-        const GLuint handle = tfb_buffer.handle;
-        const std::size_t size = binding.buffer_size;
-        glNamedBufferData(handle, static_cast<GLsizeiptr>(size), nullptr, GL_STREAM_COPY);
-        glBindBufferRange(GL_TRANSFORM_FEEDBACK_BUFFER, static_cast<GLuint>(index), handle, 0,
-                          static_cast<GLsizeiptr>(size));
-    }
+    UNIMPLEMENTED_IF(primitive_mode != GL_POINTS);
 
     // We may have to call BeginTransformFeedbackNV here since they seem to call different
     // implementations on Nvidia's driver (the pointer is different) but we are using
@@ -1692,23 +1369,7 @@ void RasterizerOpenGL::EndTransformFeedback() {
     if (regs.tfb_enabled == 0) {
         return;
     }
-
     glEndTransformFeedback();
-
-    for (std::size_t index = 0; index < Maxwell::NumTransformFeedbackBuffers; ++index) {
-        const auto& binding = regs.tfb_bindings[index];
-        if (!binding.buffer_enable) {
-            continue;
-        }
-        UNIMPLEMENTED_IF(binding.buffer_offset != 0);
-
-        const GLuint handle = transform_feedback_buffers[index].handle;
-        const GPUVAddr gpu_addr = binding.Address();
-        const std::size_t size = binding.buffer_size;
-        const auto info = buffer_cache.UploadMemory(gpu_addr, size, 4, true);
-        glCopyNamedBufferSubData(handle, info.handle, 0, info.offset,
-                                 static_cast<GLsizeiptr>(size));
-    }
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -30,7 +30,6 @@
 #include "video_core/renderer_opengl/gl_shader_decompiler.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
 #include "video_core/renderer_opengl/gl_state_tracker.h"
-#include "video_core/renderer_opengl/gl_stream_buffer.h"
 #include "video_core/renderer_opengl/gl_texture_cache.h"
 #include "video_core/shader/async_shaders.h"
 #include "video_core/textures/texture.h"
@@ -72,6 +71,7 @@ public:
     void DispatchCompute(GPUVAddr code_addr) override;
     void ResetCounter(VideoCore::QueryType type) override;
     void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
     bool MustFlushRegion(VAddr addr, u64 size) override;
@@ -119,27 +119,6 @@ private:
     void BindTextures(const ShaderEntries& entries, GLuint base_texture, GLuint base_image,
                       size_t& image_view_index, size_t& texture_index, size_t& image_index);
 
-    /// Configures the current constbuffers to use for the draw command.
-    void SetupDrawConstBuffers(std::size_t stage_index, Shader* shader);
-
-    /// Configures the current constbuffers to use for the kernel invocation.
-    void SetupComputeConstBuffers(Shader* kernel);
-
-    /// Configures a constant buffer.
-    void SetupConstBuffer(GLenum stage, u32 binding, const Tegra::Engines::ConstBufferInfo& buffer,
-                          const ConstBufferEntry& entry, bool use_unified,
-                          std::size_t unified_offset);
-
-    /// Configures the current global memory entries to use for the draw command.
-    void SetupDrawGlobalMemory(std::size_t stage_index, Shader* shader);
-
-    /// Configures the current global memory entries to use for the kernel invocation.
-    void SetupComputeGlobalMemory(Shader* kernel);
-
-    /// Configures a global memory buffer.
-    void SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& entry, GPUVAddr gpu_addr,
-                           size_t size, BindlessSSBO* ssbo);
-
     /// Configures the current textures to use for the draw command.
     void SetupDrawTextures(const Shader* shader, size_t stage_index);
 
@@ -152,6 +131,9 @@ private:
     /// Configures images in a compute shader.
     void SetupComputeImages(const Shader* shader);
 
+    /// Syncs state to match guest's
+    void SyncState();
+
     /// Syncs the viewport and depth range to match the guest state
     void SyncViewport();
 
@@ -215,6 +197,12 @@ private:
     /// Syncs the framebuffer sRGB state to match the guest state
     void SyncFramebufferSRGB();
 
+    /// Syncs vertex formats to match the guest state
+    void SyncVertexFormats();
+
+    /// Syncs vertex instances to match the guest state
+    void SyncVertexInstances();
+
     /// Syncs transform feedback state to match guest state
     /// @note Only valid on assembly shaders
     void SyncTransformFeedback();
@@ -225,19 +213,7 @@ private:
     /// End a transform feedback
     void EndTransformFeedback();
 
-    std::size_t CalculateVertexArraysSize() const;
-
-    std::size_t CalculateIndexBufferSize() const;
-
-    /// Updates the current vertex format
-    void SetupVertexFormat();
-
-    void SetupVertexBuffer();
-    void SetupVertexInstances();
-
-    GLintptr SetupIndexBuffer();
-
-    void SetupShaders();
+    void SetupShaders(bool is_indexed);
 
     Tegra::GPU& gpu;
     Tegra::Engines::Maxwell3D& maxwell3d;
@@ -249,12 +225,12 @@ private:
     ProgramManager& program_manager;
     StateTracker& state_tracker;
 
-    OGLStreamBuffer stream_buffer;
     TextureCacheRuntime texture_cache_runtime;
     TextureCache texture_cache;
+    BufferCacheRuntime buffer_cache_runtime;
+    BufferCache buffer_cache;
     ShaderCacheOpenGL shader_cache;
     QueryCache query_cache;
-    OGLBufferCache buffer_cache;
     FenceManagerOpenGL fence_manager;
 
     VideoCommon::Shader::AsyncShaders async_shaders;
@@ -262,20 +238,8 @@ private:
     boost::container::static_vector<u32, MAX_IMAGE_VIEWS> image_view_indices;
     std::array<ImageViewId, MAX_IMAGE_VIEWS> image_view_ids;
     boost::container::static_vector<GLuint, MAX_TEXTURES> sampler_handles;
-    std::array<GLuint, MAX_TEXTURES> texture_handles;
-    std::array<GLuint, MAX_IMAGES> image_handles;
-
-    std::array<OGLBuffer, Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers>
-        transform_feedback_buffers;
-    std::bitset<Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers>
-        enabled_transform_feedback_buffers;
-
-    static constexpr std::size_t NUM_CONSTANT_BUFFERS =
-        Tegra::Engines::Maxwell3D::Regs::MaxConstBuffers *
-        Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram;
-    std::array<GLuint, NUM_CONSTANT_BUFFERS> staging_cbufs{};
-    std::size_t current_cbuf = 0;
-    OGLBuffer unified_uniform_buffer;
+    std::array<GLuint, MAX_TEXTURES> texture_handles{};
+    std::array<GLuint, MAX_IMAGES> image_handles{};
 
     /// Number of commands queued to the OpenGL driver. Resetted on flush.
     std::size_t num_queued_commands = 0;

src/video_core/renderer_opengl/gl_resource_manager.cpp

@@ -171,12 +171,6 @@ void OGLBuffer::Release() {
     handle = 0;
 }
 
-void OGLBuffer::MakeStreamCopy(std::size_t buffer_size) {
-    ASSERT_OR_EXECUTE((handle != 0 && buffer_size != 0), { return; });
-
-    glNamedBufferData(handle, buffer_size, nullptr, GL_STREAM_COPY);
-}
-
 void OGLSync::Create() {
     if (handle != 0)
         return;

src/video_core/renderer_opengl/gl_resource_manager.h

@@ -234,9 +234,6 @@ public:
     /// Deletes the internal OpenGL resource
     void Release();
 
-    // Converts the buffer into a stream copy buffer with a fixed size
-    void MakeStreamCopy(std::size_t buffer_size);
-
     GLuint handle = 0;
 };
 

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -64,7 +64,7 @@ using TextureIR = std::variant<TextureOffset, TextureDerivates, TextureArgument>
 constexpr u32 MAX_CONSTBUFFER_SCALARS = static_cast<u32>(Maxwell::MaxConstBufferSize) / sizeof(u32);
 constexpr u32 MAX_CONSTBUFFER_ELEMENTS = MAX_CONSTBUFFER_SCALARS / sizeof(u32);
 
-constexpr std::string_view CommonDeclarations = R"(#define ftoi floatBitsToInt
+constexpr std::string_view COMMON_DECLARATIONS = R"(#define ftoi floatBitsToInt
 #define ftou floatBitsToUint
 #define itof intBitsToFloat
 #define utof uintBitsToFloat
@@ -77,10 +77,6 @@ bvec2 HalfFloatNanComparison(bvec2 comparison, vec2 pair1, vec2 pair2) {{
 
 const float fswzadd_modifiers_a[] = float[4](-1.0f,  1.0f, -1.0f,  0.0f );
 const float fswzadd_modifiers_b[] = float[4](-1.0f, -1.0f,  1.0f, -1.0f );
-
-layout (std140, binding = {}) uniform vs_config {{
-    float y_direction;
-}};
 )";
 
 class ShaderWriter final {
@@ -402,13 +398,6 @@ std::string FlowStackTopName(MetaStackClass stack) {
     return fmt::format("{}_flow_stack_top", GetFlowStackPrefix(stack));
 }
 
-bool UseUnifiedUniforms(const Device& device, const ShaderIR& ir, ShaderType stage) {
-    const u32 num_ubos = static_cast<u32>(ir.GetConstantBuffers().size());
-    // We waste one UBO for emulation
-    const u32 num_available_ubos = device.GetMaxUniformBuffers(stage) - 1;
-    return num_ubos > num_available_ubos;
-}
-
 struct GenericVaryingDescription {
     std::string name;
     u8 first_element = 0;
@@ -420,9 +409,8 @@ public:
     explicit GLSLDecompiler(const Device& device_, const ShaderIR& ir_, const Registry& registry_,
                             ShaderType stage_, std::string_view identifier_,
                             std::string_view suffix_)
-        : device{device_}, ir{ir_}, registry{registry_}, stage{stage_}, identifier{identifier_},
-          suffix{suffix_}, header{ir.GetHeader()}, use_unified_uniforms{
-                                                       UseUnifiedUniforms(device_, ir_, stage_)} {
+        : device{device_}, ir{ir_}, registry{registry_}, stage{stage_},
+          identifier{identifier_}, suffix{suffix_}, header{ir.GetHeader()} {
         if (stage != ShaderType::Compute) {
             transform_feedback = BuildTransformFeedback(registry.GetGraphicsInfo());
         }
@@ -516,7 +504,8 @@ private:
         if (!identifier.empty()) {
             code.AddLine("// {}", identifier);
         }
-        code.AddLine("#version 440 {}", ir.UsesLegacyVaryings() ? "compatibility" : "core");
+        const bool use_compatibility = ir.UsesLegacyVaryings() || ir.UsesYNegate();
+        code.AddLine("#version 440 {}", use_compatibility ? "compatibility" : "core");
         code.AddLine("#extension GL_ARB_separate_shader_objects : enable");
         if (device.HasShaderBallot()) {
             code.AddLine("#extension GL_ARB_shader_ballot : require");
@@ -542,7 +531,7 @@ private:
 
         code.AddNewLine();
 
-        code.AddLine(CommonDeclarations, EmulationUniformBlockBinding);
+        code.AddLine(COMMON_DECLARATIONS);
     }
 
     void DeclareVertex() {
@@ -865,17 +854,6 @@ private:
     }
 
     void DeclareConstantBuffers() {
-        if (use_unified_uniforms) {
-            const u32 binding = device.GetBaseBindings(stage).shader_storage_buffer +
-                                static_cast<u32>(ir.GetGlobalMemory().size());
-            code.AddLine("layout (std430, binding = {}) readonly buffer UnifiedUniforms {{",
-                         binding);
-            code.AddLine("    uint cbufs[];");
-            code.AddLine("}};");
-            code.AddNewLine();
-            return;
-        }
-
         u32 binding = device.GetBaseBindings(stage).uniform_buffer;
         for (const auto& [index, info] : ir.GetConstantBuffers()) {
             const u32 num_elements = Common::DivCeil(info.GetSize(), 4 * sizeof(u32));
@@ -1081,29 +1059,17 @@ private:
 
         if (const auto cbuf = std::get_if<CbufNode>(&*node)) {
             const Node offset = cbuf->GetOffset();
-            const u32 base_unified_offset = cbuf->GetIndex() * MAX_CONSTBUFFER_SCALARS;
 
             if (const auto immediate = std::get_if<ImmediateNode>(&*offset)) {
                 // Direct access
                 const u32 offset_imm = immediate->GetValue();
                 ASSERT_MSG(offset_imm % 4 == 0, "Unaligned cbuf direct access");
-                if (use_unified_uniforms) {
-                    return {fmt::format("cbufs[{}]", base_unified_offset + offset_imm / 4),
-                            Type::Uint};
-                } else {
-                    return {fmt::format("{}[{}][{}]", GetConstBuffer(cbuf->GetIndex()),
-                                        offset_imm / (4 * 4), (offset_imm / 4) % 4),
-                            Type::Uint};
-                }
-            }
-
-            // Indirect access
-            if (use_unified_uniforms) {
-                return {fmt::format("cbufs[{} + ({} >> 2)]", base_unified_offset,
-                                    Visit(offset).AsUint()),
+                return {fmt::format("{}[{}][{}]", GetConstBuffer(cbuf->GetIndex()),
+                                    offset_imm / (4 * 4), (offset_imm / 4) % 4),
                         Type::Uint};
             }
 
+            // Indirect access
             const std::string final_offset = code.GenerateTemporary();
             code.AddLine("uint {} = {} >> 2;", final_offset, Visit(offset).AsUint());
 
@@ -2293,7 +2259,6 @@ private:
                 }
             }
         }
-
         if (header.ps.omap.depth) {
             // The depth output is always 2 registers after the last color output, and current_reg
             // already contains one past the last color register.
@@ -2337,7 +2302,8 @@ private:
     }
 
     Expression YNegate(Operation operation) {
-        return {"y_direction", Type::Float};
+        // Y_NEGATE is mapped to this uniform value
+        return {"gl_FrontMaterial.ambient.a", Type::Float};
     }
 
     template <u32 element>
@@ -2787,7 +2753,6 @@ private:
     const std::string_view identifier;
     const std::string_view suffix;
     const Header header;
-    const bool use_unified_uniforms;
     std::unordered_map<u8, VaryingTFB> transform_feedback;
 
     ShaderWriter code;
@@ -3003,8 +2968,10 @@ ShaderEntries MakeEntries(const Device& device, const ShaderIR& ir, ShaderType s
     for (std::size_t i = 0; i < std::size(clip_distances); ++i) {
         entries.clip_distances = (clip_distances[i] ? 1U : 0U) << i;
     }
+    for (const auto& buffer : entries.const_buffers) {
+        entries.enabled_uniform_buffers |= 1U << buffer.GetIndex();
+    }
     entries.shader_length = ir.GetLength();
-    entries.use_unified_uniforms = UseUnifiedUniforms(device, ir, stage);
     return entries;
 }
 

src/video_core/renderer_opengl/gl_shader_decompiler.h

@@ -55,7 +55,7 @@ struct ShaderEntries {
     std::vector<ImageEntry> images;
     std::size_t shader_length{};
     u32 clip_distances{};
-    bool use_unified_uniforms{};
+    u32 enabled_uniform_buffers{};
 };
 
 ShaderEntries MakeEntries(const Device& device, const VideoCommon::Shader::ShaderIR& ir,

src/video_core/renderer_opengl/gl_state_tracker.cpp

@@ -36,16 +36,10 @@ void SetupDirtyColorMasks(Tables& tables) {
     FillBlock(tables[1], OFF(color_mask), NUM(color_mask), ColorMasks);
 }
 
-void SetupDirtyVertexArrays(Tables& tables) {
-    static constexpr std::size_t num_array = 3;
+void SetupDirtyVertexInstances(Tables& tables) {
     static constexpr std::size_t instance_base_offset = 3;
     for (std::size_t i = 0; i < Regs::NumVertexArrays; ++i) {
         const std::size_t array_offset = OFF(vertex_array) + i * NUM(vertex_array[0]);
-        const std::size_t limit_offset = OFF(vertex_array_limit) + i * NUM(vertex_array_limit[0]);
-
-        FillBlock(tables, array_offset, num_array, VertexBuffer0 + i, VertexBuffers);
-        FillBlock(tables, limit_offset, NUM(vertex_array_limit), VertexBuffer0 + i, VertexBuffers);
-
         const std::size_t instance_array_offset = array_offset + instance_base_offset;
         tables[0][instance_array_offset] = static_cast<u8>(VertexInstance0 + i);
         tables[1][instance_array_offset] = VertexInstances;
@@ -217,11 +211,11 @@ void SetupDirtyMisc(Tables& tables) {
 StateTracker::StateTracker(Tegra::GPU& gpu) : flags{gpu.Maxwell3D().dirty.flags} {
     auto& dirty = gpu.Maxwell3D().dirty;
     auto& tables = dirty.tables;
-    SetupDirtyRenderTargets(tables);
+    SetupDirtyFlags(tables);
     SetupDirtyColorMasks(tables);
     SetupDirtyViewports(tables);
     SetupDirtyScissors(tables);
-    SetupDirtyVertexArrays(tables);
+    SetupDirtyVertexInstances(tables);
     SetupDirtyVertexFormat(tables);
     SetupDirtyShaders(tables);
     SetupDirtyPolygonModes(tables);
@@ -241,19 +235,6 @@ StateTracker::StateTracker(Tegra::GPU& gpu) : flags{gpu.Maxwell3D().dirty.flags}
     SetupDirtyClipControl(tables);
     SetupDirtyDepthClampEnabled(tables);
     SetupDirtyMisc(tables);
-
-    auto& store = dirty.on_write_stores;
-    store[VertexBuffers] = true;
-    for (std::size_t i = 0; i < Regs::NumVertexArrays; ++i) {
-        store[VertexBuffer0 + i] = true;
-    }
-}
-
-void StateTracker::InvalidateStreamBuffer() {
-    flags[Dirty::VertexBuffers] = true;
-    for (int index = Dirty::VertexBuffer0; index <= Dirty::VertexBuffer31; ++index) {
-        flags[index] = true;
-    }
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_state_tracker.h

@@ -28,10 +28,6 @@ enum : u8 {
     VertexFormat0,
     VertexFormat31 = VertexFormat0 + 31,
 
-    VertexBuffers,
-    VertexBuffer0,
-    VertexBuffer31 = VertexBuffer0 + 31,
-
     VertexInstances,
     VertexInstance0,
     VertexInstance31 = VertexInstance0 + 31,
@@ -92,8 +88,6 @@ class StateTracker {
 public:
     explicit StateTracker(Tegra::GPU& gpu);
 
-    void InvalidateStreamBuffer();
-
     void BindIndexBuffer(GLuint new_index_buffer) {
         if (index_buffer == new_index_buffer) {
             return;
@@ -110,13 +104,32 @@ public:
         glBindFramebuffer(GL_DRAW_FRAMEBUFFER, framebuffer);
     }
 
+    void ClipControl(GLenum new_origin, GLenum new_depth) {
+        if (new_origin == origin && new_depth == depth) {
+            return;
+        }
+        origin = new_origin;
+        depth = new_depth;
+        glClipControl(origin, depth);
+    }
+
+    void SetYNegate(bool new_y_negate) {
+        if (new_y_negate == y_negate) {
+            return;
+        }
+        // Y_NEGATE is mapped to gl_FrontMaterial.ambient.a
+        y_negate = new_y_negate;
+        const std::array ambient{0.0f, 0.0f, 0.0f, y_negate ? -1.0f : 1.0f};
+        glMaterialfv(GL_FRONT, GL_AMBIENT, ambient.data());
+    }
+
     void NotifyScreenDrawVertexArray() {
         flags[OpenGL::Dirty::VertexFormats] = true;
         flags[OpenGL::Dirty::VertexFormat0 + 0] = true;
         flags[OpenGL::Dirty::VertexFormat0 + 1] = true;
 
-        flags[OpenGL::Dirty::VertexBuffers] = true;
-        flags[OpenGL::Dirty::VertexBuffer0] = true;
+        flags[VideoCommon::Dirty::VertexBuffers] = true;
+        flags[VideoCommon::Dirty::VertexBuffer0] = true;
 
         flags[OpenGL::Dirty::VertexInstances] = true;
         flags[OpenGL::Dirty::VertexInstance0 + 0] = true;
@@ -202,6 +215,9 @@ private:
 
     GLuint framebuffer = 0;
     GLuint index_buffer = 0;
+    GLenum origin = GL_LOWER_LEFT;
+    GLenum depth = GL_NEGATIVE_ONE_TO_ONE;
+    bool y_negate = false;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_stream_buffer.cpp

@@ -1,70 +1,64 @@
-// Copyright 2018 Citra Emulator Project
+// Copyright 2021 yuzu Emulator Project
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include <tuple>
-#include <vector>
+#include <array>
+#include <memory>
+#include <span>
+
+#include <glad/glad.h>
 
 #include "common/alignment.h"
 #include "common/assert.h"
-#include "common/microprofile.h"
-#include "video_core/renderer_opengl/gl_device.h"
-#include "video_core/renderer_opengl/gl_state_tracker.h"
 #include "video_core/renderer_opengl/gl_stream_buffer.h"
 
-MICROPROFILE_DEFINE(OpenGL_StreamBuffer, "OpenGL", "Stream Buffer Orphaning",
-                    MP_RGB(128, 128, 192));
-
 namespace OpenGL {
 
-OGLStreamBuffer::OGLStreamBuffer(const Device& device, StateTracker& state_tracker_)
-    : state_tracker{state_tracker_} {
-    gl_buffer.Create();
-
-    static constexpr GLbitfield flags = GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT;
-    glNamedBufferStorage(gl_buffer.handle, BUFFER_SIZE, nullptr, flags);
-    mapped_ptr = static_cast<u8*>(
-        glMapNamedBufferRange(gl_buffer.handle, 0, BUFFER_SIZE, flags | GL_MAP_FLUSH_EXPLICIT_BIT));
-
-    if (device.UseAssemblyShaders() || device.HasVertexBufferUnifiedMemory()) {
-        glMakeNamedBufferResidentNV(gl_buffer.handle, GL_READ_ONLY);
-        glGetNamedBufferParameterui64vNV(gl_buffer.handle, GL_BUFFER_GPU_ADDRESS_NV, &gpu_address);
+StreamBuffer::StreamBuffer() {
+    static constexpr GLenum flags = GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT;
+    buffer.Create();
+    glObjectLabel(GL_BUFFER, buffer.handle, -1, "Stream Buffer");
+    glNamedBufferStorage(buffer.handle, STREAM_BUFFER_SIZE, nullptr, flags);
+    mapped_pointer =
+        static_cast<u8*>(glMapNamedBufferRange(buffer.handle, 0, STREAM_BUFFER_SIZE, flags));
+    for (OGLSync& sync : fences) {
+        sync.Create();
     }
 }
 
-OGLStreamBuffer::~OGLStreamBuffer() {
-    glUnmapNamedBuffer(gl_buffer.handle);
-    gl_buffer.Release();
-}
-
-std::pair<u8*, GLintptr> OGLStreamBuffer::Map(GLsizeiptr size, GLintptr alignment) {
-    ASSERT(size <= BUFFER_SIZE);
-    ASSERT(alignment <= BUFFER_SIZE);
-    mapped_size = size;
-
-    if (alignment > 0) {
-        buffer_pos = Common::AlignUp<std::size_t>(buffer_pos, alignment);
+std::pair<std::span<u8>, size_t> StreamBuffer::Request(size_t size) noexcept {
+    ASSERT(size < REGION_SIZE);
+    for (size_t region = Region(used_iterator), region_end = Region(iterator); region < region_end;
+         ++region) {
+        fences[region].Create();
     }
+    used_iterator = iterator;
 
-    if (buffer_pos + size > BUFFER_SIZE) {
-        MICROPROFILE_SCOPE(OpenGL_StreamBuffer);
-        glInvalidateBufferData(gl_buffer.handle);
-        state_tracker.InvalidateStreamBuffer();
-
-        buffer_pos = 0;
+    for (size_t region = Region(free_iterator) + 1,
+                region_end = std::min(Region(iterator + size) + 1, NUM_SYNCS);
+         region < region_end; ++region) {
+        glClientWaitSync(fences[region].handle, 0, GL_TIMEOUT_IGNORED);
+        fences[region].Release();
     }
-
-    return std::make_pair(mapped_ptr + buffer_pos, buffer_pos);
-}
-
-void OGLStreamBuffer::Unmap(GLsizeiptr size) {
-    ASSERT(size <= mapped_size);
-
-    if (size > 0) {
-        glFlushMappedNamedBufferRange(gl_buffer.handle, buffer_pos, size);
+    if (iterator + size > free_iterator) {
+        free_iterator = iterator + size;
     }
+    if (iterator + size > STREAM_BUFFER_SIZE) {
+        for (size_t region = Region(used_iterator); region < NUM_SYNCS; ++region) {
+            fences[region].Create();
+        }
+        used_iterator = 0;
+        iterator = 0;
+        free_iterator = size;
 
-    buffer_pos += size;
+        for (size_t region = 0, region_end = Region(size); region <= region_end; ++region) {
+            glClientWaitSync(fences[region].handle, 0, GL_TIMEOUT_IGNORED);
+            fences[region].Release();
+        }
+    }
+    const size_t offset = iterator;
+    iterator = Common::AlignUp(iterator + size, MAX_ALIGNMENT);
+    return {std::span(mapped_pointer + offset, size), offset};
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_stream_buffer.h

@@ -1,9 +1,12 @@
-// Copyright 2018 Citra Emulator Project
+// Copyright 2021 yuzu Emulator Project
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
 #pragma once
 
+#include <array>
+#include <memory>
+#include <span>
 #include <utility>
 
 #include <glad/glad.h>
@@ -13,48 +16,35 @@
 
 namespace OpenGL {
 
-class Device;
-class StateTracker;
+class StreamBuffer {
+    static constexpr size_t STREAM_BUFFER_SIZE = 64 * 1024 * 1024;
+    static constexpr size_t NUM_SYNCS = 16;
+    static constexpr size_t REGION_SIZE = STREAM_BUFFER_SIZE / NUM_SYNCS;
+    static constexpr size_t MAX_ALIGNMENT = 256;
+    static_assert(STREAM_BUFFER_SIZE % MAX_ALIGNMENT == 0);
+    static_assert(STREAM_BUFFER_SIZE % NUM_SYNCS == 0);
+    static_assert(REGION_SIZE % MAX_ALIGNMENT == 0);
 
-class OGLStreamBuffer : private NonCopyable {
 public:
-    explicit OGLStreamBuffer(const Device& device, StateTracker& state_tracker_);
-    ~OGLStreamBuffer();
+    explicit StreamBuffer();
 
-    /*
-     * Allocates a linear chunk of memory in the GPU buffer with at least "size" bytes
-     * and the optional alignment requirement.
-     * If the buffer is full, the whole buffer is reallocated which invalidates old chunks.
-     * The return values are the pointer to the new chunk, and the offset within the buffer.
-     * The actual used size must be specified on unmapping the chunk.
-     */
-    std::pair<u8*, GLintptr> Map(GLsizeiptr size, GLintptr alignment = 0);
+    [[nodiscard]] std::pair<std::span<u8>, size_t> Request(size_t size) noexcept;
 
-    void Unmap(GLsizeiptr size);
-
-    GLuint Handle() const {
-        return gl_buffer.handle;
-    }
-
-    u64 Address() const {
-        return gpu_address;
-    }
-
-    GLsizeiptr Size() const noexcept {
-        return BUFFER_SIZE;
+    [[nodiscard]] GLuint Handle() const noexcept {
+        return buffer.handle;
     }
 
 private:
-    static constexpr GLsizeiptr BUFFER_SIZE = 256 * 1024 * 1024;
+    [[nodiscard]] static size_t Region(size_t offset) noexcept {
+        return offset / REGION_SIZE;
+    }
 
-    StateTracker& state_tracker;
-
-    OGLBuffer gl_buffer;
-
-    GLuint64EXT gpu_address = 0;
-    GLintptr buffer_pos = 0;
-    GLsizeiptr mapped_size = 0;
-    u8* mapped_ptr = nullptr;
+    size_t iterator = 0;
+    size_t used_iterator = 0;
+    size_t free_iterator = 0;
+    u8* mapped_pointer = nullptr;
+    OGLBuffer buffer;
+    std::array<OGLSync, NUM_SYNCS> fences;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -398,9 +398,6 @@ void AttachTexture(GLuint fbo, GLenum attachment, const ImageView* image_view) {
 
 } // Anonymous namespace
 
-ImageBufferMap::ImageBufferMap(GLuint handle_, u8* map, size_t size, OGLSync* sync_)
-    : span(map, size), sync{sync_}, handle{handle_} {}
-
 ImageBufferMap::~ImageBufferMap() {
     if (sync) {
         sync->Create();
@@ -487,11 +484,11 @@ void TextureCacheRuntime::Finish() {
     glFinish();
 }
 
-ImageBufferMap TextureCacheRuntime::MapUploadBuffer(size_t size) {
+ImageBufferMap TextureCacheRuntime::UploadStagingBuffer(size_t size) {
     return upload_buffers.RequestMap(size, true);
 }
 
-ImageBufferMap TextureCacheRuntime::MapDownloadBuffer(size_t size) {
+ImageBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
     return download_buffers.RequestMap(size, false);
 }
 
@@ -553,15 +550,14 @@ void TextureCacheRuntime::BlitFramebuffer(Framebuffer* dst, Framebuffer* src,
 }
 
 void TextureCacheRuntime::AccelerateImageUpload(Image& image, const ImageBufferMap& map,
-                                                size_t buffer_offset,
                                                 std::span<const SwizzleParameters> swizzles) {
     switch (image.info.type) {
     case ImageType::e2D:
-        return util_shaders.BlockLinearUpload2D(image, map, buffer_offset, swizzles);
+        return util_shaders.BlockLinearUpload2D(image, map, swizzles);
     case ImageType::e3D:
-        return util_shaders.BlockLinearUpload3D(image, map, buffer_offset, swizzles);
+        return util_shaders.BlockLinearUpload3D(image, map, swizzles);
     case ImageType::Linear:
-        return util_shaders.PitchUpload(image, map, buffer_offset, swizzles);
+        return util_shaders.PitchUpload(image, map, swizzles);
     default:
         UNREACHABLE();
         break;
@@ -596,7 +592,11 @@ ImageBufferMap TextureCacheRuntime::StagingBuffers::RequestMap(size_t requested_
                                                                bool insert_fence) {
     const size_t index = RequestBuffer(requested_size);
     OGLSync* const sync = insert_fence ? &syncs[index] : nullptr;
-    return ImageBufferMap(buffers[index].handle, maps[index], requested_size, sync);
+    return ImageBufferMap{
+        .mapped_span = std::span(maps[index], requested_size),
+        .sync = sync,
+        .buffer = buffers[index].handle,
+    };
 }
 
 size_t TextureCacheRuntime::StagingBuffers::RequestBuffer(size_t requested_size) {
@@ -709,10 +709,10 @@ Image::Image(TextureCacheRuntime& runtime, const VideoCommon::ImageInfo& info_,
     }
 }
 
-void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
+void Image::UploadMemory(const ImageBufferMap& map,
                          std::span<const VideoCommon::BufferImageCopy> copies) {
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, map.Handle());
-    glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, buffer_offset, unswizzled_size_bytes);
+    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, map.buffer);
+    glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, map.offset, unswizzled_size_bytes);
 
     glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 
@@ -728,23 +728,23 @@ void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
             current_image_height = copy.buffer_image_height;
             glPixelStorei(GL_UNPACK_IMAGE_HEIGHT, current_image_height);
         }
-        CopyBufferToImage(copy, buffer_offset);
+        CopyBufferToImage(copy, map.offset);
     }
 }
 
-void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
+void Image::UploadMemory(const ImageBufferMap& map,
                          std::span<const VideoCommon::BufferCopy> copies) {
     for (const VideoCommon::BufferCopy& copy : copies) {
-        glCopyNamedBufferSubData(map.Handle(), buffer.handle, copy.src_offset + buffer_offset,
+        glCopyNamedBufferSubData(map.buffer, buffer.handle, copy.src_offset + map.offset,
                                  copy.dst_offset, copy.size);
     }
 }
 
-void Image::DownloadMemory(ImageBufferMap& map, size_t buffer_offset,
+void Image::DownloadMemory(ImageBufferMap& map,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
     glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT); // TODO: Move this to its own API
 
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, map.Handle());
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, map.buffer);
     glPixelStorei(GL_PACK_ALIGNMENT, 1);
 
     u32 current_row_length = std::numeric_limits<u32>::max();
@@ -759,7 +759,7 @@ void Image::DownloadMemory(ImageBufferMap& map, size_t buffer_offset,
             current_image_height = copy.buffer_image_height;
             glPixelStorei(GL_PACK_IMAGE_HEIGHT, current_image_height);
         }
-        CopyImageToBuffer(copy, buffer_offset);
+        CopyImageToBuffer(copy, map.offset);
     }
 }
 

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -31,23 +31,13 @@ using VideoCommon::NUM_RT;
 using VideoCommon::Offset2D;
 using VideoCommon::RenderTargets;
 
-class ImageBufferMap {
-public:
-    explicit ImageBufferMap(GLuint handle, u8* map, size_t size, OGLSync* sync);
+struct ImageBufferMap {
     ~ImageBufferMap();
 
-    GLuint Handle() const noexcept {
-        return handle;
-    }
-
-    std::span<u8> Span() const noexcept {
-        return span;
-    }
-
-private:
-    std::span<u8> span;
+    std::span<u8> mapped_span;
+    size_t offset = 0;
     OGLSync* sync;
-    GLuint handle;
+    GLuint buffer;
 };
 
 struct FormatProperties {
@@ -69,9 +59,9 @@ public:
 
     void Finish();
 
-    ImageBufferMap MapUploadBuffer(size_t size);
+    ImageBufferMap UploadStagingBuffer(size_t size);
 
-    ImageBufferMap MapDownloadBuffer(size_t size);
+    ImageBufferMap DownloadStagingBuffer(size_t size);
 
     void CopyImage(Image& dst, Image& src, std::span<const VideoCommon::ImageCopy> copies);
 
@@ -89,7 +79,7 @@ public:
                          Tegra::Engines::Fermi2D::Filter filter,
                          Tegra::Engines::Fermi2D::Operation operation);
 
-    void AccelerateImageUpload(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+    void AccelerateImageUpload(Image& image, const ImageBufferMap& map,
                                std::span<const VideoCommon::SwizzleParameters> swizzles);
 
     void InsertUploadMemoryBarrier();
@@ -148,14 +138,12 @@ public:
     explicit Image(TextureCacheRuntime&, const VideoCommon::ImageInfo& info, GPUVAddr gpu_addr,
                    VAddr cpu_addr);
 
-    void UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
+    void UploadMemory(const ImageBufferMap& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
-    void UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
-                      std::span<const VideoCommon::BufferCopy> copies);
+    void UploadMemory(const ImageBufferMap& map, std::span<const VideoCommon::BufferCopy> copies);
 
-    void DownloadMemory(ImageBufferMap& map, size_t buffer_offset,
-                        std::span<const VideoCommon::BufferImageCopy> copies);
+    void DownloadMemory(ImageBufferMap& map, std::span<const VideoCommon::BufferImageCopy> copies);
 
     GLuint Handle() const noexcept {
         return texture.handle;

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -29,9 +29,7 @@
 #include "video_core/textures/decoders.h"
 
 namespace OpenGL {
-
 namespace {
-
 constexpr GLint PositionLocation = 0;
 constexpr GLint TexCoordLocation = 1;
 constexpr GLint ModelViewMatrixLocation = 0;
@@ -124,7 +122,6 @@ void APIENTRY DebugHandler(GLenum source, GLenum type, GLuint id, GLenum severit
         break;
     }
 }
-
 } // Anonymous namespace
 
 RendererOpenGL::RendererOpenGL(Core::TelemetrySession& telemetry_session_,
@@ -132,7 +129,17 @@ RendererOpenGL::RendererOpenGL(Core::TelemetrySession& telemetry_session_,
                                Core::Memory::Memory& cpu_memory_, Tegra::GPU& gpu_,
                                std::unique_ptr<Core::Frontend::GraphicsContext> context_)
     : RendererBase{emu_window_, std::move(context_)}, telemetry_session{telemetry_session_},
-      emu_window{emu_window_}, cpu_memory{cpu_memory_}, gpu{gpu_}, program_manager{device} {}
+      emu_window{emu_window_}, cpu_memory{cpu_memory_}, gpu{gpu_}, state_tracker{gpu},
+      program_manager{device},
+      rasterizer(emu_window, gpu, cpu_memory, device, screen_info, program_manager, state_tracker) {
+    if (Settings::values.renderer_debug && GLAD_GL_KHR_debug) {
+        glEnable(GL_DEBUG_OUTPUT);
+        glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS);
+        glDebugMessageCallback(DebugHandler, nullptr);
+    }
+    AddTelemetryFields();
+    InitOpenGLObjects();
+}
 
 RendererOpenGL::~RendererOpenGL() = default;
 
@@ -148,7 +155,7 @@ void RendererOpenGL::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
 
     ++m_current_frame;
 
-    rasterizer->TickFrame();
+    rasterizer.TickFrame();
 
     context->SwapBuffers();
     render_window.OnFrameDisplayed();
@@ -179,7 +186,7 @@ void RendererOpenGL::LoadFBToScreenInfo(const Tegra::FramebufferConfig& framebuf
     framebuffer_crop_rect = framebuffer.crop_rect;
 
     const VAddr framebuffer_addr{framebuffer.address + framebuffer.offset};
-    if (rasterizer->AccelerateDisplay(framebuffer, framebuffer_addr, framebuffer.stride)) {
+    if (rasterizer.AccelerateDisplay(framebuffer, framebuffer_addr, framebuffer.stride)) {
         return;
     }
 
@@ -267,6 +274,7 @@ void RendererOpenGL::InitOpenGLObjects() {
     // Enable unified vertex attributes and query vertex buffer address when the driver supports it
     if (device.HasVertexBufferUnifiedMemory()) {
         glEnableClientState(GL_VERTEX_ATTRIB_ARRAY_UNIFIED_NV);
+        glEnableClientState(GL_ELEMENT_ARRAY_UNIFIED_NV);
 
         glMakeNamedBufferResidentNV(vertex_buffer.handle, GL_READ_ONLY);
         glGetNamedBufferParameterui64vNV(vertex_buffer.handle, GL_BUFFER_GPU_ADDRESS_NV,
@@ -289,14 +297,6 @@ void RendererOpenGL::AddTelemetryFields() {
     telemetry_session.AddField(user_system, "GPU_OpenGL_Version", std::string(gl_version));
 }
 
-void RendererOpenGL::CreateRasterizer() {
-    if (rasterizer) {
-        return;
-    }
-    rasterizer = std::make_unique<RasterizerOpenGL>(emu_window, gpu, cpu_memory, device,
-                                                    screen_info, program_manager, state_tracker);
-}
-
 void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
                                                  const Tegra::FramebufferConfig& framebuffer) {
     texture.width = framebuffer.width;
@@ -407,6 +407,7 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
 
     program_manager.BindHostPipeline(pipeline.handle);
 
+    state_tracker.ClipControl(GL_LOWER_LEFT, GL_ZERO_TO_ONE);
     glEnable(GL_CULL_FACE);
     if (screen_info.display_srgb) {
         glEnable(GL_FRAMEBUFFER_SRGB);
@@ -425,7 +426,6 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
     glCullFace(GL_BACK);
     glFrontFace(GL_CW);
     glColorMaski(0, GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
-    glClipControl(GL_LOWER_LEFT, GL_ZERO_TO_ONE);
     glViewportIndexedf(0, 0.0f, 0.0f, static_cast<GLfloat>(layout.width),
                        static_cast<GLfloat>(layout.height));
     glDepthRangeIndexed(0, 0.0, 0.0);
@@ -497,25 +497,4 @@ void RendererOpenGL::RenderScreenshot() {
     renderer_settings.screenshot_requested = false;
 }
 
-bool RendererOpenGL::Init() {
-    if (Settings::values.renderer_debug && GLAD_GL_KHR_debug) {
-        glEnable(GL_DEBUG_OUTPUT);
-        glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS);
-        glDebugMessageCallback(DebugHandler, nullptr);
-    }
-
-    AddTelemetryFields();
-
-    if (!GLAD_GL_VERSION_4_6) {
-        return false;
-    }
-
-    InitOpenGLObjects();
-    CreateRasterizer();
-
-    return true;
-}
-
-void RendererOpenGL::ShutDown() {}
-
 } // namespace OpenGL

src/video_core/renderer_opengl/renderer_opengl.h

@@ -10,6 +10,7 @@
 #include "common/math_util.h"
 #include "video_core/renderer_base.h"
 #include "video_core/renderer_opengl/gl_device.h"
+#include "video_core/renderer_opengl/gl_rasterizer.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
 #include "video_core/renderer_opengl/gl_state_tracker.h"
@@ -63,18 +64,18 @@ public:
                             std::unique_ptr<Core::Frontend::GraphicsContext> context_);
     ~RendererOpenGL() override;
 
-    bool Init() override;
-    void ShutDown() override;
     void SwapBuffers(const Tegra::FramebufferConfig* framebuffer) override;
 
+    VideoCore::RasterizerInterface* ReadRasterizer() override {
+        return &rasterizer;
+    }
+
 private:
     /// Initializes the OpenGL state and creates persistent objects.
     void InitOpenGLObjects();
 
     void AddTelemetryFields();
 
-    void CreateRasterizer();
-
     void ConfigureFramebufferTexture(TextureInfo& texture,
                                      const Tegra::FramebufferConfig& framebuffer);
 
@@ -98,8 +99,10 @@ private:
     Core::Memory::Memory& cpu_memory;
     Tegra::GPU& gpu;
 
-    const Device device;
-    StateTracker state_tracker{gpu};
+    Device device;
+    StateTracker state_tracker;
+    ProgramManager program_manager;
+    RasterizerOpenGL rasterizer;
 
     // OpenGL object IDs
     OGLSampler present_sampler;
@@ -115,9 +118,6 @@ private:
     /// Display information for Switch screen
     ScreenInfo screen_info;
 
-    /// Global dummy shader pipeline
-    ProgramManager program_manager;
-
     /// OpenGL framebuffer data
     std::vector<u8> gl_framebuffer_data;
 

src/video_core/renderer_opengl/util_shaders.cpp

@@ -63,7 +63,7 @@ UtilShaders::UtilShaders(ProgramManager& program_manager_)
 
 UtilShaders::~UtilShaders() = default;
 
-void UtilShaders::BlockLinearUpload2D(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+void UtilShaders::BlockLinearUpload2D(Image& image, const ImageBufferMap& map,
                                       std::span<const SwizzleParameters> swizzles) {
     static constexpr Extent3D WORKGROUP_SIZE{32, 32, 1};
     static constexpr GLuint BINDING_SWIZZLE_BUFFER = 0;
@@ -71,13 +71,13 @@ void UtilShaders::BlockLinearUpload2D(Image& image, const ImageBufferMap& map, s
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
 
     program_manager.BindHostCompute(block_linear_unswizzle_2d_program.handle);
-    glFlushMappedNamedBufferRange(map.Handle(), buffer_offset, image.guest_size_bytes);
+    glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_SWIZZLE_BUFFER, swizzle_table_buffer.handle);
 
     const GLenum store_format = StoreFormat(BytesPerBlock(image.info.format));
     for (const SwizzleParameters& swizzle : swizzles) {
         const Extent3D num_tiles = swizzle.num_tiles;
-        const size_t input_offset = swizzle.buffer_offset + buffer_offset;
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
 
         const u32 num_dispatches_x = Common::DivCeil(num_tiles.width, WORKGROUP_SIZE.width);
         const u32 num_dispatches_y = Common::DivCeil(num_tiles.height, WORKGROUP_SIZE.height);
@@ -91,8 +91,8 @@ void UtilShaders::BlockLinearUpload2D(Image& image, const ImageBufferMap& map, s
         glUniform1ui(5, params.x_shift);
         glUniform1ui(6, params.block_height);
         glUniform1ui(7, params.block_height_mask);
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.Handle(),
-                          input_offset, image.guest_size_bytes - swizzle.buffer_offset);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
+                          image.guest_size_bytes - swizzle.buffer_offset);
         glBindImageTexture(BINDING_OUTPUT_IMAGE, image.Handle(), swizzle.level, GL_TRUE, 0,
                            GL_WRITE_ONLY, store_format);
         glDispatchCompute(num_dispatches_x, num_dispatches_y, image.info.resources.layers);
@@ -100,7 +100,7 @@ void UtilShaders::BlockLinearUpload2D(Image& image, const ImageBufferMap& map, s
     program_manager.RestoreGuestCompute();
 }
 
-void UtilShaders::BlockLinearUpload3D(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+void UtilShaders::BlockLinearUpload3D(Image& image, const ImageBufferMap& map,
                                       std::span<const SwizzleParameters> swizzles) {
     static constexpr Extent3D WORKGROUP_SIZE{16, 8, 8};
 
@@ -108,14 +108,14 @@ void UtilShaders::BlockLinearUpload3D(Image& image, const ImageBufferMap& map, s
     static constexpr GLuint BINDING_INPUT_BUFFER = 1;
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
 
-    glFlushMappedNamedBufferRange(map.Handle(), buffer_offset, image.guest_size_bytes);
+    glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     program_manager.BindHostCompute(block_linear_unswizzle_3d_program.handle);
     glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_SWIZZLE_BUFFER, swizzle_table_buffer.handle);
 
     const GLenum store_format = StoreFormat(BytesPerBlock(image.info.format));
     for (const SwizzleParameters& swizzle : swizzles) {
         const Extent3D num_tiles = swizzle.num_tiles;
-        const size_t input_offset = swizzle.buffer_offset + buffer_offset;
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
 
         const u32 num_dispatches_x = Common::DivCeil(num_tiles.width, WORKGROUP_SIZE.width);
         const u32 num_dispatches_y = Common::DivCeil(num_tiles.height, WORKGROUP_SIZE.height);
@@ -132,8 +132,8 @@ void UtilShaders::BlockLinearUpload3D(Image& image, const ImageBufferMap& map, s
         glUniform1ui(7, params.block_height_mask);
         glUniform1ui(8, params.block_depth);
         glUniform1ui(9, params.block_depth_mask);
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.Handle(),
-                          input_offset, image.guest_size_bytes - swizzle.buffer_offset);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
+                          image.guest_size_bytes - swizzle.buffer_offset);
         glBindImageTexture(BINDING_OUTPUT_IMAGE, image.Handle(), swizzle.level, GL_TRUE, 0,
                            GL_WRITE_ONLY, store_format);
         glDispatchCompute(num_dispatches_x, num_dispatches_y, num_dispatches_z);
@@ -141,7 +141,7 @@ void UtilShaders::BlockLinearUpload3D(Image& image, const ImageBufferMap& map, s
     program_manager.RestoreGuestCompute();
 }
 
-void UtilShaders::PitchUpload(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+void UtilShaders::PitchUpload(Image& image, const ImageBufferMap& map,
                               std::span<const SwizzleParameters> swizzles) {
     static constexpr Extent3D WORKGROUP_SIZE{32, 32, 1};
     static constexpr GLuint BINDING_INPUT_BUFFER = 0;
@@ -159,7 +159,7 @@ void UtilShaders::PitchUpload(Image& image, const ImageBufferMap& map, size_t bu
                          "Non-power of two images are not implemented");
 
     program_manager.BindHostCompute(pitch_unswizzle_program.handle);
-    glFlushMappedNamedBufferRange(map.Handle(), buffer_offset, image.guest_size_bytes);
+    glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     glUniform2ui(LOC_ORIGIN, 0, 0);
     glUniform2i(LOC_DESTINATION, 0, 0);
     glUniform1ui(LOC_BYTES_PER_BLOCK, bytes_per_block);
@@ -167,13 +167,13 @@ void UtilShaders::PitchUpload(Image& image, const ImageBufferMap& map, size_t bu
     glBindImageTexture(BINDING_OUTPUT_IMAGE, image.Handle(), 0, GL_FALSE, 0, GL_WRITE_ONLY, format);
     for (const SwizzleParameters& swizzle : swizzles) {
         const Extent3D num_tiles = swizzle.num_tiles;
-        const size_t input_offset = swizzle.buffer_offset + buffer_offset;
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
 
         const u32 num_dispatches_x = Common::DivCeil(num_tiles.width, WORKGROUP_SIZE.width);
         const u32 num_dispatches_y = Common::DivCeil(num_tiles.height, WORKGROUP_SIZE.height);
 
-        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.Handle(),
-                          input_offset, image.guest_size_bytes - swizzle.buffer_offset);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
+                          image.guest_size_bytes - swizzle.buffer_offset);
         glDispatchCompute(num_dispatches_x, num_dispatches_y, 1);
     }
     program_manager.RestoreGuestCompute();

src/video_core/renderer_opengl/util_shaders.h

@@ -15,21 +15,22 @@
 namespace OpenGL {
 
 class Image;
-class ImageBufferMap;
 class ProgramManager;
 
+struct ImageBufferMap;
+
 class UtilShaders {
 public:
     explicit UtilShaders(ProgramManager& program_manager);
     ~UtilShaders();
 
-    void BlockLinearUpload2D(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+    void BlockLinearUpload2D(Image& image, const ImageBufferMap& map,
                              std::span<const VideoCommon::SwizzleParameters> swizzles);
 
-    void BlockLinearUpload3D(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+    void BlockLinearUpload3D(Image& image, const ImageBufferMap& map,
                              std::span<const VideoCommon::SwizzleParameters> swizzles);
 
-    void PitchUpload(Image& image, const ImageBufferMap& map, size_t buffer_offset,
+    void PitchUpload(Image& image, const ImageBufferMap& map,
                      std::span<const VideoCommon::SwizzleParameters> swizzles);
 
     void CopyBC4(Image& dst_image, Image& src_image,

src/video_core/renderer_vulkan/maxwell_to_vk.cpp

@@ -531,13 +531,9 @@ VkCompareOp ComparisonOp(Maxwell::ComparisonOp comparison) {
     return {};
 }
 
-VkIndexType IndexFormat(const Device& device, Maxwell::IndexFormat index_format) {
+VkIndexType IndexFormat(Maxwell::IndexFormat index_format) {
     switch (index_format) {
     case Maxwell::IndexFormat::UnsignedByte:
-        if (!device.IsExtIndexTypeUint8Supported()) {
-            UNIMPLEMENTED_MSG("Native uint8 indices are not supported on this device");
-            return VK_INDEX_TYPE_UINT16;
-        }
         return VK_INDEX_TYPE_UINT8_EXT;
     case Maxwell::IndexFormat::UnsignedShort:
         return VK_INDEX_TYPE_UINT16;

src/video_core/renderer_vulkan/maxwell_to_vk.h

@@ -53,7 +53,7 @@ VkFormat VertexFormat(Maxwell::VertexAttribute::Type type, Maxwell::VertexAttrib
 
 VkCompareOp ComparisonOp(Maxwell::ComparisonOp comparison);
 
-VkIndexType IndexFormat(const Device& device, Maxwell::IndexFormat index_format);
+VkIndexType IndexFormat(Maxwell::IndexFormat index_format);
 
 VkStencilOp StencilOp(Maxwell::StencilOp stencil_op);
 

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -80,17 +80,50 @@ std::string BuildCommaSeparatedExtensions(std::vector<std::string> available_ext
     return separated_extensions;
 }
 
+Device CreateDevice(const vk::Instance& instance, const vk::InstanceDispatch& dld,
+                    VkSurfaceKHR surface) {
+    const std::vector<VkPhysicalDevice> devices = instance.EnumeratePhysicalDevices();
+    const s32 device_index = Settings::values.vulkan_device.GetValue();
+    if (device_index < 0 || device_index >= static_cast<s32>(devices.size())) {
+        LOG_ERROR(Render_Vulkan, "Invalid device index {}!", device_index);
+        throw vk::Exception(VK_ERROR_INITIALIZATION_FAILED);
+    }
+    const vk::PhysicalDevice physical_device(devices[device_index], dld);
+    return Device(*instance, physical_device, surface, dld);
+}
 } // Anonymous namespace
 
 RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
                                Core::Frontend::EmuWindow& emu_window,
                                Core::Memory::Memory& cpu_memory_, Tegra::GPU& gpu_,
-                               std::unique_ptr<Core::Frontend::GraphicsContext> context_)
-    : RendererBase{emu_window, std::move(context_)}, telemetry_session{telemetry_session_},
-      cpu_memory{cpu_memory_}, gpu{gpu_} {}
+                               std::unique_ptr<Core::Frontend::GraphicsContext> context_) try
+    : RendererBase(emu_window, std::move(context_)),
+      telemetry_session(telemetry_session_),
+      cpu_memory(cpu_memory_),
+      gpu(gpu_),
+      library(OpenLibrary()),
+      instance(CreateInstance(library, dld, VK_API_VERSION_1_1, render_window.GetWindowInfo().type,
+                              true, Settings::values.renderer_debug)),
+      debug_callback(Settings::values.renderer_debug ? CreateDebugCallback(instance) : nullptr),
+      surface(CreateSurface(instance, render_window)),
+      device(CreateDevice(instance, dld, *surface)),
+      memory_allocator(device, false),
+      state_tracker(gpu),
+      scheduler(device, state_tracker),
+      swapchain(*surface, device, scheduler, render_window.GetFramebufferLayout().width,
+                render_window.GetFramebufferLayout().height, false),
+      blit_screen(cpu_memory, render_window, device, memory_allocator, swapchain, scheduler,
+                  screen_info),
+      rasterizer(render_window, gpu, gpu.MemoryManager(), cpu_memory, screen_info, device,
+                 memory_allocator, state_tracker, scheduler) {
+    Report();
+} catch (const vk::Exception& exception) {
+    LOG_ERROR(Render_Vulkan, "Vulkan initialization failed with error: {}", exception.what());
+    throw std::runtime_error{fmt::format("Vulkan initialization error {}", exception.what())};
+}
 
 RendererVulkan::~RendererVulkan() {
-    ShutDown();
+    void(device.GetLogical().WaitIdle());
 }
 
 void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
@@ -101,101 +134,38 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
     if (layout.width > 0 && layout.height > 0 && render_window.IsShown()) {
         const VAddr framebuffer_addr = framebuffer->address + framebuffer->offset;
         const bool use_accelerated =
-            rasterizer->AccelerateDisplay(*framebuffer, framebuffer_addr, framebuffer->stride);
+            rasterizer.AccelerateDisplay(*framebuffer, framebuffer_addr, framebuffer->stride);
         const bool is_srgb = use_accelerated && screen_info.is_srgb;
-        if (swapchain->HasFramebufferChanged(layout) || swapchain->GetSrgbState() != is_srgb) {
-            swapchain->Create(layout.width, layout.height, is_srgb);
-            blit_screen->Recreate();
+        if (swapchain.HasFramebufferChanged(layout) || swapchain.GetSrgbState() != is_srgb) {
+            swapchain.Create(layout.width, layout.height, is_srgb);
+            blit_screen.Recreate();
         }
 
-        scheduler->WaitWorker();
+        scheduler.WaitWorker();
 
-        swapchain->AcquireNextImage();
-        const VkSemaphore render_semaphore = blit_screen->Draw(*framebuffer, use_accelerated);
+        swapchain.AcquireNextImage();
+        const VkSemaphore render_semaphore = blit_screen.Draw(*framebuffer, use_accelerated);
 
-        scheduler->Flush(render_semaphore);
+        scheduler.Flush(render_semaphore);
 
-        if (swapchain->Present(render_semaphore)) {
-            blit_screen->Recreate();
+        if (swapchain.Present(render_semaphore)) {
+            blit_screen.Recreate();
         }
-
-        rasterizer->TickFrame();
+        rasterizer.TickFrame();
     }
 
     render_window.OnFrameDisplayed();
 }
 
-bool RendererVulkan::Init() try {
-    library = OpenLibrary();
-    instance = CreateInstance(library, dld, VK_API_VERSION_1_1, render_window.GetWindowInfo().type,
-                              true, Settings::values.renderer_debug);
-    if (Settings::values.renderer_debug) {
-        debug_callback = CreateDebugCallback(instance);
-    }
-    surface = CreateSurface(instance, render_window);
-
-    InitializeDevice();
-    Report();
-
-    memory_allocator = std::make_unique<MemoryAllocator>(*device);
-
-    state_tracker = std::make_unique<StateTracker>(gpu);
-
-    scheduler = std::make_unique<VKScheduler>(*device, *state_tracker);
-
-    const auto& framebuffer = render_window.GetFramebufferLayout();
-    swapchain = std::make_unique<VKSwapchain>(*surface, *device, *scheduler);
-    swapchain->Create(framebuffer.width, framebuffer.height, false);
-
-    rasterizer = std::make_unique<RasterizerVulkan>(render_window, gpu, gpu.MemoryManager(),
-                                                    cpu_memory, screen_info, *device,
-                                                    *memory_allocator, *state_tracker, *scheduler);
-
-    blit_screen =
-        std::make_unique<VKBlitScreen>(cpu_memory, render_window, *rasterizer, *device,
-                                       *memory_allocator, *swapchain, *scheduler, screen_info);
-    return true;
-
-} catch (const vk::Exception& exception) {
-    LOG_ERROR(Render_Vulkan, "Vulkan initialization failed with error: {}", exception.what());
-    return false;
-}
-
-void RendererVulkan::ShutDown() {
-    if (!device) {
-        return;
-    }
-    if (const auto& dev = device->GetLogical()) {
-        dev.WaitIdle();
-    }
-    rasterizer.reset();
-    blit_screen.reset();
-    scheduler.reset();
-    swapchain.reset();
-    memory_allocator.reset();
-    device.reset();
-}
-
-void RendererVulkan::InitializeDevice() {
-    const std::vector<VkPhysicalDevice> devices = instance.EnumeratePhysicalDevices();
-    const s32 device_index = Settings::values.vulkan_device.GetValue();
-    if (device_index < 0 || device_index >= static_cast<s32>(devices.size())) {
-        LOG_ERROR(Render_Vulkan, "Invalid device index {}!", device_index);
-        throw vk::Exception(VK_ERROR_INITIALIZATION_FAILED);
-    }
-    const vk::PhysicalDevice physical_device(devices[static_cast<size_t>(device_index)], dld);
-    device = std::make_unique<Device>(*instance, physical_device, *surface, dld);
-}
-
 void RendererVulkan::Report() const {
-    const std::string vendor_name{device->GetVendorName()};
-    const std::string model_name{device->GetModelName()};
-    const std::string driver_version = GetDriverVersion(*device);
+    const std::string vendor_name{device.GetVendorName()};
+    const std::string model_name{device.GetModelName()};
+    const std::string driver_version = GetDriverVersion(device);
     const std::string driver_name = fmt::format("{} {}", vendor_name, driver_version);
 
-    const std::string api_version = GetReadableVersion(device->ApiVersion());
+    const std::string api_version = GetReadableVersion(device.ApiVersion());
 
-    const std::string extensions = BuildCommaSeparatedExtensions(device->GetAvailableExtensions());
+    const std::string extensions = BuildCommaSeparatedExtensions(device.GetAvailableExtensions());
 
     LOG_INFO(Render_Vulkan, "Driver: {}", driver_name);
     LOG_INFO(Render_Vulkan, "Device: {}", model_name);
@@ -209,21 +179,4 @@ void RendererVulkan::Report() const {
     telemetry_session.AddField(field, "GPU_Vulkan_Extensions", extensions);
 }
 
-std::vector<std::string> RendererVulkan::EnumerateDevices() try {
-    vk::InstanceDispatch dld;
-    const Common::DynamicLibrary library = OpenLibrary();
-    const vk::Instance instance = CreateInstance(library, dld, VK_API_VERSION_1_0);
-    const std::vector<VkPhysicalDevice> physical_devices = instance.EnumeratePhysicalDevices();
-    std::vector<std::string> names;
-    names.reserve(physical_devices.size());
-    for (const VkPhysicalDevice device : physical_devices) {
-        names.push_back(vk::PhysicalDevice(device, dld).GetProperties().deviceName);
-    }
-    return names;
-
-} catch (const vk::Exception& exception) {
-    LOG_ERROR(Render_Vulkan, "Failed to enumerate devices with error: {}", exception.what());
-    return {};
-}
-
 } // namespace Vulkan

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -9,8 +9,14 @@
 #include <vector>
 
 #include "common/dynamic_library.h"
-
 #include "video_core/renderer_base.h"
+#include "video_core/renderer_vulkan/vk_blit_screen.h"
+#include "video_core/renderer_vulkan/vk_rasterizer.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_state_tracker.h"
+#include "video_core/renderer_vulkan/vk_swapchain.h"
+#include "video_core/vulkan_common/vulkan_device.h"
+#include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Core {
@@ -27,20 +33,6 @@ class GPU;
 
 namespace Vulkan {
 
-class Device;
-class StateTracker;
-class MemoryAllocator;
-class VKBlitScreen;
-class VKSwapchain;
-class VKScheduler;
-
-struct VKScreenInfo {
-    VkImageView image_view{};
-    u32 width{};
-    u32 height{};
-    bool is_srgb{};
-};
-
 class RendererVulkan final : public VideoCore::RendererBase {
 public:
     explicit RendererVulkan(Core::TelemetrySession& telemtry_session,
@@ -49,15 +41,13 @@ public:
                             std::unique_ptr<Core::Frontend::GraphicsContext> context_);
     ~RendererVulkan() override;
 
-    bool Init() override;
-    void ShutDown() override;
     void SwapBuffers(const Tegra::FramebufferConfig* framebuffer) override;
 
-    static std::vector<std::string> EnumerateDevices();
+    VideoCore::RasterizerInterface* ReadRasterizer() override {
+        return &rasterizer;
+    }
 
 private:
-    void InitializeDevice();
-
     void Report() const;
 
     Core::TelemetrySession& telemetry_session;
@@ -68,18 +58,18 @@ private:
     vk::InstanceDispatch dld;
 
     vk::Instance instance;
-
+    vk::DebugUtilsMessenger debug_callback;
     vk::SurfaceKHR surface;
 
     VKScreenInfo screen_info;
 
-    vk::DebugUtilsMessenger debug_callback;
-    std::unique_ptr<Device> device;
-    std::unique_ptr<MemoryAllocator> memory_allocator;
-    std::unique_ptr<StateTracker> state_tracker;
-    std::unique_ptr<VKScheduler> scheduler;
-    std::unique_ptr<VKSwapchain> swapchain;
-    std::unique_ptr<VKBlitScreen> blit_screen;
+    Device device;
+    MemoryAllocator memory_allocator;
+    StateTracker state_tracker;
+    VKScheduler scheduler;
+    VKSwapchain swapchain;
+    VKBlitScreen blit_screen;
+    RasterizerVulkan rasterizer;
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_blit_screen.cpp

@@ -18,7 +18,6 @@
 #include "video_core/gpu.h"
 #include "video_core/host_shaders/vulkan_present_frag_spv.h"
 #include "video_core/host_shaders/vulkan_present_vert_spv.h"
-#include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_vulkan/renderer_vulkan.h"
 #include "video_core/renderer_vulkan/vk_blit_screen.h"
 #include "video_core/renderer_vulkan/vk_master_semaphore.h"
@@ -113,13 +112,12 @@ struct VKBlitScreen::BufferData {
 };
 
 VKBlitScreen::VKBlitScreen(Core::Memory::Memory& cpu_memory_,
-                           Core::Frontend::EmuWindow& render_window_,
-                           VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
+                           Core::Frontend::EmuWindow& render_window_, const Device& device_,
                            MemoryAllocator& memory_allocator_, VKSwapchain& swapchain_,
                            VKScheduler& scheduler_, const VKScreenInfo& screen_info_)
-    : cpu_memory{cpu_memory_}, render_window{render_window_}, rasterizer{rasterizer_},
-      device{device_}, memory_allocator{memory_allocator_}, swapchain{swapchain_},
-      scheduler{scheduler_}, image_count{swapchain.GetImageCount()}, screen_info{screen_info_} {
+    : cpu_memory{cpu_memory_}, render_window{render_window_}, device{device_},
+      memory_allocator{memory_allocator_}, swapchain{swapchain_}, scheduler{scheduler_},
+      image_count{swapchain.GetImageCount()}, screen_info{screen_info_} {
     resource_ticks.resize(image_count);
 
     CreateStaticResources();
@@ -150,8 +148,8 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
     SetUniformData(data, framebuffer);
     SetVertexData(data, framebuffer);
 
-    const std::span<u8> map = buffer_commit.Map();
-    std::memcpy(map.data(), &data, sizeof(data));
+    const std::span<u8> mapped_span = buffer_commit.Map();
+    std::memcpy(mapped_span.data(), &data, sizeof(data));
 
     if (!use_accelerated) {
         const u64 image_offset = GetRawImageOffset(framebuffer, image_index);
@@ -159,14 +157,13 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
         const VAddr framebuffer_addr = framebuffer.address + framebuffer.offset;
         const u8* const host_ptr = cpu_memory.GetPointer(framebuffer_addr);
         const size_t size_bytes = GetSizeInBytes(framebuffer);
-        rasterizer.FlushRegion(ToCacheAddr(host_ptr), size_bytes);
 
         // TODO(Rodrigo): Read this from HLE
         constexpr u32 block_height_log2 = 4;
         const u32 bytes_per_pixel = GetBytesPerPixel(framebuffer);
         Tegra::Texture::UnswizzleTexture(
-            map.subspan(image_offset, size_bytes), std::span(host_ptr, size_bytes), bytes_per_pixel,
-            framebuffer.width, framebuffer.height, 1, block_height_log2, 0);
+            mapped_span.subspan(image_offset, size_bytes), std::span(host_ptr, size_bytes),
+            bytes_per_pixel, framebuffer.width, framebuffer.height, 1, block_height_log2, 0);
 
         const VkBufferImageCopy copy{
             .bufferOffset = image_offset,
@@ -266,7 +263,6 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer, bool
         cmdbuf.Draw(4, 1, 0, 0);
         cmdbuf.EndRenderPass();
     });
-
     return *semaphores[image_index];
 }
 

src/video_core/renderer_vulkan/vk_blit_screen.h

@@ -38,12 +38,18 @@ class RasterizerVulkan;
 class VKScheduler;
 class VKSwapchain;
 
-class VKBlitScreen final {
+struct VKScreenInfo {
+    VkImageView image_view{};
+    u32 width{};
+    u32 height{};
+    bool is_srgb{};
+};
+
+class VKBlitScreen {
 public:
     explicit VKBlitScreen(Core::Memory::Memory& cpu_memory,
-                          Core::Frontend::EmuWindow& render_window,
-                          VideoCore::RasterizerInterface& rasterizer, const Device& device,
-                          MemoryAllocator& memory_allocator, VKSwapchain& swapchain,
+                          Core::Frontend::EmuWindow& render_window, const Device& device,
+                          MemoryAllocator& memory_manager, VKSwapchain& swapchain,
                           VKScheduler& scheduler, const VKScreenInfo& screen_info);
     ~VKBlitScreen();
 
@@ -84,7 +90,6 @@ private:
 
     Core::Memory::Memory& cpu_memory;
     Core::Frontend::EmuWindow& render_window;
-    VideoCore::RasterizerInterface& rasterizer;
     const Device& device;
     MemoryAllocator& memory_allocator;
     VKSwapchain& swapchain;

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -3,188 +3,308 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <array>
 #include <cstring>
-#include <memory>
+#include <span>
+#include <vector>
 
-#include "core/core.h"
 #include "video_core/buffer_cache/buffer_cache.h"
+#include "video_core/renderer_vulkan/maxwell_to_vk.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
-#include "video_core/renderer_vulkan/vk_stream_buffer.h"
+#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
+#include "video_core/renderer_vulkan/vk_update_descriptor.h"
 #include "video_core/vulkan_common/vulkan_device.h"
+#include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
-
 namespace {
+VkBufferCopy MakeBufferCopy(const VideoCommon::BufferCopy& copy) {
+    return VkBufferCopy{
+        .srcOffset = copy.src_offset,
+        .dstOffset = copy.dst_offset,
+        .size = copy.size,
+    };
+}
 
-constexpr VkBufferUsageFlags BUFFER_USAGE =
-    VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
-    VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+VkIndexType IndexTypeFromNumElements(const Device& device, u32 num_elements) {
+    if (num_elements <= 0xff && device.IsExtIndexTypeUint8Supported()) {
+        return VK_INDEX_TYPE_UINT8_EXT;
+    }
+    if (num_elements <= 0xffff) {
+        return VK_INDEX_TYPE_UINT16;
+    }
+    return VK_INDEX_TYPE_UINT32;
+}
 
-constexpr VkPipelineStageFlags UPLOAD_PIPELINE_STAGE =
-    VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
-    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-    VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
-
-constexpr VkAccessFlags UPLOAD_ACCESS_BARRIERS =
-    VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_UNIFORM_READ_BIT |
-    VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_INDEX_READ_BIT;
-
-constexpr VkAccessFlags TRANSFORM_FEEDBACK_WRITE_ACCESS =
-    VK_ACCESS_TRANSFORM_FEEDBACK_WRITE_BIT_EXT | VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_WRITE_BIT_EXT;
+size_t BytesPerIndex(VkIndexType index_type) {
+    switch (index_type) {
+    case VK_INDEX_TYPE_UINT8_EXT:
+        return 1;
+    case VK_INDEX_TYPE_UINT16:
+        return 2;
+    case VK_INDEX_TYPE_UINT32:
+        return 4;
+    default:
+        UNREACHABLE_MSG("Invalid index type={}", index_type);
+        return 1;
+    }
+}
 
+template <typename T>
+std::array<T, 6> MakeQuadIndices(u32 quad, u32 first) {
+    std::array<T, 6> indices{0, 1, 2, 0, 2, 3};
+    std::ranges::transform(indices, indices.begin(),
+                           [quad, first](u32 index) { return first + index + quad * 4; });
+    return indices;
+}
 } // Anonymous namespace
 
-Buffer::Buffer(const Device& device_, MemoryAllocator& memory_allocator, VKScheduler& scheduler_,
-               StagingBufferPool& staging_pool_, VAddr cpu_addr_, std::size_t size_)
-    : BufferBlock{cpu_addr_, size_}, device{device_}, scheduler{scheduler_}, staging_pool{
-                                                                                 staging_pool_} {
-    buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
+Buffer::Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params)
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params) {}
+
+Buffer::Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rasterizer_,
+               VAddr cpu_addr_, u64 size_bytes_)
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(rasterizer_, cpu_addr_, size_bytes_) {
+    buffer = runtime.device.GetLogical().CreateBuffer(VkBufferCreateInfo{
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
-        .size = static_cast<VkDeviceSize>(size_),
-        .usage = BUFFER_USAGE | VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+        .size = SizeBytes(),
+        .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+                 VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
+                 VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
+                 VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
+                 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
         .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
         .queueFamilyIndexCount = 0,
         .pQueueFamilyIndices = nullptr,
     });
-    commit = memory_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
+    if (runtime.device.HasDebuggingToolAttached()) {
+        buffer.SetObjectNameEXT(fmt::format("Buffer 0x{:x}", CpuAddr()).c_str());
+    }
+    commit = runtime.memory_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
 }
 
-Buffer::~Buffer() = default;
+BufferCacheRuntime::BufferCacheRuntime(const Device& device_, MemoryAllocator& memory_allocator_,
+                                       VKScheduler& scheduler_, StagingBufferPool& staging_pool_,
+                                       VKUpdateDescriptorQueue& update_descriptor_queue_,
+                                       VKDescriptorPool& descriptor_pool)
+    : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
+      staging_pool{staging_pool_}, update_descriptor_queue{update_descriptor_queue_},
+      uint8_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
+      quad_index_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue) {}
 
-void Buffer::Upload(std::size_t offset, std::size_t data_size, const u8* data) {
-    const auto& staging = staging_pool.Request(data_size, MemoryUsage::Upload);
-    std::memcpy(staging.mapped_span.data(), data, data_size);
+StagingBufferRef BufferCacheRuntime::UploadStagingBuffer(size_t size) {
+    return staging_pool.Request(size, MemoryUsage::Upload);
+}
 
+StagingBufferRef BufferCacheRuntime::DownloadStagingBuffer(size_t size) {
+    return staging_pool.Request(size, MemoryUsage::Download);
+}
+
+void BufferCacheRuntime::Finish() {
+    scheduler.Finish();
+}
+
+void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
+                                    std::span<const VideoCommon::BufferCopy> copies) {
+    static constexpr VkMemoryBarrier READ_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+    };
+    static constexpr VkMemoryBarrier WRITE_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+    };
+    // Measuring a popular game, this number never exceeds the specified size once data is warmed up
+    boost::container::small_vector<VkBufferCopy, 3> vk_copies(copies.size());
+    std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
     scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([src_buffer, dst_buffer, vk_copies](vk::CommandBuffer cmdbuf) {
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, READ_BARRIER);
+        cmdbuf.CopyBuffer(src_buffer, dst_buffer, vk_copies);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, WRITE_BARRIER);
+    });
+}
 
-    const VkBuffer handle = Handle();
-    scheduler.Record([staging = staging.buffer, handle, offset, data_size,
-                      &device = device](vk::CommandBuffer cmdbuf) {
-        const VkBufferMemoryBarrier read_barrier{
-            .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-            .pNext = nullptr,
-            .srcAccessMask =
-                VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT |
-                VK_ACCESS_HOST_WRITE_BIT |
-                (device.IsExtTransformFeedbackSupported() ? TRANSFORM_FEEDBACK_WRITE_ACCESS : 0),
-            .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
-            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .buffer = handle,
-            .offset = offset,
-            .size = data_size,
+void BufferCacheRuntime::BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format,
+                                         u32 base_vertex, u32 num_indices, VkBuffer buffer,
+                                         u32 offset, [[maybe_unused]] u32 size) {
+    VkIndexType vk_index_type = MaxwellToVK::IndexFormat(index_format);
+    VkDeviceSize vk_offset = offset;
+    VkBuffer vk_buffer = buffer;
+    if (topology == PrimitiveTopology::Quads) {
+        vk_index_type = VK_INDEX_TYPE_UINT32;
+        std::tie(vk_buffer, vk_offset) =
+            quad_index_pass.Assemble(index_format, num_indices, base_vertex, buffer, offset);
+    } else if (vk_index_type == VK_INDEX_TYPE_UINT8_EXT && !device.IsExtIndexTypeUint8Supported()) {
+        vk_index_type = VK_INDEX_TYPE_UINT16;
+        std::tie(vk_buffer, vk_offset) = uint8_pass.Assemble(num_indices, buffer, offset);
+    }
+    if (vk_buffer == VK_NULL_HANDLE) {
+        // Vulkan doesn't support null index buffers. Replace it with our own null buffer.
+        ReserveNullIndexBuffer();
+        vk_buffer = *null_index_buffer;
+    }
+    scheduler.Record([vk_buffer, vk_offset, vk_index_type](vk::CommandBuffer cmdbuf) {
+        cmdbuf.BindIndexBuffer(vk_buffer, vk_offset, vk_index_type);
+    });
+}
+
+void BufferCacheRuntime::BindQuadArrayIndexBuffer(u32 first, u32 count) {
+    ReserveQuadArrayLUT(first + count, true);
+
+    // The LUT has the indices 0, 1, 2, and 3 copied as an array
+    // To apply these 'first' offsets we can apply an offset based on the modulus.
+    const VkIndexType index_type = quad_array_lut_index_type;
+    const size_t sub_first_offset = static_cast<size_t>(first % 4) * (current_num_indices / 4);
+    const size_t offset = (sub_first_offset + first / 4) * 6ULL * BytesPerIndex(index_type);
+    scheduler.Record([buffer = *quad_array_lut, index_type, offset](vk::CommandBuffer cmdbuf) {
+        cmdbuf.BindIndexBuffer(buffer, offset, index_type);
+    });
+}
+
+void BufferCacheRuntime::BindVertexBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size,
+                                          u32 stride) {
+    if (device.IsExtExtendedDynamicStateSupported()) {
+        scheduler.Record([index, buffer, offset, size, stride](vk::CommandBuffer cmdbuf) {
+            const VkDeviceSize vk_offset = offset;
+            const VkDeviceSize vk_size = buffer != VK_NULL_HANDLE ? size : VK_WHOLE_SIZE;
+            const VkDeviceSize vk_stride = stride;
+            cmdbuf.BindVertexBuffers2EXT(index, 1, &buffer, &vk_offset, &vk_size, &vk_stride);
+        });
+    } else {
+        scheduler.Record([index, buffer, offset](vk::CommandBuffer cmdbuf) {
+            cmdbuf.BindVertexBuffer(index, buffer, offset);
+        });
+    }
+}
+
+void BufferCacheRuntime::BindTransformFeedbackBuffer(u32 index, VkBuffer buffer, u32 offset,
+                                                     u32 size) {
+    if (!device.IsExtTransformFeedbackSupported()) {
+        // Already logged in the rasterizer
+        return;
+    }
+    scheduler.Record([index, buffer, offset, size](vk::CommandBuffer cmdbuf) {
+        const VkDeviceSize vk_offset = offset;
+        const VkDeviceSize vk_size = size;
+        cmdbuf.BindTransformFeedbackBuffersEXT(index, 1, &buffer, &vk_offset, &vk_size);
+    });
+}
+
+void BufferCacheRuntime::BindBuffer(VkBuffer buffer, u32 offset, u32 size) {
+    update_descriptor_queue.AddBuffer(buffer, offset, size);
+}
+
+void BufferCacheRuntime::ReserveQuadArrayLUT(u32 num_indices, bool wait_for_idle) {
+    if (num_indices <= current_num_indices) {
+        return;
+    }
+    if (wait_for_idle) {
+        scheduler.Finish();
+    }
+    current_num_indices = num_indices;
+    quad_array_lut_index_type = IndexTypeFromNumElements(device, num_indices);
+
+    const u32 num_quads = num_indices / 4;
+    const u32 num_triangle_indices = num_quads * 6;
+    const u32 num_first_offset_copies = 4;
+    const size_t bytes_per_index = BytesPerIndex(quad_array_lut_index_type);
+    const size_t size_bytes = num_triangle_indices * bytes_per_index * num_first_offset_copies;
+    quad_array_lut = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .size = size_bytes,
+        .usage = VK_BUFFER_USAGE_INDEX_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    });
+    if (device.HasDebuggingToolAttached()) {
+        quad_array_lut.SetObjectNameEXT("Quad LUT");
+    }
+    quad_array_lut_commit = memory_allocator.Commit(quad_array_lut, MemoryUsage::DeviceLocal);
+
+    const StagingBufferRef staging = staging_pool.Request(size_bytes, MemoryUsage::Upload);
+    u8* staging_data = staging.mapped_span.data();
+    const size_t quad_size = bytes_per_index * 6;
+    for (u32 first = 0; first < num_first_offset_copies; ++first) {
+        for (u32 quad = 0; quad < num_quads; ++quad) {
+            switch (quad_array_lut_index_type) {
+            case VK_INDEX_TYPE_UINT8_EXT:
+                std::memcpy(staging_data, MakeQuadIndices<u8>(quad, first).data(), quad_size);
+                break;
+            case VK_INDEX_TYPE_UINT16:
+                std::memcpy(staging_data, MakeQuadIndices<u16>(quad, first).data(), quad_size);
+                break;
+            case VK_INDEX_TYPE_UINT32:
+                std::memcpy(staging_data, MakeQuadIndices<u32>(quad, first).data(), quad_size);
+                break;
+            default:
+                UNREACHABLE();
+                break;
+            }
+            staging_data += quad_size;
+        }
+    }
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([src_buffer = staging.buffer, src_offset = staging.offset,
+                      dst_buffer = *quad_array_lut, size_bytes](vk::CommandBuffer cmdbuf) {
+        const VkBufferCopy copy{
+            .srcOffset = src_offset,
+            .dstOffset = 0,
+            .size = size_bytes,
         };
         const VkBufferMemoryBarrier write_barrier{
             .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
             .pNext = nullptr,
             .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-            .dstAccessMask = UPLOAD_ACCESS_BARRIERS,
+            .dstAccessMask = VK_ACCESS_INDEX_READ_BIT,
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-            .buffer = handle,
-            .offset = offset,
-            .size = data_size,
+            .buffer = dst_buffer,
+            .offset = 0,
+            .size = size_bytes,
         };
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                               0, read_barrier);
-        cmdbuf.CopyBuffer(staging, handle, VkBufferCopy{0, offset, data_size});
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, UPLOAD_PIPELINE_STAGE, 0,
-                               write_barrier);
+        cmdbuf.CopyBuffer(src_buffer, dst_buffer, copy);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
+                               0, write_barrier);
     });
 }
 
-void Buffer::Download(std::size_t offset, std::size_t data_size, u8* data) {
-    auto staging = staging_pool.Request(data_size, MemoryUsage::Download);
-    scheduler.RequestOutsideRenderPassOperationContext();
-
-    const VkBuffer handle = Handle();
-    scheduler.Record(
-        [staging = staging.buffer, handle, offset, data_size](vk::CommandBuffer cmdbuf) {
-            const VkBufferMemoryBarrier barrier{
-                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
-                .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
-                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-                .buffer = handle,
-                .offset = offset,
-                .size = data_size,
-            };
-
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_VERTEX_SHADER_BIT |
-                                       VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                                       VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                                   VK_PIPELINE_STAGE_TRANSFER_BIT, 0, {}, barrier, {});
-            cmdbuf.CopyBuffer(handle, staging, VkBufferCopy{offset, 0, data_size});
-        });
-    scheduler.Finish();
-
-    std::memcpy(data, staging.mapped_span.data(), data_size);
-}
-
-void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
-                      std::size_t copy_size) {
-    scheduler.RequestOutsideRenderPassOperationContext();
-
-    const VkBuffer dst_buffer = Handle();
-    scheduler.Record([src_buffer = src.Handle(), dst_buffer, src_offset, dst_offset,
-                      copy_size](vk::CommandBuffer cmdbuf) {
-        cmdbuf.CopyBuffer(src_buffer, dst_buffer, VkBufferCopy{src_offset, dst_offset, copy_size});
-
-        std::array<VkBufferMemoryBarrier, 2> barriers;
-        barriers[0].sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
-        barriers[0].pNext = nullptr;
-        barriers[0].srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
-        barriers[0].dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
-        barriers[0].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barriers[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barriers[0].buffer = src_buffer;
-        barriers[0].offset = src_offset;
-        barriers[0].size = copy_size;
-        barriers[1].sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
-        barriers[1].pNext = nullptr;
-        barriers[1].srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
-        barriers[1].dstAccessMask = UPLOAD_ACCESS_BARRIERS;
-        barriers[1].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barriers[1].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barriers[1].buffer = dst_buffer;
-        barriers[1].offset = dst_offset;
-        barriers[1].size = copy_size;
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, UPLOAD_PIPELINE_STAGE, 0, {},
-                               barriers, {});
+void BufferCacheRuntime::ReserveNullIndexBuffer() {
+    if (null_index_buffer) {
+        return;
+    }
+    null_index_buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .size = 4,
+        .usage = VK_BUFFER_USAGE_INDEX_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
     });
-}
+    if (device.HasDebuggingToolAttached()) {
+        null_index_buffer.SetObjectNameEXT("Null index buffer");
+    }
+    null_index_buffer_commit = memory_allocator.Commit(null_index_buffer, MemoryUsage::DeviceLocal);
 
-VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer_,
-                             Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_,
-                             const Device& device_, MemoryAllocator& memory_allocator_,
-                             VKScheduler& scheduler_, VKStreamBuffer& stream_buffer_,
-                             StagingBufferPool& staging_pool_)
-    : VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer>{rasterizer_, gpu_memory_,
-                                                                 cpu_memory_, stream_buffer_},
-      device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
-      staging_pool{staging_pool_} {}
-
-VKBufferCache::~VKBufferCache() = default;
-
-std::shared_ptr<Buffer> VKBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
-    return std::make_shared<Buffer>(device, memory_allocator, scheduler, staging_pool, cpu_addr,
-                                    size);
-}
-
-VKBufferCache::BufferInfo VKBufferCache::GetEmptyBuffer(std::size_t size) {
-    size = std::max(size, std::size_t(4));
-    const auto& empty = staging_pool.Request(size, MemoryUsage::DeviceLocal);
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([size, buffer = empty.buffer](vk::CommandBuffer cmdbuf) {
-        cmdbuf.FillBuffer(buffer, 0, size, 0);
+    scheduler.Record([buffer = *null_index_buffer](vk::CommandBuffer cmdbuf) {
+        cmdbuf.FillBuffer(buffer, 0, VK_WHOLE_SIZE, 0);
     });
-    return {empty.buffer, 0, 0};
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -4,69 +4,124 @@
 
 #pragma once
 
-#include <memory>
-
-#include "common/common_types.h"
 #include "video_core/buffer_cache/buffer_cache.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
-#include "video_core/renderer_vulkan/vk_stream_buffer.h"
 #include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
 
 class Device;
+class VKDescriptorPool;
 class VKScheduler;
+class VKUpdateDescriptorQueue;
 
-class Buffer final : public VideoCommon::BufferBlock {
+class BufferCacheRuntime;
+
+class Buffer : public VideoCommon::BufferBase<VideoCore::RasterizerInterface> {
 public:
-    explicit Buffer(const Device& device, MemoryAllocator& memory_allocator, VKScheduler& scheduler,
-                    StagingBufferPool& staging_pool, VAddr cpu_addr_, std::size_t size_);
-    ~Buffer();
+    explicit Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params);
+    explicit Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rasterizer_,
+                    VAddr cpu_addr_, u64 size_bytes_);
 
-    void Upload(std::size_t offset, std::size_t data_size, const u8* data);
-
-    void Download(std::size_t offset, std::size_t data_size, u8* data);
-
-    void CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
-                  std::size_t copy_size);
-
-    VkBuffer Handle() const {
+    [[nodiscard]] VkBuffer Handle() const noexcept {
         return *buffer;
     }
 
-    u64 Address() const {
-        return 0;
+    operator VkBuffer() const noexcept {
+        return *buffer;
     }
 
 private:
-    const Device& device;
-    VKScheduler& scheduler;
-    StagingBufferPool& staging_pool;
-
     vk::Buffer buffer;
     MemoryCommit commit;
 };
 
-class VKBufferCache final : public VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer> {
+class BufferCacheRuntime {
+    friend Buffer;
+
+    using PrimitiveTopology = Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology;
+    using IndexFormat = Tegra::Engines::Maxwell3D::Regs::IndexFormat;
+
 public:
-    explicit VKBufferCache(VideoCore::RasterizerInterface& rasterizer,
-                           Tegra::MemoryManager& gpu_memory, Core::Memory::Memory& cpu_memory,
-                           const Device& device, MemoryAllocator& memory_allocator,
-                           VKScheduler& scheduler, VKStreamBuffer& stream_buffer,
-                           StagingBufferPool& staging_pool);
-    ~VKBufferCache();
+    explicit BufferCacheRuntime(const Device& device_, MemoryAllocator& memory_manager_,
+                                VKScheduler& scheduler_, StagingBufferPool& staging_pool_,
+                                VKUpdateDescriptorQueue& update_descriptor_queue_,
+                                VKDescriptorPool& descriptor_pool);
 
-    BufferInfo GetEmptyBuffer(std::size_t size) override;
+    void Finish();
 
-protected:
-    std::shared_ptr<Buffer> CreateBlock(VAddr cpu_addr, std::size_t size) override;
+    [[nodiscard]] StagingBufferRef UploadStagingBuffer(size_t size);
+
+    [[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size);
+
+    void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
+                    std::span<const VideoCommon::BufferCopy> copies);
+
+    void BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format, u32 num_indices,
+                         u32 base_vertex, VkBuffer buffer, u32 offset, u32 size);
+
+    void BindQuadArrayIndexBuffer(u32 first, u32 count);
+
+    void BindVertexBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size, u32 stride);
+
+    void BindTransformFeedbackBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size);
+
+    std::span<u8> BindMappedUniformBuffer([[maybe_unused]] size_t stage,
+                                          [[maybe_unused]] u32 binding_index, u32 size) {
+        const StagingBufferRef ref = staging_pool.Request(size, MemoryUsage::Upload);
+        BindBuffer(ref.buffer, static_cast<u32>(ref.offset), size);
+        return ref.mapped_span;
+    }
+
+    void BindUniformBuffer(VkBuffer buffer, u32 offset, u32 size) {
+        BindBuffer(buffer, offset, size);
+    }
+
+    void BindStorageBuffer(VkBuffer buffer, u32 offset, u32 size,
+                           [[maybe_unused]] bool is_written) {
+        BindBuffer(buffer, offset, size);
+    }
 
 private:
+    void BindBuffer(VkBuffer buffer, u32 offset, u32 size);
+
+    void ReserveQuadArrayLUT(u32 num_indices, bool wait_for_idle);
+
+    void ReserveNullIndexBuffer();
+
     const Device& device;
     MemoryAllocator& memory_allocator;
     VKScheduler& scheduler;
     StagingBufferPool& staging_pool;
+    VKUpdateDescriptorQueue& update_descriptor_queue;
+
+    vk::Buffer quad_array_lut;
+    MemoryCommit quad_array_lut_commit;
+    VkIndexType quad_array_lut_index_type{};
+    u32 current_num_indices = 0;
+
+    vk::Buffer null_index_buffer;
+    MemoryCommit null_index_buffer_commit;
+
+    Uint8Pass uint8_pass;
+    QuadIndexedPass quad_index_pass;
 };
 
+struct BufferCacheParams {
+    using Runtime = Vulkan::BufferCacheRuntime;
+    using Buffer = Vulkan::Buffer;
+
+    static constexpr bool IS_OPENGL = false;
+    static constexpr bool HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS = false;
+    static constexpr bool HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT = false;
+    static constexpr bool NEEDS_BIND_UNIFORM_INDEX = false;
+    static constexpr bool NEEDS_BIND_STORAGE_INDEX = false;
+    static constexpr bool USE_MEMORY_MAPS = true;
+};
+
+using BufferCache = VideoCommon::BufferCache<BufferCacheParams>;
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -10,7 +10,7 @@
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common_types.h"
-#include "video_core/host_shaders/vulkan_quad_array_comp_spv.h"
+#include "common/div_ceil.h"
 #include "video_core/host_shaders/vulkan_quad_indexed_comp_spv.h"
 #include "video_core/host_shaders/vulkan_uint8_comp_spv.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
@@ -22,30 +22,7 @@
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
-
 namespace {
-
-VkDescriptorSetLayoutBinding BuildQuadArrayPassDescriptorSetLayoutBinding() {
-    return {
-        .binding = 0,
-        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-        .descriptorCount = 1,
-        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
-        .pImmutableSamplers = nullptr,
-    };
-}
-
-VkDescriptorUpdateTemplateEntryKHR BuildQuadArrayPassDescriptorUpdateTemplateEntry() {
-    return {
-        .dstBinding = 0,
-        .dstArrayElement = 0,
-        .descriptorCount = 1,
-        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-        .offset = 0,
-        .stride = sizeof(DescriptorUpdateEntry),
-    };
-}
-
 VkPushConstantRange BuildComputePushConstantRange(std::size_t size) {
     return {
         .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
@@ -162,55 +139,6 @@ VkDescriptorSet VKComputePass::CommitDescriptorSet(
     return set;
 }
 
-QuadArrayPass::QuadArrayPass(const Device& device_, VKScheduler& scheduler_,
-                             VKDescriptorPool& descriptor_pool_,
-                             StagingBufferPool& staging_buffer_pool_,
-                             VKUpdateDescriptorQueue& update_descriptor_queue_)
-    : VKComputePass(device_, descriptor_pool_, BuildQuadArrayPassDescriptorSetLayoutBinding(),
-                    BuildQuadArrayPassDescriptorUpdateTemplateEntry(),
-                    BuildComputePushConstantRange(sizeof(u32)), VULKAN_QUAD_ARRAY_COMP_SPV),
-      scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
-      update_descriptor_queue{update_descriptor_queue_} {}
-
-QuadArrayPass::~QuadArrayPass() = default;
-
-std::pair<VkBuffer, VkDeviceSize> QuadArrayPass::Assemble(u32 num_vertices, u32 first) {
-    const u32 num_triangle_vertices = (num_vertices / 4) * 6;
-    const std::size_t staging_size = num_triangle_vertices * sizeof(u32);
-    const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
-
-    update_descriptor_queue.Acquire();
-    update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
-    const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
-
-    scheduler.RequestOutsideRenderPassOperationContext();
-
-    ASSERT(num_vertices % 4 == 0);
-    const u32 num_quads = num_vertices / 4;
-    scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer,
-                      num_quads, first, set](vk::CommandBuffer cmdbuf) {
-        constexpr u32 dispatch_size = 1024;
-        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
-        cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {});
-        cmdbuf.PushConstants(layout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(first), &first);
-        cmdbuf.Dispatch(Common::AlignUp(num_quads, dispatch_size) / dispatch_size, 1, 1);
-
-        VkBufferMemoryBarrier barrier;
-        barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
-        barrier.pNext = nullptr;
-        barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
-        barrier.dstAccessMask = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
-        barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.buffer = buffer;
-        barrier.offset = 0;
-        barrier.size = static_cast<VkDeviceSize>(num_quads) * 6 * sizeof(u32);
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                               VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, {barrier}, {});
-    });
-    return {staging_ref.buffer, 0};
-}
-
 Uint8Pass::Uint8Pass(const Device& device, VKScheduler& scheduler_,
                      VKDescriptorPool& descriptor_pool, StagingBufferPool& staging_buffer_pool_,
                      VKUpdateDescriptorQueue& update_descriptor_queue_)
@@ -221,38 +149,33 @@ Uint8Pass::Uint8Pass(const Device& device, VKScheduler& scheduler_,
 
 Uint8Pass::~Uint8Pass() = default;
 
-std::pair<VkBuffer, u64> Uint8Pass::Assemble(u32 num_vertices, VkBuffer src_buffer,
-                                             u64 src_offset) {
+std::pair<VkBuffer, VkDeviceSize> Uint8Pass::Assemble(u32 num_vertices, VkBuffer src_buffer,
+                                                      u32 src_offset) {
     const u32 staging_size = static_cast<u32>(num_vertices * sizeof(u16));
-    const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
+    const auto staging = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
 
     update_descriptor_queue.Acquire();
     update_descriptor_queue.AddBuffer(src_buffer, src_offset, num_vertices);
-    update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
+    update_descriptor_queue.AddBuffer(staging.buffer, staging.offset, staging_size);
     const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer, set,
+    scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging.buffer, set,
                       num_vertices](vk::CommandBuffer cmdbuf) {
-        constexpr u32 dispatch_size = 1024;
+        static constexpr u32 DISPATCH_SIZE = 1024;
+        static constexpr VkMemoryBarrier WRITE_BARRIER{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
+        };
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
         cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {});
-        cmdbuf.Dispatch(Common::AlignUp(num_vertices, dispatch_size) / dispatch_size, 1, 1);
-
-        VkBufferMemoryBarrier barrier;
-        barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
-        barrier.pNext = nullptr;
-        barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
-        barrier.dstAccessMask = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
-        barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.buffer = buffer;
-        barrier.offset = 0;
-        barrier.size = static_cast<VkDeviceSize>(num_vertices * sizeof(u16));
+        cmdbuf.Dispatch(Common::DivCeil(num_vertices, DISPATCH_SIZE), 1, 1);
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                               VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {});
+                               VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, WRITE_BARRIER);
     });
-    return {staging_ref.buffer, 0};
+    return {staging.buffer, staging.offset};
 }
 
 QuadIndexedPass::QuadIndexedPass(const Device& device_, VKScheduler& scheduler_,
@@ -267,9 +190,9 @@ QuadIndexedPass::QuadIndexedPass(const Device& device_, VKScheduler& scheduler_,
 
 QuadIndexedPass::~QuadIndexedPass() = default;
 
-std::pair<VkBuffer, u64> QuadIndexedPass::Assemble(
+std::pair<VkBuffer, VkDeviceSize> QuadIndexedPass::Assemble(
     Tegra::Engines::Maxwell3D::Regs::IndexFormat index_format, u32 num_vertices, u32 base_vertex,
-    VkBuffer src_buffer, u64 src_offset) {
+    VkBuffer src_buffer, u32 src_offset) {
     const u32 index_shift = [index_format] {
         switch (index_format) {
         case Tegra::Engines::Maxwell3D::Regs::IndexFormat::UnsignedByte:
@@ -286,38 +209,33 @@ std::pair<VkBuffer, u64> QuadIndexedPass::Assemble(
     const u32 num_tri_vertices = (num_vertices / 4) * 6;
 
     const std::size_t staging_size = num_tri_vertices * sizeof(u32);
-    const auto staging_ref = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
+    const auto staging = staging_buffer_pool.Request(staging_size, MemoryUsage::DeviceLocal);
 
     update_descriptor_queue.Acquire();
     update_descriptor_queue.AddBuffer(src_buffer, src_offset, input_size);
-    update_descriptor_queue.AddBuffer(staging_ref.buffer, 0, staging_size);
+    update_descriptor_queue.AddBuffer(staging.buffer, staging.offset, staging_size);
     const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging_ref.buffer, set,
+    scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = staging.buffer, set,
                       num_tri_vertices, base_vertex, index_shift](vk::CommandBuffer cmdbuf) {
-        static constexpr u32 dispatch_size = 1024;
+        static constexpr u32 DISPATCH_SIZE = 1024;
+        static constexpr VkMemoryBarrier WRITE_BARRIER{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
+        };
         const std::array push_constants = {base_vertex, index_shift};
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
         cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {});
         cmdbuf.PushConstants(layout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(push_constants),
                              &push_constants);
-        cmdbuf.Dispatch(Common::AlignUp(num_tri_vertices, dispatch_size) / dispatch_size, 1, 1);
-
-        VkBufferMemoryBarrier barrier;
-        barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
-        barrier.pNext = nullptr;
-        barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
-        barrier.dstAccessMask = VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT;
-        barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.buffer = buffer;
-        barrier.offset = 0;
-        barrier.size = static_cast<VkDeviceSize>(num_tri_vertices * sizeof(u32));
+        cmdbuf.Dispatch(Common::DivCeil(num_tri_vertices, DISPATCH_SIZE), 1, 1);
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                               VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, {}, barrier, {});
+                               VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, WRITE_BARRIER);
     });
-    return {staging_ref.buffer, 0};
+    return {staging.buffer, staging.offset};
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -41,22 +41,6 @@ private:
     vk::ShaderModule module;
 };
 
-class QuadArrayPass final : public VKComputePass {
-public:
-    explicit QuadArrayPass(const Device& device_, VKScheduler& scheduler_,
-                           VKDescriptorPool& descriptor_pool_,
-                           StagingBufferPool& staging_buffer_pool_,
-                           VKUpdateDescriptorQueue& update_descriptor_queue_);
-    ~QuadArrayPass();
-
-    std::pair<VkBuffer, VkDeviceSize> Assemble(u32 num_vertices, u32 first);
-
-private:
-    VKScheduler& scheduler;
-    StagingBufferPool& staging_buffer_pool;
-    VKUpdateDescriptorQueue& update_descriptor_queue;
-};
-
 class Uint8Pass final : public VKComputePass {
 public:
     explicit Uint8Pass(const Device& device_, VKScheduler& scheduler_,
@@ -64,7 +48,10 @@ public:
                        VKUpdateDescriptorQueue& update_descriptor_queue_);
     ~Uint8Pass();
 
-    std::pair<VkBuffer, u64> Assemble(u32 num_vertices, VkBuffer src_buffer, u64 src_offset);
+    /// Assemble uint8 indices into an uint16 index buffer
+    /// Returns a pair with the staging buffer, and the offset where the assembled data is
+    std::pair<VkBuffer, VkDeviceSize> Assemble(u32 num_vertices, VkBuffer src_buffer,
+                                               u32 src_offset);
 
 private:
     VKScheduler& scheduler;
@@ -80,9 +67,9 @@ public:
                              VKUpdateDescriptorQueue& update_descriptor_queue_);
     ~QuadIndexedPass();
 
-    std::pair<VkBuffer, u64> Assemble(Tegra::Engines::Maxwell3D::Regs::IndexFormat index_format,
-                                      u32 num_vertices, u32 base_vertex, VkBuffer src_buffer,
-                                      u64 src_offset);
+    std::pair<VkBuffer, VkDeviceSize> Assemble(
+        Tegra::Engines::Maxwell3D::Regs::IndexFormat index_format, u32 num_vertices,
+        u32 base_vertex, VkBuffer src_buffer, u32 src_offset);
 
 private:
     VKScheduler& scheduler;

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -45,8 +45,8 @@ void InnerFence::Wait() {
 }
 
 VKFenceManager::VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
-                               Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_,
-                               VKBufferCache& buffer_cache_, VKQueryCache& query_cache_,
+                               TextureCache& texture_cache_, BufferCache& buffer_cache_,
+                               VKQueryCache& query_cache_, const Device& device_,
                                VKScheduler& scheduler_)
     : GenericFenceManager{rasterizer_, gpu_, texture_cache_, buffer_cache_, query_cache_},
       scheduler{scheduler_} {}

src/video_core/renderer_vulkan/vk_fence_manager.h

@@ -22,7 +22,6 @@ class RasterizerInterface;
 namespace Vulkan {
 
 class Device;
-class VKBufferCache;
 class VKQueryCache;
 class VKScheduler;
 
@@ -45,14 +44,14 @@ private:
 using Fence = std::shared_ptr<InnerFence>;
 
 using GenericFenceManager =
-    VideoCommon::FenceManager<Fence, TextureCache, VKBufferCache, VKQueryCache>;
+    VideoCommon::FenceManager<Fence, TextureCache, BufferCache, VKQueryCache>;
 
 class VKFenceManager final : public GenericFenceManager {
 public:
-    explicit VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
-                            Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_,
-                            VKBufferCache& buffer_cache_, VKQueryCache& query_cache_,
-                            VKScheduler& scheduler_);
+    explicit VKFenceManager(VideoCore::RasterizerInterface& rasterizer, Tegra::GPU& gpu,
+                            TextureCache& texture_cache, BufferCache& buffer_cache,
+                            VKQueryCache& query_cache, const Device& device,
+                            VKScheduler& scheduler);
 
 protected:
     Fence CreateFence(u32 value, bool is_stubbed) override;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -8,8 +8,6 @@
 #include <mutex>
 #include <vector>
 
-#include <boost/container/static_vector.hpp>
-
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
@@ -24,7 +22,6 @@
 #include "video_core/renderer_vulkan/maxwell_to_vk.h"
 #include "video_core/renderer_vulkan/renderer_vulkan.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
-#include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_compute_pipeline.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_graphics_pipeline.h"
@@ -50,15 +47,16 @@ MICROPROFILE_DEFINE(Vulkan_WaitForWorker, "Vulkan", "Wait for worker", MP_RGB(25
 MICROPROFILE_DEFINE(Vulkan_Drawing, "Vulkan", "Record drawing", MP_RGB(192, 128, 128));
 MICROPROFILE_DEFINE(Vulkan_Compute, "Vulkan", "Record compute", MP_RGB(192, 128, 128));
 MICROPROFILE_DEFINE(Vulkan_Clearing, "Vulkan", "Record clearing", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_Geometry, "Vulkan", "Setup geometry", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_ConstBuffers, "Vulkan", "Setup constant buffers", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_GlobalBuffers, "Vulkan", "Setup global buffers", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_RenderTargets, "Vulkan", "Setup render targets", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_Textures, "Vulkan", "Setup textures", MP_RGB(192, 128, 128));
-MICROPROFILE_DEFINE(Vulkan_Images, "Vulkan", "Setup images", MP_RGB(192, 128, 128));
 MICROPROFILE_DEFINE(Vulkan_PipelineCache, "Vulkan", "Pipeline cache", MP_RGB(192, 128, 128));
 
 namespace {
+struct DrawParams {
+    u32 base_instance;
+    u32 num_instances;
+    u32 base_vertex;
+    u32 num_vertices;
+    bool is_indexed;
+};
 
 constexpr auto COMPUTE_SHADER_INDEX = static_cast<size_t>(Tegra::Engines::ShaderType::Compute);
 
@@ -67,7 +65,6 @@ VkViewport GetViewportState(const Device& device, const Maxwell& regs, size_t in
     const float width = src.scale_x * 2.0f;
     const float height = src.scale_y * 2.0f;
     const float reduce_z = regs.depth_mode == Maxwell::DepthMode::MinusOneToOne ? 1.0f : 0.0f;
-
     VkViewport viewport{
         .x = src.translate_x - src.scale_x,
         .y = src.translate_y - src.scale_y,
@@ -76,12 +73,10 @@ VkViewport GetViewportState(const Device& device, const Maxwell& regs, size_t in
         .minDepth = src.translate_z - src.scale_z * reduce_z,
         .maxDepth = src.translate_z + src.scale_z,
     };
-
     if (!device.IsExtDepthRangeUnrestrictedSupported()) {
         viewport.minDepth = std::clamp(viewport.minDepth, 0.0f, 1.0f);
         viewport.maxDepth = std::clamp(viewport.maxDepth, 0.0f, 1.0f);
     }
-
     return viewport;
 }
 
@@ -146,13 +141,6 @@ TextureHandle GetTextureInfo(const Engine& engine, bool via_header_index, const
     return TextureHandle(engine.AccessConstBuffer32(shader_type, buffer, offset), via_header_index);
 }
 
-template <size_t N>
-std::array<VkDeviceSize, N> ExpandStrides(const std::array<u16, N>& strides) {
-    std::array<VkDeviceSize, N> expanded;
-    std::copy(strides.begin(), strides.end(), expanded.begin());
-    return expanded;
-}
-
 ImageViewType ImageViewTypeFromEntry(const SamplerEntry& entry) {
     if (entry.is_buffer) {
         return ImageViewType::e2D;
@@ -221,190 +209,25 @@ void PushImageDescriptors(const ShaderEntries& entries, TextureCache& texture_ca
     }
 }
 
-} // Anonymous namespace
-
-class BufferBindings final {
-public:
-    void AddVertexBinding(VkBuffer buffer, VkDeviceSize offset, VkDeviceSize size, u32 stride) {
-        vertex.buffers[vertex.num_buffers] = buffer;
-        vertex.offsets[vertex.num_buffers] = offset;
-        vertex.sizes[vertex.num_buffers] = size;
-        vertex.strides[vertex.num_buffers] = static_cast<u16>(stride);
-        ++vertex.num_buffers;
-    }
-
-    void SetIndexBinding(VkBuffer buffer, VkDeviceSize offset, VkIndexType type) {
-        index.buffer = buffer;
-        index.offset = offset;
-        index.type = type;
-    }
-
-    void Bind(const Device& device, VKScheduler& scheduler) const {
-        // Use this large switch case to avoid dispatching more memory in the record lambda than
-        // what we need. It looks horrible, but it's the best we can do on standard C++.
-        switch (vertex.num_buffers) {
-        case 0:
-            return BindStatic<0>(device, scheduler);
-        case 1:
-            return BindStatic<1>(device, scheduler);
-        case 2:
-            return BindStatic<2>(device, scheduler);
-        case 3:
-            return BindStatic<3>(device, scheduler);
-        case 4:
-            return BindStatic<4>(device, scheduler);
-        case 5:
-            return BindStatic<5>(device, scheduler);
-        case 6:
-            return BindStatic<6>(device, scheduler);
-        case 7:
-            return BindStatic<7>(device, scheduler);
-        case 8:
-            return BindStatic<8>(device, scheduler);
-        case 9:
-            return BindStatic<9>(device, scheduler);
-        case 10:
-            return BindStatic<10>(device, scheduler);
-        case 11:
-            return BindStatic<11>(device, scheduler);
-        case 12:
-            return BindStatic<12>(device, scheduler);
-        case 13:
-            return BindStatic<13>(device, scheduler);
-        case 14:
-            return BindStatic<14>(device, scheduler);
-        case 15:
-            return BindStatic<15>(device, scheduler);
-        case 16:
-            return BindStatic<16>(device, scheduler);
-        case 17:
-            return BindStatic<17>(device, scheduler);
-        case 18:
-            return BindStatic<18>(device, scheduler);
-        case 19:
-            return BindStatic<19>(device, scheduler);
-        case 20:
-            return BindStatic<20>(device, scheduler);
-        case 21:
-            return BindStatic<21>(device, scheduler);
-        case 22:
-            return BindStatic<22>(device, scheduler);
-        case 23:
-            return BindStatic<23>(device, scheduler);
-        case 24:
-            return BindStatic<24>(device, scheduler);
-        case 25:
-            return BindStatic<25>(device, scheduler);
-        case 26:
-            return BindStatic<26>(device, scheduler);
-        case 27:
-            return BindStatic<27>(device, scheduler);
-        case 28:
-            return BindStatic<28>(device, scheduler);
-        case 29:
-            return BindStatic<29>(device, scheduler);
-        case 30:
-            return BindStatic<30>(device, scheduler);
-        case 31:
-            return BindStatic<31>(device, scheduler);
-        case 32:
-            return BindStatic<32>(device, scheduler);
-        }
-        UNREACHABLE();
-    }
-
-private:
-    // Some of these fields are intentionally left uninitialized to avoid initializing them twice.
-    struct {
-        size_t num_buffers = 0;
-        std::array<VkBuffer, Maxwell::NumVertexArrays> buffers;
-        std::array<VkDeviceSize, Maxwell::NumVertexArrays> offsets;
-        std::array<VkDeviceSize, Maxwell::NumVertexArrays> sizes;
-        std::array<u16, Maxwell::NumVertexArrays> strides;
-    } vertex;
-
-    struct {
-        VkBuffer buffer = nullptr;
-        VkDeviceSize offset;
-        VkIndexType type;
-    } index;
-
-    template <size_t N>
-    void BindStatic(const Device& device, VKScheduler& scheduler) const {
-        if (device.IsExtExtendedDynamicStateSupported()) {
-            if (index.buffer) {
-                BindStatic<N, true, true>(scheduler);
-            } else {
-                BindStatic<N, false, true>(scheduler);
-            }
-        } else {
-            if (index.buffer) {
-                BindStatic<N, true, false>(scheduler);
-            } else {
-                BindStatic<N, false, false>(scheduler);
-            }
-        }
-    }
-
-    template <size_t N, bool is_indexed, bool has_extended_dynamic_state>
-    void BindStatic(VKScheduler& scheduler) const {
-        static_assert(N <= Maxwell::NumVertexArrays);
-        if constexpr (N == 0) {
-            return;
-        }
-
-        std::array<VkBuffer, N> buffers;
-        std::array<VkDeviceSize, N> offsets;
-        std::copy(vertex.buffers.begin(), vertex.buffers.begin() + N, buffers.begin());
-        std::copy(vertex.offsets.begin(), vertex.offsets.begin() + N, offsets.begin());
-
-        if constexpr (has_extended_dynamic_state) {
-            // With extended dynamic states we can specify the length and stride of a vertex buffer
-            std::array<VkDeviceSize, N> sizes;
-            std::array<u16, N> strides;
-            std::copy(vertex.sizes.begin(), vertex.sizes.begin() + N, sizes.begin());
-            std::copy(vertex.strides.begin(), vertex.strides.begin() + N, strides.begin());
-
-            if constexpr (is_indexed) {
-                scheduler.Record(
-                    [buffers, offsets, sizes, strides, index = index](vk::CommandBuffer cmdbuf) {
-                        cmdbuf.BindIndexBuffer(index.buffer, index.offset, index.type);
-                        cmdbuf.BindVertexBuffers2EXT(0, static_cast<u32>(N), buffers.data(),
-                                                     offsets.data(), sizes.data(),
-                                                     ExpandStrides(strides).data());
-                    });
-            } else {
-                scheduler.Record([buffers, offsets, sizes, strides](vk::CommandBuffer cmdbuf) {
-                    cmdbuf.BindVertexBuffers2EXT(0, static_cast<u32>(N), buffers.data(),
-                                                 offsets.data(), sizes.data(),
-                                                 ExpandStrides(strides).data());
-                });
-            }
-            return;
-        }
-
-        if constexpr (is_indexed) {
-            // Indexed draw
-            scheduler.Record([buffers, offsets, index = index](vk::CommandBuffer cmdbuf) {
-                cmdbuf.BindIndexBuffer(index.buffer, index.offset, index.type);
-                cmdbuf.BindVertexBuffers(0, static_cast<u32>(N), buffers.data(), offsets.data());
-            });
-        } else {
-            // Array draw
-            scheduler.Record([buffers, offsets](vk::CommandBuffer cmdbuf) {
-                cmdbuf.BindVertexBuffers(0, static_cast<u32>(N), buffers.data(), offsets.data());
-            });
-        }
-    }
-};
-
-void RasterizerVulkan::DrawParameters::Draw(vk::CommandBuffer cmdbuf) const {
-    if (is_indexed) {
-        cmdbuf.DrawIndexed(num_vertices, num_instances, 0, base_vertex, base_instance);
-    } else {
-        cmdbuf.Draw(num_vertices, num_instances, base_vertex, base_instance);
+DrawParams MakeDrawParams(const Maxwell& regs, u32 num_instances, bool is_instanced,
+                          bool is_indexed) {
+    DrawParams params{
+        .base_instance = regs.vb_base_instance,
+        .num_instances = is_instanced ? num_instances : 1,
+        .base_vertex = is_indexed ? regs.vb_element_base : regs.vertex_buffer.first,
+        .num_vertices = is_indexed ? regs.index_array.count : regs.vertex_buffer.count,
+        .is_indexed = is_indexed,
+    };
+    if (regs.draw.topology == Maxwell::PrimitiveTopology::Quads) {
+        // 6 triangle vertices per quad, base vertex is part of the index
+        // See BindQuadArrayIndexBuffer for more details
+        params.num_vertices = (params.num_vertices / 4) * 6;
+        params.base_vertex = 0;
+        params.is_indexed = true;
     }
+    return params;
 }
+} // Anonymous namespace
 
 RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra::GPU& gpu_,
                                    Tegra::MemoryManager& gpu_memory_,
@@ -414,21 +237,19 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
     : RasterizerAccelerated{cpu_memory_}, gpu{gpu_},
       gpu_memory{gpu_memory_}, maxwell3d{gpu.Maxwell3D()}, kepler_compute{gpu.KeplerCompute()},
       screen_info{screen_info_}, device{device_}, memory_allocator{memory_allocator_},
-      state_tracker{state_tracker_}, scheduler{scheduler_}, stream_buffer(device, scheduler),
+      state_tracker{state_tracker_}, scheduler{scheduler_},
       staging_pool(device, memory_allocator, scheduler), descriptor_pool(device, scheduler),
       update_descriptor_queue(device, scheduler),
       blit_image(device, scheduler, state_tracker, descriptor_pool),
-      quad_array_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
-      quad_indexed_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
-      uint8_pass(device, scheduler, descriptor_pool, staging_pool, update_descriptor_queue),
       texture_cache_runtime{device, scheduler, memory_allocator, staging_pool, blit_image},
       texture_cache(texture_cache_runtime, *this, maxwell3d, kepler_compute, gpu_memory),
+      buffer_cache_runtime(device, memory_allocator, scheduler, staging_pool,
+                           update_descriptor_queue, descriptor_pool),
+      buffer_cache(*this, maxwell3d, kepler_compute, gpu_memory, cpu_memory_, buffer_cache_runtime),
       pipeline_cache(*this, gpu, maxwell3d, kepler_compute, gpu_memory, device, scheduler,
                      descriptor_pool, update_descriptor_queue),
-      buffer_cache(*this, gpu_memory, cpu_memory_, device, memory_allocator, scheduler,
-                   stream_buffer, staging_pool),
       query_cache{*this, maxwell3d, gpu_memory, device, scheduler},
-      fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, scheduler),
+      fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) {
     scheduler.SetQueryCache(query_cache);
     if (device.UseAsynchronousShaders()) {
@@ -449,22 +270,14 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
     GraphicsPipelineCacheKey key;
     key.fixed_state.Fill(maxwell3d.regs, device.IsExtExtendedDynamicStateSupported());
 
-    buffer_cache.Map(CalculateGraphicsStreamBufferSize(is_indexed));
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
 
-    BufferBindings buffer_bindings;
-    const DrawParameters draw_params =
-        SetupGeometry(key.fixed_state, buffer_bindings, is_indexed, is_instanced);
-
-    auto lock = texture_cache.AcquireLock();
     texture_cache.SynchronizeGraphicsDescriptors();
-
     texture_cache.UpdateRenderTargets(false);
 
     const auto shaders = pipeline_cache.GetShaders();
     key.shaders = GetShaderAddresses(shaders);
-    SetupShaderDescriptors(shaders);
-
-    buffer_cache.Unmap();
+    SetupShaderDescriptors(shaders, is_indexed);
 
     const Framebuffer* const framebuffer = texture_cache.GetFramebuffer();
     key.renderpass = framebuffer->RenderPass();
@@ -476,22 +289,29 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
         return;
     }
 
-    buffer_bindings.Bind(device, scheduler);
-
     BeginTransformFeedback();
 
     scheduler.RequestRenderpass(framebuffer);
     scheduler.BindGraphicsPipeline(pipeline->GetHandle());
     UpdateDynamicStates();
 
-    const auto pipeline_layout = pipeline->GetLayout();
-    const auto descriptor_set = pipeline->CommitDescriptorSet();
+    const auto& regs = maxwell3d.regs;
+    const u32 num_instances = maxwell3d.mme_draw.instance_count;
+    const DrawParams draw_params = MakeDrawParams(regs, num_instances, is_instanced, is_indexed);
+    const VkPipelineLayout pipeline_layout = pipeline->GetLayout();
+    const VkDescriptorSet descriptor_set = pipeline->CommitDescriptorSet();
     scheduler.Record([pipeline_layout, descriptor_set, draw_params](vk::CommandBuffer cmdbuf) {
         if (descriptor_set) {
             cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline_layout,
-                                      DESCRIPTOR_SET, descriptor_set, {});
+                                      DESCRIPTOR_SET, descriptor_set, nullptr);
+        }
+        if (draw_params.is_indexed) {
+            cmdbuf.DrawIndexed(draw_params.num_vertices, draw_params.num_instances, 0,
+                               draw_params.base_vertex, draw_params.base_instance);
+        } else {
+            cmdbuf.Draw(draw_params.num_vertices, draw_params.num_instances,
+                        draw_params.base_vertex, draw_params.base_instance);
         }
-        draw_params.Draw(cmdbuf);
     });
 
     EndTransformFeedback();
@@ -515,7 +335,7 @@ void RasterizerVulkan::Clear() {
         return;
     }
 
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{texture_cache.mutex};
     texture_cache.UpdateRenderTargets(true);
     const Framebuffer* const framebuffer = texture_cache.GetFramebuffer();
     const VkExtent2D render_area = framebuffer->RenderArea();
@@ -559,7 +379,6 @@ void RasterizerVulkan::Clear() {
     if (use_stencil) {
         aspect_flags |= VK_IMAGE_ASPECT_STENCIL_BIT;
     }
-
     scheduler.Record([clear_depth = regs.clear_depth, clear_stencil = regs.clear_stencil,
                       clear_rect, aspect_flags](vk::CommandBuffer cmdbuf) {
         VkClearAttachment attachment;
@@ -580,12 +399,11 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
     auto& pipeline = pipeline_cache.GetComputePipeline({
         .shader = code_addr,
         .shared_memory_size = launch_desc.shared_alloc,
-        .workgroup_size =
-            {
-                launch_desc.block_dim_x,
-                launch_desc.block_dim_y,
-                launch_desc.block_dim_z,
-            },
+        .workgroup_size{
+            launch_desc.block_dim_x,
+            launch_desc.block_dim_y,
+            launch_desc.block_dim_z,
+        },
     });
 
     // Compute dispatches can't be executed inside a renderpass
@@ -594,10 +412,21 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
     image_view_indices.clear();
     sampler_handles.clear();
 
-    auto lock = texture_cache.AcquireLock();
-    texture_cache.SynchronizeComputeDescriptors();
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
 
     const auto& entries = pipeline.GetEntries();
+    buffer_cache.SetEnabledComputeUniformBuffers(entries.enabled_uniform_buffers);
+    buffer_cache.UnbindComputeStorageBuffers();
+    u32 ssbo_index = 0;
+    for (const auto& buffer : entries.global_buffers) {
+        buffer_cache.BindComputeStorageBuffer(ssbo_index, buffer.cbuf_index, buffer.cbuf_offset,
+                                              buffer.is_written);
+        ++ssbo_index;
+    }
+    buffer_cache.UpdateComputeBuffers();
+
+    texture_cache.SynchronizeComputeDescriptors();
+
     SetupComputeUniformTexels(entries);
     SetupComputeTextures(entries);
     SetupComputeStorageTexels(entries);
@@ -606,20 +435,15 @@ void RasterizerVulkan::DispatchCompute(GPUVAddr code_addr) {
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
     texture_cache.FillComputeImageViews(indices_span, image_view_ids);
 
-    buffer_cache.Map(CalculateComputeStreamBufferSize());
-
     update_descriptor_queue.Acquire();
 
-    SetupComputeConstBuffers(entries);
-    SetupComputeGlobalBuffers(entries);
+    buffer_cache.BindHostComputeBuffers();
 
     ImageViewId* image_view_id_ptr = image_view_ids.data();
     VkSampler* sampler_ptr = sampler_handles.data();
     PushImageDescriptors(entries, texture_cache, update_descriptor_queue, image_view_id_ptr,
                          sampler_ptr);
 
-    buffer_cache.Unmap();
-
     const VkPipeline pipeline_handle = pipeline.GetHandle();
     const VkPipelineLayout pipeline_layout = pipeline.GetLayout();
     const VkDescriptorSet descriptor_set = pipeline.CommitDescriptorSet();
@@ -644,6 +468,11 @@ void RasterizerVulkan::Query(GPUVAddr gpu_addr, VideoCore::QueryType type,
     query_cache.Query(gpu_addr, type, timestamp);
 }
 
+void RasterizerVulkan::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
+                                                 u32 size) {
+    buffer_cache.BindGraphicsUniformBuffer(stage, index, gpu_addr, size);
+}
+
 void RasterizerVulkan::FlushAll() {}
 
 void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
@@ -651,19 +480,23 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
         return;
     }
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.DownloadMemory(addr, size);
     }
-    buffer_cache.FlushRegion(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.DownloadMemory(addr, size);
+    }
     query_cache.FlushRegion(addr, size);
 }
 
 bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size) {
+    std::scoped_lock lock{texture_cache.mutex, buffer_cache.mutex};
     if (!Settings::IsGPULevelHigh()) {
-        return buffer_cache.MustFlushRegion(addr, size);
+        return buffer_cache.IsRegionGpuModified(addr, size);
     }
     return texture_cache.IsRegionGpuModified(addr, size) ||
-           buffer_cache.MustFlushRegion(addr, size);
+           buffer_cache.IsRegionGpuModified(addr, size);
 }
 
 void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
@@ -671,11 +504,14 @@ void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) {
         return;
     }
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
     }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
     pipeline_cache.InvalidateRegion(addr, size);
-    buffer_cache.InvalidateRegion(addr, size);
     query_cache.InvalidateRegion(addr, size);
 }
 
@@ -683,25 +519,34 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
     if (addr == 0 || size == 0) {
         return;
     }
+    pipeline_cache.OnCPUWrite(addr, size);
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
     }
-    pipeline_cache.OnCPUWrite(addr, size);
-    buffer_cache.OnCPUWrite(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.CachedWriteMemory(addr, size);
+    }
 }
 
 void RasterizerVulkan::SyncGuestHost() {
-    buffer_cache.SyncGuestHost();
     pipeline_cache.SyncGuestHost();
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.FlushCachedWrites();
+    }
 }
 
 void RasterizerVulkan::UnmapMemory(VAddr addr, u64 size) {
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.UnmapMemory(addr, size);
     }
-    buffer_cache.OnCPUWrite(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
     pipeline_cache.OnCPUWrite(addr, size);
 }
 
@@ -774,18 +619,21 @@ void RasterizerVulkan::TickFrame() {
     draw_counter = 0;
     update_descriptor_queue.TickFrame();
     fence_manager.TickFrame();
-    buffer_cache.TickFrame();
     staging_pool.TickFrame();
     {
-        auto lock = texture_cache.AcquireLock();
+        std::scoped_lock lock{texture_cache.mutex};
         texture_cache.TickFrame();
     }
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.TickFrame();
+    }
 }
 
 bool RasterizerVulkan::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Surface& src,
                                              const Tegra::Engines::Fermi2D::Surface& dst,
                                              const Tegra::Engines::Fermi2D::Config& copy_config) {
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{texture_cache.mutex};
     texture_cache.BlitImage(dst, src, copy_config);
     return true;
 }
@@ -795,13 +643,11 @@ bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config,
     if (!framebuffer_addr) {
         return false;
     }
-
-    auto lock = texture_cache.AcquireLock();
+    std::scoped_lock lock{texture_cache.mutex};
     ImageView* const image_view = texture_cache.TryFindFramebufferImageView(framebuffer_addr);
     if (!image_view) {
         return false;
     }
-
     screen_info.image_view = image_view->Handle(VideoCommon::ImageViewType::e2D);
     screen_info.width = image_view->size.width;
     screen_info.height = image_view->size.height;
@@ -830,29 +676,8 @@ void RasterizerVulkan::FlushWork() {
     draw_counter = 0;
 }
 
-RasterizerVulkan::DrawParameters RasterizerVulkan::SetupGeometry(FixedPipelineState& fixed_state,
-                                                                 BufferBindings& buffer_bindings,
-                                                                 bool is_indexed,
-                                                                 bool is_instanced) {
-    MICROPROFILE_SCOPE(Vulkan_Geometry);
-
-    const auto& regs = maxwell3d.regs;
-
-    SetupVertexArrays(buffer_bindings);
-
-    const u32 base_instance = regs.vb_base_instance;
-    const u32 num_instances = is_instanced ? maxwell3d.mme_draw.instance_count : 1;
-    const u32 base_vertex = is_indexed ? regs.vb_element_base : regs.vertex_buffer.first;
-    const u32 num_vertices = is_indexed ? regs.index_array.count : regs.vertex_buffer.count;
-
-    DrawParameters params{base_instance, num_instances, base_vertex, num_vertices, is_indexed};
-    SetupIndexBuffer(buffer_bindings, params, is_indexed);
-
-    return params;
-}
-
 void RasterizerVulkan::SetupShaderDescriptors(
-    const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders) {
+    const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders, bool is_indexed) {
     image_view_indices.clear();
     sampler_handles.clear();
     for (size_t stage = 0; stage < Maxwell::MaxShaderStage; ++stage) {
@@ -860,15 +685,27 @@ void RasterizerVulkan::SetupShaderDescriptors(
         if (!shader) {
             continue;
         }
-        const auto& entries = shader->GetEntries();
+        const ShaderEntries& entries = shader->GetEntries();
         SetupGraphicsUniformTexels(entries, stage);
         SetupGraphicsTextures(entries, stage);
         SetupGraphicsStorageTexels(entries, stage);
         SetupGraphicsImages(entries, stage);
+
+        buffer_cache.SetEnabledUniformBuffers(stage, entries.enabled_uniform_buffers);
+        buffer_cache.UnbindGraphicsStorageBuffers(stage);
+        u32 ssbo_index = 0;
+        for (const auto& buffer : entries.global_buffers) {
+            buffer_cache.BindGraphicsStorageBuffer(stage, ssbo_index, buffer.cbuf_index,
+                                                   buffer.cbuf_offset, buffer.is_written);
+            ++ssbo_index;
+        }
     }
     const std::span indices_span(image_view_indices.data(), image_view_indices.size());
+    buffer_cache.UpdateGraphicsBuffers(is_indexed);
     texture_cache.FillGraphicsImageViews(indices_span, image_view_ids);
 
+    buffer_cache.BindHostGeometryBuffers(is_indexed);
+
     update_descriptor_queue.Acquire();
 
     ImageViewId* image_view_id_ptr = image_view_ids.data();
@@ -879,11 +716,9 @@ void RasterizerVulkan::SetupShaderDescriptors(
         if (!shader) {
             continue;
         }
-        const auto& entries = shader->GetEntries();
-        SetupGraphicsConstBuffers(entries, stage);
-        SetupGraphicsGlobalBuffers(entries, stage);
-        PushImageDescriptors(entries, texture_cache, update_descriptor_queue, image_view_id_ptr,
-                             sampler_ptr);
+        buffer_cache.BindHostStageBuffers(stage);
+        PushImageDescriptors(shader->GetEntries(), texture_cache, update_descriptor_queue,
+                             image_view_id_ptr, sampler_ptr);
     }
 }
 
@@ -916,27 +751,11 @@ void RasterizerVulkan::BeginTransformFeedback() {
         LOG_ERROR(Render_Vulkan, "Transform feedbacks used but not supported");
         return;
     }
-
     UNIMPLEMENTED_IF(regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationControl) ||
                      regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::TesselationEval) ||
                      regs.IsShaderConfigEnabled(Maxwell::ShaderProgram::Geometry));
-
-    UNIMPLEMENTED_IF(regs.tfb_bindings[1].buffer_enable);
-    UNIMPLEMENTED_IF(regs.tfb_bindings[2].buffer_enable);
-    UNIMPLEMENTED_IF(regs.tfb_bindings[3].buffer_enable);
-
-    const auto& binding = regs.tfb_bindings[0];
-    UNIMPLEMENTED_IF(binding.buffer_enable == 0);
-    UNIMPLEMENTED_IF(binding.buffer_offset != 0);
-
-    const GPUVAddr gpu_addr = binding.Address();
-    const VkDeviceSize size = static_cast<VkDeviceSize>(binding.buffer_size);
-    const auto info = buffer_cache.UploadMemory(gpu_addr, size, 4, true);
-
-    scheduler.Record([buffer = info.handle, offset = info.offset, size](vk::CommandBuffer cmdbuf) {
-        cmdbuf.BindTransformFeedbackBuffersEXT(0, 1, &buffer, &offset, &size);
-        cmdbuf.BeginTransformFeedbackEXT(0, 0, nullptr, nullptr);
-    });
+    scheduler.Record(
+        [](vk::CommandBuffer cmdbuf) { cmdbuf.BeginTransformFeedbackEXT(0, 0, nullptr, nullptr); });
 }
 
 void RasterizerVulkan::EndTransformFeedback() {
@@ -947,104 +766,11 @@ void RasterizerVulkan::EndTransformFeedback() {
     if (!device.IsExtTransformFeedbackSupported()) {
         return;
     }
-
     scheduler.Record(
         [](vk::CommandBuffer cmdbuf) { cmdbuf.EndTransformFeedbackEXT(0, 0, nullptr, nullptr); });
 }
 
-void RasterizerVulkan::SetupVertexArrays(BufferBindings& buffer_bindings) {
-    const auto& regs = maxwell3d.regs;
-
-    for (size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
-        const auto& vertex_array = regs.vertex_array[index];
-        if (!vertex_array.IsEnabled()) {
-            continue;
-        }
-        const GPUVAddr start{vertex_array.StartAddress()};
-        const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()};
-
-        ASSERT(end >= start);
-        const size_t size = end - start;
-        if (size == 0) {
-            buffer_bindings.AddVertexBinding(DefaultBuffer(), 0, DEFAULT_BUFFER_SIZE, 0);
-            continue;
-        }
-        const auto info = buffer_cache.UploadMemory(start, size);
-        buffer_bindings.AddVertexBinding(info.handle, info.offset, size, vertex_array.stride);
-    }
-}
-
-void RasterizerVulkan::SetupIndexBuffer(BufferBindings& buffer_bindings, DrawParameters& params,
-                                        bool is_indexed) {
-    if (params.num_vertices == 0) {
-        return;
-    }
-    const auto& regs = maxwell3d.regs;
-    switch (regs.draw.topology) {
-    case Maxwell::PrimitiveTopology::Quads: {
-        if (!params.is_indexed) {
-            const auto [buffer, offset] =
-                quad_array_pass.Assemble(params.num_vertices, params.base_vertex);
-            buffer_bindings.SetIndexBinding(buffer, offset, VK_INDEX_TYPE_UINT32);
-            params.base_vertex = 0;
-            params.num_vertices = params.num_vertices * 6 / 4;
-            params.is_indexed = true;
-            break;
-        }
-        const GPUVAddr gpu_addr = regs.index_array.IndexStart();
-        const auto info = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
-        VkBuffer buffer = info.handle;
-        u64 offset = info.offset;
-        std::tie(buffer, offset) = quad_indexed_pass.Assemble(
-            regs.index_array.format, params.num_vertices, params.base_vertex, buffer, offset);
-
-        buffer_bindings.SetIndexBinding(buffer, offset, VK_INDEX_TYPE_UINT32);
-        params.num_vertices = (params.num_vertices / 4) * 6;
-        params.base_vertex = 0;
-        break;
-    }
-    default: {
-        if (!is_indexed) {
-            break;
-        }
-        const GPUVAddr gpu_addr = regs.index_array.IndexStart();
-        const auto info = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
-        VkBuffer buffer = info.handle;
-        u64 offset = info.offset;
-
-        auto format = regs.index_array.format;
-        const bool is_uint8 = format == Maxwell::IndexFormat::UnsignedByte;
-        if (is_uint8 && !device.IsExtIndexTypeUint8Supported()) {
-            std::tie(buffer, offset) = uint8_pass.Assemble(params.num_vertices, buffer, offset);
-            format = Maxwell::IndexFormat::UnsignedShort;
-        }
-
-        buffer_bindings.SetIndexBinding(buffer, offset, MaxwellToVK::IndexFormat(device, format));
-        break;
-    }
-    }
-}
-
-void RasterizerVulkan::SetupGraphicsConstBuffers(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_ConstBuffers);
-    const auto& shader_stage = maxwell3d.state.shader_stages[stage];
-    for (const auto& entry : entries.const_buffers) {
-        SetupConstBuffer(entry, shader_stage.const_buffers[entry.GetIndex()]);
-    }
-}
-
-void RasterizerVulkan::SetupGraphicsGlobalBuffers(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_GlobalBuffers);
-    const auto& cbufs{maxwell3d.state.shader_stages[stage]};
-
-    for (const auto& entry : entries.global_buffers) {
-        const auto addr = cbufs.const_buffers[entry.GetCbufIndex()].address + entry.GetCbufOffset();
-        SetupGlobalBuffer(entry, addr);
-    }
-}
-
 void RasterizerVulkan::SetupGraphicsUniformTexels(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const auto& regs = maxwell3d.regs;
     const bool via_header_index = regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex;
     for (const auto& entry : entries.uniform_texels) {
@@ -1054,7 +780,6 @@ void RasterizerVulkan::SetupGraphicsUniformTexels(const ShaderEntries& entries,
 }
 
 void RasterizerVulkan::SetupGraphicsTextures(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const auto& regs = maxwell3d.regs;
     const bool via_header_index = regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex;
     for (const auto& entry : entries.samplers) {
@@ -1070,7 +795,6 @@ void RasterizerVulkan::SetupGraphicsTextures(const ShaderEntries& entries, size_
 }
 
 void RasterizerVulkan::SetupGraphicsStorageTexels(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const auto& regs = maxwell3d.regs;
     const bool via_header_index = regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex;
     for (const auto& entry : entries.storage_texels) {
@@ -1080,7 +804,6 @@ void RasterizerVulkan::SetupGraphicsStorageTexels(const ShaderEntries& entries,
 }
 
 void RasterizerVulkan::SetupGraphicsImages(const ShaderEntries& entries, size_t stage) {
-    MICROPROFILE_SCOPE(Vulkan_Images);
     const auto& regs = maxwell3d.regs;
     const bool via_header_index = regs.sampler_index == Maxwell::SamplerIndex::ViaHeaderIndex;
     for (const auto& entry : entries.images) {
@@ -1089,32 +812,7 @@ void RasterizerVulkan::SetupGraphicsImages(const ShaderEntries& entries, size_t
     }
 }
 
-void RasterizerVulkan::SetupComputeConstBuffers(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_ConstBuffers);
-    const auto& launch_desc = kepler_compute.launch_description;
-    for (const auto& entry : entries.const_buffers) {
-        const auto& config = launch_desc.const_buffer_config[entry.GetIndex()];
-        const std::bitset<8> mask = launch_desc.const_buffer_enable_mask.Value();
-        const Tegra::Engines::ConstBufferInfo info{
-            .address = config.Address(),
-            .size = config.size,
-            .enabled = mask[entry.GetIndex()],
-        };
-        SetupConstBuffer(entry, info);
-    }
-}
-
-void RasterizerVulkan::SetupComputeGlobalBuffers(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_GlobalBuffers);
-    const auto& cbufs{kepler_compute.launch_description.const_buffer_config};
-    for (const auto& entry : entries.global_buffers) {
-        const auto addr{cbufs[entry.GetCbufIndex()].Address() + entry.GetCbufOffset()};
-        SetupGlobalBuffer(entry, addr);
-    }
-}
-
 void RasterizerVulkan::SetupComputeUniformTexels(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const bool via_header_index = kepler_compute.launch_description.linked_tsc;
     for (const auto& entry : entries.uniform_texels) {
         const TextureHandle handle =
@@ -1124,7 +822,6 @@ void RasterizerVulkan::SetupComputeUniformTexels(const ShaderEntries& entries) {
 }
 
 void RasterizerVulkan::SetupComputeTextures(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const bool via_header_index = kepler_compute.launch_description.linked_tsc;
     for (const auto& entry : entries.samplers) {
         for (size_t index = 0; index < entry.size; ++index) {
@@ -1139,7 +836,6 @@ void RasterizerVulkan::SetupComputeTextures(const ShaderEntries& entries) {
 }
 
 void RasterizerVulkan::SetupComputeStorageTexels(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_Textures);
     const bool via_header_index = kepler_compute.launch_description.linked_tsc;
     for (const auto& entry : entries.storage_texels) {
         const TextureHandle handle =
@@ -1149,7 +845,6 @@ void RasterizerVulkan::SetupComputeStorageTexels(const ShaderEntries& entries) {
 }
 
 void RasterizerVulkan::SetupComputeImages(const ShaderEntries& entries) {
-    MICROPROFILE_SCOPE(Vulkan_Images);
     const bool via_header_index = kepler_compute.launch_description.linked_tsc;
     for (const auto& entry : entries.images) {
         const TextureHandle handle =
@@ -1158,42 +853,6 @@ void RasterizerVulkan::SetupComputeImages(const ShaderEntries& entries) {
     }
 }
 
-void RasterizerVulkan::SetupConstBuffer(const ConstBufferEntry& entry,
-                                        const Tegra::Engines::ConstBufferInfo& buffer) {
-    if (!buffer.enabled) {
-        // Set values to zero to unbind buffers
-        update_descriptor_queue.AddBuffer(DefaultBuffer(), 0, DEFAULT_BUFFER_SIZE);
-        return;
-    }
-    // Align the size to avoid bad std140 interactions
-    const size_t size = Common::AlignUp(CalculateConstBufferSize(entry, buffer), 4 * sizeof(float));
-    ASSERT(size <= MaxConstbufferSize);
-
-    const u64 alignment = device.GetUniformBufferAlignment();
-    const auto info = buffer_cache.UploadMemory(buffer.address, size, alignment);
-    update_descriptor_queue.AddBuffer(info.handle, info.offset, size);
-}
-
-void RasterizerVulkan::SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAddr address) {
-    const u64 actual_addr = gpu_memory.Read<u64>(address);
-    const u32 size = gpu_memory.Read<u32>(address + 8);
-
-    if (size == 0) {
-        // Sometimes global memory pointers don't have a proper size. Upload a dummy entry
-        // because Vulkan doesn't like empty buffers.
-        // Note: Do *not* use DefaultBuffer() here, storage buffers can be written breaking the
-        // default buffer.
-        static constexpr size_t dummy_size = 4;
-        const auto info = buffer_cache.GetEmptyBuffer(dummy_size);
-        update_descriptor_queue.AddBuffer(info.handle, info.offset, dummy_size);
-        return;
-    }
-
-    const auto info = buffer_cache.UploadMemory(
-        actual_addr, size, device.GetStorageBufferAlignment(), entry.IsWritten());
-    update_descriptor_queue.AddBuffer(info.handle, info.offset, size);
-}
-
 void RasterizerVulkan::UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& regs) {
     if (!state_tracker.TouchViewports()) {
         return;
@@ -1206,7 +865,8 @@ void RasterizerVulkan::UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& reg
         GetViewportState(device, regs, 8),  GetViewportState(device, regs, 9),
         GetViewportState(device, regs, 10), GetViewportState(device, regs, 11),
         GetViewportState(device, regs, 12), GetViewportState(device, regs, 13),
-        GetViewportState(device, regs, 14), GetViewportState(device, regs, 15)};
+        GetViewportState(device, regs, 14), GetViewportState(device, regs, 15),
+    };
     scheduler.Record([viewports](vk::CommandBuffer cmdbuf) { cmdbuf.SetViewport(0, viewports); });
 }
 
@@ -1214,13 +874,14 @@ void RasterizerVulkan::UpdateScissorsState(Tegra::Engines::Maxwell3D::Regs& regs
     if (!state_tracker.TouchScissors()) {
         return;
     }
-    const std::array scissors = {
+    const std::array scissors{
         GetScissorState(regs, 0),  GetScissorState(regs, 1),  GetScissorState(regs, 2),
         GetScissorState(regs, 3),  GetScissorState(regs, 4),  GetScissorState(regs, 5),
         GetScissorState(regs, 6),  GetScissorState(regs, 7),  GetScissorState(regs, 8),
         GetScissorState(regs, 9),  GetScissorState(regs, 10), GetScissorState(regs, 11),
         GetScissorState(regs, 12), GetScissorState(regs, 13), GetScissorState(regs, 14),
-        GetScissorState(regs, 15)};
+        GetScissorState(regs, 15),
+    };
     scheduler.Record([scissors](vk::CommandBuffer cmdbuf) { cmdbuf.SetScissor(0, scissors); });
 }
 
@@ -1385,73 +1046,4 @@ void RasterizerVulkan::UpdateStencilTestEnable(Tegra::Engines::Maxwell3D::Regs&
     });
 }
 
-size_t RasterizerVulkan::CalculateGraphicsStreamBufferSize(bool is_indexed) const {
-    size_t size = CalculateVertexArraysSize();
-    if (is_indexed) {
-        size = Common::AlignUp(size, 4) + CalculateIndexBufferSize();
-    }
-    size += Maxwell::MaxConstBuffers * (MaxConstbufferSize + device.GetUniformBufferAlignment());
-    return size;
-}
-
-size_t RasterizerVulkan::CalculateComputeStreamBufferSize() const {
-    return Tegra::Engines::KeplerCompute::NumConstBuffers *
-           (Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
-}
-
-size_t RasterizerVulkan::CalculateVertexArraysSize() const {
-    const auto& regs = maxwell3d.regs;
-
-    size_t size = 0;
-    for (u32 index = 0; index < Maxwell::NumVertexArrays; ++index) {
-        // This implementation assumes that all attributes are used in the shader.
-        const GPUVAddr start{regs.vertex_array[index].StartAddress()};
-        const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()};
-        DEBUG_ASSERT(end >= start);
-
-        size += (end - start) * regs.vertex_array[index].enable;
-    }
-    return size;
-}
-
-size_t RasterizerVulkan::CalculateIndexBufferSize() const {
-    return static_cast<size_t>(maxwell3d.regs.index_array.count) *
-           static_cast<size_t>(maxwell3d.regs.index_array.FormatSizeInBytes());
-}
-
-size_t RasterizerVulkan::CalculateConstBufferSize(
-    const ConstBufferEntry& entry, const Tegra::Engines::ConstBufferInfo& buffer) const {
-    if (entry.IsIndirect()) {
-        // Buffer is accessed indirectly, so upload the entire thing
-        return buffer.size;
-    } else {
-        // Buffer is accessed directly, upload just what we use
-        return entry.GetSize();
-    }
-}
-
-VkBuffer RasterizerVulkan::DefaultBuffer() {
-    if (default_buffer) {
-        return *default_buffer;
-    }
-    default_buffer = device.GetLogical().CreateBuffer({
-        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .pNext = nullptr,
-        .flags = 0,
-        .size = DEFAULT_BUFFER_SIZE,
-        .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
-                 VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
-        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
-        .queueFamilyIndexCount = 0,
-        .pQueueFamilyIndices = nullptr,
-    });
-    default_buffer_commit = memory_allocator.Commit(default_buffer, MemoryUsage::DeviceLocal);
-
-    scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([buffer = *default_buffer](vk::CommandBuffer cmdbuf) {
-        cmdbuf.FillBuffer(buffer, 0, DEFAULT_BUFFER_SIZE, 0);
-    });
-    return *default_buffer;
-}
-
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -18,14 +18,12 @@
 #include "video_core/renderer_vulkan/blit_image.h"
 #include "video_core/renderer_vulkan/fixed_pipeline_state.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
-#include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_fence_manager.h"
 #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
 #include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
-#include "video_core/renderer_vulkan/vk_stream_buffer.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/renderer_vulkan/vk_update_descriptor.h"
 #include "video_core/shader/async_shaders.h"
@@ -49,7 +47,6 @@ namespace Vulkan {
 struct VKScreenInfo;
 
 class StateTracker;
-class BufferBindings;
 
 class RasterizerVulkan final : public VideoCore::RasterizerAccelerated {
 public:
@@ -65,6 +62,7 @@ public:
     void DispatchCompute(GPUVAddr code_addr) override;
     void ResetCounter(VideoCore::QueryType type) override;
     void Query(GPUVAddr gpu_addr, VideoCore::QueryType type, std::optional<u64> timestamp) override;
+    void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
     bool MustFlushRegion(VAddr addr, u64 size) override;
@@ -107,24 +105,11 @@ private:
 
     static constexpr VkDeviceSize DEFAULT_BUFFER_SIZE = 4 * sizeof(float);
 
-    struct DrawParameters {
-        void Draw(vk::CommandBuffer cmdbuf) const;
-
-        u32 base_instance = 0;
-        u32 num_instances = 0;
-        u32 base_vertex = 0;
-        u32 num_vertices = 0;
-        bool is_indexed = 0;
-    };
-
     void FlushWork();
 
-    /// Setups geometry buffers and state.
-    DrawParameters SetupGeometry(FixedPipelineState& fixed_state, BufferBindings& buffer_bindings,
-                                 bool is_indexed, bool is_instanced);
-
     /// Setup descriptors in the graphics pipeline.
-    void SetupShaderDescriptors(const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders);
+    void SetupShaderDescriptors(const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders,
+                                bool is_indexed);
 
     void UpdateDynamicStates();
 
@@ -132,16 +117,6 @@ private:
 
     void EndTransformFeedback();
 
-    void SetupVertexArrays(BufferBindings& buffer_bindings);
-
-    void SetupIndexBuffer(BufferBindings& buffer_bindings, DrawParameters& params, bool is_indexed);
-
-    /// Setup constant buffers in the graphics pipeline.
-    void SetupGraphicsConstBuffers(const ShaderEntries& entries, std::size_t stage);
-
-    /// Setup global buffers in the graphics pipeline.
-    void SetupGraphicsGlobalBuffers(const ShaderEntries& entries, std::size_t stage);
-
     /// Setup uniform texels in the graphics pipeline.
     void SetupGraphicsUniformTexels(const ShaderEntries& entries, std::size_t stage);
 
@@ -154,12 +129,6 @@ private:
     /// Setup images in the graphics pipeline.
     void SetupGraphicsImages(const ShaderEntries& entries, std::size_t stage);
 
-    /// Setup constant buffers in the compute pipeline.
-    void SetupComputeConstBuffers(const ShaderEntries& entries);
-
-    /// Setup global buffers in the compute pipeline.
-    void SetupComputeGlobalBuffers(const ShaderEntries& entries);
-
     /// Setup texel buffers in the compute pipeline.
     void SetupComputeUniformTexels(const ShaderEntries& entries);
 
@@ -172,11 +141,6 @@ private:
     /// Setup images in the compute pipeline.
     void SetupComputeImages(const ShaderEntries& entries);
 
-    void SetupConstBuffer(const ConstBufferEntry& entry,
-                          const Tegra::Engines::ConstBufferInfo& buffer);
-
-    void SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAddr address);
-
     void UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateScissorsState(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs);
@@ -193,19 +157,6 @@ private:
     void UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateStencilTestEnable(Tegra::Engines::Maxwell3D::Regs& regs);
 
-    size_t CalculateGraphicsStreamBufferSize(bool is_indexed) const;
-
-    size_t CalculateComputeStreamBufferSize() const;
-
-    size_t CalculateVertexArraysSize() const;
-
-    size_t CalculateIndexBufferSize() const;
-
-    size_t CalculateConstBufferSize(const ConstBufferEntry& entry,
-                                    const Tegra::Engines::ConstBufferInfo& buffer) const;
-
-    VkBuffer DefaultBuffer();
-
     Tegra::GPU& gpu;
     Tegra::MemoryManager& gpu_memory;
     Tegra::Engines::Maxwell3D& maxwell3d;
@@ -217,24 +168,19 @@ private:
     StateTracker& state_tracker;
     VKScheduler& scheduler;
 
-    VKStreamBuffer stream_buffer;
     StagingBufferPool staging_pool;
     VKDescriptorPool descriptor_pool;
     VKUpdateDescriptorQueue update_descriptor_queue;
     BlitImageHelper blit_image;
-    QuadArrayPass quad_array_pass;
-    QuadIndexedPass quad_indexed_pass;
-    Uint8Pass uint8_pass;
 
     TextureCacheRuntime texture_cache_runtime;
     TextureCache texture_cache;
+    BufferCacheRuntime buffer_cache_runtime;
+    BufferCache buffer_cache;
     VKPipelineCache pipeline_cache;
-    VKBufferCache buffer_cache;
     VKQueryCache query_cache;
     VKFenceManager fence_manager;
 
-    vk::Buffer default_buffer;
-    MemoryCommit default_buffer_commit;
     vk::Event wfi_event;
     VideoCommon::Shader::AsyncShaders async_shaders;
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -52,18 +52,6 @@ VKScheduler::~VKScheduler() {
     worker_thread.join();
 }
 
-u64 VKScheduler::CurrentTick() const noexcept {
-    return master_semaphore->CurrentTick();
-}
-
-bool VKScheduler::IsFree(u64 tick) const noexcept {
-    return master_semaphore->IsFree(tick);
-}
-
-void VKScheduler::Wait(u64 tick) {
-    master_semaphore->Wait(tick);
-}
-
 void VKScheduler::Flush(VkSemaphore semaphore) {
     SubmitExecution(semaphore);
     AllocateNewContext();
@@ -269,7 +257,7 @@ void VKScheduler::EndRenderPass() {
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
                                    VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
                                    VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
-                               VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, 0, nullptr, nullptr,
+                               VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, nullptr, nullptr,
                                vk::Span(barriers.data(), num_images));
     });
     state.renderpass = nullptr;

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -14,6 +14,7 @@
 #include "common/alignment.h"
 #include "common/common_types.h"
 #include "common/threadsafe_queue.h"
+#include "video_core/renderer_vulkan/vk_master_semaphore.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
@@ -21,7 +22,6 @@ namespace Vulkan {
 class CommandPool;
 class Device;
 class Framebuffer;
-class MasterSemaphore;
 class StateTracker;
 class VKQueryCache;
 
@@ -32,15 +32,6 @@ public:
     explicit VKScheduler(const Device& device, StateTracker& state_tracker);
     ~VKScheduler();
 
-    /// Returns the current command buffer tick.
-    [[nodiscard]] u64 CurrentTick() const noexcept;
-
-    /// Returns true when a tick has been triggered by the GPU.
-    [[nodiscard]] bool IsFree(u64 tick) const noexcept;
-
-    /// Waits for the given tick to trigger on the GPU.
-    void Wait(u64 tick);
-
     /// Sends the current execution context to the GPU.
     void Flush(VkSemaphore semaphore = nullptr);
 
@@ -82,6 +73,21 @@ public:
         (void)chunk->Record(command);
     }
 
+    /// Returns the current command buffer tick.
+    [[nodiscard]] u64 CurrentTick() const noexcept {
+        return master_semaphore->CurrentTick();
+    }
+
+    /// Returns true when a tick has been triggered by the GPU.
+    [[nodiscard]] bool IsFree(u64 tick) const noexcept {
+        return master_semaphore->IsFree(tick);
+    }
+
+    /// Waits for the given tick to trigger on the GPU.
+    void Wait(u64 tick) {
+        master_semaphore->Wait(tick);
+    }
+
     /// Returns the master timeline semaphore.
     [[nodiscard]] MasterSemaphore& GetMasterSemaphore() const noexcept {
         return *master_semaphore;

src/video_core/renderer_vulkan/vk_shader_decompiler.cpp

@@ -3106,7 +3106,11 @@ ShaderEntries GenerateShaderEntries(const VideoCommon::Shader::ShaderIR& ir) {
         entries.const_buffers.emplace_back(cbuf.second, cbuf.first);
     }
     for (const auto& [base, usage] : ir.GetGlobalMemory()) {
-        entries.global_buffers.emplace_back(base.cbuf_index, base.cbuf_offset, usage.is_written);
+        entries.global_buffers.emplace_back(GlobalBufferEntry{
+            .cbuf_index = base.cbuf_index,
+            .cbuf_offset = base.cbuf_offset,
+            .is_written = usage.is_written,
+        });
     }
     for (const auto& sampler : ir.GetSamplers()) {
         if (sampler.is_buffer) {
@@ -3127,6 +3131,9 @@ ShaderEntries GenerateShaderEntries(const VideoCommon::Shader::ShaderIR& ir) {
             entries.attributes.insert(GetGenericAttributeLocation(attribute));
         }
     }
+    for (const auto& buffer : entries.const_buffers) {
+        entries.enabled_uniform_buffers |= 1U << buffer.GetIndex();
+    }
     entries.clip_distances = ir.GetClipDistances();
     entries.shader_length = ir.GetLength();
     entries.uses_warps = ir.UsesWarps();

src/video_core/renderer_vulkan/vk_shader_decompiler.h

@@ -39,24 +39,7 @@ private:
     u32 index{};
 };
 
-class GlobalBufferEntry {
-public:
-    constexpr explicit GlobalBufferEntry(u32 cbuf_index_, u32 cbuf_offset_, bool is_written_)
-        : cbuf_index{cbuf_index_}, cbuf_offset{cbuf_offset_}, is_written{is_written_} {}
-
-    constexpr u32 GetCbufIndex() const {
-        return cbuf_index;
-    }
-
-    constexpr u32 GetCbufOffset() const {
-        return cbuf_offset;
-    }
-
-    constexpr bool IsWritten() const {
-        return is_written;
-    }
-
-private:
+struct GlobalBufferEntry {
     u32 cbuf_index{};
     u32 cbuf_offset{};
     bool is_written{};
@@ -78,6 +61,7 @@ struct ShaderEntries {
     std::set<u32> attributes;
     std::array<bool, Maxwell::NumClipDistances> clip_distances{};
     std::size_t shader_length{};
+    u32 enabled_uniform_buffers{};
     bool uses_warps{};
 };
 

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -8,6 +8,7 @@
 
 #include <fmt/format.h>
 
+#include "common/alignment.h"
 #include "common/assert.h"
 #include "common/bit_util.h"
 #include "common/common_types.h"
@@ -17,18 +18,119 @@
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
+namespace {
+// Maximum potential alignment of a Vulkan buffer
+constexpr VkDeviceSize MAX_ALIGNMENT = 256;
+// Maximum size to put elements in the stream buffer
+constexpr VkDeviceSize MAX_STREAM_BUFFER_REQUEST_SIZE = 8 * 1024 * 1024;
+// Stream buffer size in bytes
+constexpr VkDeviceSize STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
+constexpr VkDeviceSize REGION_SIZE = STREAM_BUFFER_SIZE / StagingBufferPool::NUM_SYNCS;
+
+constexpr VkMemoryPropertyFlags HOST_FLAGS =
+    VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
+constexpr VkMemoryPropertyFlags STREAM_FLAGS = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT | HOST_FLAGS;
+
+bool IsStreamHeap(VkMemoryHeap heap) noexcept {
+    return STREAM_BUFFER_SIZE < (heap.size * 2) / 3;
+}
+
+std::optional<u32> FindMemoryTypeIndex(const VkPhysicalDeviceMemoryProperties& props, u32 type_mask,
+                                       VkMemoryPropertyFlags flags) noexcept {
+    for (u32 type_index = 0; type_index < props.memoryTypeCount; ++type_index) {
+        if (((type_mask >> type_index) & 1) == 0) {
+            // Memory type is incompatible
+            continue;
+        }
+        const VkMemoryType& memory_type = props.memoryTypes[type_index];
+        if ((memory_type.propertyFlags & flags) != flags) {
+            // Memory type doesn't have the flags we want
+            continue;
+        }
+        if (!IsStreamHeap(props.memoryHeaps[memory_type.heapIndex])) {
+            // Memory heap is not suitable for streaming
+            continue;
+        }
+        // Success!
+        return type_index;
+    }
+    return std::nullopt;
+}
+
+u32 FindMemoryTypeIndex(const VkPhysicalDeviceMemoryProperties& props, u32 type_mask) {
+    // Try to find a DEVICE_LOCAL_BIT type, Nvidia and AMD have a dedicated heap for this
+    std::optional<u32> type = FindMemoryTypeIndex(props, type_mask, STREAM_FLAGS);
+    if (type) {
+        return *type;
+    }
+    // Otherwise try without the DEVICE_LOCAL_BIT
+    type = FindMemoryTypeIndex(props, type_mask, HOST_FLAGS);
+    if (type) {
+        return *type;
+    }
+    // This should never happen, and in case it does, signal it as an out of memory situation
+    throw vk::Exception(VK_ERROR_OUT_OF_DEVICE_MEMORY);
+}
+
+size_t Region(size_t iterator) noexcept {
+    return iterator / REGION_SIZE;
+}
+} // Anonymous namespace
 
 StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& memory_allocator_,
                                      VKScheduler& scheduler_)
-    : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_} {}
+    : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_} {
+    const vk::Device& dev = device.GetLogical();
+    stream_buffer = dev.CreateBuffer(VkBufferCreateInfo{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .size = STREAM_BUFFER_SIZE,
+        .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
+                 VK_BUFFER_USAGE_INDEX_BUFFER_BIT,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    });
+    if (device.HasDebuggingToolAttached()) {
+        stream_buffer.SetObjectNameEXT("Stream Buffer");
+    }
+    VkMemoryDedicatedRequirements dedicated_reqs{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
+        .pNext = nullptr,
+        .prefersDedicatedAllocation = VK_FALSE,
+        .requiresDedicatedAllocation = VK_FALSE,
+    };
+    const auto requirements = dev.GetBufferMemoryRequirements(*stream_buffer, &dedicated_reqs);
+    const bool make_dedicated = dedicated_reqs.prefersDedicatedAllocation == VK_TRUE ||
+                                dedicated_reqs.requiresDedicatedAllocation == VK_TRUE;
+    const VkMemoryDedicatedAllocateInfo dedicated_info{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
+        .pNext = nullptr,
+        .image = nullptr,
+        .buffer = *stream_buffer,
+    };
+    const auto memory_properties = device.GetPhysical().GetMemoryProperties();
+    stream_memory = dev.AllocateMemory(VkMemoryAllocateInfo{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+        .pNext = make_dedicated ? &dedicated_info : nullptr,
+        .allocationSize = requirements.size,
+        .memoryTypeIndex = FindMemoryTypeIndex(memory_properties, requirements.memoryTypeBits),
+    });
+    if (device.HasDebuggingToolAttached()) {
+        stream_memory.SetObjectNameEXT("Stream Buffer Memory");
+    }
+    stream_buffer.BindMemory(*stream_memory, 0);
+    stream_pointer = stream_memory.Map(0, STREAM_BUFFER_SIZE);
+}
 
 StagingBufferPool::~StagingBufferPool() = default;
 
 StagingBufferRef StagingBufferPool::Request(size_t size, MemoryUsage usage) {
-    if (const std::optional<StagingBufferRef> ref = TryGetReservedBuffer(size, usage)) {
-        return *ref;
+    if (usage == MemoryUsage::Upload && size <= MAX_STREAM_BUFFER_REQUEST_SIZE) {
+        return GetStreamBuffer(size);
     }
-    return CreateStagingBuffer(size, usage);
+    return GetStagingBuffer(size, usage);
 }
 
 void StagingBufferPool::TickFrame() {
@@ -39,6 +141,51 @@ void StagingBufferPool::TickFrame() {
     ReleaseCache(MemoryUsage::Download);
 }
 
+StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
+    if (AreRegionsActive(Region(free_iterator) + 1,
+                         std::min(Region(iterator + size) + 1, NUM_SYNCS))) {
+        // Avoid waiting for the previous usages to be free
+        return GetStagingBuffer(size, MemoryUsage::Upload);
+    }
+    const u64 current_tick = scheduler.CurrentTick();
+    std::fill(sync_ticks.begin() + Region(used_iterator), sync_ticks.begin() + Region(iterator),
+              current_tick);
+    used_iterator = iterator;
+    free_iterator = std::max(free_iterator, iterator + size);
+
+    if (iterator + size > STREAM_BUFFER_SIZE) {
+        std::fill(sync_ticks.begin() + Region(used_iterator), sync_ticks.begin() + NUM_SYNCS,
+                  current_tick);
+        used_iterator = 0;
+        iterator = 0;
+        free_iterator = size;
+
+        if (AreRegionsActive(0, Region(size) + 1)) {
+            // Avoid waiting for the previous usages to be free
+            return GetStagingBuffer(size, MemoryUsage::Upload);
+        }
+    }
+    const size_t offset = iterator;
+    iterator = Common::AlignUp(iterator + size, MAX_ALIGNMENT);
+    return StagingBufferRef{
+        .buffer = *stream_buffer,
+        .offset = static_cast<VkDeviceSize>(offset),
+        .mapped_span = std::span<u8>(stream_pointer + offset, size),
+    };
+}
+
+bool StagingBufferPool::AreRegionsActive(size_t region_begin, size_t region_end) const {
+    return std::any_of(sync_ticks.begin() + region_begin, sync_ticks.begin() + region_end,
+                       [this](u64 sync_tick) { return !scheduler.IsFree(sync_tick); });
+};
+
+StagingBufferRef StagingBufferPool::GetStagingBuffer(size_t size, MemoryUsage usage) {
+    if (const std::optional<StagingBufferRef> ref = TryGetReservedBuffer(size, usage)) {
+        return *ref;
+    }
+    return CreateStagingBuffer(size, usage);
+}
+
 std::optional<StagingBufferRef> StagingBufferPool::TryGetReservedBuffer(size_t size,
                                                                         MemoryUsage usage) {
     StagingBuffers& cache_level = GetCache(usage)[Common::Log2Ceil64(size)];

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -19,11 +19,14 @@ class VKScheduler;
 
 struct StagingBufferRef {
     VkBuffer buffer;
+    VkDeviceSize offset;
     std::span<u8> mapped_span;
 };
 
 class StagingBufferPool {
 public:
+    static constexpr size_t NUM_SYNCS = 16;
+
     explicit StagingBufferPool(const Device& device, MemoryAllocator& memory_allocator,
                                VKScheduler& scheduler);
     ~StagingBufferPool();
@@ -33,6 +36,11 @@ public:
     void TickFrame();
 
 private:
+    struct StreamBufferCommit {
+        size_t upper_bound;
+        u64 tick;
+    };
+
     struct StagingBuffer {
         vk::Buffer buffer;
         MemoryCommit commit;
@@ -42,6 +50,7 @@ private:
         StagingBufferRef Ref() const noexcept {
             return {
                 .buffer = *buffer,
+                .offset = 0,
                 .mapped_span = mapped_span,
             };
         }
@@ -56,6 +65,12 @@ private:
     static constexpr size_t NUM_LEVELS = sizeof(size_t) * CHAR_BIT;
     using StagingBuffersCache = std::array<StagingBuffers, NUM_LEVELS>;
 
+    StagingBufferRef GetStreamBuffer(size_t size);
+
+    bool AreRegionsActive(size_t region_begin, size_t region_end) const;
+
+    StagingBufferRef GetStagingBuffer(size_t size, MemoryUsage usage);
+
     std::optional<StagingBufferRef> TryGetReservedBuffer(size_t size, MemoryUsage usage);
 
     StagingBufferRef CreateStagingBuffer(size_t size, MemoryUsage usage);
@@ -70,6 +85,15 @@ private:
     MemoryAllocator& memory_allocator;
     VKScheduler& scheduler;
 
+    vk::Buffer stream_buffer;
+    vk::DeviceMemory stream_memory;
+    u8* stream_pointer = nullptr;
+
+    size_t iterator = 0;
+    size_t used_iterator = 0;
+    size_t free_iterator = 0;
+    std::array<u64, NUM_SYNCS> sync_ticks{};
+
     StagingBuffersCache device_local_cache;
     StagingBuffersCache upload_cache;
     StagingBuffersCache download_cache;

src/video_core/renderer_vulkan/vk_state_tracker.cpp

@@ -30,15 +30,18 @@ using Table = Maxwell3D::DirtyState::Table;
 using Flags = Maxwell3D::DirtyState::Flags;
 
 Flags MakeInvalidationFlags() {
-    static constexpr std::array INVALIDATION_FLAGS{
+    static constexpr int INVALIDATION_FLAGS[]{
         Viewports,         Scissors,  DepthBias,         BlendConstants,    DepthBounds,
         StencilProperties, CullMode,  DepthBoundsEnable, DepthTestEnable,   DepthWriteEnable,
-        DepthCompareOp,    FrontFace, StencilOp,         StencilTestEnable,
+        DepthCompareOp,    FrontFace, StencilOp,         StencilTestEnable, VertexBuffers,
     };
     Flags flags{};
     for (const int flag : INVALIDATION_FLAGS) {
         flags[flag] = true;
     }
+    for (int index = VertexBuffer0; index <= VertexBuffer31; ++index) {
+        flags[index] = true;
+    }
     return flags;
 }
 
@@ -130,7 +133,7 @@ void SetupDirtyStencilTestEnable(Tables& tables) {
 StateTracker::StateTracker(Tegra::GPU& gpu)
     : flags{gpu.Maxwell3D().dirty.flags}, invalidation_flags{MakeInvalidationFlags()} {
     auto& tables = gpu.Maxwell3D().dirty.tables;
-    SetupDirtyRenderTargets(tables);
+    SetupDirtyFlags(tables);
     SetupDirtyViewports(tables);
     SetupDirtyScissors(tables);
     SetupDirtyDepthBias(tables);

src/video_core/renderer_vulkan/vk_swapchain.cpp

@@ -56,8 +56,11 @@ VkExtent2D ChooseSwapExtent(const VkSurfaceCapabilitiesKHR& capabilities, u32 wi
 
 } // Anonymous namespace
 
-VKSwapchain::VKSwapchain(VkSurfaceKHR surface_, const Device& device_, VKScheduler& scheduler_)
-    : surface{surface_}, device{device_}, scheduler{scheduler_} {}
+VKSwapchain::VKSwapchain(VkSurfaceKHR surface_, const Device& device_, VKScheduler& scheduler_,
+                         u32 width, u32 height, bool srgb)
+    : surface{surface_}, device{device_}, scheduler{scheduler_} {
+    Create(width, height, srgb);
+}
 
 VKSwapchain::~VKSwapchain() = default;
 

src/video_core/renderer_vulkan/vk_swapchain.h

@@ -20,7 +20,8 @@ class VKScheduler;
 
 class VKSwapchain {
 public:
-    explicit VKSwapchain(VkSurfaceKHR surface, const Device& device, VKScheduler& scheduler);
+    explicit VKSwapchain(VkSurfaceKHR surface, const Device& device, VKScheduler& scheduler,
+                         u32 width, u32 height, bool srgb);
     ~VKSwapchain();
 
     /// Creates (or recreates) the swapchain with a given size.

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -426,46 +426,47 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
 void CopyBufferToImage(vk::CommandBuffer cmdbuf, VkBuffer src_buffer, VkImage image,
                        VkImageAspectFlags aspect_mask, bool is_initialized,
                        std::span<const VkBufferImageCopy> copies) {
-    static constexpr VkAccessFlags ACCESS_FLAGS = VK_ACCESS_SHADER_WRITE_BIT |
-                                                  VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                                                  VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+    static constexpr VkAccessFlags WRITE_ACCESS_FLAGS =
+        VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+        VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+    static constexpr VkAccessFlags READ_ACCESS_FLAGS = VK_ACCESS_SHADER_READ_BIT |
+                                                       VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+                                                       VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
     const VkImageMemoryBarrier read_barrier{
         .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
         .pNext = nullptr,
-        .srcAccessMask = ACCESS_FLAGS,
+        .srcAccessMask = WRITE_ACCESS_FLAGS,
         .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
         .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
         .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
         .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
         .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
         .image = image,
-        .subresourceRange =
-            {
-                .aspectMask = aspect_mask,
-                .baseMipLevel = 0,
-                .levelCount = VK_REMAINING_MIP_LEVELS,
-                .baseArrayLayer = 0,
-                .layerCount = VK_REMAINING_ARRAY_LAYERS,
-            },
+        .subresourceRange{
+            .aspectMask = aspect_mask,
+            .baseMipLevel = 0,
+            .levelCount = VK_REMAINING_MIP_LEVELS,
+            .baseArrayLayer = 0,
+            .layerCount = VK_REMAINING_ARRAY_LAYERS,
+        },
     };
     const VkImageMemoryBarrier write_barrier{
         .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
         .pNext = nullptr,
         .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-        .dstAccessMask = ACCESS_FLAGS,
+        .dstAccessMask = WRITE_ACCESS_FLAGS | READ_ACCESS_FLAGS,
         .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
         .newLayout = VK_IMAGE_LAYOUT_GENERAL,
         .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
         .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
         .image = image,
-        .subresourceRange =
-            {
-                .aspectMask = aspect_mask,
-                .baseMipLevel = 0,
-                .levelCount = VK_REMAINING_MIP_LEVELS,
-                .baseArrayLayer = 0,
-                .layerCount = VK_REMAINING_ARRAY_LAYERS,
-            },
+        .subresourceRange{
+            .aspectMask = aspect_mask,
+            .baseMipLevel = 0,
+            .levelCount = VK_REMAINING_MIP_LEVELS,
+            .baseArrayLayer = 0,
+            .layerCount = VK_REMAINING_ARRAY_LAYERS,
+        },
     };
     cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
                            read_barrier);
@@ -569,20 +570,12 @@ void TextureCacheRuntime::Finish() {
     scheduler.Finish();
 }
 
-ImageBufferMap TextureCacheRuntime::MapUploadBuffer(size_t size) {
-    const auto staging_ref = staging_buffer_pool.Request(size, MemoryUsage::Upload);
-    return {
-        .handle = staging_ref.buffer,
-        .span = staging_ref.mapped_span,
-    };
+StagingBufferRef TextureCacheRuntime::UploadStagingBuffer(size_t size) {
+    return staging_buffer_pool.Request(size, MemoryUsage::Upload);
 }
 
-ImageBufferMap TextureCacheRuntime::MapDownloadBuffer(size_t size) {
-    const auto staging_ref = staging_buffer_pool.Request(size, MemoryUsage::Download);
-    return {
-        .handle = staging_ref.buffer,
-        .span = staging_ref.mapped_span,
-    };
+StagingBufferRef TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
+    return staging_buffer_pool.Request(size, MemoryUsage::Download);
 }
 
 void TextureCacheRuntime::BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src,
@@ -754,7 +747,7 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
                 .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
                                  VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
                                  VK_ACCESS_TRANSFER_WRITE_BIT,
-                .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
                 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
@@ -765,12 +758,9 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
             VkImageMemoryBarrier{
                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
                 .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT |
-                                 VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
-                                 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                                 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
+                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
                                  VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
-                                 VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+                                 VK_ACCESS_TRANSFER_WRITE_BIT,
                 .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
                 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                 .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
@@ -828,12 +818,11 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
     }
 }
 
-void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
-                         std::span<const BufferImageCopy> copies) {
+void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
     // TODO: Move this to another API
     scheduler->RequestOutsideRenderPassOperationContext();
-    std::vector vk_copies = TransformBufferImageCopies(copies, buffer_offset, aspect_mask);
-    const VkBuffer src_buffer = map.handle;
+    std::vector vk_copies = TransformBufferImageCopies(copies, map.offset, aspect_mask);
+    const VkBuffer src_buffer = map.buffer;
     const VkImage vk_image = *image;
     const VkImageAspectFlags vk_aspect_mask = aspect_mask;
     const bool is_initialized = std::exchange(initialized, true);
@@ -843,12 +832,12 @@ void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
     });
 }
 
-void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
+void Image::UploadMemory(const StagingBufferRef& map,
                          std::span<const VideoCommon::BufferCopy> copies) {
     // TODO: Move this to another API
     scheduler->RequestOutsideRenderPassOperationContext();
-    std::vector vk_copies = TransformBufferCopies(copies, buffer_offset);
-    const VkBuffer src_buffer = map.handle;
+    std::vector vk_copies = TransformBufferCopies(copies, map.offset);
+    const VkBuffer src_buffer = map.buffer;
     const VkBuffer dst_buffer = *buffer;
     scheduler->Record([src_buffer, dst_buffer, vk_copies](vk::CommandBuffer cmdbuf) {
         // TODO: Barriers
@@ -856,13 +845,57 @@ void Image::UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
     });
 }
 
-void Image::DownloadMemory(const ImageBufferMap& map, size_t buffer_offset,
-                           std::span<const BufferImageCopy> copies) {
-    std::vector vk_copies = TransformBufferImageCopies(copies, buffer_offset, aspect_mask);
-    scheduler->Record([buffer = map.handle, image = *image, aspect_mask = aspect_mask,
+void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
+    std::vector vk_copies = TransformBufferImageCopies(copies, map.offset, aspect_mask);
+    scheduler->Record([buffer = map.buffer, image = *image, aspect_mask = aspect_mask,
                        vk_copies](vk::CommandBuffer cmdbuf) {
-        // TODO: Barriers
-        cmdbuf.CopyImageToBuffer(image, VK_IMAGE_LAYOUT_GENERAL, buffer, vk_copies);
+        const VkImageMemoryBarrier read_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        const VkImageMemoryBarrier image_write_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = 0,
+            .dstAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        const VkMemoryBarrier memory_write_barrier{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, read_barrier);
+        cmdbuf.CopyImageToBuffer(image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, buffer, vk_copies);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, memory_write_barrier, nullptr, image_write_barrier);
     });
 }
 
@@ -1127,7 +1160,7 @@ Framebuffer::Framebuffer(TextureCacheRuntime& runtime, std::span<ImageView*, NUM
         .pAttachments = attachments.data(),
         .width = key.size.width,
         .height = key.size.height,
-        .layers = static_cast<u32>(num_layers),
+        .layers = static_cast<u32>(std::max(num_layers, 1)),
     });
     if (runtime.device.HasDebuggingToolAttached()) {
         framebuffer.SetObjectNameEXT(VideoCommon::Name(key).c_str());

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -7,6 +7,7 @@
 #include <compare>
 #include <span>
 
+#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
 #include "video_core/texture_cache/texture_cache.h"
 #include "video_core/vulkan_common/vulkan_memory_allocator.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -53,19 +54,6 @@ struct hash<Vulkan::RenderPassKey> {
 
 namespace Vulkan {
 
-struct ImageBufferMap {
-    [[nodiscard]] VkBuffer Handle() const noexcept {
-        return handle;
-    }
-
-    [[nodiscard]] std::span<u8> Span() const noexcept {
-        return span;
-    }
-
-    VkBuffer handle;
-    std::span<u8> span;
-};
-
 struct TextureCacheRuntime {
     const Device& device;
     VKScheduler& scheduler;
@@ -76,9 +64,9 @@ struct TextureCacheRuntime {
 
     void Finish();
 
-    [[nodiscard]] ImageBufferMap MapUploadBuffer(size_t size);
+    [[nodiscard]] StagingBufferRef UploadStagingBuffer(size_t size);
 
-    [[nodiscard]] ImageBufferMap MapDownloadBuffer(size_t size);
+    [[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size);
 
     void BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src,
                    const std::array<Offset2D, 2>& dst_region,
@@ -94,7 +82,7 @@ struct TextureCacheRuntime {
         return false;
     }
 
-    void AccelerateImageUpload(Image&, const ImageBufferMap&, size_t,
+    void AccelerateImageUpload(Image&, const StagingBufferRef&,
                                std::span<const VideoCommon::SwizzleParameters>) {
         UNREACHABLE();
     }
@@ -112,13 +100,12 @@ public:
     explicit Image(TextureCacheRuntime&, const VideoCommon::ImageInfo& info, GPUVAddr gpu_addr,
                    VAddr cpu_addr);
 
-    void UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
+    void UploadMemory(const StagingBufferRef& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
-    void UploadMemory(const ImageBufferMap& map, size_t buffer_offset,
-                      std::span<const VideoCommon::BufferCopy> copies);
+    void UploadMemory(const StagingBufferRef& map, std::span<const VideoCommon::BufferCopy> copies);
 
-    void DownloadMemory(const ImageBufferMap& map, size_t buffer_offset,
+    void DownloadMemory(const StagingBufferRef& map,
                         std::span<const VideoCommon::BufferImageCopy> copies);
 
     [[nodiscard]] VkImage Handle() const noexcept {

src/video_core/shader/async_shaders.h

@@ -9,16 +9,7 @@
 #include <shared_mutex>
 #include <thread>
 
-// This header includes both Vulkan and OpenGL headers, this has to be fixed
-// Unfortunately, including OpenGL will include Windows.h that defines macros that can cause issues.
-// Forcefully include glad early and undefine macros
 #include <glad/glad.h>
-#ifdef CreateEvent
-#undef CreateEvent
-#endif
-#ifdef CreateSemaphore
-#undef CreateSemaphore
-#endif
 
 #include "common/common_types.h"
 #include "video_core/renderer_opengl/gl_device.h"

src/video_core/shader/decode/other.cpp

@@ -76,6 +76,7 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
             case SystemVariable::InvocationId:
                 return Operation(OperationCode::InvocationId);
             case SystemVariable::Ydirection:
+                uses_y_negate = true;
                 return Operation(OperationCode::YNegate);
             case SystemVariable::InvocationInfo:
                 LOG_WARNING(HW_GPU, "S2R instruction with InvocationInfo is incomplete");

src/video_core/shader/shader_ir.h

@@ -139,6 +139,10 @@ public:
         return uses_legacy_varyings;
     }
 
+    bool UsesYNegate() const {
+        return uses_y_negate;
+    }
+
     bool UsesWarps() const {
         return uses_warps;
     }
@@ -465,6 +469,7 @@ private:
     bool uses_instance_id{};
     bool uses_vertex_id{};
     bool uses_legacy_varyings{};
+    bool uses_y_negate{};
     bool uses_warps{};
     bool uses_indexed_samplers{};
 

src/video_core/texture_cache/texture_cache.h

@@ -103,9 +103,6 @@ public:
     /// Notify the cache that a new frame has been queued
     void TickFrame();
 
-    /// Return an unique mutually exclusive lock for the cache
-    [[nodiscard]] std::unique_lock<std::mutex> AcquireLock();
-
     /// Return a constant reference to the given image view id
     [[nodiscard]] const ImageView& GetImageView(ImageViewId id) const noexcept;
 
@@ -179,6 +176,8 @@ public:
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
+    std::mutex mutex;
+
 private:
     /// Iterate over all page indices in a range
     template <typename Func>
@@ -212,8 +211,8 @@ private:
     void RefreshContents(Image& image);
 
     /// Upload data from guest to an image
-    template <typename MapBuffer>
-    void UploadImageContents(Image& image, MapBuffer& map, size_t buffer_offset);
+    template <typename StagingBuffer>
+    void UploadImageContents(Image& image, StagingBuffer& staging_buffer);
 
     /// Find or create an image view from a guest descriptor
     [[nodiscard]] ImageViewId FindImageView(const TICEntry& config);
@@ -325,8 +324,6 @@ private:
 
     RenderTargets render_targets;
 
-    std::mutex mutex;
-
     std::unordered_map<TICEntry, ImageViewId> image_views;
     std::unordered_map<TSCEntry, SamplerId> samplers;
     std::unordered_map<RenderTargets, FramebufferId> framebuffers;
@@ -385,11 +382,6 @@ void TextureCache<P>::TickFrame() {
     ++frame_tick;
 }
 
-template <class P>
-std::unique_lock<std::mutex> TextureCache<P>::AcquireLock() {
-    return std::unique_lock{mutex};
-}
-
 template <class P>
 const typename P::ImageView& TextureCache<P>::GetImageView(ImageViewId id) const noexcept {
     return slot_image_views[id];
@@ -598,11 +590,11 @@ void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
     });
     for (const ImageId image_id : images) {
         Image& image = slot_images[image_id];
-        auto map = runtime.MapDownloadBuffer(image.unswizzled_size_bytes);
+        auto map = runtime.DownloadStagingBuffer(image.unswizzled_size_bytes);
         const auto copies = FullDownloadCopies(image.info);
-        image.DownloadMemory(map, 0, copies);
+        image.DownloadMemory(map, copies);
         runtime.Finish();
-        SwizzleImage(gpu_memory, image.gpu_addr, image.info, copies, map.Span());
+        SwizzleImage(gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span);
     }
 }
 
@@ -757,25 +749,25 @@ void TextureCache<P>::PopAsyncFlushes() {
     for (const ImageId image_id : download_ids) {
         total_size_bytes += slot_images[image_id].unswizzled_size_bytes;
     }
-    auto download_map = runtime.MapDownloadBuffer(total_size_bytes);
-    size_t buffer_offset = 0;
+    auto download_map = runtime.DownloadStagingBuffer(total_size_bytes);
+    const size_t original_offset = download_map.offset;
     for (const ImageId image_id : download_ids) {
         Image& image = slot_images[image_id];
         const auto copies = FullDownloadCopies(image.info);
-        image.DownloadMemory(download_map, buffer_offset, copies);
-        buffer_offset += image.unswizzled_size_bytes;
+        image.DownloadMemory(download_map, copies);
+        download_map.offset += image.unswizzled_size_bytes;
     }
     // Wait for downloads to finish
     runtime.Finish();
 
-    buffer_offset = 0;
-    const std::span<u8> download_span = download_map.Span();
+    download_map.offset = original_offset;
+    std::span<u8> download_span = download_map.mapped_span;
     for (const ImageId image_id : download_ids) {
         const ImageBase& image = slot_images[image_id];
         const auto copies = FullDownloadCopies(image.info);
-        const std::span<u8> image_download_span = download_span.subspan(buffer_offset);
-        SwizzleImage(gpu_memory, image.gpu_addr, image.info, copies, image_download_span);
-        buffer_offset += image.unswizzled_size_bytes;
+        SwizzleImage(gpu_memory, image.gpu_addr, image.info, copies, download_span);
+        download_map.offset += image.unswizzled_size_bytes;
+        download_span = download_span.subspan(image.unswizzled_size_bytes);
     }
     committed_downloads.pop();
 }
@@ -806,32 +798,32 @@ void TextureCache<P>::RefreshContents(Image& image) {
         LOG_WARNING(HW_GPU, "MSAA image uploads are not implemented");
         return;
     }
-    auto map = runtime.MapUploadBuffer(MapSizeBytes(image));
-    UploadImageContents(image, map, 0);
+    auto staging = runtime.UploadStagingBuffer(MapSizeBytes(image));
+    UploadImageContents(image, staging);
     runtime.InsertUploadMemoryBarrier();
 }
 
 template <class P>
-template <typename MapBuffer>
-void TextureCache<P>::UploadImageContents(Image& image, MapBuffer& map, size_t buffer_offset) {
-    const std::span<u8> mapped_span = map.Span().subspan(buffer_offset);
+template <typename StagingBuffer>
+void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging) {
+    const std::span<u8> mapped_span = staging.mapped_span;
     const GPUVAddr gpu_addr = image.gpu_addr;
 
     if (True(image.flags & ImageFlagBits::AcceleratedUpload)) {
         gpu_memory.ReadBlockUnsafe(gpu_addr, mapped_span.data(), mapped_span.size_bytes());
         const auto uploads = FullUploadSwizzles(image.info);
-        runtime.AccelerateImageUpload(image, map, buffer_offset, uploads);
+        runtime.AccelerateImageUpload(image, staging, uploads);
     } else if (True(image.flags & ImageFlagBits::Converted)) {
         std::vector<u8> unswizzled_data(image.unswizzled_size_bytes);
         auto copies = UnswizzleImage(gpu_memory, gpu_addr, image.info, unswizzled_data);
         ConvertImage(unswizzled_data, image.info, mapped_span, copies);
-        image.UploadMemory(map, buffer_offset, copies);
+        image.UploadMemory(staging, copies);
     } else if (image.info.type == ImageType::Buffer) {
         const std::array copies{UploadBufferCopy(gpu_memory, gpu_addr, image, mapped_span)};
-        image.UploadMemory(map, buffer_offset, copies);
+        image.UploadMemory(staging, copies);
     } else {
         const auto copies = UnswizzleImage(gpu_memory, gpu_addr, image.info, mapped_span);
-        image.UploadMemory(map, buffer_offset, copies);
+        image.UploadMemory(staging, copies);
     }
 }
 

src/video_core/video_core.cpp

@@ -38,19 +38,18 @@ namespace VideoCore {
 
 std::unique_ptr<Tegra::GPU> CreateGPU(Core::Frontend::EmuWindow& emu_window, Core::System& system) {
     const bool use_nvdec = Settings::values.use_nvdec_emulation.GetValue();
-    std::unique_ptr<Tegra::GPU> gpu = std::make_unique<Tegra::GPU>(
-        system, Settings::values.use_asynchronous_gpu_emulation.GetValue(), use_nvdec);
-
+    const bool use_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
+    auto gpu = std::make_unique<Tegra::GPU>(system, use_async, use_nvdec);
     auto context = emu_window.CreateSharedContext();
-    const auto scope = context->Acquire();
-
-    auto renderer = CreateRenderer(system, emu_window, *gpu, std::move(context));
-    if (!renderer->Init()) {
+    auto scope = context->Acquire();
+    try {
+        auto renderer = CreateRenderer(system, emu_window, *gpu, std::move(context));
+        gpu->BindRenderer(std::move(renderer));
+        return gpu;
+    } catch (const std::runtime_error& exception) {
+        LOG_ERROR(HW_GPU, "Failed to initialize GPU: {}", exception.what());
         return nullptr;
     }
-
-    gpu->BindRenderer(std::move(renderer));
-    return gpu;
 }
 
 u16 GetResolutionScaleFactor(const RendererBase& renderer) {

src/video_core/vulkan_common/vulkan_device.cpp

@@ -18,27 +18,22 @@
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 namespace Vulkan {
-
 namespace {
-
 namespace Alternatives {
-
-constexpr std::array Depth24UnormS8_UINT{
+constexpr std::array DEPTH24_UNORM_STENCIL8_UINT{
     VK_FORMAT_D32_SFLOAT_S8_UINT,
     VK_FORMAT_D16_UNORM_S8_UINT,
-    VkFormat{},
+    VK_FORMAT_UNDEFINED,
 };
 
-constexpr std::array Depth16UnormS8_UINT{
+constexpr std::array DEPTH16_UNORM_STENCIL8_UINT{
     VK_FORMAT_D24_UNORM_S8_UINT,
     VK_FORMAT_D32_SFLOAT_S8_UINT,
-    VkFormat{},
+    VK_FORMAT_UNDEFINED,
 };
-
 } // namespace Alternatives
 
 constexpr std::array REQUIRED_EXTENSIONS{
-    VK_KHR_SWAPCHAIN_EXTENSION_NAME,
     VK_KHR_MAINTENANCE1_EXTENSION_NAME,
     VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME,
     VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
@@ -52,6 +47,12 @@ constexpr std::array REQUIRED_EXTENSIONS{
     VK_EXT_SHADER_SUBGROUP_BALLOT_EXTENSION_NAME,
     VK_EXT_SHADER_SUBGROUP_VOTE_EXTENSION_NAME,
     VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
+#ifdef _WIN32
+    VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME,
+#endif
+#ifdef __linux__
+    VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
+#endif
 };
 
 template <typename T>
@@ -63,9 +64,9 @@ void SetNext(void**& next, T& data) {
 constexpr const VkFormat* GetFormatAlternatives(VkFormat format) {
     switch (format) {
     case VK_FORMAT_D24_UNORM_S8_UINT:
-        return Alternatives::Depth24UnormS8_UINT.data();
+        return Alternatives::DEPTH24_UNORM_STENCIL8_UINT.data();
     case VK_FORMAT_D16_UNORM_S8_UINT:
-        return Alternatives::Depth16UnormS8_UINT.data();
+        return Alternatives::DEPTH16_UNORM_STENCIL8_UINT.data();
     default:
         return nullptr;
     }
@@ -195,78 +196,77 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
                const vk::InstanceDispatch& dld_)
     : instance{instance_}, dld{dld_}, physical{physical_}, properties{physical.GetProperties()},
       format_properties{GetFormatProperties(physical)} {
-    CheckSuitability();
+    CheckSuitability(surface != nullptr);
     SetupFamilies(surface);
     SetupFeatures();
 
     const auto queue_cis = GetDeviceQueueCreateInfos();
-    const std::vector extensions = LoadExtensions();
+    const std::vector extensions = LoadExtensions(surface != nullptr);
 
     VkPhysicalDeviceFeatures2 features2{
         .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
         .pNext = nullptr,
-        .features{},
+        .features{
+            .robustBufferAccess = true,
+            .fullDrawIndexUint32 = false,
+            .imageCubeArray = true,
+            .independentBlend = true,
+            .geometryShader = true,
+            .tessellationShader = true,
+            .sampleRateShading = false,
+            .dualSrcBlend = false,
+            .logicOp = false,
+            .multiDrawIndirect = false,
+            .drawIndirectFirstInstance = false,
+            .depthClamp = true,
+            .depthBiasClamp = true,
+            .fillModeNonSolid = false,
+            .depthBounds = false,
+            .wideLines = false,
+            .largePoints = true,
+            .alphaToOne = false,
+            .multiViewport = true,
+            .samplerAnisotropy = true,
+            .textureCompressionETC2 = false,
+            .textureCompressionASTC_LDR = is_optimal_astc_supported,
+            .textureCompressionBC = false,
+            .occlusionQueryPrecise = true,
+            .pipelineStatisticsQuery = false,
+            .vertexPipelineStoresAndAtomics = true,
+            .fragmentStoresAndAtomics = true,
+            .shaderTessellationAndGeometryPointSize = false,
+            .shaderImageGatherExtended = true,
+            .shaderStorageImageExtendedFormats = false,
+            .shaderStorageImageMultisample = is_shader_storage_image_multisample,
+            .shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
+            .shaderStorageImageWriteWithoutFormat = true,
+            .shaderUniformBufferArrayDynamicIndexing = false,
+            .shaderSampledImageArrayDynamicIndexing = false,
+            .shaderStorageBufferArrayDynamicIndexing = false,
+            .shaderStorageImageArrayDynamicIndexing = false,
+            .shaderClipDistance = false,
+            .shaderCullDistance = false,
+            .shaderFloat64 = false,
+            .shaderInt64 = false,
+            .shaderInt16 = false,
+            .shaderResourceResidency = false,
+            .shaderResourceMinLod = false,
+            .sparseBinding = false,
+            .sparseResidencyBuffer = false,
+            .sparseResidencyImage2D = false,
+            .sparseResidencyImage3D = false,
+            .sparseResidency2Samples = false,
+            .sparseResidency4Samples = false,
+            .sparseResidency8Samples = false,
+            .sparseResidency16Samples = false,
+            .sparseResidencyAliased = false,
+            .variableMultisampleRate = false,
+            .inheritedQueries = false,
+        },
     };
     const void* first_next = &features2;
     void** next = &features2.pNext;
 
-    features2.features = {
-        .robustBufferAccess = false,
-        .fullDrawIndexUint32 = false,
-        .imageCubeArray = true,
-        .independentBlend = true,
-        .geometryShader = true,
-        .tessellationShader = true,
-        .sampleRateShading = false,
-        .dualSrcBlend = false,
-        .logicOp = false,
-        .multiDrawIndirect = false,
-        .drawIndirectFirstInstance = false,
-        .depthClamp = true,
-        .depthBiasClamp = true,
-        .fillModeNonSolid = false,
-        .depthBounds = false,
-        .wideLines = false,
-        .largePoints = true,
-        .alphaToOne = false,
-        .multiViewport = true,
-        .samplerAnisotropy = true,
-        .textureCompressionETC2 = false,
-        .textureCompressionASTC_LDR = is_optimal_astc_supported,
-        .textureCompressionBC = false,
-        .occlusionQueryPrecise = true,
-        .pipelineStatisticsQuery = false,
-        .vertexPipelineStoresAndAtomics = true,
-        .fragmentStoresAndAtomics = true,
-        .shaderTessellationAndGeometryPointSize = false,
-        .shaderImageGatherExtended = true,
-        .shaderStorageImageExtendedFormats = false,
-        .shaderStorageImageMultisample = is_shader_storage_image_multisample,
-        .shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
-        .shaderStorageImageWriteWithoutFormat = true,
-        .shaderUniformBufferArrayDynamicIndexing = false,
-        .shaderSampledImageArrayDynamicIndexing = false,
-        .shaderStorageBufferArrayDynamicIndexing = false,
-        .shaderStorageImageArrayDynamicIndexing = false,
-        .shaderClipDistance = false,
-        .shaderCullDistance = false,
-        .shaderFloat64 = false,
-        .shaderInt64 = false,
-        .shaderInt16 = false,
-        .shaderResourceResidency = false,
-        .shaderResourceMinLod = false,
-        .sparseBinding = false,
-        .sparseResidencyBuffer = false,
-        .sparseResidencyImage2D = false,
-        .sparseResidencyImage3D = false,
-        .sparseResidency2Samples = false,
-        .sparseResidency4Samples = false,
-        .sparseResidency8Samples = false,
-        .sparseResidency16Samples = false,
-        .sparseResidencyAliased = false,
-        .variableMultisampleRate = false,
-        .inheritedQueries = false,
-    };
     VkPhysicalDeviceTimelineSemaphoreFeaturesKHR timeline_semaphore{
         .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR,
         .pNext = nullptr,
@@ -384,7 +384,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         robustness2 = {
             .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT,
             .pNext = nullptr,
-            .robustBufferAccess2 = false,
+            .robustBufferAccess2 = true,
             .robustImageAccess2 = true,
             .nullDescriptor = true,
         };
@@ -535,16 +535,18 @@ bool Device::IsFormatSupported(VkFormat wanted_format, VkFormatFeatureFlags want
     return (supported_usage & wanted_usage) == wanted_usage;
 }
 
-void Device::CheckSuitability() const {
+void Device::CheckSuitability(bool requires_swapchain) const {
     std::bitset<REQUIRED_EXTENSIONS.size()> available_extensions;
+    bool has_swapchain = false;
     for (const VkExtensionProperties& property : physical.EnumerateDeviceExtensionProperties()) {
-        for (std::size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
+        const std::string_view name{property.extensionName};
+        for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
             if (available_extensions[i]) {
                 continue;
             }
-            const std::string_view name{property.extensionName};
             available_extensions[i] = name == REQUIRED_EXTENSIONS[i];
         }
+        has_swapchain = has_swapchain || name == VK_KHR_SWAPCHAIN_EXTENSION_NAME;
     }
     for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
         if (available_extensions[i]) {
@@ -553,6 +555,11 @@ void Device::CheckSuitability() const {
         LOG_ERROR(Render_Vulkan, "Missing required extension: {}", REQUIRED_EXTENSIONS[i]);
         throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
     }
+    if (requires_swapchain && !has_swapchain) {
+        LOG_ERROR(Render_Vulkan, "Missing required extension: VK_KHR_swapchain");
+        throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
+    }
+
     struct LimitTuple {
         u32 minimum;
         u32 value;
@@ -574,7 +581,9 @@ void Device::CheckSuitability() const {
     }
     const VkPhysicalDeviceFeatures features{physical.GetFeatures()};
     const std::array feature_report{
+        std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
         std::make_pair(features.vertexPipelineStoresAndAtomics, "vertexPipelineStoresAndAtomics"),
+        std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
         std::make_pair(features.imageCubeArray, "imageCubeArray"),
         std::make_pair(features.independentBlend, "independentBlend"),
         std::make_pair(features.depthClamp, "depthClamp"),
@@ -599,10 +608,13 @@ void Device::CheckSuitability() const {
     }
 }
 
-std::vector<const char*> Device::LoadExtensions() {
+std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
     std::vector<const char*> extensions;
-    extensions.reserve(7 + REQUIRED_EXTENSIONS.size());
+    extensions.reserve(8 + REQUIRED_EXTENSIONS.size());
     extensions.insert(extensions.begin(), REQUIRED_EXTENSIONS.begin(), REQUIRED_EXTENSIONS.end());
+    if (requires_surface) {
+        extensions.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
+    }
 
     bool has_khr_shader_float16_int8{};
     bool has_ext_subgroup_size_control{};
@@ -743,7 +755,8 @@ std::vector<const char*> Device::LoadExtensions() {
         robustness2.pNext = nullptr;
         features.pNext = &robustness2;
         physical.GetFeatures2KHR(features);
-        if (robustness2.nullDescriptor && robustness2.robustImageAccess2) {
+        if (robustness2.nullDescriptor && robustness2.robustBufferAccess2 &&
+            robustness2.robustImageAccess2) {
             extensions.push_back(VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
             ext_robustness2 = true;
         }

src/video_core/vulkan_common/vulkan_device.h

@@ -23,7 +23,7 @@ enum class FormatType { Linear, Optimal, Buffer };
 const u32 GuestWarpSize = 32;
 
 /// Handles data specific to a physical device.
-class Device final {
+class Device {
 public:
     explicit Device(VkInstance instance, vk::PhysicalDevice physical, VkSurfaceKHR surface,
                     const vk::InstanceDispatch& dld);
@@ -227,10 +227,10 @@ public:
 
 private:
     /// Checks if the physical device is suitable.
-    void CheckSuitability() const;
+    void CheckSuitability(bool requires_swapchain) const;
 
     /// Loads extensions into a vector and stores available ones in this object.
-    std::vector<const char*> LoadExtensions();
+    std::vector<const char*> LoadExtensions(bool requires_surface);
 
     /// Sets up queue families.
     void SetupFamilies(VkSurfaceKHR surface);

src/video_core/vulkan_common/vulkan_instance.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <future>
 #include <optional>
 #include <span>
 #include <utility>
@@ -140,7 +141,10 @@ vk::Instance CreateInstance(const Common::DynamicLibrary& library, vk::InstanceD
                   VK_VERSION_MAJOR(required_version), VK_VERSION_MINOR(required_version));
         throw vk::Exception(VK_ERROR_INCOMPATIBLE_DRIVER);
     }
-    vk::Instance instance = vk::Instance::Create(required_version, layers, extensions, dld);
+    vk::Instance instance =
+        std::async([&] {
+            return vk::Instance::Create(required_version, layers, extensions, dld);
+        }).get();
     if (!vk::Load(*instance, dld)) {
         LOG_ERROR(Render_Vulkan, "Failed to load Vulkan instance function pointers");
         throw vk::Exception(VK_ERROR_INITIALIZATION_FAILED);

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -7,6 +7,8 @@
 #include <optional>
 #include <vector>
 
+#include <glad/glad.h>
+
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common_types.h"
@@ -55,10 +57,24 @@ struct Range {
 
 class MemoryAllocation {
 public:
-    explicit MemoryAllocation(const Device& device_, vk::DeviceMemory memory_,
-                              VkMemoryPropertyFlags properties, u64 allocation_size_, u32 type)
-        : device{device_}, memory{std::move(memory_)}, allocation_size{allocation_size_},
-          property_flags{properties}, shifted_memory_type{1U << type} {}
+    explicit MemoryAllocation(vk::DeviceMemory memory_, VkMemoryPropertyFlags properties,
+                              u64 allocation_size_, u32 type)
+        : memory{std::move(memory_)}, allocation_size{allocation_size_}, property_flags{properties},
+          shifted_memory_type{1U << type} {}
+
+#if defined(_WIN32) || defined(__linux__)
+    ~MemoryAllocation() {
+        if (owning_opengl_handle != 0) {
+            glDeleteMemoryObjectsEXT(1, &owning_opengl_handle);
+        }
+    }
+#endif
+
+    MemoryAllocation& operator=(const MemoryAllocation&) = delete;
+    MemoryAllocation(const MemoryAllocation&) = delete;
+
+    MemoryAllocation& operator=(MemoryAllocation&&) = delete;
+    MemoryAllocation(MemoryAllocation&&) = delete;
 
     [[nodiscard]] std::optional<MemoryCommit> Commit(VkDeviceSize size, VkDeviceSize alignment) {
         const std::optional<u64> alloc = FindFreeRegion(size, alignment);
@@ -88,6 +104,31 @@ public:
         return memory_mapped_span;
     }
 
+#ifdef _WIN32
+    [[nodiscard]] u32 ExportOpenGLHandle() {
+        if (!owning_opengl_handle) {
+            glCreateMemoryObjectsEXT(1, &owning_opengl_handle);
+            glImportMemoryWin32HandleEXT(owning_opengl_handle, allocation_size,
+                                         GL_HANDLE_TYPE_OPAQUE_WIN32_EXT,
+                                         memory.GetMemoryWin32HandleKHR());
+        }
+        return owning_opengl_handle;
+    }
+#elif __linux__
+    [[nodiscard]] u32 ExportOpenGLHandle() {
+        if (!owning_opengl_handle) {
+            glCreateMemoryObjectsEXT(1, &owning_opengl_handle);
+            glImportMemoryFdEXT(owning_opengl_handle, allocation_size, GL_HANDLE_TYPE_OPAQUE_FD_EXT,
+                                memory.GetMemoryFdKHR());
+        }
+        return owning_opengl_handle;
+    }
+#else
+    [[nodiscard]] u32 ExportOpenGLHandle() {
+        return 0;
+    }
+#endif
+
     /// Returns whether this allocation is compatible with the arguments.
     [[nodiscard]] bool IsCompatible(VkMemoryPropertyFlags flags, u32 type_mask) const {
         return (flags & property_flags) && (type_mask & shifted_memory_type) != 0;
@@ -118,13 +159,15 @@ private:
         return candidate;
     }
 
-    const Device& device;                       ///< Vulkan device.
     const vk::DeviceMemory memory;              ///< Vulkan memory allocation handler.
     const u64 allocation_size;                  ///< Size of this allocation.
     const VkMemoryPropertyFlags property_flags; ///< Vulkan memory property flags.
     const u32 shifted_memory_type;              ///< Shifted Vulkan memory type.
     std::vector<Range> commits;                 ///< All commit ranges done from this allocation.
     std::span<u8> memory_mapped_span; ///< Memory mapped span. Empty if not queried before.
+#if defined(_WIN32) || defined(__linux__)
+    u32 owning_opengl_handle{}; ///< Owning OpenGL memory object handle.
+#endif
 };
 
 MemoryCommit::MemoryCommit(MemoryAllocation* allocation_, VkDeviceMemory memory_, u64 begin_,
@@ -156,14 +199,19 @@ std::span<u8> MemoryCommit::Map() {
     return span;
 }
 
+u32 MemoryCommit::ExportOpenGLHandle() const {
+    return allocation->ExportOpenGLHandle();
+}
+
 void MemoryCommit::Release() {
     if (allocation) {
         allocation->Free(begin);
     }
 }
 
-MemoryAllocator::MemoryAllocator(const Device& device_)
-    : device{device_}, properties{device_.GetPhysical().GetMemoryProperties()} {}
+MemoryAllocator::MemoryAllocator(const Device& device_, bool export_allocations_)
+    : device{device_}, properties{device_.GetPhysical().GetMemoryProperties()},
+      export_allocations{export_allocations_} {}
 
 MemoryAllocator::~MemoryAllocator() = default;
 
@@ -196,14 +244,24 @@ MemoryCommit MemoryAllocator::Commit(const vk::Image& image, MemoryUsage usage)
 
 void MemoryAllocator::AllocMemory(VkMemoryPropertyFlags flags, u32 type_mask, u64 size) {
     const u32 type = FindType(flags, type_mask).value();
+    const VkExportMemoryAllocateInfo export_allocate_info{
+        .sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO,
+        .pNext = nullptr,
+#ifdef _WIN32
+        .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT,
+#elif __linux__
+        .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT,
+#else
+        .handleTypes = 0,
+#endif
+    };
     vk::DeviceMemory memory = device.GetLogical().AllocateMemory({
         .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
-        .pNext = nullptr,
+        .pNext = export_allocations ? &export_allocate_info : nullptr,
         .allocationSize = size,
         .memoryTypeIndex = type,
     });
-    allocations.push_back(
-        std::make_unique<MemoryAllocation>(device, std::move(memory), flags, size, type));
+    allocations.push_back(std::make_unique<MemoryAllocation>(std::move(memory), flags, size, type));
 }
 
 std::optional<MemoryCommit> MemoryAllocator::TryCommit(const VkMemoryRequirements& requirements,

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -43,6 +43,9 @@ public:
     /// It will map the backing allocation if it hasn't been mapped before.
     std::span<u8> Map();
 
+    /// Returns an non-owning OpenGL handle, creating one if it doesn't exist.
+    u32 ExportOpenGLHandle() const;
+
     /// Returns the Vulkan memory handler.
     VkDeviceMemory Memory() const {
         return memory;
@@ -67,7 +70,15 @@ private:
 /// Allocates and releases memory allocations on demand.
 class MemoryAllocator {
 public:
-    explicit MemoryAllocator(const Device& device_);
+    /**
+     * Construct memory allocator
+     *
+     * @param device_             Device to allocate from
+     * @param export_allocations_ True when allocations have to be exported
+     *
+     * @throw vk::Exception on failure
+     */
+    explicit MemoryAllocator(const Device& device_, bool export_allocations_);
     ~MemoryAllocator();
 
     MemoryAllocator& operator=(const MemoryAllocator&) = delete;
@@ -106,8 +117,9 @@ private:
     /// Returns index to the fastest memory type compatible with the passed requirements.
     std::optional<u32> FindType(VkMemoryPropertyFlags flags, u32 type_mask) const;
 
-    const Device& device;                                       ///< Device handle.
-    const VkPhysicalDeviceMemoryProperties properties;          ///< Physical device properties.
+    const Device& device;                              ///< Device handle.
+    const VkPhysicalDeviceMemoryProperties properties; ///< Physical device properties.
+    const bool export_allocations; ///< True when memory allocations have to be exported.
     std::vector<std::unique_ptr<MemoryAllocation>> allocations; ///< Current allocations.
 };
 

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -168,11 +168,15 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkFreeCommandBuffers);
     X(vkFreeDescriptorSets);
     X(vkFreeMemory);
-    X(vkGetBufferMemoryRequirements);
+    X(vkGetBufferMemoryRequirements2);
     X(vkGetDeviceQueue);
     X(vkGetEventStatus);
     X(vkGetFenceStatus);
     X(vkGetImageMemoryRequirements);
+    X(vkGetMemoryFdKHR);
+#ifdef _WIN32
+    X(vkGetMemoryWin32HandleKHR);
+#endif
     X(vkGetQueryPoolResults);
     X(vkGetSemaphoreCounterValueKHR);
     X(vkMapMemory);
@@ -505,6 +509,32 @@ void ImageView::SetObjectNameEXT(const char* name) const {
     SetObjectName(dld, owner, handle, VK_OBJECT_TYPE_IMAGE_VIEW, name);
 }
 
+int DeviceMemory::GetMemoryFdKHR() const {
+    const VkMemoryGetFdInfoKHR get_fd_info{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
+        .pNext = nullptr,
+        .memory = handle,
+        .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
+    };
+    int fd;
+    Check(dld->vkGetMemoryFdKHR(owner, &get_fd_info, &fd));
+    return fd;
+}
+
+#ifdef _WIN32
+HANDLE DeviceMemory::GetMemoryWin32HandleKHR() const {
+    const VkMemoryGetWin32HandleInfoKHR get_win32_handle_info{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_GET_WIN32_HANDLE_INFO_KHR,
+        .pNext = nullptr,
+        .memory = handle,
+        .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT_KHR,
+    };
+    HANDLE win32_handle;
+    Check(dld->vkGetMemoryWin32HandleKHR(owner, &get_win32_handle_info, &win32_handle));
+    return win32_handle;
+}
+#endif
+
 void DeviceMemory::SetObjectNameEXT(const char* name) const {
     SetObjectName(dld, owner, handle, VK_OBJECT_TYPE_DEVICE_MEMORY, name);
 }
@@ -756,10 +786,20 @@ DeviceMemory Device::AllocateMemory(const VkMemoryAllocateInfo& ai) const {
     return DeviceMemory(memory, handle, *dld);
 }
 
-VkMemoryRequirements Device::GetBufferMemoryRequirements(VkBuffer buffer) const noexcept {
-    VkMemoryRequirements requirements;
-    dld->vkGetBufferMemoryRequirements(handle, buffer, &requirements);
-    return requirements;
+VkMemoryRequirements Device::GetBufferMemoryRequirements(VkBuffer buffer,
+                                                         void* pnext) const noexcept {
+    const VkBufferMemoryRequirementsInfo2 info{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
+        .pNext = nullptr,
+        .buffer = buffer,
+    };
+    VkMemoryRequirements2 requirements{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
+        .pNext = pnext,
+        .memoryRequirements{},
+    };
+    dld->vkGetBufferMemoryRequirements2(handle, &info, &requirements);
+    return requirements.memoryRequirements;
 }
 
 VkMemoryRequirements Device::GetImageMemoryRequirements(VkImage image) const noexcept {

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -15,8 +15,19 @@
 #include <vector>
 
 #define VK_NO_PROTOTYPES
+#ifdef _WIN32
+#define VK_USE_PLATFORM_WIN32_KHR
+#endif
 #include <vulkan/vulkan.h>
 
+// Sanitize macros
+#ifdef CreateEvent
+#undef CreateEvent
+#endif
+#ifdef CreateSemaphore
+#undef CreateSemaphore
+#endif
+
 #include "common/common_types.h"
 
 #ifdef _MSC_VER
@@ -174,7 +185,7 @@ struct InstanceDispatch {
 };
 
 /// Table holding Vulkan device function pointers.
-struct DeviceDispatch : public InstanceDispatch {
+struct DeviceDispatch : InstanceDispatch {
     PFN_vkAcquireNextImageKHR vkAcquireNextImageKHR{};
     PFN_vkAllocateCommandBuffers vkAllocateCommandBuffers{};
     PFN_vkAllocateDescriptorSets vkAllocateDescriptorSets{};
@@ -272,11 +283,15 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkFreeCommandBuffers vkFreeCommandBuffers{};
     PFN_vkFreeDescriptorSets vkFreeDescriptorSets{};
     PFN_vkFreeMemory vkFreeMemory{};
-    PFN_vkGetBufferMemoryRequirements vkGetBufferMemoryRequirements{};
+    PFN_vkGetBufferMemoryRequirements2 vkGetBufferMemoryRequirements2{};
     PFN_vkGetDeviceQueue vkGetDeviceQueue{};
     PFN_vkGetEventStatus vkGetEventStatus{};
     PFN_vkGetFenceStatus vkGetFenceStatus{};
     PFN_vkGetImageMemoryRequirements vkGetImageMemoryRequirements{};
+    PFN_vkGetMemoryFdKHR vkGetMemoryFdKHR{};
+#ifdef _WIN32
+    PFN_vkGetMemoryWin32HandleKHR vkGetMemoryWin32HandleKHR{};
+#endif
     PFN_vkGetQueryPoolResults vkGetQueryPoolResults{};
     PFN_vkGetSemaphoreCounterValueKHR vkGetSemaphoreCounterValueKHR{};
     PFN_vkMapMemory vkMapMemory{};
@@ -344,6 +359,9 @@ public:
     /// Construct an empty handle.
     Handle() = default;
 
+    /// Construct an empty handle.
+    Handle(std::nullptr_t) {}
+
     /// Copying Vulkan objects is not supported and will never be.
     Handle(const Handle&) = delete;
     Handle& operator=(const Handle&) = delete;
@@ -659,6 +677,12 @@ class DeviceMemory : public Handle<VkDeviceMemory, VkDevice, DeviceDispatch> {
     using Handle<VkDeviceMemory, VkDevice, DeviceDispatch>::Handle;
 
 public:
+    int GetMemoryFdKHR() const;
+
+#ifdef _WIN32
+    HANDLE GetMemoryWin32HandleKHR() const;
+#endif
+
     /// Set object name.
     void SetObjectNameEXT(const char* name) const;
 
@@ -847,7 +871,8 @@ public:
 
     DeviceMemory AllocateMemory(const VkMemoryAllocateInfo& ai) const;
 
-    VkMemoryRequirements GetBufferMemoryRequirements(VkBuffer buffer) const noexcept;
+    VkMemoryRequirements GetBufferMemoryRequirements(VkBuffer buffer,
+                                                     void* pnext = nullptr) const noexcept;
 
     VkMemoryRequirements GetImageMemoryRequirements(VkImage image) const noexcept;
 
@@ -1031,6 +1056,12 @@ public:
         PipelineBarrier(src_stage_mask, dst_stage_mask, dependency_flags, {}, {}, {});
     }
 
+    void PipelineBarrier(VkPipelineStageFlags src_stage_mask, VkPipelineStageFlags dst_stage_mask,
+                         VkDependencyFlags dependency_flags,
+                         const VkMemoryBarrier& memory_barrier) const noexcept {
+        PipelineBarrier(src_stage_mask, dst_stage_mask, dependency_flags, memory_barrier, {}, {});
+    }
+
     void PipelineBarrier(VkPipelineStageFlags src_stage_mask, VkPipelineStageFlags dst_stage_mask,
                          VkDependencyFlags dependency_flags,
                          const VkBufferMemoryBarrier& buffer_barrier) const noexcept {

src/yuzu/bootmanager.cpp

@@ -64,7 +64,7 @@ void EmuThread::run() {
 
     emit LoadProgress(VideoCore::LoadCallbackStage::Prepare, 0, 0);
 
-    system.Renderer().Rasterizer().LoadDiskResources(
+    system.Renderer().ReadRasterizer()->LoadDiskResources(
         system.CurrentProcess()->GetTitleID(), stop_run,
         [this](VideoCore::LoadCallbackStage stage, std::size_t value, std::size_t total) {
             emit LoadProgress(stage, value, total);

src/yuzu/configuration/config.cpp

@@ -782,14 +782,14 @@ void Config::ReadRendererValues() {
     ReadSettingGlobal(Settings::values.frame_limit, QStringLiteral("frame_limit"), 100);
     ReadSettingGlobal(Settings::values.use_disk_shader_cache,
                       QStringLiteral("use_disk_shader_cache"), true);
-    ReadSettingGlobal(Settings::values.gpu_accuracy, QStringLiteral("gpu_accuracy"), 0);
+    ReadSettingGlobal(Settings::values.gpu_accuracy, QStringLiteral("gpu_accuracy"), 1);
     ReadSettingGlobal(Settings::values.use_asynchronous_gpu_emulation,
                       QStringLiteral("use_asynchronous_gpu_emulation"), true);
     ReadSettingGlobal(Settings::values.use_nvdec_emulation, QStringLiteral("use_nvdec_emulation"),
                       true);
     ReadSettingGlobal(Settings::values.use_vsync, QStringLiteral("use_vsync"), true);
     ReadSettingGlobal(Settings::values.use_assembly_shaders, QStringLiteral("use_assembly_shaders"),
-                      true);
+                      false);
     ReadSettingGlobal(Settings::values.use_asynchronous_shaders,
                       QStringLiteral("use_asynchronous_shaders"), false);
     ReadSettingGlobal(Settings::values.use_fast_gpu_time, QStringLiteral("use_fast_gpu_time"),
@@ -1351,14 +1351,14 @@ void Config::SaveRendererValues() {
                        Settings::values.use_disk_shader_cache, true);
     WriteSettingGlobal(QStringLiteral("gpu_accuracy"),
                        static_cast<int>(Settings::values.gpu_accuracy.GetValue(global)),
-                       Settings::values.gpu_accuracy.UsingGlobal(), 0);
+                       Settings::values.gpu_accuracy.UsingGlobal(), 1);
     WriteSettingGlobal(QStringLiteral("use_asynchronous_gpu_emulation"),
                        Settings::values.use_asynchronous_gpu_emulation, true);
     WriteSettingGlobal(QStringLiteral("use_nvdec_emulation"), Settings::values.use_nvdec_emulation,
                        true);
     WriteSettingGlobal(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
     WriteSettingGlobal(QStringLiteral("use_assembly_shaders"),
-                       Settings::values.use_assembly_shaders, true);
+                       Settings::values.use_assembly_shaders, false);
     WriteSettingGlobal(QStringLiteral("use_asynchronous_shaders"),
                        Settings::values.use_asynchronous_shaders, false);
     WriteSettingGlobal(QStringLiteral("use_fast_gpu_time"), Settings::values.use_fast_gpu_time,

src/yuzu/configuration/configure_graphics.cpp

@@ -2,6 +2,9 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+// Include this early to include Vulkan headers how we want to
+#include "video_core/vulkan_common/vulkan_wrapper.h"
+
 #include <QColorDialog>
 #include <QComboBox>
 #include <QVulkanInstance>
@@ -11,7 +14,8 @@
 #include "core/core.h"
 #include "core/settings.h"
 #include "ui_configure_graphics.h"
-#include "video_core/renderer_vulkan/renderer_vulkan.h"
+#include "video_core/vulkan_common/vulkan_instance.h"
+#include "video_core/vulkan_common/vulkan_library.h"
 #include "yuzu/configuration/configuration_shared.h"
 #include "yuzu/configuration/configure_graphics.h"
 
@@ -212,11 +216,23 @@ void ConfigureGraphics::UpdateDeviceComboBox() {
     ui->device->setEnabled(enabled && !Core::System::GetInstance().IsPoweredOn());
 }
 
-void ConfigureGraphics::RetrieveVulkanDevices() {
+void ConfigureGraphics::RetrieveVulkanDevices() try {
+    using namespace Vulkan;
+
+    vk::InstanceDispatch dld;
+    const Common::DynamicLibrary library = OpenLibrary();
+    const vk::Instance instance = CreateInstance(library, dld, VK_API_VERSION_1_0);
+    const std::vector<VkPhysicalDevice> physical_devices = instance.EnumeratePhysicalDevices();
+
     vulkan_devices.clear();
-    for (const auto& name : Vulkan::RendererVulkan::EnumerateDevices()) {
+    vulkan_devices.reserve(physical_devices.size());
+    for (const VkPhysicalDevice device : physical_devices) {
+        const char* const name = vk::PhysicalDevice(device, dld).GetProperties().deviceName;
         vulkan_devices.push_back(QString::fromStdString(name));
     }
+
+} catch (const Vulkan::vk::Exception& exception) {
+    LOG_ERROR(Frontend, "Failed to enumerate devices with error: {}", exception.what());
 }
 
 Settings::RendererBackend ConfigureGraphics::GetCurrentGraphicsBackend() const {

src/yuzu_cmd/config.cpp

@@ -388,7 +388,7 @@ void Config::ReadValues() {
         static_cast<u16>(sdl2_config->GetInteger("Renderer", "frame_limit", 100)));
     Settings::values.use_disk_shader_cache.SetValue(
         sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false));
-    const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 0);
+    const int gpu_accuracy_level = sdl2_config->GetInteger("Renderer", "gpu_accuracy", 1);
     Settings::values.gpu_accuracy.SetValue(static_cast<Settings::GPUAccuracy>(gpu_accuracy_level));
     Settings::values.use_asynchronous_gpu_emulation.SetValue(
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", true));

src/yuzu_cmd/yuzu.cpp

@@ -215,7 +215,7 @@ int main(int argc, char** argv) {
     // Core is loaded, start the GPU (makes the GPU contexts current to this thread)
     system.GPU().Start();
 
-    system.Renderer().Rasterizer().LoadDiskResources(
+    system.Renderer().ReadRasterizer()->LoadDiskResources(
         system.CurrentProcess()->GetTitleID(), false,
         [](VideoCore::LoadCallbackStage, size_t value, size_t total) {});